Introduction To Prescriptive AI: A Primer For Decision Intelligence Solutioning With Python

Introduction to
Prescriptive AI
A Primer for Decision Intelligence
Solutioning with Python
—
Akshay Kulkarni
Adarsha Shivananda
Avinash Manure
Introduction to
Prescriptive AI
A Primer for Decision
Intelligence Solutioning
with Python
Akshay Kulkarni
Adarsha Shivananda
Avinash Manure
Introduction to Prescriptive AI: A Primer for Decision Intelligence
Solutioning with Python
Akshay Kulkarni Adarsha Shivananda
Bangalore, Karnataka, India Hosanagara, Karnataka, India
Avinash Manure
Bangalore, Karnataka, India
ISBN-13 (pbk): 978-1-4842-9567-0 ISBN-13 (electronic): 978-1-4842-9568-7

https://doi.org/10.1007/978-1-4842-9568-7
Copyright © 2023 by Akshay Kulkarni, Adarsha Shivananda, Avinash Manure

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way,
and transmission or information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed.
Trademarked names, logos, and images may appear in this book. Rather than use a trademark
symbol with every occurrence of a trademarked name, logo, or image we use the names, logos,
and images only in an editorial fashion and to the benefit of the trademark owner, with no
intention of infringement of the trademark.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if
they are not identified as such, is not to be taken as an expression of opinion as to whether or not
they are subject to proprietary rights.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal
responsibility for any errors or omissions that may be made. The publisher makes no warranty,
express or implied, with respect to the material contained herein.
Managing Director, Apress Media LLC: Welmoed Spahr
Acquisitions Editor: Celestin Suresh John
Development Editor: Laura Berendson
Coordinating Editor: Mark Powers
Cover designed by eStudioCalamar
Cover image by Anna Tsukanova on Unsplash (www.unsplash.com)
Distributed to the book trade worldwide by Apress Media, LLC, 1 New York Plaza, New York, NY
10004, U.S.A. Phone 1-800-SPRINGER, fax (201) 348-4505, e-mail [email protected],
or visit www.springeronline.com. Apress Media, LLC is a California LLC and the sole member
(owner) is Springer Science + Business Media Finance Inc (SSBM Finance Inc). SSBM Finance
Inc is a Delaware corporation.
For information on translations, please e-mail [email protected]; for
reprint, paperback, or audio rights, please e-mail [email protected].
Apress titles may be purchased in bulk for academic, corporate, or promotional use. eBook
versions and licenses are also available for most titles. For more information, reference our Print
and eBook Bulk Sales web page at www.apress.com/bulk-sales.
Any source code or other supplementary material referenced by the author in this book is
available to readers on GitHub (https://github.com/Apress). For more detailed information,
please visit www.apress.com/source-code.
Printed on acid-free paper
This book is dedicated to my 2-year-old daughter,
Siri, who has been a constant source of inspiration,
joy, and wonder in my life since the day she was born.
Siri, you have shown me the true meaning of
unconditional love, and watching you grow and
learn has been the greatest adventure of my life.
This book is a tribute to the amazing person
you are becoming and a reminder of the infinite
potential that lies within you. Thank you for being
my daughter, my teacher, and my greatest blessing.
—Avinash Manure
Table of Contents
About the Authors��ix
About the Technical Reviewer��xi
Acknowledgments��xiii
Introduction��xv
Chapter 1: Decision Intelligence Overview��1

Types of AI��2
Decision Intelligence��5
Decision Intelligence History��6
Challenges in AI Adoption��7
How Can DI Help Bridge the Gap Between AI and Business?��8
The Need for Decision Intelligence��9
The Evolution of Decision-Making��11
Challenges��13
Applications��14
Understanding Where Decision Intelligence Fits Within the AI Life Cycle��16
Decision Intelligence Methodologies��18
Some Potential Pros and Cons of DI��22
Examples of How Companies Are Leveraging DI��24
Conclusion��25
v
Table of Contents
Chapter 2: Decision Intelligence Requirements��27

Why Do AI Projects Fail?��27
DI Requirements Framework��30
Planning��31
Approach��33
Approval Mechanism/Organization Alignment��36
Key Performance Indicators��37
Define Clear Metrics��38
Value��40
Return on Investment��41
Value per Decision��42
Consumption of the AI Predictions��43
Conclusion��46
Chapter 3: Decision Intelligence Methodologies��47

Decision-Making��47
Types of Decision-Making��48
Decision-Making Process��51
Decision-Making Methodologies��57
Human-Only Decision-Making��58
Human-Machine Decision-Making��73
Machine-Only Decision-Making��81
Conclusion��83
Chapter 4: Interpreting Results from Different Methodologies��85

Decision Intelligence Methodology: Mathematical Models��85
Linear Models��86
Nonlinear Models��90
vi
Table of Contents
Decision Intelligence Methodology: Probabilistic Models��94

Markov Chain��95
Decision Intelligence Methodology: AI/ML Models��103
Conclusion��108
Chapter 5: Augmenting Decision Intelligence Results into the

Business Workflow��109
Challenges��110
Workflow��112
Decision Intelligence Apps��114
How and Why?��115
User-Friendly Interfaces��116
Augmenting AI Predictions to Business Workflow��118
Connect to Business Tools��120
Map the Data��121
Conclusion��122
Chapter 6: Actions, Biases, and Human-in-the-Loop��125

Key Ethical Considerations in AI��125
Actions, Biases, and Human-in-the-Loop��127
Cognitive Biases��128
Why Is Detecting Bias Important?��129
Types��130
What Happens If Bias Is Ignored?��132
Bias Detection��133
What Do Bias Tools Do?��134
Incorporation of Feedback Through Human Intervention��136
How to Build HITL Systems?��139
vii
Table of Contents
Example: Customer Churn��140

Conclusion��141
Chapter 7: Case Studies��143

Case Study 1: Telecom Customer Churn Management��143
Case Study 2: Mobile Phone Pricing/Configuration Strategy��173
Conclusion��188
Index��189
viii
About the Authors
Akshay Kulkarni is an artificial intelligence
(AI) and machine learning (ML) evangelist
and a thought leader. As a consultant, he has
worked with several Fortune 500 and global
enterprises to drive AI and data science–led
strategic transformations. He is a Google
developer, an author, and a regular speaker
at major AI and data science conferences
(including the O’Reilly Strata Data & AI
Conference and Great International Developer
Summit [GIDS]). He is a visiting faculty member at some of the top
graduate institutes in India. In 2019, he was featured as one of India’s “top
40 under 40” data scientists. In his spare time, Akshay enjoys reading,
writing, coding, and helping aspiring data scientists.
Adarsha Shivananda is a data science and

MLOps leader. He is working on creating
world-class MLOps capabilities to ensure
continuous value delivery from AI. He aims to
build a pool of exceptional data scientists within
and outside organizations to solve problems
through training programs. He always wants to
stay ahead of the curve. Adarsha has worked
extensively in the pharma, healthcare, CPG, retail,
and marketing domains. He lives in Bangalore
and loves to read and teach data science.
ix
About the Authors
Avinash Manure is a seasoned machine

learning professional with 10+ years of
experience building, deploying, and
maintaining state-of-the-art machine learning
solutions across different industries. He has
six years of experience leading and mentoring
high-performance teams in developing
ML systems catering to different business
requirements. He is proficient in deploying
complex machine learning and statistical
modeling algorithms/techniques for identifying patterns and extracting
valuable insights for key stakeholders and organizational leadership.
x
About the Technical Reviewer
Nitin Ranjan Sharma is a manager at Novartis.
He leads a team that develops products using
multimodal techniques. As a consultant,
he has developed solutions for Fortune 500
companies and has been involved in solving
complex business problems using machine
learning and deep learning frameworks.
His major focus area and core expertise is
computer vision, including solving challenging
business problems dealing with images and
video data. Before Novartis, he was part of the data science team at Publicis
Sapient, EY, and TekSystems Global Services. He is a regular speaker at
data science community meetups and an open-source contributor. He also
enjoys training and mentoring data science enthusiasts.
xi
Acknowledgments
We would like to thank our families who have always made sure we had
the right environment at home to concentrate on this book and complete
it on time. We would also like to thank the publishing team—Mark Powers
and Celestin Suresh John, and our technical reviewer, Nitin Sharma—who
helped us make sure this book was the best it could be. We would also
like to thank our mentors who made sure we grew professionally and
personally by always supporting us in our dreams and guiding us toward
achieving our goals. Last but not least, we thank our parents, our friends,
and our colleagues who were always there in tough times and motivated us
to chase our dreams.
xiii
Introduction
This book will introduce you to the concept of decision intelligence,
including its history and current and future trends. It will help you
evaluate different decision intelligence techniques and guide you on
how to them through different prescriptive AI methods and incorporate
them into business workflows through different domain-specific use case
implementations.
This book is for data scientists, AI/machine learning engineers, and
deep learning professionals who are working toward building advanced
intelligent AI/ML applications. This book is also for business professionals
and nontechnical stakeholders who want to understand how decision
intelligence can help a business grow.
This book will take you through the journey of decision-making
in companies with key milestones, key statistics, and benefits. It will
provide insights on where decision intelligence fits within the AI life
cycle. This book will provide insights on how to prepare for prescriptive
AI (a key requirement to decision intelligence) with the help of a business
requirement document. It will then deep dive into different decision
intelligence methodologies, their advantages, and their limitations. Next,
you will learn how to perform different simulations and interpret the
results from them. Then you will be guided on how to enable and embed
the decision intelligence process into the business workflow through
prescriptive AI. You will learn about different cognitive biases that humans
make and how that can be lowered/eliminated through the combination
of machine and human intelligence. Finally, you will find different cases
studies by domain through tailored use cases.
xv
Introduction
By the end of the book, you will have a solid understanding of

decision intelligence and prescriptive AI tools and techniques and how
to incorporate them within the business workflow for greater productivity
and profit to the business.
The source code for this book is available on GitHub (github.com/
apress/introduction-prescriptive-ai).
xvi
CHAPTER 1
Decision Intelligence
Overview
Prescriptive AI is a type of artificial intelligence that is designed to provide
recommendations, solutions, or actions to optimize or improve a specific
process or outcome. It is unlike descriptive AI, which describes “what”
has happened in the past/present; inferential/diagnostic AI, which
helps us understand “why” something has happened; and predictive AI,
which helps predict what might happen in the future. Once we have the
predictions, prescriptive AI focuses on what actions should be taken to
achieve a particular goal or outcome.
Prescriptive AI can be used in a variety of applications, such as in
healthcare to help doctors diagnose diseases and prescribe treatments, in
finance to make investment decisions, and in manufacturing to optimize
production processes. It typically uses machine learning algorithms
and other advanced technologies to analyze large datasets and generate
recommendations based on the data.
Overall, the goal of prescriptive AI is to help humans make better
decisions by providing them with accurate and actionable insights based
on data-driven analysis.
© Akshay Kulkarni, Adarsha Shivananda, Avinash Manure 2023 1

A. Kulkarni et al., Introduction to Prescriptive AI,
https://doi.org/10.1007/978-1-4842-9568-7_1
Chapter 1 Decision Intelligence Overview
As mentioned, prescriptive AI systems differ from descriptive and

predictive AI systems in that they not only provide insights and predictions
based on data analysis but also offer specific recommendations or actions.
Prescriptive AI goes a step further by suggesting optimal solutions and
strategies to achieve desired outcomes.
In this chapter, we will start with the types of AI and then deep dive
into prescriptive AI and decision intelligence.
Types of AI
Descriptive, diagnostic, predictive, and prescriptive AI are four different
types of artificial intelligence that are used in various applications. Here’s
an overview of each:
• Descriptive AI: Descriptive AI, also known as
descriptive analytics, is used to describe what has
happened in the past or is currently happening. It
involves analyzing historical data to identify patterns,
trends, and insights. Descriptive AI is often used for
reporting and business intelligence applications,
such as dashboards and scorecards that summarize
key performance metrics. It can also be used for
exploratory data analysis to uncover hidden insights
and correlations.
For example, a retail company may use descriptive
AI to analyze its sales data and identify which
products are selling the most, which stores are
performing the best, and which customers are
buying the most. This information can then be
used to make data-driven decisions on inventory
management, pricing strategies, and marketing
campaigns.
2
• Diagnostic AI: Diagnostic AI, also known as diagnostic

analytics, is used to identify the root cause of a
problem or issue. It involves analyzing data to identify
factors that contribute to a specific outcome or event.
Diagnostic AI is often used in healthcare, finance, and
manufacturing applications.
For example, a hospital may use diagnostic AI to

analyze patient data and identify the factors that
contribute to re-admissions. This could involve
analyzing patient demographics, medical histories,
and treatment plans to identify common risk factors
and develop targeted interventions to reduce re-
admissions.
• Predictive AI: Predictive AI, also known as predictive

analytics, is used to predict what is likely to happen in
the future. It involves using statistical modeling and
machine learning techniques to analyze historical
data and make predictions about future outcomes.
Predictive AI is often used in marketing, finance, and
customer service applications.
For example, a bank may use predictive AI to
analyze customer data and predict which customers
are most likely to default on their loans. This
information can then be used to proactively reach
out to those customers and offer them personalized
loan modification options to avoid default.
• Prescriptive AI: Prescriptive AI, also known

as prescriptive analytics, is used to provide
recommendations on what actions to take based on
the predictions made by predictive AI. It involves
3
using optimization algorithms to identify the best

course of action based on a set of constraints and
objectives. Prescriptive AI is often used in supply
chain management, logistics, and resource-planning
applications.
For example, a logistics company may use

prescriptive AI to optimize its delivery routes based
on real-time traffic data, weather conditions, and
delivery deadlines. This information can then be
used to automatically reroute delivery trucks to
avoid traffic jams and ensure on-time delivery.
In summary, as shown in Figure 1-1, descriptive AI is used to describe

what has happened, diagnostic AI is used to identify the root cause of a
problem, predictive AI is used to predict what is likely to happen, and
prescriptive AI is used to recommend what actions to take based on those
predictions. Each type of AI has different applications and can be used to
solve different types of problems.
Figure 1-1. Types of AI
4
Prescriptive AI typically involves analyzing large amounts of data

using machine learning algorithms to identify patterns and relationships
between variables. It then uses this analysis to provide recommendations
or decisions that can optimize business processes, improve customer
experiences, or achieve other desired outcomes.
Examples of prescriptive AI applications include personalized
medicine, where AI systems provide doctors with recommendations for
personalized treatments based on a patient’s genetic and medical history,
and supply chain management, where AI systems optimize logistics and
inventory management to minimize costs and improve efficiency.
We try to achieve preceptive AI through decision intelligence.
Decision intelligence (DI) is a discipline that combines techniques from
artificial intelligence, decision theory, and behavioral science to support
and improve human decision-making. It aims to provide a more structured
and rigorous approach to decision-making by using data and analytics
to help people make better decisions. It involves using a combination of
human expertise and machine intelligence to make better decisions in
complex and uncertain situations.
DI involves a variety of techniques, including machine learning,
optimization, simulation, and behavioral science, to provide decision-
makers with insights and recommendations. It typically involves collecting
and analyzing data from various sources, identifying decision criteria, and
generating possible alternatives based on various constraints and objectives.
The ultimate goal of DI is to enable organizations and individuals
to make better decisions that are more aligned with their goals, values,
and preferences. This can lead to improved efficiency, effectiveness,
and competitiveness in a wide range of domains, including business,
healthcare, government, and education.
5
Decision Intelligence History

Decision intelligence has its roots in a variety of fields, including decision
theory, game theory, and operations research. Let us understand a brief
history of DI.
• Decision theory: Decision theory is a field of study
that has been around for more than a century. In
the early 1900s, economists and mathematicians
began to develop formal models for decision-making,
which included elements such as utility theory and
expected value. These models provided a framework
for understanding how decisions are made and how
outcomes can be optimized.
• Operations research: Operations research (OR)

emerged during World War II as a way to solve military
logistics problems. OR involves using mathematical
and statistical methods to analyze complex systems
and make decisions. OR has been applied to a
wide range of industries, including manufacturing,
transportation, and healthcare.
• Game theory: Game theory is a branch of mathematics

that studies how people interact in strategic
situations. It was developed in the mid-20th century
by mathematicians such as John Nash and John von
Neumann and has been used to study decision-making
in fields such as economics, political science, and
psychology.
• Business analytics: In recent years, with the rise of
big data and data analytics, decision-making has
become increasingly data-driven. Business analytics
6
involves using data to inform decision-making and has

been used to optimize processes, improve customer
experiences, and drive revenue growth.
• Decision intelligence: DI emerged in the early 2000s

as a way to integrate the various disciplines involved
in decision-making, including decision theory, game
theory, and analytics. DI involves using a range of tools
and techniques to support decision-making, including
machine learning, optimization, and simulation.
Challenges in AI Adoption
There are several challenges that businesses face when adopting AI
technology. These are some of the key challenges:
• Data quality and availability: AI systems rely heavily

on data, so poor data quality and a lack of available
data can hinder the effectiveness of the system.
• Technical expertise: Implementing/maintaining AI

systems often requires specialized technical expertise,
which may not be readily available within the business.
• Cost: AI systems can be expensive to implement and

maintain, especially if the business does not have the
necessary infrastructure in place.
• Resistance to change: There may be resistance to

change from employees who are used to working with
traditional systems and processes.
• Ethical considerations: AI systems can raise ethical

concerns related to privacy, bias, and transparency,
which must be addressed to ensure ethical
decision-making.
7
• Integration with existing systems: Integrating AI with

existing systems can be challenging, particularly if
those systems are outdated or incompatible with the
new technology.
• Regulation and compliance: AI systems may need

to comply with regulations such as GDPR or HIPAA,
which can add an extra layer of complexity to their
implementation.
• Lack of understanding: There may be a lack of

understanding among business leaders and employees
about the potential benefits and limitations of AI
technology, which can hinder adoption.
To overcome these challenges, businesses need to carefully evaluate

the potential benefits and risks of AI adoption, identify the necessary
resources and expertise required, and develop a comprehensive plan for
implementing and integrating the technology into their existing systems
and processes. They also need to prioritize ethical considerations and
work closely with employees to ensure that they are comfortable and
confident using the new technology.
ow Can DI Help Bridge the Gap Between AI

H
and Business?
One of the challenges with AI is that it can be difficult to understand how
it arrives at its decisions. DI can help by providing insights into how AI
models work, allowing business users to understand and interpret the
decisions that are being made.
8
• Integration: DI can help integrate AI technology into

existing business processes and workflows, making
it easier for business users to take advantage of AI
capabilities without having to learn complex new
systems.
• Customization: DI can be customized to fit the specific

needs of a business, allowing for a more personalized
approach to decision-making that takes into account
the unique challenges and opportunities facing the
organization.
• Risk management: DI can help identify and mitigate

risks associated with AI technology, such as bias,
privacy concerns, and data quality issues, ensuring
that the benefits of AI are realized while minimizing
the risks.
• Transparency: DI can help improve transparency

in decision-making by providing clear and
understandable explanations of the decisions being
made, helping to build trust and confidence in AI
technology among business users and stakeholders.
By leveraging the power of AI in a way that is tailored to the specific
needs of businesses, DI can help bridge the gap between AI and business,
enabling organizations to take advantage of the benefits of AI technology
while minimizing the risks and ensuring that decisions are made in a way
that is transparent, ethical, and effective.
The Need for Decision Intelligence

There are several reasons why DI is becoming increasingly important in
today’s world.
9
• Information overload: With the explosion of data and

information, decision-makers are often overwhelmed
with too much information, making it difficult to
make informed decisions. DI can help cut through the
noise by analyzing and synthesizing relevant data into
actionable insights.
• Complex and uncertain environments: Many

decisions today are made in complex and uncertain
environments, where there are multiple factors and
variables to consider. DI can help decision-makers
better understand the interdependencies and
trade-offs between different options and provide
recommendations based on probabilistic modeling
and simulation.
• Cognitive biases: Human decision-making is often

subject to cognitive biases, such as overconfidence,
anchoring, and confirmation bias. DI can help mitigate
these biases by providing a more objective and data-
driven approach to decision-making.
• Competitive advantage: In today’s competitive

environment, organizations that can make better
decisions faster are more likely to succeed. DI
can provide a competitive advantage by enabling
organizations to make better decisions based on data-
driven insights.
• Improved decision quality: DI can help decision-

makers make better, more informed decisions by
providing insights and recommendations based on
data and analytics. This can lead to improved outcomes
and increased efficiency.
10
• Risk management: Many decisions involve some level

of risk. DI can help decision-makers understand and
manage this risk by providing probabilistic assessments
and scenario analysis.
• Agility and flexibility: In today’s fast-paced and

rapidly changing environment, organizations need to
be agile and flexible in their decision-making. DI can
help organizations respond quickly and effectively to
changing circumstances by providing real-time data
and insights.
• Transparency and accountability: DI can help ensure

that decisions are transparent and accountable by
providing a clear audit trail and documentation of the
decision-making process.
• Personalization: DI can help personalize decisions

based on individual preferences and goals, leading to a
more tailored and effective decision-making process.
Overall, DI can help organizations and individuals make better

decisions that are more aligned with their goals, values, and preferences,
leading to improved outcomes and performance.
The Evolution of Decision-Making

Companies have been shaped by various factors, including advancements
in technology, changes in organizational structures, and the growing
importance of data and analytics. Here are some key milestones in the
evolution of decision-making in companies:
11
• Intuition-based decision-making: In the early days

of business, decision-making was often based on the
intuition and experience of leaders and managers.
Decisions were made quickly and informally, based on
a hunch or gut feeling.
• Rule-based decision-making: As companies grew

and became more complex, decision-making became
more structured and rule-based. Managers developed
standardized processes and procedures for decision-
making, based on predetermined criteria and rules.
• Data-driven decision-making: With the rise of

information technology and data analytics, decision-
making in companies shifted toward a more data-
driven approach. Managers began to use data and
metrics to inform their decisions, leveraging tools
such as business intelligence, data visualization, and
predictive analytics.
• Collaborative decision-making: In recent years,

decision-making in companies has become more
collaborative, with a focus on involving stakeholders
from across the organization in the decision-making
process. This approach helps to ensure that decisions
are made with input from diverse perspectives and that
all relevant information and expertise are taken into
account.
• AI-powered decision-making: As artificial intelligence

(AI) and machine learning technologies continue to
advance, companies are beginning to explore how they
can be used to support decision-making. AI-powered
decision-making systems can help companies process
12
large amounts of data quickly and identify patterns

and insights that may not be immediately apparent
to humans.
Overall, the evolution of decision-making in companies has been

marked by a shift toward a more data-driven, collaborative, and
technology-enabled approach, with a growing emphasis on using data and
analytics to support decision-making processes.
Challenges
Decision intelligence can be challenging to implement because it involves
several complex and interconnected processes. Here are a few reasons why
DI is hard:
• Complexity of data: In many cases, the data that is

required for DI is complex, voluminous, and unstructured,
making it difficult to collect, clean, and process.
• Interpretability of AI models: AI models can be

complex and difficult to interpret, which can make
it challenging to understand how they arrive at their
decisions.
• Expertise required: Implementing DI requires

specialized expertise in areas such as data science,
machine learning, and decision-making theory.
Finding and hiring the right talent can be challenging,
especially given the high demand for these skills.
• Integration with existing systems: DI often requires

integration with existing systems, which can be complex
and time-consuming, especially if those systems are
outdated or incompatible with new technology.
13
• Change management: Implementing DI can require

significant changes to existing business processes
and workflows, which can be difficult to manage and
communicate to employees.
Despite these challenges, DI can provide significant benefits to

businesses by enabling more informed and effective decision-making. By
carefully planning and managing the implementation process, businesses
can overcome these challenges and realize the full potential of DI.
Applications
There are numerous applications of DI across different domains. Here are
some examples:
• Business: DI can help businesses make better

decisions in areas such as pricing, product
development, supply chain management, and
customer service. For example, DI can help retailers
optimize pricing strategies based on customer behavior
and competitor analysis.
• Healthcare: DI can be used to help healthcare
professionals make more informed decisions about
patient care, diagnosis, and treatment. For example, DI
can help doctors determine the best treatment plan for
a patient based on their medical history and symptoms.
• Government: DI can be used by government agencies

to improve decision-making in areas such as policy
development, resource allocation, and emergency
response. For example, DI can help emergency
responders make faster and more effective decisions
during a natural disaster.
14
• Finance: DI can help financial institutions make better

decisions in areas such as investment management,
risk management, and fraud detection. For example,
DI can help investment managers optimize portfolio
allocation based on market trends and risk profiles.
• Education: DI can be used to improve decision-making

in areas such as student performance evaluation,
curriculum design, and resource allocation. For
example, DI can help teachers personalize learning
for students based on their individual needs and
learning styles.
• Marketing: DI can help marketers make better

decisions in areas such as customer segmentation,
campaign optimization, and product positioning.
For example, DI can help marketers identify the most
effective marketing channels for different customer
segments.
• Manufacturing: DI can be used to optimize

manufacturing processes, improve quality control, and
reduce waste. For example, DI can help manufacturers
identify the most efficient production methods
based on variables such as raw material cost, energy
consumption, and product quality.
• Transportation: DI can be used to optimize

transportation systems, improve route planning,
and reduce congestion. For example, DI can help
transportation planners determine the optimal bus
routes based on passenger demand and traffic patterns.
15
• Agriculture: DI can be used to improve crop yields,

reduce water usage, and optimize fertilizer application.
For example, DI can help farmers determine the
optimal planting and harvesting times based on
weather patterns and soil conditions.
• Energy: DI can be used to optimize energy

consumption, reduce emissions, and improve energy
efficiency. For example, DI can help energy companies
determine the most cost-effective sources of renewable
energy based on factors such as location and weather
patterns.
Overall, DI has the potential to improve decision-making across a wide

range of domains, leading to better outcomes and increased efficiency.
nderstanding Where Decision Intelligence

U
Fits Within the AI Life Cycle
Decision intelligence plays a critical role in the AI life cycle, particularly in
the later stages of the process. Here’s how DI fits within the AI life cycle:
• Data collection: The first stage of the AI life cycle
involves collecting and cleaning data. This stage is
important for decision intelligence because the quality
of the data collected will influence the quality of the
decisions made.
• Data preparation: Once data is collected, it must be

processed and prepared for use in AI models. This stage
may involve data transformation, feature engineering,
and data augmentation. These steps can impact the
accuracy and effectiveness of decision intelligence.
16
• Model training: This stage involves training AI

models using machine learning algorithms. Decision
intelligence is critical at this stage, as it involves
selecting the right algorithms, defining the problem
and objectives, and validating the accuracy and
reliability of the models.
• Model deployment: Once models are trained, they

need to be deployed in production systems. Decision
intelligence is important at this stage to ensure that
the models are deployed in a way that maximizes their
impact and effectiveness while minimizing risk.
• Monitoring and optimization: AI models must be

continuously monitored and optimized to ensure that
they are performing as intended. Decision intelligence
plays a critical role in this stage, as it involves
analyzing performance metrics, identifying areas
for improvement, and making decisions to optimize
the models.
• Decision-making: Finally, decision intelligence is

used to inform decision-making based on the insights
and predictions generated by the AI models. This may
involve selecting the best course of action based on
the predictions generated by the models or using the
models to inform strategic decision-making processes.
Overall, decision intelligence is a critical component of the AI life

cycle, as it helps to ensure that AI models are developed, deployed, and
used in a way that maximizes their impact and effectiveness.
17
Decision Intelligence Methodologies

There are several methodologies that can be used in decision intelligence,
depending on the specific problem being addressed and the data available.
Here are some commonly used methodologies:
• Decision analysis: This methodology involves
structuring a decision problem into a decision
tree, identifying possible alternatives, evaluating
the outcomes of each alternative, and assigning
probabilities to each outcome. This helps decision-
makers understand the trade-offs between different
options and choose the best course of action.
• Machine learning: This methodology involves

using algorithms to learn patterns in data and make
predictions or decisions based on those patterns.
Machine learning can be used for tasks such as
predictive modeling, classification, and clustering.
• Optimization: This methodology involves finding

the best solution to a problem that maximizes or
minimizes a particular objective function, subject to
certain constraints. Optimization can be used for tasks
such as resource allocation, production planning, and
scheduling.
• Simulation: This methodology involves creating a

computer model of a system or process and using it to
generate data and test different scenarios. Simulation
can be used for tasks such as risk assessment,
performance evaluation, and process improvement.
18
• Game theory: This methodology involves modeling

decision-making in competitive situations, such as
auctions, negotiations, and pricing. Game theory can
be used to predict the behavior of other players and
choose the best strategy based on that prediction.
• Bayesian inference: This methodology involves

updating prior beliefs based on new data to make
probabilistic predictions or decisions. Bayesian
inference can be used for tasks such as fraud detection,
customer segmentation, and anomaly detection.
• Multicriteria decision analysis (MCDA): This

methodology involves evaluating alternatives based
on multiple criteria or attributes and weighting each
criterion based on its importance. MCDA can be
used to support complex decisions involving multiple
stakeholders with different preferences.
• Fuzzy logic: This methodology involves dealing with

imprecise or uncertain data by using fuzzy sets and
fuzzy rules. Fuzzy logic can be used to model decision-
making in situations where data is incomplete or vague,
such as in natural language processing.
• Expert systems: This methodology involves encoding

human expertise and knowledge into a computer
system to provide decision support. Expert systems can
be used to diagnose problems, recommend solutions,
and provide explanations for their reasoning.
• Heuristics: This methodology involves using rules

of thumb or shortcuts to make decisions quickly and
efficiently, based on past experience or common
sense. Heuristics can be used to support decisions in
situations where time or resources are limited.
19
These methodologies can be combined and customized to fit specific

decision problems and datasets, providing a more rigorous and structured
approach to decision-making in complex and uncertain environments.
Let us understand how DI systems are used in a real-world scenario
once a model has made its predictions.
Suppose the financial institution from the previous example has used
DI to develop a predictive model to identify which customers are most
likely to use a new credit card product. The model has predicted that
customers who frequently shop at certain retailers are the most likely to
use the new credit card.
• Develop a targeted marketing campaign: Using the

predictions from the model, the financial institution
would develop a targeted marketing campaign to
promote the new credit card to customers who
frequently shop at the identified retailers. This
campaign might include personalized offers, targeted
advertisements, and incentives to encourage customers
to sign up for the new credit card. The financial
institution could use data analytics tools to segment the
identified customers based on their shopping behavior,
demographics, and credit history to tailor the campaign
messaging and offers to their specific needs.
• Monitor the effectiveness of the campaign: The

financial institution would monitor the effectiveness
of the marketing campaign by collecting and analyzing
customer data such as response rates, conversion
rates, and revenue generated by the new credit card
product. By using this data, the institution can refine
the campaign messaging, improve the targeting of the
campaign, and optimize its decision-making processes.
20
• Continuously update the model: As new data

becomes available, the financial institution would
continuously update the predictive model to improve
its accuracy and effectiveness. This might involve
incorporating new data sources such as customer
feedback, social media data, or external economic
indicators. The institution might also adjust the model’s
algorithms, adjust the weighting of different variables,
or use ensemble modeling techniques to improve its
performance.
• Evaluate the impact on the business: The financial

institution would evaluate the impact of the new
credit card product on the business by collecting and
analyzing data such as revenue generated, customer
satisfaction, and retention rates. By using this data, the
institution can adjust its strategy, refine its models, and
optimize its decision-making processes to maximize its
business performance.
By using DI to develop a targeted marketing campaign based on the

predictions of a predictive model, the financial institution can maximize
the effectiveness of its marketing efforts and increase the chances of
success for the new credit card product. By continuously updating the
model and evaluating its impact on the business, the institution can
further optimize its decision-making processes and improve its overall
performance.
21
Some Potential Pros and Cons of DI

Here are some pros of DI:
• Improved decision-making: DI can help businesses

make better decisions by providing them with data-
driven insights that would be difficult to obtain using
traditional decision-making methods.
• Increased efficiency: DI can automate certain

decision-making processes, which can save businesses
time and money.
• Enhanced accuracy: DI algorithms can analyze large

amounts of data and identify patterns that humans may
not be able to see, which can lead to more accurate
predictions and better decision-making.
• Customization: DI can be tailored to fit the specific

needs of a business, allowing for a more personalized
approach to decision-making.
• Improved risk management: DI can help businesses

identify and mitigate risks by analyzing data and
predicting potential outcomes.
• Scalability: DI can scale to handle large amounts of

data and decision-making tasks, which can be difficult
or impossible to do manually.
• Real-time decision-making: DI can provide real-time

insights and decision-making, which can be critical in
time-sensitive situations.
• Competitive advantage: DI can provide businesses
with a competitive advantage by enabling them to
make better decisions faster than their competitors.
22
• Continuous improvement: DI algorithms can

continuously learn and improve over time, allowing
businesses to refine their decision-making processes
and stay ahead of the competition.
• Cost-effective: In the long run, DI can be cost-effective

compared to traditional decision-making methods,
especially when dealing with large amounts of data.
Here are some cons of DI:
• Data quality issues: DI relies heavily on the quality

of the data it uses, and if the data is inaccurate or
incomplete, it can lead to flawed decisions.
• Privacy concerns: DI often involves collecting and

analyzing large amounts of data, which can raise
privacy concerns among customers and stakeholders.
• Bias: DI algorithms can be biased if they are trained on

biased data or if the algorithm itself is biased.
• Overreliance on technology: DI can lead to

overreliance on technology, which can result in a lack
of human input and oversight.
• Complexity: DI can be complex and difficult to
understand, which can make it challenging for
businesses to implement and use effectively.
• Technical expertise: Implementing DI requires

technical expertise and resources that may be lacking
in some businesses.
• Ethical concerns: DI raises ethical concerns related to

privacy, bias, and transparency, which businesses must
address to ensure ethical decision-making.
23
• Dependence on data: DI is dependent on having

enough data to make informed decisions, which may
not always be available.
• Integration challenges: Integrating DI into existing
decision-making processes can be challenging and
require significant effort.
• Adoption barriers: DI adoption may face barriers due
to resistance to change or lack of understanding of its
potential benefits.
E xamples of How Companies Are

Leveraging DI
It is difficult to estimate the number of companies that have successfully
implemented DI, as the adoption of DI is still relatively new and varies
by industry and organization. However, there are several examples
of companies that have successfully implemented DI and are seeing
significant benefits. Here are a few examples:
• Capital One: Capital One, a financial services

company, has implemented DI to improve its credit
decision-making process. By combining machine
learning models with human expertise, Capital One has
been able to increase the accuracy of credit decisions
and reduce the risk of default.
• Boeing: Boeing, an aerospace company, has
implemented DI to optimize its aircraft maintenance
process. By analyzing data from sensors on aircraft
engines, Boeing can predict when maintenance is
required and proactively schedule repairs, reducing
downtime and improving safety.
24
• Procter & Gamble: Procter & Gamble, a consumer

goods company, has implemented DI to improve
its supply chain operations. By analyzing data on
inventory, production, and logistics, Procter & Gamble
can optimize its supply chain to reduce costs and
improve delivery times.
• UPS: UPS, a logistics company, has implemented

DI to optimize its package delivery process. By using
machine learning models to predict package demand
and optimize delivery routes, UPS has been able to
reduce fuel consumption, lower costs, and improve
delivery times.
While these are just a few examples, they demonstrate the potential
of DI to improve decision-making and drive business value. As the
adoption of DI continues to grow, we can expect to see more companies
implementing DI and realizing its benefits.
Conclusion
You now understand what prescriptive AI is, its history, how it works,
its methodologies, its applications, and its pros and cons. In the coming
chapter, we will dig deeper into each of these aspects along with building
an end-to-end DI solution.
25
CHAPTER 2
Requirements
Lots of AI projects will fail across industries because AI projects can be
complex and challenging, and many potential factors can contribute
to their failure. Some of the reasons for failure are lack of clear goals,
data quality issues, and difficulty integrating the AI system into existing
processes. Additionally, there can be organizational challenges such as a
lack of support or buy-in from stakeholders, unrealistic expectations, and a
lack of understanding of the capabilities and limitations of AI.
In this chapter, we are going to discuss how we might reduce the
failures by planning, having clear requirements, and understanding the DI
requirements framework.
Why Do AI Projects Fail?

AI projects can fail for a variety of reasons, including lack of clearly defined
goals, lack of data or poor-quality data, lack of expertise or resources, and
difficulty in integrating the AI system into existing processes or systems.
Additionally, ethical considerations, such as bias in the data or the
algorithms themselves, can contribute to the failure of an AI project.

Chapter 2 Decision Intelligence Requirements
Decision intelligence (DI) is a field that combines artificial intelligence

(AI) and decision-making techniques to help organizations make better
decisions. Decision intelligence requirements refer to the specific needs
and criteria that must be met for a DI system to be effective and valuable to
an organization.
Let’s take a step back and understand some of the questions we need
to ask before we start any AI projects.
• “Can AI perform better than the current processes?”
• “Is the data good for the AI models?”
• “Are the goals/objectives clearly defined for an AI

project?”
• “Is this AI project creating any value?”
• “Did we select the right key performance indicators to

measure the value of the AI projects?”
• “Are AI results being consumed (through taking

action)?”
• “Are business users getting benefits from AI

predictions?”
These are some of the myths surrounding AI:
• Myth 1: The right data scientists and the right

algorithms create the perfect AI.
• Myth 2: To win at AI, you need a lot of data, and you

need the best algorithms.
• Myth 3: The more accurate the AI model, the better it is.
• Myth 4: AI pilots work in one region/domain, and the

same AI models work in enterprise rollout.
• Myth 5: If the AI model works today, it works forever.
28
In answer to the previous questions and to overcome these myths, a

DI requirements framework is essential. A framework acts as a bible to
AI projects because it contains every detail about the project, from the
ideation phase to the operationalize phase.
To deliver value through AI, it’s important to follow some of the
guidelines/paths that can help ensure the successful implementation and
adoption of decision intelligence solutions.
• Clearly define the problem: Before implementing any

AI and decision intelligence solution, it’s important to
have a clear understanding of the problem that needs
to be solved. This means identifying the stakeholders,
their needs and requirements, the scope of the
problem, and the expected outcomes.
• Gather and analyze data: AI solutions rely heavily on

data. Therefore, it’s important to collect and analyze
the required data to ensure that the solution is based
on accurate and reliable information.
• Identify decision-making criteria: AI should be

designed to support decision-making by providing
criteria for evaluating alternatives. These criteria
should be based on the problem definition, stakeholder
requirements, and relevant data.
• Choose the appropriate decision-making method:

Different decision-making methods can be used
depending on the problem and the available data.
It’s important to choose the appropriate method that
will provide the best outcome based on the available
information. This will be discussed in detail in
Chapter 3.
29
• Build the AI model: Once the criteria and method

have been identified, a model can be built. The model
should be transparent, interpretable, and user-friendly.
• Validate the model: The AI model should be validated

to ensure that it’s accurate and reliable. This can
be done through testing, simulation, and real-time
analysis.
• Implement the solution: After the model has been

validated, the solution can be implemented. It’s
important to ensure that the solution is integrated with
existing systems and that stakeholders are trained and
supported to use the solution effectively.
• Monitor and evaluate the solution: AI solutions

should be continuously monitored and evaluated to
ensure that they are delivering the expected value. This
can be done through regular performance checks and
feedback from stakeholders.
Keeping all these guidelines in mind, let’s build a DI requirements

framework that will act as a handbook to build requirement documents
before even starting a project.
DI Requirements Framework

To make AI projects successful, building fancy machine learning and deep
learning models and deploying them won’t do the trick. That’s just the tip
of the iceberg. The real effort needs to be put into planning, defining clear
objectives, building the approach framework, and evaluating the potential
value that could be generated from these projects, as shown in Figure 2-1.
30
Figure 2-1. Hidden steps in the AI life cycle
So, let’s discuss a framework in detail. The first step is proper planning.
Planning
Planning is a crucial aspect of any AI project, as it helps to ensure that the
project is well-defined, is well-organized, and has a clear path to success.
These are some of the specific reasons why planning is important for AI
projects:
• Identify the problem: Start by identifying a specific

problem or opportunity that AI can help to solve
or address. This could be anything from improving
customer service to automating repetitive tasks to
identifying new revenue streams. Lots of AI projects are
force-fitting AI into a business flow even though there is
no problem in the first place.
31
• Define project goals and objectives: Planning helps

to ensure that the project has clear, measurable, and
attainable goals and objectives. This helps to guide the
development of the AI system and ensure that it aligns
with the organization’s overall strategy.
• Identify and address potential risks: Planning

allows teams to identify and assess potential risks that
may impact the project, such as lack of data, lack of
expertise, or difficulty integrating the AI system into
existing processes. Addressing these risks early on can
help to mitigate them and ensure the project’s success.
• Build the right team: Assemble a team with the right

mix of skills, including data scientists, engineers,
and domain experts. This will help to ensure that the
project is executed effectively and that the AI system is
tailored to the specific needs of the organization.
• Manage expectations: Planning helps to manage

expectations by clearly communicating the project’s
goals, objectives, timeline, and budget to stakeholders.
This helps to ensure that everyone is on the same page
and that the project is delivered as expected.
• Identify and address ethical considerations:

Planning allows teams to identify and address ethical
considerations, such as the potential for bias in the data
or the algorithms, that may impact the project.
• Understand the data: Make sure you have access to the

necessary data and that it is of high quality. Understand
how the data is collected, how it is stored, and how it
can be used to train and operate the AI system.
32
• Measure the results: Plan and define the right KPIs

and metrics to measure the results of the AI project.
This will help to determine the system’s effectiveness
and avoid confusion later.
• Communicate the value: Communicate the value of

the AI project to stakeholders, both inside and outside
of the organization. This will help to build support and
ensure that the project is seen as a success.
Overall, planning is important for AI projects because it helps to

ensure that the project is well-defined, is well-organized, and has a clear
path to success.
Approach
Approaching an AI project requires careful consideration of several key
factors including the type of problem being addressed, the available data,
the computing resources available, and the deployment requirements.
Here are some general guidelines for approaching an AI project:
• Model type: Depending on the problem you are trying

to solve, choose an appropriate machine learning or
deep learning model. There are many different models
to choose from, such as neural network–based models,
decision trees, and support vector machines. Your
choice of model will depend on the specific problem
you are trying to solve, as well as the amount and
quality of your data.
• Classification: Use classification when you have

labeled data and want to predict a categorical
outcome. This is often used in scenarios where
you want to identify which class or category a new
33
data point belongs to. For example, you might use

classification to predict whether a customer will
buy a product or not based on their demographic
information.
• Regression: Use regression when you have labeled

data and want to predict a continuous outcome.
This is often used in scenarios where you want
to predict a numerical value, such as the price of
a house or the amount of rainfall in a particular
region. Regression models are useful when you
want to make predictions that are not limited to a
discrete set of values.
• Unsupervised learning: Use unsupervised

learning when you have unlabeled data and want
to identify patterns or groupings in the data. This
is often used in scenarios where you want to
discover hidden relationships or structures in your
data, without having any prior knowledge about
what you are looking for. Clustering is a common
unsupervised learning technique that can be used
to group similar data points based on their features.
It is also important to consider the size and

quality of your data when deciding between these
different types of learning. If you have a small
dataset with limited features, you may want to use
a simpler model like logistic regression or k-nearest
neighbors. On the other hand, if you have a large
dataset with complex features, you may want to
use a more sophisticated model like a deep neural
network.
34
Ultimately, the choice of learning algorithm will depend

on the specifics of your project and your data.
• Framework: Select a framework for your AI project

based on your choice of model. There are many
popular frameworks available for building machine
learning and deep learning models, including
TensorFlow, PyTorch, and Scikit-learn. Choose a
framework that is well-suited to your model and that
has the necessary features and functionality you need.
• Deployment details: Deploying an AI model is an

important part of the overall project and should be
planned ahead of time to ensure that the model can be
integrated smoothly into a production environment.
Here are some steps you can take to plan the
deployment of your AI model:
• Identify the deployment requirements: Begin by

identifying the requirements for deploying your
model in a production environment. Consider
factors such as the hardware and software
infrastructure needed to run the model, the data
inputs and outputs, and any security or compliance
considerations.
• Select a deployment strategy: Once you have

identified the deployment requirements, choose
a deployment strategy that is appropriate for your
model and project goals. Some common strategies
include deploying the model as a web service,
integrating it into an existing application, or
deploying it to an edge device.
35
• Choose a deployment platform: Depending on

your deployment strategy, choose a deployment
platform that is well-suited to your model and
project goals. There are many platforms available
for deploying AI models, including cloud platforms
such as AWS, Azure, and Google Cloud, as well
as edge devices such as the Raspberry Pi and
Nvidia Jetson.
With a proper approach before even implementation, we can ensure

that the AI projects will deliver value over the long term and don’t end up
as toy projects.
Approval Mechanism/
Organization Alignment
Establishing an approval mechanism for AI projects is an important step to
ensure that the projects are aligned with organizational goals and values.
Here are some steps to follow to establish an approval mechanism for AI
projects:
1. Identify stakeholders: Start by identifying the
stakeholders who should be involved in the
approval process. This may include executive
leadership, legal counsel, data privacy and
security experts, and representatives from affected
business units.
2. Establish evaluation criteria: Define the criteria

that will be used to evaluate AI projects. This may
include factors such as data quality, the potential
impact on stakeholders, the value generated, and
alignment with organizational goals.
36
3. Develop an approval process: Create a formal

approval process that outlines the steps and
stakeholders involved in the decision-making
process. This may include submitting a project
proposal, evaluating the proposal against the
established criteria, and making a final decision.
4. Implement a review board: Establish a review

board that is responsible for reviewing and
approving or rejecting AI projects. The review board
should include representation from the identified
stakeholders and be responsible for ensuring that
the project meets the established evaluation criteria.
5. Conduct a risk assessment: Before approving an AI

project, conduct a thorough risk assessment to identify
potential risks and develop mitigation strategies.
This may involve assessing the potential impact on
stakeholders, identifying potential biases in the data, and
evaluating the accuracy and reliability of the AI model.
By establishing an approval mechanism for AI projects, you can ensure

that AI projects are evaluated responsibly and ethically and that they align
with organizational goals and values.
Key Performance Indicators

Measuring the results of an AI project is crucial for understanding its
effectiveness and identifying areas for improvement. By using clear
metrics, collecting and analyzing data, comparing data to baselines,
identifying areas for improvement, and communicating the results,
organizations can ensure that the AI system is aligned with the project’s
goals and objectives and continues to generate value over time.
37
Define Clear Metrics

Define clear and specific metrics that will be used to measure the results
of the AI project. These metrics should align with the project’s goals and
objectives and be easy to understand and communicate.
Some examples of metrics that can be used to measure the results
of an AI project include accurate churn rate for churn prediction model,
customer satisfaction score, average delivery time for route optimization
algorithm, ROI, etc.
Collecting the right key performance indicators (KPIs) is an important
step in measuring the results of an AI project. KPIs are specific metrics that
are used to measure the performance of a system or process, and they help
to provide a clear and objective way to measure the results of an AI project.
When selecting KPIs for an AI project, it’s important to ensure that
they align with the project’s goals and objectives and that they are easy
to understand and communicate. It’s also important to consider the data
that will be required to measure the KPIs and make sure that the data is
available and of high quality.
It’s crucial to do the following:
• Track business KPI performance against benchmarks

over time.
• Analyze business KPIs versus model KPIs for
insights into how model performance is affecting the
bottom line.
• Calculate realized value (in dollars) and the ROI.
Obtaining business KPIs from stakeholders is an important step in

measuring the results of an AI project. Here are a few steps for getting
these KPIs from stakeholders:
38
1. Identify the stakeholders: Identify the key

stakeholders in the organization, including those
who will be affected by the AI project, those
who will be using the system, and those who
will be responsible for its implementation and
maintenance.
2. Communicate the goals and objectives: Clearly

communicate the goals and objectives of the AI
project to stakeholders. This will help to ensure that
everyone is on the same page and that the project is
aligned with the organization’s overall strategy.
3. Involve stakeholders in the planning process:

Involve stakeholders in the planning process, such
as identifying the problem, understanding the data,
and defining the use case. This will help to ensure
that the project is tailored to the specific needs of
the organization and that stakeholders are invested
in its success.
4. Identify the KPIs: Work with stakeholders to

identify the specific KPIs that will be used to
measure the results of the AI project. This will help
to ensure that the KPIs align with the project’s goals
and objectives and are relevant to the stakeholders.
5. Communicate the process: Communicate the

process for collecting and analyzing data, as well
as the schedule for measuring the results. This will
help to ensure that stakeholders understand the
process and can provide the necessary data and
support.
39
6. Measure the KPIs: Measure the KPIs for AI projects

as planned once models are in production and
perform impact analysis based on controllable
factors. Evaluate if AI is truly adding value and
helping to make appropriate decisions.
7. Keep stakeholders informed: Keep stakeholders
informed of the results of the AI project,
including the performance of the system, areas
for improvement, and any changes that have
been made.
By following these strategies, organizations can obtain the necessary

business KPIs from stakeholders and ensure that the AI project is aligned
with the organization’s goals and objectives.
Value
One of the biggest challenges that organizations face is to quantify the
benefits of their AI initiatives. Value realization involves effort, and until
one knows what benefits they are getting through their AI initiatives, they
cannot improve them. No wonder most of the AI use cases fail to get any
adoption at all due to the complexities in realizing the business value of
the use cases.
As shown in Figure 2-2, processing data, building models with good
accuracy, and deploying the models seems easier than tracking and
realizing the value of AI. It seems difficult because we have come from a
value standpoint when building AI projects and have never tracked the
KPIs and return on investment.
40
Figure 2-2. Data to value
Let’s discuss what return on investment is and how to calculate it for

an AI project.
Return on Investment
The return on investment (ROI) of the project should consider the costs
of implementing the project and the potential value it will bring to the
business.
The ROI for an AI project refers to the expected or realized financial
return on investment for implementing an AI solution or initiative. It is
determined by comparing the cost of implementing the project, including
the cost of data acquisition, hardware and software, and personnel, to the
expected or realized benefits.
The benefits of an AI project can be difficult to quantify but may
include increased efficiency, improved accuracy, better decision-making,
and reduced labor costs. The ROI for an AI project can be calculated by
41
dividing the net financial benefits, such as increased revenue or reduced

costs, by the total cost of the project and then multiplying the result by 100
to get a percentage.
But how do we calculate the potential value? Let’s discuss that.
Value per Decision

Calculate the value per decision by estimating the potential impact of each
decision made by the AI project.
Example 1:
Let’s take the example of churn prediction again. If we successfully
predict a customer is going to leave your business, the user can take some
actions to retain the customer. So, the value from one customer would be X
dollars.
In this scenario,
Value from one prediction = One generated from one customer
Example 2:
In some cases, we might want to know the overall value generated, not
at prediction level.
The value per prediction can be calculated by dividing the total value
of the project by the number of decisions it will make.
By considering these factors, you can derive the potential value from AI
projects or calculate the value per decision.
Keep in mind that the actual value may vary based on various other
factors, such as the following:
• Probabilistic or uncertain outcomes: Since the

predictions are probabilistic, every prediction won’t
give us 100 percent value.
• Cost of errors: It’s the cost of errors made by

the model.
42
So, the overall ROI is as follows:
ROI = Return / Investment
where Return = Benefits from the model – Uncertainty of the

predictions
Investments = Resources * Cost of resources
Keep in mind that deriving the dollar value from an AI prediction can
be challenging, especially when the benefits are not directly financial.
However, by identifying the problem, defining the use case, identifying
the costs and benefits, calculating the ROI, and communicating the value,
organizations can get a better understanding of the financial impact of the
AI system.
Consumption of the AI Predictions

A major part of the AI life cycle is making sure the business consumes
the AI predictions. Otherwise, the value won’t be generated. So, while
planning, it’s crucial to understand how the business is going to consume
the outputs. The following are the steps to be followed to have a clear
understanding of this:
1. Understand the prediction: Before consuming an
AI prediction, it’s important to understand what
it means and how it was generated. This includes
understanding the input data, the algorithm used,
the model’s accuracy and limitations, and any
assumptions made during the modeling process.
2. Determine how to use the prediction: Once you

understand the prediction, determine how it can be
used to make a decision or take action. This could
involve using the prediction as input to another
43
system or using it to trigger an alert or notification. It

should also include the following:
• Explainability: The system must be able to provide

clear and transparent explanations for its decisions
and recommendations, which helps to build trust
and understanding among stakeholders.
• Scalability: The system must be able to handle

large amounts of data and handle the complexity of
the problem and adapt to changing conditions.
• Flexibility: The system should be able to handle

different types of problems and be used in different
areas of the organization.
• Integration: The system should be able to integrate

with existing systems and processes to ensure a
smooth and efficient workflow.
3. Consider the confidence level: AI models typically

provide a level of confidence or probability
associated with a prediction. It’s important to
consider this confidence level when deciding how
to use the prediction. For example, if the confidence
level is low, it may be necessary to verify the
prediction with other data or methods before acting.
4. Determine the appropriate storage solution:

Depending on the type of AI prediction and the
use case, you may need to use different storage
solutions. For example, if the prediction is an image
or video, blob storage may be appropriate, while for
text data, API calls may be more suitable.
44
5. Set up the storage solution: Once you have

determined the appropriate storage solution, you
will need to set it up and configure it appropriately.
This may involve creating storage containers,
configuring access policies and permissions, or
integrating with other systems.
6. Establish data flows: To consume AI predictions,

you will need to establish data flows between the AI
model and the storage solution. This may involve
setting up event triggers, establishing data transfer
protocols, or configuring API endpoints.
7. Ensure data security: When working with sensitive

data, it’s important to ensure that appropriate
security measures are in place. This may involve
encrypting data in transit and at rest, configuring
access controls and permissions, or integrating with
other security systems.
8. Identify user groups: Once the problem has been

defined, identify the different user groups who
will be using the AI predictions. This may include
different departments, teams, or individuals.
9. Understand user needs: For each user group, it’s

important to understand their specific needs and
requirements for using the AI predictions. This
includes understanding what decisions they need to
make, what data they need to access, and what level
of detail and accuracy they require.
10. Provide training and support: Once the AI

predictions have been made available to users, it’s
important to provide training and support to ensure
45
that users can effectively consume and act on the

predictions. This may involve providing user guides,
training sessions, or support resources.
By following these steps, you can identify users for AI predictions and
ensure that they are provided with the information and resources they
need to effectively use the predictions to support decision-making and
improve outcomes.
Conclusion
In this chapter, we discussed the potential reasons for AI project failure,
the importance of decision intelligence in making better decisions using
AI, and the guidelines for the successful implementation of AI solutions.
It highlighted the need to ask critical questions before starting any AI
projects and debunked some of the myths surrounding AI. We also
introduced a DI requirements framework that covers planning, approach,
approval mechanism, KPIs, potential value, and consumption of AI
predictions, which are fundamental to AI projects.
46
CHAPTER 3
Methodologies
The idea of decision-making is the main topic of this chapter, along
with its types. We will begin with the definition of decision-making and
examine its various forms. Then, we will understand the decision-making
process through an example. We will next go into greater detail about
the development of decision-making, also known as decision intelligence
methodologies, employed by humans over time, as well as their benefits
and drawbacks. Our discussion will end with suggestions for decision
intelligence approaches for various scenarios.
Decision-Making
To get to know the different decision intelligence methodologies, you have
to understand what decision-making is.
Decision-making can be defined as the mental process of choosing
among different options, culminating in the selection of a belief or action.
Thus, decision-making consists of two major components.
• An objective function
• Alternatives/options available to reach the objective

Chapter 3 Decision Intelligence Methodologies
The objective function is, in simple terms, the goal to achieve.

Objectives can vary in complexity, from something as simple as selecting
an outfit for a party to something as complex as finding a cure for cancer.
The objective function typically has multiple ways to achieve it, known as
alternatives or options, which must be evaluated to arrive at a decision.
The decision-making process involves evaluating these alternatives to
achieve the given objective. For example, when choosing an outfit for a
party, a gentleman might have two options: a blue tuxedo and a black
tuxedo. He might consider a variety of factors such as personal preference,
what looks good, the theme of the party, etc., to make his decision, or he
might randomly pick something without giving much thought to it. The
thought process behind the decision can be complex or simple, and the
decision-maker’s reasoning, whether rational or irrational, is an important
part of the process and impacts the outcome.
Decision-making can be classified in many ways, depending upon
the factors taken into consideration such as single- and multiple-criterion
decision-making, the number of people involved (individual and group
decision-making), and the level of decision-making (strategic, tactical, and
operational decision-making).
Types of Decision-Making
There are basically three types of decision-making. Let’s examine each
in detail.
Individual vs. Group Decision-Making

As the name suggests, individual decision-making is something that is the
responsibility of a single person. This type of decision-making is mostly
seen in the personal lives of individuals wherein they are responsible for
making their own life decisions. It is a type of problem-solving that is used
to identify and choose a course of action from among several alternatives.
48
In individual decision-making, the person is solely responsible for the

outcome of the decision, and the decision is based on their own personal
values, beliefs, and experiences.
Group decision-making, on the other hand, refers to the process of
making a decision with a group of individuals. This type of decision-
making can be used in various settings, such as in the workplace, in
government, or in social organizations. Group decision-making is a way
of problem-solving that is used to identify and choose a course of action
from among several alternatives. The group members work together to
identify and evaluate the different options and then make a decision that
is collectively agreed upon. Group decision-making can be more effective
than individual decision-making because it allows for the sharing of
diverse perspectives and ideas and can increase the likelihood of better
decisions being made.
Single- vs. Multiple-Criterion Decision-Making

Single-criterion decision-making (SCDM) is a decision-making approach
where only one criterion is used to evaluate alternatives. In this type, the
decision-maker selects the alternative that maximizes the value of the
chosen criterion. This approach is useful when the decision is based on a
single objective, such as maximizing profit/revenue or minimizing cost/
time. Examples of SCDM include choosing an apartment for rent based on
the price or buying shoes based on the brand.
Multiple-criterion decision-making (MCDM) is a decision-making
approach where, unlike SCDM, multiple criteria are used to evaluate
alternatives. This type of decision-making involves a trade-off between
different criteria and often requires the decision-maker to assign weights
to each criterion to reflect their relative importance. MCDM is useful
when the decision involves multiple objectives, such as, for the previous
example, choosing an apartment for rent not only based on the price but
also based on location, size, view, distance to public transportation, etc.
49
One of the main advantages of MCDM over SCDM is that it provides

a more comprehensive evaluation of alternatives. In MCDM, decision-
makers can consider a range of factors and evaluate trade-offs between
them. Additionally, MCDM can help to ensure that other important criteria
are not overlooked, which can happen in SCDM when only one criterion
is used to make the decision and the rest of them are ignored. However,
MCDM can also be more complex and time-consuming than SCDM,
and it requires more information and analysis. Additionally, assigning
weights to criteria can be difficult, and the results may be sensitive to the
weights chosen.
In summary, SCDM is appropriate when the decision is based on a
single objective, while MCDM is more appropriate when the decision
involves multiple objectives that need to be traded off against each other.
Strategic, Tactical, and Operational Decision-Making

Strategic, tactical, and operational decision-making is commonly observed
in organizations. Figure 3-1 is a high-level overview of the decision-making
process under strategic, tactical, and operational conditions. It explains
what questions these types of decision-making answer, their execution
time frame, and their frequency. Let’s look at each one of them in detail.
Figure 3-1. Strategic, tactical, and operational decision-making
50
Strategic decision-making is the process of determining the long-term

goals and objectives of an individual/organization and the actions needed
to achieve them. In an organizational setting, these types of decisions are
usually made by top-level management and typically involve significant
investments and changes to the organization’s overall strategy and
structure. Hence, they usually are nonroutine decisions. These types of
decisions are usually slow moving, and changes happen over a long period
of time.
Tactical decision-making involves the implementation and execution
of medium-term plans and objectives. In an organizational setting, these
types of decisions are usually formulated by middle-level management.
These types of decisions often involve the management of day-to-day
operations and the allocation of resources to achieve medium-term
goals that align with the organization’s overall strategy. These types of
decisions are a mix of routine and nonroutine tasks and move faster than
the strategic ones but are slower than that of the operational ones and are
planned for medium to long term.
Operational decision-making is the process of making decisions
related to the day-to-day management and operations of an individual/
organization. These decisions are usually made by front-line managers
and employees and include scheduling, production, and other activities
that are necessary to maintain the smooth functioning of the organization.
These types of decisions are the fastest moving ones and are planned for
the short term and involve routine tasks.
Decision-Making Process
As per the research by Baker et al. (2001), before making a decision, it
is important to identify the individuals or groups that are responsible
for making the decision and those who will be affected by it. This
helps to reduce any possible disagreement about problem definition,
51
requirements, goals, and criteria. Once the individual/group identification

is done, it is recommended to follow a structured approach for decision-
making, as shown in Figure 3-2.
Figure 3-2. Decision-making process
Step 1: Problem Definition

This is the most critical and often ignored step in the decision-making
process. Usually, framing the problem statement is a more iterative and
time-consuming process than the solution itself. It requires input from
the stakeholders, understanding the environment under which a decision
has to be made, and what will happen if the problem is left as is, i.e., no
decision is made. Once the problem definition is reviewed and agreed
on by all will this step be considered done, and one can move to the
subsequent steps.
The problem definition stage also requires the “desirable” state to be
defined, i.e., what the ideal solution to the problem looks like once the
decision is made.
A wrong problem definition will have wrong decisions made, and the
actual problem will still remain the same, causing negative effects on the
individuals/groups.
52
Step 2: Listing Constraints

Decision-making is, almost all the time, never done without constraints.
These constraints can be financial, ethical/moral, environmental, time
related, etc. The constraints help to filter out the alternatives and bring them
from a possibly infinite number to a finite number. Listing all the constraints
for a given problem is as important as the problem definition. Wrong
constraints can lead to no or wrong decisions.
The constraints can be categorized into hard and soft ones. The
hard constraints are the ones that are bound to be fulfilled and are non-
negotiable in any condition. These constraints cannot be overlooked. The
soft constraints, on the other hand, are the ones that are not “compulsory”
but “good to have.” They add more value to the decision-making but do not
act as a blocker, if not fulfilled.
Step 3: Possible Decisions and Their Effects Evaluation
Once the problem is defined and constraints are listed, one needs
to evaluate the possible decisions/solutions and their closeness to the
desired state mentioned in step 1. Listing the pros and cons of every
possible decision/solution within the scope of the constraints is a good
practice to choose the right one.
Step 4: Decision Execution
Once all the possible solutions/alternatives are evaluated and the
best one is chosen, it is time to execute the decision under the given
constraints.
Step 5: Outcome Evaluation and Process Update
This step enables you to monitor and compare the outcome of the
decision made to that of the expected one shown in step 3. This helps to
re-evaluate the whole process and change the problem, constraints, and
decisions if required.
These steps continue in a cyclic fashion until the right decision with
the right outcome emerges. Let’s go through the decision-making process
with an example.
53
Decision-Making Process Example

In this example, we will go through the decision-making process of a
person wanting to purchase a car. We will see how the person navigates
the process, as listed in Figure 3-2, to arrive at the decision of which car to
purchase based on their needs.
Step 1: Problem Definition
Amit, a 25-year-old unmarried software professional from Bangalore,
is planning to purchase a car. It will be his first car, so he is super excited
about it. He wants to ensure he does all the due diligence in the process to
purchase a car so that he doesn’t regret it later.
Step 2: Listing Constraints
Here are the hard constraints:
• Car condition: Brand new (not considering used cars)
• Budget: INR 8-10 lakhs ($9.6K–$12K)
• Size: Hatchback, 13.7 feet long (164.4 inches) × 6

feet wide (72 inches) × 5.4 feet tall (64.8 inches),
comfortable for four people
• Safety Ratings: 4 or 5 (NCAP)
• Purpose: Travel to work and for short trips with family

during weekends
• Engine: Diesel (manual), for better mileage
• Boot space: Minimum 250 liters
• Waiting Period: 1 to 3 months
• After sales service charges: Economical (INR 25–35K

per year)
• Authorized service centers: Maximum coverage, even
in remote areas
54
• Comfort: Comfortable and smooth driving quality
Here are the soft constraints:
• Premium interiors
• Cheap spare parts cost
• Wireless charger
• Top-notch music system
• Touchscreen display
• Alloy wheels
• Wide variety of color options
• Hill assist feature
• Good resale value
Step 3: Possible Decisions and Their Effects Evaluation

As this is the first time Amit is buying a car, he reaches out to his family
and close friends, relatives, etc., who have a hatchback or have had one
in the past. He wants to ensure at least the after service and authorized
service center details are known through their experience. He also does
secondary research with details available on social media, brand websites,
forums, etc., that extend his scope of information beyond family/peers and
validate their claims. He prepares a list of the constraints and the values/
parameters according to the requirements (as shown in Figure 3-3).
He also visits the showrooms of the brands to test-drive the vehicles and
find out details such as drive quality, comfort, and waiting period.
55
Figure 3-3. Possible decisions and their effects evaluation
Step 4: Decision Execution

From Table 3-3, it seems like Brand A and Brand C are the closest in
terms of what Amit needs. Brand A is a little costlier than Brand C, has less
trunk space, has a longer waiting period, and has an average maintenance
cost. Brand C’s safety rating is lesser than that of Brand A, with slightly
less mileage, fewer authorized service centers, fewer color options, and
less resale value. After doing the cost-benefit analysis and immediate
requirements, Amit goes ahead and buys Brand C. The price and waiting
period played a bigger role in his decision-making than some of the other
factors.
Step 5: Outcome Evaluation and Process Update
Ideally, for a once-in-a-lifetime type of decision, this step is not
required as the decision-maker often makes the decision only after
thorough research. However, Amit might face issues after buying the
car, such as loan issues, hidden charges by the showroom, a sudden
extension of the waiting period, etc. So, Amit might cancel his purchase
and go for Brand A or re-evaluate all other options. Amit could also face
issues post-delivery of the car, such as lower mileage than claimed, faulty
56
car components, higher than expected maintenance cost, etc., which

could force him to sell his car and consider buying a new one or look into
used cars.
Decision-Making Methodologies
Humans have evolved over the years, and along with them, the way they
make the decisions has also seen lots of changes. From making decisions
as an individual randomly to group decision-makers who are rational
to combining random and process-oriented and individual and group
decision-makers, the decision-making process has been constantly
increasing in its complexity.
The decision-making methodologies can be broadly divided into three
major categories, human-only, human-machine, and machine-only, as
shown in Figure 3-4.
57
Figure 3-4. Decision-making methodologies
This type of categorization is purely based on the share of decision-

making between humans and machines.
Let’s take a closer look at decision-making methodologies in detail.
Human-Only Decision-Making
Human-only decision-making refers to the process of making choices
or decisions solely on the judgment of human beings, without the
involvement of artificial intelligence, machine learning, or any automated
decision-making systems.
Before the advent of machines (from the abacus to computers and AI
robots), humans have been making decisions either individually or in a
group for personal, social, or organizational needs. The type of decisions
that humans make vary depending upon the situation at hand, criticality of
58
the decision, mental state of the people making the decision etc., ranging
from totally random to outcome based to a more systematic complex
decision-making.
The history of human decision-making can be traced back to the earliest
civilizations, where decisions were made by leaders, elders, or individuals
in positions of authority. Over time, the development of democratic
institutions, laws, and norms has led to an increased emphasis on individual
autonomy and the ability of individuals to make decisions for themselves.
Let’s further evaluate the different types of human-only decision-
making techniques.
Random Decisions
Random decision-making refers to the process of selecting an option or
making a choice randomly, rather than relying on reasoning, intuition,
or personal preference. This type of decision-making technique comes
naturally to us and has been used in situations where the stakes are not
high, there is an urgency to decide, there is no historical data to support
the decision-making process, or all or any combination of these situations.
Examples of random decision-making are flipping a coin or rolling a die.
Here are the advantages of random decisions:
• Removes confirmation bias: By relying on chance
instead of personal preference, random decision-
making can help to remove personal bias and increase
objectivity.
• Increases creativity: Random selection can lead to

unexpected and creative solutions.
• Enhances exploration: By forcing individuals to

consider options they may not have otherwise
considered, random decision-making can encourage
exploration and increase the chances of discovering
new and innovative solutions.
59
These are the disadvantages:
• Lack of control: Random decision-making leaves

little room for control or direction, which can lead to
suboptimal or even negative outcomes.
• Inefficient: It can be time-consuming to generate

enough random outcomes to make an informed
decision.
• Unpredictable: The outcomes of random decision-

making can be unpredictable and difficult to control,
leading to a lack of confidence in the decision-making
process.
Morality/Ethics Based
Ethics-based or moral-based decision-making refers to the process of
making choices and taking actions that align with a set of moral or ethical
principles and values. The outcomes of these type of decisions may or may
not be favorable to the decision-maker; however, they are of the highest
ethics. The opposite of this is unethical decision-making. The decision-
maker deliberately makes a decision that is not ethical and is most of the
time favorable to them. There are several types of ethics-based decision-
making, including the following:
• Utilitarianism: This is a theory of morality that

advocates actions that foster happiness or pleasure and
oppose actions that cause unhappiness or harm. An
example of utilitarianism would be to choose playing
with your kid over completing an office report, as you
find more happiness in spending time with your kid.
60
• Deontology: This approach focuses on following a

set of moral duties and obligations, regardless of the
outcomes. A classic example of deontology would be
“Don’t lie. Don’t steal. Don’t cheat.”
• Virtue ethics: This approach emphasizes the character

and habits of the decision-maker, rather than rules
or consequences, in determining the morality of a
decision. An example of virtue ethics can be a customer
going back to a store to settle the payment for an item
that was unintentionally left unpaid or notifying the
cashier if they issued a refund for an amount greater
than what was owed.
Here are the advantages:
• Consistency: By following a set of ethical principles,

individuals can make decisions that are consistent with
their values and beliefs.
• Increased trust: Ethical decision-making can increase

trust in organizations and relationships, as individuals
perceive decisions as fair and just.
• Personal fulfillment: Making decisions that align
with personal values and beliefs can lead to a sense of
personal fulfillment and satisfaction.
Here are the disadvantages:
• Subjectivity: Ethics and morality can be subjective and

open to interpretation, leading to disagreements and
conflicts about the correct course of action.
• Time-consuming: Making decisions based on ethics

can be time-consuming and require extensive research
and consideration.
61
• Potential for negative consequences: Ethics-based

decision-making may prioritize moral principles over
practical considerations, leading to decisions with
negative outcomes.
Experience Based
Experience-based decision-making refers to the process of making choices
and making actions based on past experiences, knowledge, and expertise.
There are several types of experience-based decision-making, including
the following:
• Rule-based: This involves following established

protocols, procedures, and guidelines based on past
experiences.
• Intuition-based: This involves relying on gut feelings

or instinct, based on past experiences and knowledge,
to make decisions.
• Evidence-based: This involves making decisions based

on data, research, and past experiences to inform and
support the decision-making process.
• Speed: Experience-based decision-making can be

quicker, as individuals can rely on past experiences to
inform their choices.
• Improved accuracy: Experience-based decision-

making can lead to more accurate outcomes, as
individuals can draw on past experiences to inform
their choices.
62
• Personalized: Experience-based decision-making

can be personalized, as individuals can tailor their
decisions based on their own experiences and
knowledge.
• Bias: Past experiences can introduce bias into the

decision-making process, leading to suboptimal or
even negative outcomes.
• Lack of innovation: Relying solely on past experiences

can limit the scope for innovation and limit the
exploration of new solutions.
• Inflexibility: Experience-based decision-making

can lead to inflexibility and a resistance to change,
as individuals may be unwilling to deviate from
established protocols or approaches.
Authority Based
Authority-based decision-making, aka “just follow orders” decision-
making, is a process in which decisions are made by individuals in
positions of authority, and subordinates are expected to carry out those
decisions without question. This approach values obedience and hierarchy
and is often used in situations where time is limited or immediate action
is required. In such scenarios, the responsibility for decision-making is
centralized, and individuals lower in the hierarchy are expected to carry
out the decisions of their superiors. This approach can lead to efficient
implementation of decisions, but it can also discourage initiative,
creativity, and critical thinking and can create a lack of accountability
for the outcomes of decisions. There are several types of authority-based
decision-making, including the following:
63
• Hierarchical: This involves making decisions based on

the chain of command within an organization or group.
• Legal: This involves making decisions based on laws,

regulations, and legal mandates.
• Expert-based: This involves making decisions

based on the advice or guidance of individuals with
specialized knowledge or expertise.
• Efficiency: Authority-based decision-making can be

efficient, as individuals can rely on the guidance of
higher authorities or experts to make decisions.
• Clarity: Authority-based decision-making can provide

clarity, as individuals have a clear understanding of
who is responsible for making decisions and what their
role is.
• Reduced responsibility: Authority-based decision-

making can reduce personal responsibility, as
individuals can defer to the decisions of higher
authorities or experts.
• Lack of autonomy: Authority-based decision-

making can limit personal autonomy and creativity,
as individuals are limited by the decisions of higher
authorities.
• Potential for abuse: Authority-based decision-making

can lead to abuse of power, as individuals in positions
of authority may abuse their power to make decisions
that serve their own interests.
64
• Inflexibility: Authority-based decision-making can

lead to inflexibility, as individuals may be limited in
their ability to adapt to changing circumstances and
make decisions that are best for the situation.
Consensus Based
Consensus-based decision-making is a group decision-making process
in which all members of a group participate in reaching a decision
that everyone can support. This approach values collaboration,
communication, and agreement, rather than relying on a single person or
a majority vote. Participants work together to identify and explore options
and seek to find a solution that is satisfactory to all. The goal of consensus
decision-making is to find a solution that addresses the concerns and
needs of all parties involved, leading to more committed and effective
implementation of the decision. There are several types of consensus-
based decision-making, including the following:
• Unanimity: This involves making decisions based on

the agreement of all individuals in a group.
• Collaborative: This involves making decisions

through a collaborative process, in which individuals
work together to reach a shared understanding and
agreement.
These are the advantages:
• Inclusiveness: Consensus-based decision-making can

promote inclusiveness, as all individuals in a group
have a voice and a role in the decision-making process.
• Improved outcomes: Consensus-based decision-

making can lead to improved outcomes, as individuals
can draw on the diverse perspectives and experiences
of the group to inform their choices.
65
• Increased commitment: Consensus-based decision-

making can increase commitment to decisions, as
individuals are more likely to support and implement
decisions that they have helped to shape.
• Time-consuming: Consensus-based decision-making

can be time-consuming, as individuals may need to
engage in extensive discussions and negotiations to
reach a shared agreement.
• Potential for groupthink: Consensus-based decision-

making can lead to groupthink, in which individuals
conform to the opinions of the group and suppress
dissenting views.
• Difficulty reaching agreement: Consensus-based

decision-making can be difficult, as individuals may
have differing opinions, values, and experiences that
make it challenging to reach a shared agreement.
Voting Based
Voting-based decision-making is a process in which a group of people
cast ballots to determine a course of action. The decision is based on a
tally of the votes, with a majority or supermajority needed to determine
the outcome. This method is often used when time is limited or when
a consensus cannot be reached. Voting can take various forms, such as
a simple majority, a two-thirds majority, or a weighted vote based on
individual member’s stake or influence.
This approach values efficiency and fairness and ensures that the
decision reflects the will of the group. However, it can also lead to minority
views being ignored or overlooked. There are several types of voting-based
66
• Simple majority: This involves making decisions based

on the will of the majority of individuals in a group.
• Supermajority: This involves making decisions

based on the agreement of a specified percentage of
individuals in a group.
• Plurality: This involves making decisions based on the

candidate or option with the most votes, even if they do
not have a majority.
• Fairness: Voting-based decision-making can promote

fairness, as all individuals in a group have an equal
opportunity to influence the outcome.
• Democracy: Voting-based decision-making is a

cornerstone of democratic systems, as it allows for the
will of the people to be expressed and implemented.
• Transparency: Voting-based decision-making can

promote transparency, as the results of a vote are clear
and easily understood.
• Limited representation: Voting-based decision-

making may not accurately represent the views of all
individuals in a group, as some may not participate
in the vote or may have their views diluted by the
majority.
• Biased outcomes: Voting-based decision-making

can lead to biased outcomes, as the results may be
influenced by factors such as voter turnout or unequal
representation.
67
• Lack of accountability: Voting-based decision-making

can lead to a lack of accountability, as individuals may
be less likely to take responsibility for the outcome of a
vote or election.
Threshold Based
This type of decision-making takes certain thresholds into consideration
for given activities to perform an action. This happens when the outcome
sample space is very limited and there is a lot of historical information
on the decisions taken based on the outcome. There are several types of
threshold-based decision-making, including the following:
• Quorum: This involves making decisions based on the

minimum number of individuals who must be present
or represented in a group to make decisions.
• Approval threshold: This involves making decisions

based on a minimum level of approval or support that
must be met before a decision can be taken.
• Performance threshold: This involves making

decisions based on a minimum level of performance
or achievement that must be met before a decision can
be taken.
• Objectivity: Threshold-based decision-making can

provide objective standards for making decisions, as it
is based on specific, measurable criteria.
• Transparency: Threshold-based decision-making

can promote transparency, as the criteria for making
decisions are clear and easily understood.
68
• Consistency: Threshold-based decision-making can

lead to consistency in decision-making, as individuals
are held to the same standards and expectations.
• Rigidity: Threshold-based decision-making can be

rigid and inflexible, as it may not allow for consideration
of contextual factors or individual circumstances.
• Lack of creativity: Threshold-based decision-making

can limit creativity and innovation, as individuals may
focus solely on meeting the minimum criteria rather
than exploring new and innovative solutions.
• Inadequate representation: Threshold-based

decision-making may not accurately represent the
views and needs of all individuals in a group, as some
may not meet the minimum criteria or may have their
views diluted by others who do.
First Acceptable Match Based

First acceptable match, aka satisficing, decision-making is a process
in which individuals choose the first option that meets a minimum
threshold of acceptability, rather than searching for the best possible
solution. This approach values efficiency and speed and is often used
when decision-makers face time constraints or limited resources. In
a satisficing approach, individuals set a minimum acceptable level of
quality or performance and then choose the first option that meets that
criterion. This method can lead to quick decision-making, but it may result
in suboptimal outcomes, as it may overlook better options that require
additional effort or resources to identify. Satisficing can also result in a
69
failure to fully explore all available options and to consider the long-term
consequences of decisions. There are several types of satisficing-based
• Bounded rationality: This involves making decisions

based on limited information, resources, and time, and
choosing the first option that meets a minimum level of
acceptability.
• Satisficing heuristics: This involves making

decisions based on rules of thumb or shortcuts that
help individuals quickly identify options that meet a
minimum level of acceptability.
• Efficiency: Satisficing-based decision-making can

be an efficient way to make decisions, as it allows
individuals to quickly identify options that meet a
minimum level of acceptability.
• Realism: Satisficing-based decision-making can promote

realism, as it acknowledges that individuals have limited
information, resources, and time, and that it is not always
possible to achieve the best possible solution.
• Practicality: Satisficing-based decision-making can be a

practical way to make decisions, as it focuses on what is
achievable and realistic, rather than what is ideal or optimal
• Lack of quality: Satisficing-based decision-making

may result in lower-quality outcomes, as individuals
may settle for an option that meets a minimum level of
acceptability, rather than striving for the best possible
solution.
70
• Limited perspective: Satisficing-based decision-

making may be based on limited information,
resources, and time, and may not consider all relevant
factors or perspectives.
• Missed opportunities: Satisficing-based decision-

making may result in missed opportunities, as
individuals may settle for an option that meets a
minimum level of acceptability, rather than exploring
alternative or better options.
Optimization/Maximization Based
Optimizing or maximizing decision-making is a process in which
individuals strive to make the best possible decision based on all
available information and constraints. This approach values finding the
best solution and is often used when the consequences of a decision
are significant or long-lasting. In an optimizing approach, individuals
consider all available options, weigh the pros and cons, and choose the
option that provides the greatest benefits or the least drawbacks. This
method can lead to the best possible outcome, but it can also be time-
consuming and resource-intensive, as it requires a thorough analysis of all
options. Maximizing decision-making can also result in decision paralysis
if individuals become overwhelmed by the available information and
are unable to make a decision. There are several types of optimization/
maximization-based decision-making, including these:
• Linear programming: This involves using

mathematical models to optimize decisions based on
constraints and objectives.
• Nonlinear programming: This involves using

complex, nonlinear relationships between variables
and constraints.
71
• Quality: Optimization/maximization-based decision-

making can result in higher-quality outcomes, as
individuals strive to find the best possible solution.
• Efficiency: Optimization/maximization-based
decision-making can be an efficient way to make
decisions, as it involves evaluating all options and
choosing the best one.
• Complete perspective: Optimization/maximization-

based decision-making can consider all relevant
factors and perspectives, providing a complete and
comprehensive perspective on the decision-making
process.
• Complexity: Optimization/maximization-based
decision-making can be complex, as it involves
evaluating all options and considering multiple factors.
• Resource-intensive: Optimization/maximization-
based decision-making can be resource-intensive,
requiring significant time, information, and
computational resources to find the best possible
solution.
• Unrealistic expectations: Optimization/

maximization-based decision-making can result in
unrealistic expectations, as individuals may strive for
the best possible solution, even if it is not achievable or
practical.
72
With human-only decision-making, a lot of different cognitive biases

tend to exist. Let’s deep dive into the concept of cognitive bias and its
different types.
Cognitive Bias Due to Human-Only Decision-Making

Cognitive bias refers to systematic errors in our thinking that can lead
to flawed decisions. These biases arise from our emotions, experiences,
and beliefs, and they can distort our perception of reality. Some common
cognitive biases include confirmation bias, where we tend to favor
information that supports our existing beliefs, and negativity bias, where
we give more weight to negative information. Overconfidence bias, where
we overestimate our abilities and the accuracy of our judgments, is another
example. Understanding these biases is important, as they can impact
our decision-making and lead to suboptimal outcomes. By recognizing
and correcting for our biases, we can make more informed and objective
decisions.
We will study cognitive biases in detail in the following chapters.
Human-Machine Decision-Making
As humans evolved, their decision-making techniques evolved too. Just
like we invented tools/techniques to help ease our day-to-day life (e.g.,
wheels for traveling, tools for farming, etc.), we also started to create tools
and techniques to aid our decision-making process. This led to human-
machine interactions that helped us make decisions quickly and with
more accuracy and to some extent mitigate the cognitive bias effects. Let’s
evaluate the different types of human-machine decision-making, aka
machine assisted decision-making.
73
Instruction/Rule-Based Systems
In this technique, there is a computerized system that uses a set of
explicitly stated rules to support decision-making activities. These rules
are derived from domain expertise, regulations, or best practices, and
they are encoded in the system to assist with decision-making processes.
The system operates by matching the facts of a particular case against
the rules and suggesting the course of action that best fits the rules. Rule-
based human-machine decision-making is useful in situations where the
decision-making process is well defined and the rules are well understood,
as it provides a consistent and repeatable approach to decision-making.
However, it may not be well suited to situations where the rules are
complex or subject to change, as this requires manual updates to the
system’s rule base. Before the advent of computer systems, there were
other tools used for decision-making. For example, the abacus was used
for calculating finances. There are several types of instruction/rule-based
decision systems, including the following:
• Expert systems: These are computer programs that

use a set of rules and knowledge to mimic the decision-
making abilities of human experts.
• Decision trees: These are graphical representations of

a set of rules or conditions that specify how decisions
should be made based on a set of variables or factors.
• Consistency: Instruction/rule-based decision systems

can provide consistent outcomes, as they rely on a set
of predefined rules that specify how decisions should
be made.
74
• Speed: Instruction/rule-based decision systems can

make decisions quickly, as they rely on predefined
rules and algorithms.
• Accuracy: Instruction/rule-based decision systems

can provide accurate outcomes, as they rely on well-
defined rules and algorithms that have been tested and
validated.
• Rigidity: Instruction/rule-based decision systems

can be rigid, as they rely on a set of predefined rules
that may not adapt to changing circumstances or new
information.
• Lack of flexibility: Instruction/rule-based decision

systems may not be flexible, as they rely on predefined
rules that may not account for all possible scenarios or
exceptions.
• Limited perspective: Instruction/rule-based decision

systems may have a limited perspective, as they rely
on predefined rules that may not consider all relevant
factors or perspectives.
Mathematical Models
In this technique, there is a computer-based system that uses
mathematical and statistical techniques to support decision-making. It
integrates mathematical models, data, and algorithms to analyze complex
problems and provide recommendations or solutions. The models used in
a mathematical decision system can be linear or nonlinear programming,
dynamic programming, or any other mathematical method suitable for
a particular problem. The system can also include data visualization
75
tools to help users understand the results of the analysis. The goal of a
mathematical model is to provide objective and quantifiable insights
into decision problems and to help users make more informed decisions.
Mathematical modeling is commonly used in fields such as finance,
engineering, and healthcare to support complex decision-making
processes. There are several types of mathematical model-based decision
systems, including the following:
• Linear programming: This involves using

constraints and objectives.
• Nonlinear programming: This involves using

complex, nonlinear relationships between variables
and constraints.
• Decision analysis: This involves using mathematical

models to evaluate different courses of action based on
their expected outcomes, risks, and uncertainties.
• Accuracy: Mathematical models can provide accurate

predictions about the outcomes of different decisions.
• Speed: Mathematical models can make predictions

quickly, as they rely on algorithms and mathematical
equations.
• Consistency: Mathematical models can provide

consistent outcomes, as they rely on well-defined
algorithms and equations that have been tested and
validated.
76
• Complexity: Mathematical models can be complex, as

they involve mathematical equations and algorithms
that may not be easy to understand or interpret.
• Limited perspective: Mathematical models may

have a limited perspective, as they rely on predefined
relationships between variables that may not account
for all relevant factors or perspectives.
• Model assumptions: Mathematical models rely on

certain assumptions about the relationships between
variables, which may not always hold true in reality.
This can lead to inaccuracies in the predictions made
by mathematical models.
Probabilistic Models
In this technique, there is a computerized system that uses probability
theory to support decision-making. It is based on the idea that many
real-world problems involve uncertainty and that a probabilistic approach
is more appropriate than a deterministic one in such cases. The system
uses probabilistic models, such as Bayesian networks or Markov decision
processes, to represent the relationships between variables and the
uncertainty associated with them. The models are used to calculate the
probability of different outcomes given certain inputs and to support
decision-making by providing recommendations or suggesting actions
with the highest expected value. Probabilistic decision-making is
commonly used in fields such as finance, insurance, and healthcare, where
decisions must be made in the face of uncertainty. By explicitly modeling
and quantifying uncertainty, a probabilistic model can provide more
robust and flexible support for decision-making. There are several types of
probability-based decision systems, including the following:
77
• Bayesian networks: These are probabilistic graphical

models that represent the relationships between
variables and their probabilities.
• Monte Carlo simulation: This involves generating

many random samples from a probabilistic model to
estimate the probabilities of different outcomes.
• Decision analysis: This involves using probability

models to evaluate different courses of action based on
their expected outcomes and risks.
• Consideration of uncertainty: Probability-based

decision systems account for uncertainty by using
probabilistic models to estimate the likelihood of
different outcomes.
• Accuracy: Probability-based decision systems can

provide accurate predictions about the likelihood of
different outcomes, as they rely on statistical methods
and algorithms.
• Consistency: Probability-based decision systems can

provide consistent outcomes, as they rely on well-
defined algorithms and equations that have been tested
and validated.
• Complexity: Probability-based decision systems can

be complex, as they involve statistical methods and
algorithms that may not be easy to understand or
interpret.
78
• Limited data: Probability-based decision systems

rely on historical data or other relevant information to
estimate the probabilities of different outcomes. If this
data is limited or not representative, the predictions
made by probability-based decision systems may not
be accurate.
• Model assumptions: Probability-based decision

systems rely on certain assumptions about the
relationships between variables, which may not always
hold true in reality. This can lead to inaccuracies in the
probabilities estimated by probability-based decision
systems.
AI-Based Models
In this technique, there is a computer system that uses artificial
intelligence and machine learning algorithms to support decision-making.
It analyzes data, predicts outcomes, and makes recommendations to
help decision-makers make informed choices. It can be used in various
applications such as assessing risk, optimizing business processes, and
predicting future trends. The goal is to provide accurate and relevant
information to decision-makers in real time, enhancing the speed and
quality of decision-making. There are several types of AI or machine
learning–based decision systems, including the following:
• Supervised learning: This involves training machine

learning algorithms on labeled data to predict
outcomes based on that data.
• Unsupervised learning: This involves training

machine learning algorithms on unlabeled data to
identify patterns and relationships within the segments
of the data.
79
• Automation: AI or machine learning–based decision

systems can automate decision-making processes,
reducing the time and effort required for manual
decision-making.
• Accuracy: AI or machine learning–based decision

systems can provide accurate predictions and decisions
based on the data they have been trained on.
• Scalability: AI or machine learning–based decision

systems can be easily scaled up to handle large
amounts of data, making them well-suited for large,
complex decision-making processes.
• Bias: AI or machine learning–based decision systems

can be biased if the data they have been trained on is
biased or unrepresentative.
• Explainability: AI or machine learning–based decision

systems can be difficult to interpret or explain, as
their decisions are based on complex algorithms and
mathematical models.
• Limitations of training data: AI or machine learning–

based decision systems rely on the data they have
been trained on to make predictions and decisions.
If the training data is limited or not representative,
the decisions made by AI or machine learning–based
decision systems may not be accurate.
80
Machine-Only Decision-Making
The human-machine decision-making, over the years, led to humans
realizing that some of the decision-making can be fully automated through
machines, with little to almost no oversight needed. This led to a new,
more advanced type of decision-making known as machine-only decision-
making, and the machines that make those type of decisions are called
autonomous systems. Let’s go through a few examples of machine-only
decision-making.
Autonomous Systems
Autonomous systems can make decisions and take actions without human
intervention. They use artificial intelligence, machine learning, and
other advanced technologies to analyze data, identify patterns, and make
predictions. Autonomous systems are designed to operate independently
and make decisions based on their understanding of the data and the
situation. These systems can be used in various applications, such as
robotics, unmanned aerial vehicles (UAVs), autonomous vehicles, and
smart homes. The goal of autonomous systems is to provide real-time,
accurate, and relevant information to support decision-making, while
reducing the need for human intervention and increasing efficiency and
speed. There are several types of fully autonomous decision systems,
including the following:
• Rule-based systems: These systems make decisions

based on preprogrammed rules, such as if-then
statements.
• Model-based systems: These systems use

mathematical, probabilistic, or AI-based models
to make decisions, such as decision trees or neural
networks.
81
• Reinforcement learning systems: These systems

make decisions based on trial and error, with the goal
of maximizing a reward signal.
• Efficiency: Fully autonomous decision systems can

make decisions faster and more accurately than
humans, as they are not subject to human biases or
limitations.
• 24/7 operation: Fully autonomous decision systems

can operate 24/7 without the need for rest, making
them ideal for tasks that require continuous operation.
• Increased safety: In some cases, fully autonomous

decision systems can make safer decisions than
humans, as they can process more information and
make decisions based on that information in real time.
• Lack of flexibility: Fully autonomous decision systems

can be limited by the rules, algorithms, or models they
have been programmed with and may not be able to
adapt to new situations or changing conditions.
• Responsibility: It can be difficult to determine

who is responsible for the decisions made by fully
autonomous decision systems, as they are not subject
to human control or oversight.
• Lack of transparency: Fully autonomous decision

systems can be difficult to understand, interpret,
or explain, as their decisions are based on complex
82
algorithms and mathematical models. This can make

it difficult for humans to understand the reasoning
behind the decisions made by fully autonomous
decision systems.
Conclusion
As we have seen, the decision-making process and its different
methodologies are complex. The structure of human society, comprising
individuals, groups, organizations, governments, etc., add to the
complexity. Hence, there usually is no single way of making decisions.
It can involve a combination of different types of decision-making
techniques with different methodologies. For example, a person planning
to purchase an apartment can do so just by doing the following:
• Buying an apartment in the same society that they

might be staying in, because they had a good past
experience: strategic humans-only individual decision-
making with single criterion, i.e., society
• Buying an apartment after researching online, asking

friends/family/peers, and exploring different localities
based on price, area, connectivity to public transport
and workplace: strategic human-machine group
decision-making with multiple criteria
This complexity of decision-making is what makes us human and has

helped us to be on the top of the food chain. We can expect the decision-
making process to become more complex over time, ensuring humans
evolve further.
83
CHAPTER 4
Interpreting Results
from Different
Methodologies
In the previous chapter, we saw different decision intelligence
methodologies, with their advantages and limitations. In this chapter, we will
implement various human-machine decision-making methodologies using
the Python programming language. Because covering all the methodologies
is not possible in a single chapter, or even book, we will focus on the most
commonly used ones: the mathematical, probabilistic, and AI-based
techniques. Also, we will look at how to interpret the results generated from
those methodologies and how to use them to make decisions.
ecision Intelligence Methodology:

D
Mathematical Models
Mathematical models have long been used in various processes. As
discussed in the previous chapter, these types of models fare better than
the others because of their accuracy, execution speed, and consistency
in the outcomes produced. These types of models perform very well in
situations where the factors/variables have a consistent behavior and there

Chapter 4 Interpreting Results from Different Methodologies
is less uncertainty involved. Hence, we can find a lot of use cases in the
physics and engineering domains for mathematical models. Let’s look at
some examples of the linear and nonlinear mathematical models.
Linear Models
Linear models, as the name implies, are mathematical equations that
describe a linear relationship between two or more variables. They are
represented by the following equation:
y = β0 + β1x1 + β2x2 +….+ βnxn
where:
y is the dependent variable (a variable that needs to

be calculated/predicted based on the input).
x1, x2, …., xn are the n independent variables used

as input.
β0 is a constant value. It is equal to y when x1, x2, …., xn

are all 0.
β1, β2, …., βn are beta coefficients of the independent

variables x1, x2, …., xn, respectively.
The linear model equations usually operate under certain constraints

for the independent variables. Let’s explore linear models through a real-
world use case.
Problem Statement:
A factory produces four different products: diapers, sanitary pads,
facial tissue, and toilet paper. The daily produced number of these four
products is x1, x2, x3, and x4, respectively.
Conditions:
Based on the following conditions, the factory management wants to
come up with a production schedule to maximize profit:
86
• The profit per unit of product is $10, $11, $8, and $12.5
for diapers, sanitary pads, facial tissue, and toilet paper,
respectively.
• Because of workforce constraints, the total number of

units produced per day can’t exceed 45.
• The raw materials required for producing the

aforementioned products are wood pulp and fiber. To
produce a single unit of a diaper, 5 units of wood pulp
are consumed. Each unit of sanitary pads requires 7
units of wood pulp and 2 units of fiber. Each unit of
facial tissue needs 2 units of wood pulp and 1 unit of
fiber. Finally, each unit of toilet paper requires 3 units
of fiber.
• Because of the transportation and storage constraints,

the factory can consume up to 100 units of wood pulp
and 75 units of fiber per day.
Solution:
As the problem is related to production optimization, the environment
can be safely assumed to be a controlled one with little to no uncertainty.
Hence, we can use a mathematical model as a solution.
Based on the given problem statement and associated conditions, the
mathematical model can be defined like this:
Maximize: 10x1 + 11x2 + 8x3 + 12.5x4 (Profit)
Constraints:
x1 + x2 + x3 + x4 ≤ 45 (Manpower)
5x1 + 7x2 + 2x3 ≤ 100 (Wood pulp)
2x2 + 1x3 + 3x4 ≤ 75 (Fiber)
x1, x2, x3, x4 ≥ 0
87
We will be using a general-purpose and open-source linear programming

modeling package in Python named PuLP. It is a highly suitable option for
a diverse set of optimization problems due to its exceptional qualities such
as user friendliness, versatility, compatibility with other Python libraries,
continual support from a thriving community, and open-source availability.
Step 1: Importing the PuLP library and defining the mathematical

model object
[In]: from pulp import LpProblem, LpMaximize, LpVariable,

LpStatus, lpSum
[In]: opt_model = LpProblem(name="production-planning",
sense=LpMaximize)
Here the model object is defined using the LpProblem module with
sense set to LpMaximize as we want to maximize the profit made from the
production.
Step 2: Defining the decision variables
[In]: x = {i: LpVariable(name=f"x{i}", lowBound=0) for i in

range(1, 5)}
[In]: print(x)
[Out]: {1: x1, 2: x2, 3: x3, 4: x4}
Step 3: Adding constraints
[In]: opt_model += (lpSum(x.values()) <= 45, "Manpower")

[In]: opt_model += (6 * x[1] + 4 * x[2] + 2 * x[3] <= 100,
"Wood Pulp")
[In]: opt_model += (2 * x[2] + 4 * x[3] + 6 * x[4] <= 75,
"Fiber")
Step 4: Setting the objective function
[In]: opt_model += (10 * x[1] + 11 * x[2] + 8 * x[3] + 12.5 *

x[4], "Objective Function")
88
Step 5: Utilizing the default solver method for solving the given
optimization problem
[In]: status = opt_model.solve()
Step 6: Getting the results
[In]: print(f"status: {opt_model.status}, {LpStatus[opt_model.

status]}")
[In]: print(f"objective: {opt_model.objective.value()}")
[In]: for var in x.values():
print(f"{var.name}: {var.value()}")
[In]: for name, constraint in opt_model.constraints.items():
print(f"{name}: {constraint.value()}")
[Out]:
status: 1, Optimal
objective: 512.5
x1: 20.0
x2: 0.0
x3: 0.0
x4: 25.0
Manpower: 0.0
Wood_Pulp: 0.0
Fiber: 0.0
From the previous output, we can see that the model suggests
producing 20 units of x1, i.e., diapers, and 25 units of x4, i.e., toilet paper, for
a maximum profit ($512.5) under the given constraints. Now, in a dynamic
real-world factory setting, the constraints will change, and hence there
arises a need to build a solution that can update the constraints based
on the changes (e.g., changes in profit value of the products) happening
over time due to multiple factors. So, this solution can be built into an
89
API solution that can accept inputs (constraints and other details) either
manually through an UI or through other system outputs, and the results
can be consumed either through the same UI or through any other
application (Excel, custom applications, etc.).
Figure 4-1 shows how the high-level solution architecture would look.
Figure 4-1. High-level architecture: linear mathematical models
Nonlinear Models
Linear models assume that the relationship between the inputs and
outputs is always linear. This assumption is often not true in real-world
situations where the relationship between variables is complex and
nonlinear. This is where the need for nonlinear models arises. They
can capture the complex behavior/relationships between the inputs
and outputs more accurately. They can be represented by the following
equation:
y = f(x)
where:
y is the dependent variable (a variable that needs to

be calculated/predicted based on input).
f(x) could be any nonlinear function, such as a

polynomial, exponential function, trigonometric
function, or any other nonlinear function.
90
Let’s understand how to solve a nonlinear problem with the help of the
following use case.
Problem Statement:
After the COVID-19 pandemic, company X plans to gradually have
its employees return to work at the office while maintaining a minimum
of 25 percent occupancy in both of its office blocks to ensure efficient
operations. The government has provided a risk index to assist companies
in determining the risks associated with re-opening offices during the
current situation. The risk index (RI) is defined in the following manner:
RI = 3x12 + 3x13 +x1x2- 2x22 + 2x23
where:
x1 = Proportion of block A available capacity
x2 = Proportion of block B available capacity
What is the safest way (according to the RI) for company X to resume
work from the office?
Solution:
With the RI definition and considering company X wants to minimize
the risk, we can define the objective function as follows:
Minimize: RI = 3x12 + 3x13 +x1x2- 2x22 + 2x23
Constraints:
x1 ≥ 0.25
x2 ≥ 0.25
x1, x2 ≤ 1
For this nonlinear problem, we will use the scipy package to derive the
solution.
Step 1: Importing the required libraries (especially scipy)
[In]:
import numpy as np
91
import pandas as pd
import random
from scipy.optimize import minimize
Step 2: Create a function to generate starting points

We will use the following function to generate a tuple of starting points
within the feasible range for this problem:
[In]:
def geneate_kickoff_points(number_of_points):
'''
number_of_points [list]: how many points to generate
'''
kickoff_points = []
for point in range(number_of_points):
kickoff_points.append((random.random(), random.
random()))
return kickoff_points
Step 3: Formulating the objective function and constraints
objective_function = lambda x: (3*x[0]**2) + (3*x[0]**3) +

(x[0]*x[1]) - (2*x[1]**2) + (2*x[1]**3)
constraints = [
{'type': 'ineq', 'fun': lambda x: x[0] - 0.25}, # Block A
capacity >= 0.25
{'type': 'ineq', 'fun': lambda x: x[1] - 0.25} # Block B
capacity >= 0.25
]
boundaries = [(0,1), (0,1)]
Step 4: Running the optimal solution function
# generate a list of N potential kickoff points

[In]: kickoff_points = geneate_kickoff_points(20)
92
[In]: first_iteration = True
[In]:
for point in kickoff_points:
# for each point run the algorithm
result = minimize(
objective_function,
[point[0], point[1]],
method='SLSQP',
bounds=boundaries,
constraints=constraints
)
# first iteration always going to be the best so far
if first_iteration:
better_solution_found = False
best_solution = result
else:
# if we find a better solution, lets use it
if result.success and result.fun < best_solution.fun:
better_solution_found = True
best_solution = result
# print results if algorithim was successful

[In]:
if best_solution.success:
print(f"""Optimal solution found:
- Proportion of block A available capacity: {round
(best_solution.x[0], 3)}
- Proportion of block B available capacity: {round
(best_solution.x[1], 3)}
- Risk index value: {round(best_solution.fun, 3)}""")
else:
93
print("No solution found to problem")

[Out]:
Optimal solution found:
- Proportion of block A available capacity: 0.25
- Proportion of block B available capacity: 0.597
- Risk index value: 0.096
From the previous simulation, we find that company X can resume

work from office for its employees with block A capacity at 25 percent and
block B capacity at about 60 percent, with the optimal about 10 percent
risk for the given constraints/conditions.
This solution can be deployed the same way as how the linear model
was deployed. However, certain details such as the infrastructure might
need changes as usually nonlinear solutions require more compute power
for calculations, as compared to the linear ones.

D
Probabilistic Models
Probabilistic models employ probability theory principles, unlike
mathematical models that use mathematical equations, to model
problems. This makes the probabilistic models particularly appropriate
for situations where numerous uncertain factors influence the
outcome. Probabilistic models are often used in fields such as finance,
economics, and healthcare, where the outcomes of decisions are
uncertain and unpredictable. We will now see how we can use one of the
widely used probabilistic models, a Markov chain, to build a decision
intelligence system.
94
Markov Chain
Markov chain models are a type of probabilistic models that can be
employed to study dynamic systems, where the likelihood of transitioning
from one state to another is determined solely by the present state, without
any dependence on the system’s past history.
An example of a Markov chain model is a stock market model that
has three states: bull, bear, and flat. We can represent this as a three-state
Markov chain model, where the transition probabilities between states
depend only on the current state. For example, let’s say that the probability
of moving from a bull market to a bear market is 0.3, and it is 0.6 for staying
in bull market itself. The probability of moving from a bear market to a flat
market is 0.2, and it is 0.3 for staying in bear market itself. The probability
of moving from a flat market to a bull market is 0.2, and the probability of
staying in the flat market is 0.6. Figure 4-2 shows what the transition matrix
for the stock market model.
Figure 4-2. State transition diagram: Markov chain
This matrix tells us that there is a 60 percent chance of staying in the

bull market on a given day, a 30 percent chance of moving to the bear
market, and a 10 percent chance of moving to the flat market. Similarly,
95
there is a 30 percent chance of staying in the bear market on a given day, a

20 percent chance of moving to the flat market, and a 50 percent chance of
moving to the bull market. Finally, there is a 20 percent chance of moving
from the flat market to the bull market, a 20 percent chance of moving to
the bear market, and a 60 percent chance of staying in the flat market.
Let’s understand how Markov chain models can be used for allocating
a budget for different marketing channels/touchpoints.
Problem Statement:
Company X is launching a new Android/iOS expense manager
application and wants to create awareness among its target audience.
It plans to offer a one-month free trial of the application to people who
subscribe through the website. To achieve their marketing goal, they
need to identify effective marketing channels that can reach their target
audience and generate a high ROI through website sign-ups. While
they have data from a previous similar campaign and have been using a
heuristic model (last touch attribution) for budget allocation, they want to
leverage their ML team’s inputs to build a more data-driven solution that
can analyze and optimize their marketing channels.
Solution:
Based on the given problem, the ML team suggested using a Markov
chain model as it best suits attribution problems. The primary aim of
the team is to find out the optimal percentage budget allocation for each
marketing channel, based on the output of the Markov chain model.
The team has access to the historical user path data along with the
daily marketing spend for each of the channel.
The team proposes to build the model using the ChannelAttribution
Python package.
Step 1: Importing the required libraries
[In]:
import numpy as np
import pandas as pd
from ChannelAttribution import *
96
Step 2: Reading historical campaign data (user path and daily

marketing spend for the channels)
[In]:
attribution_data = pd.read_csv('attribution_data.csv')
attribution_data['time'] = pd.to_datetime(attribution_
data['time'])
daily_budget = pd.read_csv('daily_budget.csv')
Step 3: Data preprocessing

We create the path order for when users interact with different
channels (this is equivalent to SQL window functions).
[In]:
attribution_data['path_order'] = attribution_data.sort_
values(['time']).groupby(['cookie']).cumcount() + 1
We aggregate the channels a user interacted with into a single row.
[In]:
attribution_data_paths = attribution_data.groupby('cookie')
['channel'].agg(lambda x: x.tolist()).reset_index()
attribution_data_paths = attribution_data_paths.
rename(columns={'channel': 'path'})
We aggregate conversions and their value by cookie ID.
[In]:
attribution_data_conversions = attribution_data.
groupby('cookie', as_index=False).agg({'conversion': 'sum',
'conversion_value': 'sum'})
We append path data with that of conversion.
97
[In]:
attribution_data_final = pd.merge(attribution_data_paths,
attribution_data_conversions, how='left', on='cookie')
[In]:
print('Total conversions: {}'.format(sum(attribution_data.
conversion)))
print('Total conversion rate: {}%'.
format(round(sum(attribution_data.conversion) /
len(attribution_data)*100)))
print('Total value of conversions: ${}'.
format(round(sum(attribution_data.conversion_value))))
print('Average conversion value: ${}'.
format(round(sum(attribution_data.conversion_value) /
sum(attribution_data.conversion))))
[Out]:
Total conversions: 19613
Total conversion rate: 3%
Total value of conversions: $122529
Average conversion value: $6
We will create a variable path in the specific format required by the

Attribution Model package where the ordered channels a user interacts
with are delimited by >.
[In]:
def listToString(df):
str1 = ""
for i in df['path']:
str1 += i + ' > '
return str1[:-3]
attribution_data_final['path'] = attribution_data_final.
apply(listToString, axis=1)
98
We remove the user’s cookie and grouping by the path to see how
many times a specific combination of channels led to a conversion or null
outcome.
[In]:
attribution_data_final.drop(columns = 'cookie', inplace = True)
attribution_data_final['null'] = np.where(attribution_data_
final['conversion'] == 0,1,0)
attribution_data_final = attribution_data_final.
groupby(['path'], as_index = False).sum()
attribution_data_final.rename(columns={"conversion": "total_
conversions", "null": "total_null", "conversion_value": "total_
conversion_value"}, inplace = True)
Step 4: Building the attribution models (heuristic for comparison

and Markov models for calculations) and saving the results in a
dataframe
[In]:
heuristic_models = heuristic_models(attribution_data_
final,"path","total_conversions",var_value="total_
conversion_value")
markov_model = markov_model(attribution_data_final, "path",
"total_conversions", var_value="total_conversion_value")
[Out]:
Number of simulations: 100000 - Convergence reached:
0.85% < 5.00%
Percentage of simulated paths that successfully end before
maximum number of steps (456) is reached: 99.99%
[In]:
all_attribution_models_result = pd.merge(heuristic_
models,markov_model,on="channel_name",how="inner")
99
all_attribution_models_conversion_result = all_attribution_
models_result[["channel_name","first_touch_conversions","last_
touch_conversions",\
"linear_touch_conversions","total_conversions"]]
all_attribution_models_conversion_result.columns = ["channel_
name","first_touch","last_touch","linear_touch","markov_model"]
all_attribution_models_conversion_value_result = all_
attribution_models_result[["channel_name","first_touch_
value","last_touch_value",\
"linear_touch_value","total_conversion_value"]]
all_attribution_models_conversion_value_result.columns
= ["channel_name","first_touch","last_touch","linear_
touch","markov_model"]
all_attribution_models_conversion_value_result =
pd.melt(all_attribution_models_conversion_value_result, id_
vars="channel_name")
Step 5: Calculating the optimal budget for the marketing channels
[In]:
daily_budget_agg = daily_budget.drop(['day', 'impressions'],
axis=1).groupby('channel', as_index=False).sum('cost')
daily_budget_agg.columns = ['channel_name', 'cost']
roas_data = pd.merge(all_attribution_models_conversion_result,
daily_budget_agg)
roas_data['channel_weight'] = (roas_data['markov_model'])/
(sum(roas_data['markov_model']))
roas_data['cost_weight'] = (roas_data['cost'])/(sum(roas_
data['cost']))
roas_data['roas'] = (roas_data['channel_weight'])/(roas_
data['cost_weight'])
100
roas_data['optimal_budget'] = (roas_data['cost'])*(roas_
data['roas'])
roas_data['optimal_budget_percentage'] = (round((roas_
data['optimal_budget'])/(sum(roas_data['optimal_
budget'])),2))*100
roas_data.head()
[Out]:
101
102
Chapter 4
channel_ markov_ cost channel_ cost_weight roas optimal_ optimal_budget_

name model weight budget percentage
Facebook 5289.103801 1481.679 0.26967337 0.305076822 0.883952337 1309.733172 27

Instagram 4468.380797 641.159 0.227827502 0.1320143 1.725778963 1106.498714 23
Paid Search 3957.708706 1187.142 0.201790073 0.244431809 0.825547518 980.0417185 20
Online 2395.295285 554.937 0.12212794 0.114261236 1.068848407 593.1435285 12
Display
Online 3502.511411 991.823 0.178581115 0.204215833 0.874472427 867.3218664 18
Video
Interpreting Results from Different Methodologies
From the previous output, the team realized that Instagram and online
display channels, while performing well in driving the conversions, were
under-budgeted. They need to be given a higher share of the marketing
budget, which, in this case is 23 percent and 12 percent for Instagram and
online display, respectively.

D
AI/ML Models
AI/ML models are based on machine learning algorithms that can learn
from data and improve their performance over time. These models are
used in a wide range of applications, from image recognition and natural
language processing to predictive analytics and decision-making.
Unlike mathematical and probabilistic models, AI models do not
require explicit formulas or assumptions about the relationships between
variables. Instead, they learn patterns and relationships from large
datasets and can identify complex patterns and trends that are difficult or
impossible for humans to discern.
Let’s understand how AI/ML models can be used for decision
intelligence through an HR analytics use case.
Problem Statement:
A multinational corporation with nine distinct areas of operation faces
a challenge in selecting suitable candidates for managerial (and below
positions) promotions and providing them with timely training. The
announcement of final promotions occurs only after evaluations, resulting
in delays in transitioning to new roles. Therefore, the company seeks
assistance in identifying qualified candidates at specific checkpoints to
accelerate the promotion process as a whole.
The company has historical data comprising demographic,
educational, work experience, and skill details, along with the promotion
details.
103
Solution:
As the company has access to historical data, it makes sense to
build a machine learning model that can find the traits of employees
fit for promotion. As the goal is to find the list of employees who can be
promoted, it will come under the supervised classification category.
Step 1: Importing the libraries and reading the training and
test data
[In]:
from sklearn.preprocessing import LabelEncoder, StandardScaler,
RobustScaler
from sklearn.model_selection import train_test_split,
StratifiedKFold
from sklearn.linear_model import LogisticRegression
from sklearn.impute import SimpleImputer
from sklearn.metrics import *
import pandas as pd
import numpy as np
[In]:
training_data = pd.read_csv('hr_train.csv')
test_data = pd.read_csv('hr_test.csv')

As the employee_id column is an ID column, it serves no purpose in
the modeling process. So, we delete the employee_id column from the
training and test data.
[In]:
del training_data['employee_id']
del test_data['employee_id']
categorical_columns = [c for c in training_data.columns if
training_data[c].dtypes=='object']
104
numeric_columns = [n for n in training_data.columns if n not in

categorical_columns]
true_categorical_columns = ['department', 'region',
'education', 'gender', 'recruitment_channel','awards_won?',
'previous_year_rating','length_of_service',
'no_of_trainings']
true_numeric_columns = [c for c in training_data.columns if c
not in true_categorical_columns]
true_numeric_columns.remove('is_promoted')
There are very few employees whose work experience is 30 years or

more (0.1 percent). These people can be considered as outliers, and it is a
best practice to remove them before we feed data to the model.
[In]:
training_data = training_data[training_data.length_of_
service<30]
We apply the imputation strategy for the education and previous_

year_rating columns.
[In]:
target = training_data.pop('is_promoted')
all_data = pd.concat([training_data, test_data], axis=0)
si_most_frequent = SimpleImputer(strategy= 'most_frequent')
all_data['education'] = si_most_frequent.fit_transform(all_
data.education.values.reshape(-1, 1))
si_mean = SimpleImputer(strategy='mean')
all_data['previous_year_rating'] = si_mean.fit_transform(all_
data.previous_year_rating.values.reshape(-1, 1))
We convert categorical features into dummy ones.
105
[In]:
all_data_dummy = pd.get_dummies(all_data)
new_training_data = all_data_dummy.iloc[:len(training_data)]
new_test_data = all_data_dummy.iloc[len(training_data):]
We apply StandardScaler for standardizing the features in training

and test data.
ss = StandardScaler()
new_training_data = ss.fit_transform(new_training_data)
new_test_data = ss.transform(new_test_data)
Step 3: Model training (logistic regression) and evaluation

We perform stratified K-fold cross validation on the data and training
model and evaluate the performance. As the purpose of this activity is to
show how AI/ML model outputs can be used for decision intelligence,
we will keep the modeling process pretty simple and not train multiple
algorithms to see which ones perform the best. Instead, we will train a
logistic regression model with default/minimal changes to the parameters
and will see how the results can be interpreted.
[In]:
scores = []
probs = np.zeros(len(new_training_data))
y_le = target.values
folds = StratifiedKFold(n_splits=5, shuffle=True, random_
state=42)
for fold_, (train_ind, val_ind) in enumerate(folds.split(new_

training_data, y_le)):
print('fold:', fold_)
X_tr, X_test = new_training_data[train_ind], new_training_
data[val_ind]
y_tr, y_test = y_le[train_ind], y_le[val_ind]
106
clf = LogisticRegression(max_iter=200, random_state=2020)

clf.fit(X_tr, y_tr)
probs[val_ind]= clf.predict_proba(X_test)[:, 1]
y = clf.predict_proba(X_tr)[:,1]
print('train:',roc_auc_score(y_tr, y),'val :' , roc_auc_
score(y_test, (probs[val_ind])))
print(20 * '-')
scores.append(roc_auc_score(y_test, probs[val_ind]))
print('log reg roc_auc= ', round(np.mean(scores)*100,2))
probs_rnd = np.where(probs > 0.5, 1, 0)
print('F1 Score= ', round(f1_score(target, probs_rnd)*100,2))
print('Recall Score= ', round(recall_score(target, probs_
rnd)*100,2))
[Out]:
fold: 0
train: 0.7931202718718571 val : 0.7819824102253672
--------------------
fold: 1
train: 0.7897501709069539 val : 0.79278434937156
--------------------
fold: 2
train: 0.7868061847019592 val : 0.804318671120831
--------------------
fold: 3
train: 0.7930340720726938 val : 0.7815370354855481
--------------------
fold: 4
train: 0.7943494182608088 val : 0.7806564317616108
--------------------
log reg roc_auc= 78.83
F1 Score= 44.66
Recall Score= 29.24
107
As we can see, the model performance is satisfactory, and with

additional techniques, it can be improved. Now, the HR team can leverage
the predictions made on the employees through the ML model and create
a list of potential promotion candidates based on their probability scores.
Conclusion
In this chapter, we went through practical real-world use-case problems
and how they can be solved through different decision intelligence
methodologies. We also saw how the output from the different models can
be used for interpretation and how one can build a decision intelligence
workflow from it. We have specifically focused on some of the most
used mathematical, probabilistic, and AI/ML models, as covering all
methodologies in a single chapter/book is impossible.
108
CHAPTER 5
Augmenting Decision
Intelligence Results
into the Business
Workflow
In today’s rapidly changing business environment, organizations are under
increasing pressure to make informed decisions that drive growth and
competitiveness. Decision intelligence (DI) is a powerful tool that can help
organizations make better decisions by leveraging the power of artificial
intelligence (AI) and data analytics. However, to fully realize the benefits
of DI, it is essential to integrate the outputs of DI models into existing
business workflows.
In this chapter, we will discuss how to effectively integrate DI into
business workflows through the development of user-friendly applications
that connect AI predictions to existing business tools, the importance
of collaboration between DI experts and business teams, and some
of the technical challenges involved in integrating DI into existing
business systems. By successfully integrating DI into business workflows,
organizations can leverage the power of AI to drive better decision-making
and improve business outcomes.

Chapter 5 Augmenting Decision Intelligence Results into the Business Workflow
Challenges
Integrating decision intelligence with business workflows can be a
complex process and comes with several challenges. Here are some of the
main challenges organizations may face:
• Data quality and availability: One of the primary
challenges is ensuring that the data used by the DI
process is of high quality and readily available. Data
quality issues, such as missing or incorrect data, can
impact the accuracy and usefulness of the insights
generated by the DI process.
• Model selection and validation: Another challenge

is selecting the appropriate machine learning models
for the specific business problem and validating their
accuracy and reliability. This requires expertise in both
data science and the specific domain being analyzed.
• Integration with existing systems: Integrating the

AI outputs with existing systems can be challenging.
It requires a thorough understanding of the
organization’s IT infrastructure and security protocols
to ensure that the app can be integrated smoothly and
securely.
• User adoption and training: Even the best-designed

app may not be effective if users are not willing or
able to adopt it. Ensuring user adoption and providing
adequate training is critical to the success of the app.
• Ongoing maintenance and updates: Like all software

applications, even this require ongoing maintenance
and updates to ensure that it remains functional and
110
up-to-date with the latest technology and data. This

requires resources and expertise to ensure that the app
remains effective over time.
Especially integrating decision intelligence with existing systems can

be challenging for several reasons.
• Different data formats: Existing systems may store

data in different formats, making it difficult to integrate
with the DI system. For example, one system may use
a CSV file format, while another system may use a
database format. Converting data between formats can
be time-consuming and error prone.
• Legacy systems: Some organizations may still be using

legacy systems that are not designed to integrate with
modern software applications. These systems may
have limited capabilities and may not support modern
integration methods, such as REST APIs.
• Security considerations: Existing systems may have

strict security protocols in place to protect sensitive
data. Integrating with these systems may require
additional security measures, such as authentication
and encryption, to ensure that data is protected.
• Scalability: Existing systems may not be designed

to handle the increased workload that comes
with integrating with a DI system. This can lead to
performance issues and may require additional
hardware or infrastructure to support the integration.
• Complexity: Existing systems may be complex and

difficult to understand, especially if they have been
111
in use for many years. This can make it challenging

to identify the data required for the DI system and to
design an integration strategy that works effectively.
Overall, integrating DI with existing systems requires a thorough

understanding of the existing IT infrastructure, as well as expertise in data
integration and system architecture. Addressing these challenges is critical
to ensure that the integration is successful and that the DI system can
provide accurate and reliable insights to support decision-making.
Workflow
Let’s understand the steps involved to ensure DI outputs are properly
ingested into business workflows and overcome the previous challenges.
1. Clearly define the business problem: Before

starting the DI process, it is important to clearly
define the business problem that you are trying to
solve. This will help you to identify the relevant data
sources and develop machine learning models that
are aligned with your business objectives.
2. Involve key stakeholders: Involve key stakeholders
in the DI process, including business users, data
scientists, and IT professionals. This will help to
ensure that the DI outputs are aligned with business
needs and can be integrated into existing business
workflows.
3. Integrations: Develop user-friendly interfaces

that can be easily integrated into existing business
workflows. This may involve developing APIs that
can be called from within business applications
112
or integrating the models into existing business

processes. Business workflow could be in CRM,
marketing application, apps, etc.
4. Ensure data quality: Ensure that the data used in

the DI process is accurate, complete, and relevant.
This may require data cleansing, normalization, and
transformation.
5. Monitor performance: Monitor the performance of

the DI process and evaluate its impact on business
outcomes. This will help you to identify areas for
improvement and make adjustments as necessary.
6. Provide training and support: Provide training

and support to business users on how to use the
DI outputs in their daily work. This will help to
ensure that the DI outputs are properly ingested into
business workflows and that the organization can
derive maximum value from the DI process.
As shown in Figure 5-1, organizations can ensure that DI outputs are

properly ingested into business workflows, leading to better decisions,
improved operations, and a competitive advantage.
Figure 5-1. Workflow
113
Decision Intelligence Apps

One of the ways businesses can use predictions is through apps. So, let’s
explore how to build/integrate apps with DI and why this is important.
Integrating a DI framework through apps involves developing
applications that can leverage the insights generated by the DI process.
Here are some steps to consider when integrating DI through apps:
1. Define the business problem: Clearly define the

business problem you are trying to solve using
DI. This will help you to identify the relevant data
sources and determine the appropriate machine
learning models to use.
2. Develop the DI process: Develop the DI process to

generate insights that are relevant to the business
problem. This may involve data cleansing, feature
engineering, model training, and validation.
3. Develop the app: Develop an app that can leverage

the insights generated by the DI process. This may
involve developing APIs that can be called from
within the app or integrating the models into the
app’s functionality.
4. Test and validate: Test and validate the app to

ensure that it is working as expected. This may
involve unit testing, integration testing, and user
acceptance testing.
5. Deploy the app: Deploy the app to a production

environment where it can be used by business users.
This may involve working with IT professionals to
ensure that the app is secure, scalable, and reliable.
114
6. Monitor performance: Monitor the performance

of the app and the DI process to identify areas for
improvement and make adjustments as necessary.
How and Why?

Developing an app to integrate DI involves creating a software application
that leverages the insights generated by the DI process. Here are some
details on how and why to develop an app for DI:
1. Identify the use case: The first step is to identify

the use case for the app. What business problem are
you trying to solve with DI? What insights do you
need to generate to address this problem? Once you
have a clear understanding of the use case, you can
determine the functionality required for the app.
2. Define the user interface: The next step is to

define the user interface for the app. This involves
identifying the user personas and designing an
interface that is easy to use and understand. The
interface should provide users with access to the
insights generated by the DI process in a clear and
intuitive manner.
3. Develop the app functionality: Once you have

defined the user interface, the next step is to develop
the functionality for the app. This may involve
developing APIs that can be called from within
the app to access the insights generated by the DI
process. It may also involve integrating the machine
learning models into the app’s functionality to
generate predictions or recommendations based on
the user’s input.
115
4. Test and validate the app: Once the app has been
developed, it is important to test and validate it
to ensure that it is working as expected. This may
involve unit testing, integration testing, and user
acceptance testing. The testing phase helps to
identify any bugs or issues that need to be addressed
before the app is deployed.
5. Deploy the app: Once the app has been tested and
validated, it is ready to be deployed to a production
environment. This may involve working with IT
professionals to ensure that the app is secure,
scalable, and reliable.
By developing an app to integrate DI, organizations can provide

business users with easy access to the insights generated by the DI process.
The app can help to streamline workflows, improve decision-making, and
drive business value. However, it is important to ensure that the app is
user-friendly, reliable, and secure to maximize adoption and ensure that
the organization can derive maximum value from the DI process.
User-Friendly Interfaces
Designing user-friendly interfaces is essential when integrating decision
intelligence outputs into business workflows. Because the users are
nontechnical, it’s important to develop interfaces that are easy to use and
take actions.
Here are some reasons why a user-friendly interface matters:
• Facilitates adoption: User-friendly interfaces make

it easier for business users to adopt and use the DI
outputs in their daily work. If the interfaces are difficult
to use, it may discourage adoption, which could
undermine the effectiveness of the DI process.
116
• Enhances usability: User-friendly interfaces enhance

the usability of the DI outputs. By making the interfaces
easy to use and understand, business users can
quickly access the insights they need and make better
decisions.
• Improves productivity: User-friendly interfaces can

improve productivity by reducing the time and effort
required to access the DI outputs. This can help to
streamline business workflows, increase efficiency, and
reduce costs.
• Supports collaboration: User-friendly interfaces can

support collaboration between business users and
data scientists. By providing a common interface for
accessing DI outputs, business users and data scientists
can work together more effectively and share insights
more easily.
• Increases engagement: User-friendly interfaces can

increase engagement with the DI outputs. By providing
a visually appealing and intuitive interface, business
users are more likely to engage with the insights
provided by the DI process.
In summary, designing user-friendly interfaces is critical when

integrating DI outputs into business workflows. It can facilitate adoption,
enhance usability, improve productivity, support collaboration, and
increase engagement. By prioritizing user experience, organizations
can maximize the value derived from the DI process and achieve better
business outcomes.
117
ugmenting AI Predictions
A
to Business Workflow
Connecting AI predictions to business tools involves integrating the
AI system with existing business systems, such as enterprise resource
planning (ERP) systems, customer relationship management (CRM)
systems, and business intelligence (BI) tools. Here are the technical steps
to connect AI predictions to business tools:
1. Collect and process data: Collect and process data

required for the AI model to make predictions. This
data can be sourced from internal business systems
or external sources.
2. Build and train the AI model: Develop the AI

model that will generate predictions based on the
data. Train the model using historical data to ensure
accurate predictions.
3. Deploy the AI model: Deploy the trained AI model

to an environment where it can be accessed by
business tools. This can be done using application
programming interfaces (APIs), web services, or
other integration tools.
4. Connect to business tools: Connect the AI model

to the relevant business tools, such as ERP systems,
CRM systems, or BI tools. This can be done through
APIs, custom integrations, or prebuilt connectors.
5. Test and validate: Test the integration to ensure

that AI predictions are properly integrated into the
business tools. Validate that the predictions align
with the business goals and objectives.
118
6. Monitor and maintain: Monitor the AI system

to ensure that it continues to provide accurate
predictions. Maintenance and updates may be
required to address changes in the business
environment or to improve the accuracy of the
predictions.
As shown in Figure 5-2, in summary, connecting AI predictions to

business tools involves data collection and processing, building and
training an AI model, deploying the model, connecting to business tools,
testing and validating the integration, and ongoing monitoring and
maintenance. By following these technical steps, organizations can ensure
that AI predictions are effectively integrated into their business processes
and support informed decision-making
Figure 5-2. Augmenting predictions with business tools
119
Connect to Business Tools

Connecting AI predictions to business tools means connecting the AI
model to the relevant business tools, such as ERP systems, CRM systems,
or BI tools. This step is crucial because it enables the AI model to share its
predictions with other systems that are critical to business operations.
Connecting the AI model to business tools involves integrating the AI
model with the business tool’s API. APIs provide a standard interface for
different systems to communicate with each other, enabling data to be
exchanged seamlessly between systems. The integration process typically
involves the following steps:
1. Identify the appropriate API: Determine which

API the business tool supports and what data format
it uses.
2. Define the data requirements: Identify the specific

data that the business tool requires to make use of
the AI model’s predictions.
3. Map the data: Map the data produced by the AI

model to the data required by the business tool. This
may involve data transformation or conversion to
ensure that the data is compatible.
4. Test the integration: Test the integration to ensure

that data is being transferred correctly and that the
business tool can use the AI model’s predictions
effectively.
There are several methods for connecting AI models to business

tools, including custom integration, prebuilt connectors, and third-party
integration platforms. The choice of method depends on the specific
requirements of the organization and the systems being integrated.
120
In summary, connecting AI predictions to business tools involves

integrating the AI model with the business tool’s API, mapping data
between the two systems, and testing the integration to ensure that it
works effectively. By connecting AI models to business tools, organizations
can leverage AI predictions to support informed decision-making and
improve business outcomes.
Map the Data

Mapping the data in the context of connecting AI predictions to business
tools refers to the process of identifying the relevant data produced by the
AI model and aligning it with the data required by the business tool. The
data produced by the AI model may be in a format that is different from
what the business tool requires, so mapping is necessary to ensure that the
data can be effectively shared between the two systems.
For example, as shown in Figure 5-3, let’s say that an organization has
an AI model that predicts customer churn. The AI model produces a list of
customers who are at high risk of churning, along with a probability score
for each customer. To integrate this prediction into the organization’s CRM
system, the organization would need to map the data from the AI model to
the data required by the CRM system.
Figure 5-3. Example AI model that predicts customer churn
121
Mapping the data would involve identifying the specific fields in

the CRM system that correspond to the data produced by the AI model.
For example, the CRM system may require the customer name, account
number, and contact information, in addition to the churn probability
score. The organization would need to ensure that the data produced
by the AI model is aligned with the data required by the CRM system
and may need to perform data transformations or conversions to ensure
compatibility.
Once the data mapping is complete, the organization can then connect
the AI model to the CRM system, using the CRM system’s API or other
integration methods, to share the AI predictions with the CRM system.
This enables the organization to leverage the AI predictions to inform
its customer retention strategies, and to take targeted actions to prevent
customer churn.
In summary, mapping the data involves aligning the data produced by
the AI model with the data required by the business tool to ensure that the
data can be effectively shared and used to support business operations.
Conclusion
In this chapter, we discussed how decision intelligence can be integrated
into business workflows through the development of user-friendly
applications that connect AI predictions to existing business tools. We
highlighted the importance of collaboration between DI experts and
business teams to ensure that the outputs of DI models are properly
integrated into business workflows. We also discussed some of the
technical challenges involved in integrating DI into business workflows,
such as connecting to legacy systems and ensuring data security.
To successfully integrate DI into business workflows, it is essential to
develop user-friendly interfaces that make it easy for nontechnical users to
interact with the DI models and understand the outputs. The development
122
of custom applications that connect AI predictions to existing business

tools is an effective way to achieve this. To increase user adoption, it is
important to involve business teams in the development process and to
provide training and support to help users understand how to use the
applications effectively.
Overall, integrating DI into business workflows requires a combination
of technical expertise, collaboration between DI experts and business
teams, and a focus on developing user-friendly interfaces that make it
easy for nontechnical users to interact with the outputs of DI models.
By successfully integrating DI into business workflows, organizations
can leverage the power of AI to support informed decision-making and
improve business outcomes.
123
CHAPTER 6
Actions, Biases, and

Human-in-the-Loop
Ethical AI refers to the development and use of artificial intelligence (AI) in
a way that aligns with ethical principles and values. This includes ensuring
that AI systems are transparent, accountable, and unbiased, and that
they do not cause harm to individuals or society. In this chapter, we will
uncover the ethical aspects of AI such as bias and fairness, responsible AI,
and human in the loop.
Key Ethical Considerations in AI

Several key ethical considerations must be taken into account when
developing AI systems. This includes ensuring that AI systems are
designed and implemented in ways that respect individual privacy
and autonomy, that they are not used to create or perpetuate unjust
power structures, and that they do not contribute to social or economic
inequality.
To promote ethical AI, organizations and developers should follow
several best practices, such as involving diverse stakeholders in the design
and development process, conducting thorough testing and evaluation
of AI systems to identify potential biases and unintended consequences,
and implementing clear guidelines and standards for ethical AI across the
industry.
Chapter 6 Actions, Biases, and Human-in-the-Loop
• Bias and fairness: AI systems can perpetuate and

amplify biases, particularly if the data used to train
them is biased. Ethical AI requires addressing and
mitigating these biases to ensure fairness and equity.
• Privacy and security: AI systems often deal with

sensitive personal data, such as medical records or
financial information. Ethical AI requires ensuring that
this data is handled securely and in accordance with
relevant privacy laws and regulations.
• Transparency and explainability: AI systems can

be difficult to understand and interpret, making it
challenging to determine how decisions are made.
Ethical AI requires transparency and explainability
so that individuals can understand how and why
decisions are being made.
• Accountability: Ethical AI requires accountability

so that individuals or organizations responsible for
developing and deploying AI systems can be held
responsible for any negative outcomes.
• Human values and rights: AI systems should be
designed and used in a way that respects human values
and rights, including autonomy, privacy, and dignity.
Addressing these ethical considerations requires collaboration and

engagement across a range of stakeholders, including technologists,
policymakers, ethicists, and civil society groups. It also requires ongoing
evaluation and adaptation as AI technologies continue to evolve.
126
Actions, Biases, and Human-in-the-Loop

Actions, biases, and human in the loop are all important considerations
when implementing decision intelligence (DI) to ensure that the decisions
made are accurate, ethical, and aligned with business goals. Here’s a brief
explanation of each:
• Actions: In DI, the actions taken based on the insights

generated by the system are critical. It is important to
ensure that the actions are aligned with business goals
and are ethical. DI should be used to augment human
decision-making, not replace it entirely. Therefore, it’s
essential to ensure that the actions taken are consistent
with organizational values and principles.
• Biases: One of the most significant challenges in

AI and DI is the potential for biases in the data and
models used. Biases can result from various factors,
including the collection of data, selection of algorithms,
or training data. Therefore, it’s crucial to identify
and mitigate any biases that may exist in the data or
models used. This can be achieved by using diverse
data sources, testing and evaluating models, and
continuously monitoring and updating the models
as needed.
• Human in the loop: The concept of human in the loop

refers to the involvement of humans in the decision-
making process. Despite the advances in AI and DI,
human expertise and judgment are still crucial in many
decision-making processes. Therefore, it’s essential to
have a human in the loop to monitor and evaluate the
127
decisions made by the system. This can be achieved

by providing humans with visibility into the decision-
making process, allowing them to intervene when
necessary.
In summary, when implementing DI, it’s essential to ensure that

the actions taken are consistent with business goals, identify and
mitigate biases, and involve humans in the decision-making process. By
considering these factors, businesses can ensure that the decisions made
using DI are accurate, ethical, and aligned with organizational value.
Cognitive Biases
Cognitive biases can also play a significant role in the development and
implementation of artificial intelligence (AI) and machine learning (ML)
algorithms. These biases can arise from a variety of sources, including
the data used to train the algorithms, the assumptions and biases of the
developers and users, and the design of the algorithms themselves.
For example, if the data used to train an AI system is biased, such as
containing gender or racial biases, then the system may produce biased
results. This is because the AI system learns from the data it is trained on,
and if the data contains biased patterns, those patterns can be replicated in
the output of the system.
Another example is the bias in the selection of features used to train
the AI system. If the selected features are not representative of the problem
domain, the AI system may produce inaccurate or biased results.
In addition, cognitive biases of the developers and users can also
influence the design and implementation of AI and ML systems. For
instance, confirmation bias may lead to overlooking important factors in
the development process, while anchoring bias may result in over-reliance
on certain data or features.
128
To address these potential biases in AI and ML, it is essential to

promote diversity and inclusivity in data collection and analysis. This
can include using diverse datasets, including diverse perspectives in the
development team, and testing and validating the algorithms with diverse
groups of users.
Furthermore, there are various techniques for detecting and mitigating
cognitive biases in AI and ML, such as fairness testing and bias mitigation
methods. These techniques help to ensure that AI and ML algorithms
are free from cognitive biases and produce results that are accurate and
unbiased.
Why Is Detecting Bias Important?

It is important to detect cognitive biases because they can lead to
inaccurate and suboptimal decisions. When individuals or organizations
make decisions based on biased thinking, they may not be considering
all of the available information or may be giving undue weight to certain
factors, which can result in poor outcomes.
For example, a hiring manager who is affected by confirmation bias
may select only those job candidates who share their own background
and beliefs, rather than objectively evaluating the qualifications of all
candidates. This can result in a less diverse and less qualified workforce,
which can ultimately harm the organization’s performance.
Similarly, an investor who is affected by availability bias may make
decisions based on the most recent or easily accessible information, rather
than taking a more comprehensive and objective view of the market. This
can lead to poor investment decisions and financial losses.
By detecting and addressing cognitive biases, individuals and
organizations can make more informed and accurate decisions, which
can lead to better outcomes. It can also help to promote fairness, diversity,
and inclusivity in decision-making, which are important values in many
contexts.
129
Overall, detecting cognitive biases is important because it allows

individuals and organizations to make decisions based on a more accurate
and comprehensive understanding of the situation, which can lead to
better outcomes and more successful outcomes.
Types
Several types of biases can occur in AI and ML algorithms. Here are some
of the most common types:
• Sampling bias: This occurs when the data used to train

an algorithm is not representative of the population
it is intended to serve. For example, if an algorithm
is trained on data that over-represents a particular
demographic group, it may produce biased results
when applied to the broader population.
• Confirmation bias: This occurs when an algorithm

is designed to seek out and reinforce pre-existing
beliefs or assumptions. For example, an algorithm
used to identify potential job candidates may prioritize
certain skills or experiences that are associated with a
particular demographic group.
• Selection bias: This occurs when an algorithm

selectively considers certain features or data points
while ignoring others. For example, an algorithm used
to evaluate creditworthiness may consider factors
such as income and credit history but not consider
other important factors such as education level or job
stability.
130
• Automation bias: This occurs when humans place

too much trust in the decisions made by an algorithm,
without adequately questioning or scrutinizing the
results. For example, a medical diagnosis algorithm
may identify a particular condition, but a doctor should
still use their clinical judgment and evaluate the result
in the context of the patient’s other symptoms and
medical history.
• Group attribution bias: This occurs when an algorithm

attributes certain characteristics or behaviors to an
entire group, based on the actions of a few individuals
within that group. For example, an algorithm used
to predict crime rates may unfairly attribute certain
criminal behaviors to specific demographic groups.
• Algorithmic bias: This occurs when the design or

implementation of an algorithm itself is inherently
biased. For example, an algorithm may use certain
features or decision-making processes that unfairly
advantage or disadvantage certain groups.
Detecting and mitigating these biases requires careful attention to the

design and implementation of algorithms, as well as ongoing monitoring
and evaluation to ensure that the algorithms are producing accurate and
fair results.
131
What Happens If Bias Is Ignored?

Bias can have a significant impact on an organization in several ways.
• Decreased performance: Biased decision-making can

lead to poor outcomes and decreased performance.
When decisions are made based on flawed
assumptions or incomplete information, organizations
may miss opportunities or make suboptimal choices
that negatively impact their bottom line.
• Damage to reputation: Biased decision-making

can also damage an organization’s reputation. If
an organization is perceived as being unfair or
discriminatory, it can lose the trust and confidence of
customers, employees, and other stakeholders, which
can have long-term consequences.
• Legal and regulatory risks: Biased decision-making

can also expose organizations to legal and regulatory
risks. Discriminatory practices can lead to lawsuits,
fines, and other legal penalties, which can be costly and
damaging to an organization’s reputation.
• Reduced diversity and inclusion: Biased decision-

making can also lead to reduced diversity and inclusion
within an organization. If decisions are made based
on preconceived notions or stereotypes, it can limit
opportunities for individuals from diverse backgrounds
and create a less inclusive workplace culture.
• Inefficient use of resources: Biased decision-making

can also result in an inefficient use of resources. If
decisions are made based on flawed assumptions or
incomplete information, organizations may invest
132
resources in projects or initiatives that are unlikely

to succeed, which can waste time, money, and other
resources.
Overall, bias can have a significant impact on an organization’s

performance, reputation, and ability to attract and retain talented
employees. By detecting and addressing biases in decision-making,
organizations can promote fairness, diversity, and inclusion, and improve
their overall performance and success.
Bias Detection
Detecting bias in AI and ML algorithms can be challenging, but there are
several approaches that can be taken to identify and mitigate bias.
• Data analysis: One way to detect bias is to conduct a

comprehensive analysis of the data used to train the
algorithm. This analysis should include examining the
data for any patterns or biases that may be present, as
well as identifying any gaps or omissions in the data.
• Auditing: Another way to detect bias is to conduct

an audit of the algorithm to evaluate its performance
and accuracy. This audit should include evaluating
the algorithm’s outputs and assessing its impact on
different groups.
• Testing: Testing the algorithm on different inputs

and scenarios can help identify any biases or
inconsistencies in its outputs. This testing should
include both positive and negative examples to ensure
that the algorithm is producing accurate results across
different scenarios.
133
• Expert review: Having experts in the relevant field

review the algorithm can provide valuable insights
into potential biases and how to mitigate them. These
experts can provide feedback on the algorithm’s design
and implementation and suggest ways to improve its
performance and accuracy.
• Bias detection tools: There are also a growing number
of tools available that are designed specifically to detect
and mitigate bias in AI and ML algorithms. These tools
use a variety of techniques, including data analysis
and machine learning, to identify potential biases and
suggest ways to mitigate them.
Overall, detecting bias in AI and ML algorithms requires a combination

of data analysis, testing, expert review, and use of bias detection tools.
By employing these approaches, organizations can ensure that their
algorithms are producing accurate and fair results and promote fairness,
diversity, and inclusivity in their decision-making processes.
What Do Bias Tools Do?

Bias detection tools are software applications that use a variety of
techniques to identify and mitigate bias in AI and ML algorithms. These
tools typically work by analyzing the data used to train the algorithm,
evaluating the algorithm’s outputs, and providing feedback on potential
biases and ways to mitigate them.
These are some common techniques used by bias detection tools:
• Data analysis: Many bias detection tools use data
analysis techniques to identify patterns and biases in
the data used to train the algorithm. These tools may
look for correlations between different variables and
evaluate the quality and diversity of the data.
134
• Machine learning: Some bias detection tools use

machine learning algorithms to train models that can
detect bias in other algorithms. These models are
trained on large datasets and can identify patterns and
biases that may be difficult to detect manually.
• Statistical analysis: Statistical analysis techniques can

also be used to detect biases in AI and ML algorithms.
These techniques include regression analysis,
hypothesis testing, and other statistical methods that
can identify patterns and relationships in the data.
• Fairness metrics: Some bias detection tools use

fairness metrics to evaluate the algorithm’s outputs and
identify potential biases. Fairness metrics measure the
impact of the algorithm on different groups and can
help identify where bias may be present.
• Human review: Finally, many bias detection tools

incorporate human review into the process. This may
involve having experts in the relevant field review the
algorithm’s outputs and provide feedback on potential
biases and ways to mitigate them.
Overall, bias detection tools use a variety of techniques to identify and
mitigate bias in AI and ML algorithms. These tools can help organizations
ensure that their algorithms are producing fair and accurate results and
promote fairness, diversity, and inclusivity in their decision-making
processes.
135
Incorporation of Feedback Through

Human Intervention
Incorporating feedback through human intervention in AI and ML
algorithms can be done in several ways, depending on the specific algorithm
and the nature of the feedback. Here are some common approaches:
• User feedback: One way to incorporate feedback is

to collect feedback from users of the algorithm. This
can be done through surveys, user testing, or other
methods. User feedback can provide valuable insights
into how the algorithm is performing and where
improvements may be needed.
• Expert review: Another way to incorporate feedback

is to have experts in the relevant field review the
algorithm and provide feedback. These experts
can provide feedback on the algorithm’s design,
performance, and accuracy, as well as suggest ways to
improve its performance.
• Human-in-the-loop systems: Some AI and ML

algorithms incorporate a “human-in-the-loop”
system, where human intervention is built into the
algorithm itself. For example, the algorithm may flag
certain outputs for human review or require human
intervention for certain decisions.
• Continuous monitoring: Finally, continuous

monitoring of the algorithm’s outputs can also
provide valuable feedback. By tracking the algorithm’s
performance over time, organizations can identify
areas where the algorithm may be biased or where
improvements are needed.
136
Incorporating feedback through human intervention can be useful in

several ways.
• Improving accuracy: By incorporating feedback from

users and experts, organizations can identify areas
where the algorithm may be inaccurate or biased and
make improvements to increase its accuracy.
• Enhancing fairness: Feedback from users and experts

can also help organizations identify areas where the
algorithm may be unfair or discriminatory and make
changes to promote fairness and inclusivity.
• Increasing trust: Incorporating feedback from users

and experts can help increase trust in the algorithm
and the organization using it. When users feel that their
feedback is being listened to and incorporated into the
algorithm, they may be more likely to trust its outputs.
Overall, incorporating feedback through human intervention can help

organizations improve the accuracy, fairness, and trustworthiness of their
AI and ML algorithms, and promote fairness, diversity, and inclusivity in
their decision-making processes.
Human-in-the-loop (HITL) is a design pattern for AI and ML systems
where a human is involved in the decision-making process alongside the
algorithm. HITL systems are designed to combine the strengths of both
humans and machines, with the human providing the judgment and
expertise and the machine handling the more tedious and repetitive tasks.
In HITL systems, the algorithm provides a recommendation or output,
which is then reviewed or modified by a human operator. The human
operator can either approve the algorithm’s recommendation or modify it
based on their own judgment and expertise. The modified output is then
used to train the algorithm, improving its performance over time.
137
HITL systems can be used in a wide range of applications, including

image recognition, natural language processing, and decision-making
systems. For example, in an image recognition system, the algorithm may
identify objects in an image, which are then reviewed by a human operator
to ensure their accuracy. In a decision-making system, the algorithm may
provide a recommendation, which is then reviewed by a human operator
to ensure its fairness and ethicality.
HITL systems can be useful in several ways.
• Improved accuracy: By combining the strengths

of both humans and machines, HITL systems can
improve the accuracy of the algorithm’s outputs.
Humans can provide judgment and expertise, which
can help the algorithm make better decisions.
• Increased transparency: HITL systems can increase

transparency in the decision-making process. By
involving a human operator, the decision-making
process becomes more explainable, making it easier
to understand how the algorithm arrived at its
recommendation.
• Better adaptability: HITL systems are more adaptable

than fully automated systems. When faced with new or
unexpected situations, human operators can provide
judgment and expertise that the algorithm may not
possess, allowing the system to handle a wider range of
scenarios.
Overall, HITL systems are a powerful tool for improving the accuracy,
transparency, and adaptability of AI and ML systems and can help
organizations ensure that their algorithms are making fair and ethical
decisions.
138
How to Build HITL Systems?

Building a human-in-the-loop system involves designing an AI or ML
system that incorporates human input and feedback into its decision-
making process. Here are some steps to consider when building a
HITL system:
1. Identify the problem: The first step in building

an HITL system is to identify the problem you
want to solve. This could be anything from image
recognition to natural language processing or
decision-making.
2. Choose the right algorithm: Once you have

identified the problem, you need to choose the
right AI or ML algorithm to solve it. Consider the
strengths and weaknesses of different algorithms,
and choose one that is well-suited to your specific
problem.
3. Define the human-in-the-loop process: Next, you

need to define the process for incorporating human
input and feedback into the algorithm. This could
involve defining the types of decisions that require
human review, setting up a user interface for human
input, or defining the criteria for accepting or
rejecting human input.
4. Develop the user interface: Depending on the

specific HITL process you have defined, you may
need to develop a user interface for humans to
interact with the algorithm. This could involve
building a web application or mobile app or
integrating with an existing software tool.
139
5. Collect and incorporate human feedback: Once

your HITL system is up and running, you can begin
collecting human feedback and incorporating it
into the algorithm. This could involve collecting
feedback through surveys or user testing or
providing human operators with specific criteria for
modifying the algorithm’s output.
6. Continuously monitor and refine the system: As

you collect feedback and incorporate it into the
algorithm, it’s important to continuously monitor
the system’s performance and refine it over time.
This could involve updating the algorithm to
address common errors or biases or modifying the
HITL process to improve efficiency or accuracy.
Building an HITL system can be a complex and iterative process,

but it can be a powerful tool for improving the accuracy, transparency,
and fairness of AI and ML systems. By incorporating human input and
feedback into the decision-making process, organizations can ensure that
their algorithms are making ethical and fair decisions that align with their
values and goals.
Let’s understand this more through an example.
Example: Customer Churn

Churn prediction is a common problem in industries such as
telecommunications, where customers may cancel their services after a
certain period of time. In this context, an action could be to proactively
reach out to customers who are at risk of churning in order to retain their
business. However, there may be biases in the algorithm that determines
140
which customers are at risk of churning, such as an overemphasis on

a certain demographic or a failure to consider certain factors that may
influence customer loyalty.
To address these biases, a human-in-the-loop system could
be implemented. In this system, the algorithm would provide a
recommendation for which customers are at risk of churning, but this
recommendation would be reviewed by a human operator before any
action is taken. The human operator could review the algorithm’s output,
assess its fairness and accuracy, and modify the recommendation based
on their judgment and expertise.
For example, the algorithm may identify a certain demographic as
being at a higher risk of churning, but the human operator may notice that
this demographic is actually more loyal than the algorithm suggests. The
human operator could modify the recommendation to focus on a different
demographic or to take additional factors into account that the algorithm
did not consider.
By incorporating human input and feedback into the churn prediction
process, the HITL system can improve the accuracy, fairness, and
ethicality of the algorithm’s recommendations. This can lead to better
outcomes for the organization, such as increased customer retention and
higher revenue, while avoiding biases that could harm certain groups of
customers or lead to unfair decision-making.
Conclusion
In this chapter, we discussed how cognitive biases are a prevalent issue
in AI and ML and can have significant negative impacts on organizations
and individuals. To address this problem, human-in-the-loop systems
can be implemented to incorporate human input and feedback into the
decision-making process of algorithms. HITL systems can help detect and
mitigate biases, improve the accuracy and fairness of AI and ML systems,
141
and ensure that they align with the values and goals of the organization.
Building an HITL system involves identifying the problem, choosing
the right algorithm, defining the HITL process, developing the user
interface, collecting and incorporating human feedback, and continuously
monitoring and refining the system. By using HITL systems, organizations
can make more informed and ethical decisions that benefit both the
organization and its stakeholders.
142
CHAPTER 7
Case Studies
This chapter will explore how decision intelligence can be integrated into
organizations through a series of case studies. We will delve into various
methodologies and examine how they can produce actionable insights for
decision-makers, both independently and when used in conjunction with
other methods.
While we have already covered many methodologies in prior chapters
and demonstrated how they can be applied to decision intelligence, it
is uncommon to find a single methodology that can effectively solve a
decision intelligence problem. The problems are often so intricate that
employing a combination of methodologies is necessary to achieve the
desired outcome.
Let’s explore how AI/ML along with other advanced techniques can
help in decision-making through two case studies: one classification and
another regression.
ase Study 1: Telecom Customer

C
Churn Management
Problem Statement: A telecom company is experiencing high customer
churn rates and is looking for innovative AI solutions to reduce churn. The
company wants to leverage machine learning along with other techniques
to do the following:

Chapter 7 Case Studies
• Monitor customer behavior and identify early warning

signs of churn
• Identify the factors that lead customers to churn
• Create personalized solutions for retaining the

customers and run simulations
• Update the solutions based on the simulation output
Solution:
The AI team proposes a hybrid solution to solve the given problem.
The plan is to first build a machine learning model that takes historical
customer details such as account information, demographics, and
services used, along with churn status. The model output will be used to
find potential customers who are likely to churn in the coming time. The
team also proposes a counterfactual generation engine that can provide
various scenarios wherein the predicted outcomes change. For example, if
a customer is likely to churn and their churn probability is 90 percent, the
counterfactual engine can generate change in input features that might
bring the churn probability for the given customer to maybe 40 percent.
This change in the input parameters can be used as part of their retention
strategy. A third component of a what-if analysis solution can help
decision-makers tweak the counterfactual suggestions based on certain
constraints and check whether those changes can still retain customers.
The AI team suggests following benefits of using the proposed hybrid
solution:
• Improved accuracy: A machine learning model can

analyze large volumes of data and identify patterns and
insights that are not easily visible to human analysts.
Counterfactuals and what-if analysis can be used to
test and validate the accuracy of the machine learning
models and get ideas on retaining the customers who
are likely to churn, leading to higher customer retention.
144
• Personalized solutions: Machine learning can

analyze customer behavior and preferences to offer
personalized solutions that can help retain customers.
Counterfactuals and what-if analysis can be used to
test different scenarios and predict how customers will
respond to different solutions, leading to more effective
personalized solutions.
• Faster decision-making: Machine learning can

process data in real time and offer insights that can
help decision-makers make faster and more informed
decisions. Counterfactuals and what-if analysis can
be used to test different scenarios and predict how
different decisions will impact customer churn,
enabling decision-makers to make more informed
decisions more quickly.
• Cost-effective: The proposed solution can automate

many tasks that would otherwise require human
analysts, leading to cost savings for the company.
By automating routine tasks, such as data analysis
and report generation, human analysts can focus on
more complex tasks that require human intuition and
expertise.
• Scalability: The solution can handle large volumes of

data and scale as the customer base grows. This can be
particularly useful for a company that is experiencing
rapid growth and needs a solution that can keep up
with its expanding customer base.
145
• Continuous improvement: The solution can learn

from new data and continuously improve their
predictions and recommendations. This can lead to
better outcomes over time as the decision intelligence
system becomes more sophisticated and accurate.
• Revenue growth: Retaining current customers can

contribute to revenue growth, as satisfied customers
are more likely to purchase additional products and
services.
• Competitive advantage: Successfully predicting

and mitigating customer churn can set the company
apart from its competitors. By delivering exceptional
customer experiences and satisfying customer
demands, it can establish a devoted customer base and
stand out in the marketplace.
Dataset Details:
The dataset on customer churn pertains to a hypothetical telecom
business that operated in California during the third quarter and
encompasses information on 7,043 customers who had either churned,
retained, or acquired the company’s services. The dataset further features
various crucial demographic variables for each customer.
Demographics:
• CustomerID: A unique ID that identifies each

customer.
• Gender: The customer’s gender: Male, Female.
• Senior Citizen: Indicates if the customer is 65 or older:

Yes, No.
• Partner: Indicates if the customer has a partner:

Yes, No.
146
• Dependents: Indicates if the customer lives with any

dependents: Yes, No. Dependents could be children,
parents, grandparents, etc.
Services:
• Tenure in Months: Indicates the total number of

months that the customer has been with the company
by the end of the quarter.
• Phone Service: Indicates if the customer subscribes to

home phone service with the company: Yes, No.
• Multiple Lines: Indicates if the customer subscribes to

multiple telephone lines with the company: Yes, No.
• Internet Service: Indicates if the customer subscribes

to Internet service with the company: No, DSL, Fiber
Optic, Cable.
• Online Security: Indicates if the customer subscribes

to an additional online security service provided by the
company: Yes, No.
• Online Backup: Indicates if the customer subscribes

to an additional online backup service provided by the
company: Yes, No.
• Device Protection Plan: Indicates if the customer

subscribes to an additional device protection plan for
their Internet equipment provided by the company:
Yes, No.
• Premium Tech Support: Indicates if the customer

subscribes to an additional technical support plan from
the company with reduced wait times: Yes, No.
147
• Streaming TV: Indicates if the customer uses their

Internet service to stream television programming from
a third party provider: Yes, No. The company does not
charge an additional fee for this service.
• Streaming Movies: Indicates if the customer uses their

Internet service to stream movies from a third party
provider: Yes, No. The company does not charge an
additional fee for this service.
• Contract: Indicates the customer’s current contract

type: Month-to-Month, One Year, Two Year.
• Paperless Billing: Indicates if the customer has chosen

paperless billing: Yes, No.
• Payment Method: Indicates how the customer pays

their bill: Bank Withdrawal, Credit Card, Mailed Check.
• Monthly Charge: Indicates the customer’s current total

monthly charge for all their services from the company.
• Total Charges: Indicates the customer’s total charges,

calculated to the end of the quarter.
Churn Status:
• Churn Label: Yes = the customer left the company

this quarter. No = the customer remained with the
company. This is directly related to the churn value.
Source: https://community.ibm.com/community/user/
businessanalytics/blogs/steven-macko/2019/07/11/telco-customer-
churn-1113
Stage 1: AI Model Creation

The goal of this activity is to build a machine learning model that can
fairly predict the customer churn. Although advanced data processing
148
techniques and ML algorithms are available, we will use the ones that give
satisfactory results to ensure we are spending more efforts toward the goal,
i.e., building a decision intelligence system.
[In]:
import dice_ml
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import RandomizedSearchCV,
train_test_split
from sklearn.metrics import accuracy_score, precision_score,
recall_score, f1_score
Step 2: Loading the telecom churn dataset
[In]:
churn_data_all = pd.read_csv('telco_churn.csv')
Step 3: Checking the data characteristics to see whether there

are any discrepancies in the data (e.g., missing values, wrong data
types, etc.)
[In]:
churn_data_all.info()
[Out]:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 7043 entries, 0 to 7042
Data columns (total 21 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 customerID 7043 non-null object
149
1 gender 7043 non-null object
2 SeniorCitizen 7043 non-null int64
3 Partner 7043 non-null object
4 Dependents 7043 non-null object
5 tenure 7043 non-null int64
6 PhoneService 7043 non-null object
7 MultipleLines 7043 non-null object
8 InternetService 7043 non-null object
9 OnlineSecurity 7043 non-null object
10 OnlineBackup 7043 non-null object
11 DeviceProtection 7043 non-null object
12 TechSupport 7043 non-null object
13 StreamingTV 7043 non-null object
14 StreamingMovies 7043 non-null object
15 Contract 7043 non-null object
16 PaperlessBilling 7043 non-null object
17 PaymentMethod 7043 non-null object
18 MonthlyCharges 7043 non-null float64
19 TotalCharges 7032 non-null float64
20 Churn 7043 non-null object
dtypes: float64(2), int64(2), object(17)
memory usage: 1.1+ MB

All columns except the TotalCharges do not have any null values. The
TotalCharges column needs to be treated with some appropriate missing
value strategy. Let’s look at the rows where the column has null values.
[In]:
churn_data_all[churn_data_all['TotalCharges'].isnull()].head()
[Out]:
150
customerID gender SeniorCitizen Partner Dependents tenure Phone MultipleLines Total Churn
Service Charges
4472-LVYGI Female 0 Yes Yes 0 No No phone No

service
3115-CZMZD Male 0 No Yes 0 Yes No No
5709-LVOEQ Female 0 Yes Yes 0 Yes No No
4367-NUYAO Male 0 Yes Yes 0 Yes Yes No
1371-DWPAZ Female 0 Yes Yes 0 No No phone No
service
7644-OMVMY Male 0 Yes Yes 0 Yes No No
3213-VVOLG Male 0 Yes Yes 0 Yes Yes No
2520-SGTTA Female 0 Yes Yes 0 Yes No No
2923-ARZLG Male 0 Yes Yes 0 Yes No No
4075-WKNIU Female 0 Yes Yes 0 Yes Yes No
Chapter 7
2775-SEFEE Male 0 No Yes 0 Yes Yes No
151
Case Studies
(Some columns are not shown in the previous table to fit within the
viewable window.)
From the previous table, we find that the TotalCharges column is null
when the tenure is 0. This means the customers who have joined recently
and haven’t even completed 1 month with the telecom operator will have
their TotalCharges as null. Let’s convert the null value to 0, as that would
be the most appropriate value.
[In]:
churn_data_all['TotalCharges'] = np.where(churn_
data_all['TotalCharges'].isnull(), 0, churn_data_
all['TotalCharges'])
While the SeniorCitizen column contains binary values, its encoding

differs from that of other attributes. Therefore, we convert it to a format
that conforms to a standard representation for all variables.
[In]:
churn_data_all['SeniorCitizen'] = np.where(churn_data_all.
SeniorCitizen == 1,"Yes","No")
We create a list of column names based on their type and use that.
[In]:
all_columns = [x for x in churn_data_all.drop('customerID',
axis = 1).columns]
id_column = ['customerID']
target_column = ['Churn']
categorical_columns = [y for y in churn_data_all.
drop('customerID', axis = 1).select_dtypes(include = [object]).
columns]
numeric_columns = [z for z in all_columns if z not in
categorical_columns]
[In]:
152
get_dummies = []
label_encoding = []
for i in categorical_columns:
print('Column Name:', i, ', Unique Value Counts:',
len(churn_data_all[i].unique()), ', Values:', churn_data_
all[i].unique())
if len(churn_data_all[i].unique()) > 2:
get_dummies.append(i)
else:
label_encoding.append(i)
[Out]:
Column Name: gender , Unique Value Counts: 2 , Values:
['Female' 'Male']
Column Name: SeniorCitizen , Unique Value Counts: 2 , Values:
['No' 'Yes']
Column Name: Partner , Unique Value Counts: 2 , Values:
['Yes' 'No']
Column Name: Dependents , Unique Value Counts: 2 , Values:
['No' 'Yes']
Column Name: PhoneService , Unique Value Counts: 2 , Values:
['No' 'Yes']
Column Name: MultipleLines , Unique Value Counts: 3 , Values:
['No phone service' 'No' 'Yes']
Column Name: InternetService , Unique Value Counts: 3 , Values:
['DSL' 'Fiber optic' 'No']
Column Name: OnlineSecurity , Unique Value Counts: 3 , Values:
['No' 'Yes' 'No internet service']
Column Name: OnlineBackup , Unique Value Counts: 3 , Values:
['Yes' 'No' 'No internet service']
Column Name: DeviceProtection , Unique Value Counts: 3 ,
Values: ['No' 'Yes' 'No internet service']
153
Column Name: TechSupport , Unique Value Counts: 3 , Values:

Column Name: StreamingTV , Unique Value Counts: 3 , Values:
Column Name: StreamingMovies , Unique Value Counts: 3 , Values:
Column Name: Contract , Unique Value Counts: 3 , Values:
['Month-to-month' 'One year' 'Two year']
Column Name: PaperlessBilling , Unique Value Counts: 2 ,
Values: ['Yes' 'No']
Column Name: PaymentMethod , Unique Value Counts: 4 , Values:
['Electronic check' 'Mailed check' 'Bank transfer (automatic)'
'Credit card (automatic)']
Column Name: Churn , Unique Value Counts: 2 , Values:
['No' 'Yes']
We can see that some categorical columns have two unique values,
whereas some have more than two. To apply appropriate techniques, we
have split the columns.
We apply dummy variable creation techniques to columns having
more than two unique values.
[In]:
churn_data_all_dl = pd.get_dummies(churn_data_all, prefix=get_
dummies, columns=get_dummies)
We apply a label encoding technique to the columns with two unique

values and save the mappings.
mappings = {}
for col in label_encoding:
le = LabelEncoder()
154
churn_data_all_dl[col] = le.fit_transform(churn_data_all_
dl[col])
mappings[col] = dict(zip(le.classes_,range(len(le.
classes_))))
mappings
[Out]:
{'gender': {'Female': 0, 'Male': 1},
'SeniorCitizen': {'No': 0, 'Yes': 1},
'Partner': {'No': 0, 'Yes': 1},
'Dependents': {'No': 0, 'Yes': 1},
'PhoneService': {'No': 0, 'Yes': 1},
'PaperlessBilling': {'No': 0, 'Yes': 1},
'Churn': {'No': 0, 'Yes': 1}}
Step 5: Modeling
We split the dataset for training and inference.
[In]:
X = churn_data_all_dl.drop(['customerID', 'Churn'], axis=1)
y = churn_data_all_dl['Churn']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_
size = 0.2, stratify=y, random_state = 0)
Hyperparameter Tuning
The process of hyperparameter tuning entails identifying the best
hyperparameter combination for a learning algorithm, which can be
employed to optimize its performance on any given dataset. By minimizing
a prespecified loss function, the selected hyperparameters can reduce
errors and improve the model’s results.
155
Running the following lines of code aids in determining the ideal

hyperparameter for our machine learning algorithm:
[In]:
n_estimators = [int(x) for x in np.linspace(start = 100,
stop = 2100, num = 6)]
feature_name = list(X_test.columns)
max_depth = [int(x) for x in np.linspace(10, 100, num = 5)]
max_depth.append(None)
min_samples_split = [2, 5, 10]
min_samples_leaf = [1, 2, 4, 6, 8, 10]
random_grid = {'n_estimators':n_estimators,
'max_depth': max_depth,
'min_samples_split': min_samples_split,
'min_samples_leaf': min_samples_leaf}
print(random_grid)
rf = RandomForestClassifier()
rf_random = RandomizedSearchCV(estimator = rf, param_
distributions = random_grid, n_iter = 10, cv = 2, verbose=2,
random_state=42, n_jobs = -1)
rf_random.fit(X_train[feature_name], y_train)
print(rf_random.best_params_)
[Out]:
{'n_estimators': [100, 500, 900, 1300, 1700, 2100], 'max_
depth': [10, 32, 55, 77, 100, None], 'min_samples_split':
[2, 5, 10], 'min_samples_leaf': [1, 2, 4, 6, 8, 10]}
We fit 2 folds for each of 10 candidates, totaling 20 fits.
{'n_estimators': 100, 'min_samples_split': 2, 'min_samples_

leaf': 8, 'max_depth': 10}
156
We see that for the random forest model (the algorithm of choice in
this case), the best combination of hyperparameters is as follows:
• n_estimators = 100
• min_samples_split = 2
• min_samples_leaf = 8
• max_depth = 10
We use the previous hyperparameters to build the random

forest model.
[In]:
feature_name = list(X_test.columns)
churn_classifier=RandomForestClassifier(n_estimators=100,min_
samples_split=2,min_samples_leaf=8,max_depth=10)
churn_classifier.fit(X_train[feature_name],y_train)
[Out]:
RandomForestClassifier(bootstrap=True, ccp_alpha=0.0, class_
weight=None,
criterion='gini', max_depth=10, max_
features='auto',
max_leaf_nodes=None, max_samples=None,
min_impurity_decrease=0.0, min_impurity_
split=None,
min_samples_leaf=8, min_samples_split=2,
min_weight_fraction_leaf=0.0,
n_estimators=100,
n_jobs=None, oob_score=False, random_
state=None,
verbose=0, warm_start=False)
Now, we use the trained model for making predictions on the test data
and save them along with the prediction probabilities.
157
[In]:
pred_df = X_test.copy()
pred_df['Churn'] = y_test
pred_df['pred'] = churn_classifier.predict
(X_test[feature_name])
prediction_of_probability = churn_classifier.predict_proba
(X_test[feature_name])
pred_df['prob_0'] = prediction_of_probability[:,0]
pred_df['prob_1'] = prediction_of_probability[:,1]
We evaluate the model’s performance.
[In]:
print("Accuracy: ", (accuracy_score(pred_df['Churn'],
pred_df['pred']))*100)
print("Precision: ", (precision_score(pred_df['Churn'],
pred_df['pred']))*100)
print("Recall: ", (recall_score(pred_df['Churn'], pred_
df['pred']))*100)
print("F1 Score: ", (f1_score(pred_df['Churn'], pred_
df['pred']))*100)
[Out]:
Accuracy: 80.48261178140525
Precision: 67.1280276816609
Recall: 51.87165775401069
F1 Score: 58.521870286576174
As mentioned, we are aiming for satisfactory model performance

for the demonstration. There are multiple ways to get better model
performance; these are some of the prominent ones:
• Feature engineering: This involves creating new

features or transforming existing features to make them
more informative for the model.
158
• Regularization: This involves adding a penalty term to

the loss function to prevent overfitting of the model to
the training data.
• Ensemble methods: This involves combining multiple

models to create a more accurate prediction.
• Data augmentation: This involves increasing the size

of the training dataset by generating new examples
using techniques such as rotating, flipping, or adding
noise to the data.
• Oversampling/undersampling: This involves

balancing imbalanced datasets by either increasing
the number of instances in the minority class
(oversampling) or decreasing the number of instances
in the majority class (undersampling).
• Transfer learning: This involves leveraging a

pretrained model to solve a similar problem or using a
pretrained model as a feature extractor.
• Gradient clipping: This involves clipping the gradients

during training to prevent them from becoming too
large or too small.
• Early stopping: This involves stopping the training

process when the model’s performance on the
validation set stops improving.
Now that we have the model and predictions ready, let’s move on and
start using them for decision-making.
Stage 2: Generating Counterfactuals

We will be using a combination of counterfactual and what-if analysis
to drive decision intelligence. Let’s look into what they mean and how they
can be used.
159
Counterfactual Analysis:
Counterfactuals refer to a type of analysis that involves identifying
the minimal set of changes required to an input instance such that the
machine learning model output changes to a desired outcome. In simpler
terms, counterfactuals are hypothetical examples of inputs that would lead
to a different output from a machine learning model.
Counterfactual analysis can be used to explain the behavior of a
machine learning model and to provide insights into how the model
can be improved as well as what actions can be taken so that favorable
outcomes can be generated. It can also be used for various applications
such as fairness analysis, where the minimal set of changes required to
achieve a desired outcome can be used to identify potential sources of bias
in the model. Additionally, counterfactual analysis can be used for causal
inference, where the minimal set of changes can be used to estimate the
causal effect of a treatment on an outcome.
We will be using the DiCEML Python package for generating
counterfactuals.
DiCELML
DiCEML (Diverse Counterfactual Explanations through Mixed Integer
Linear Programming) is a framework that enables the generation of
counterfactual examples for machine learning models. DiCEML employs
mixed-integer linear programming to find a minimal set of changes to
input instances required to change the model output to a desired outcome.
It is a state-of-the-art method for generating diverse counterfactual
explanations that take into account the constraints and features of the
input data and can be used for various applications such as explainable AI,
fairness analysis, and causal inference.
Step 6: Building a counterfactual explainer object
[In]:
def initialize_counterfactuals(
160
train_df, model, feature_list, continuous_features, target,

model_type
):
"""
Initialize Counterfactual explainer object for the given
input model and also the feature range dictionary
(both are used in calculating Counterfactuals in Local
Explainability)
Args:
train_df (dataframe) : train dataframe
model (object) : input model (to explain)
feature_list (list) : list of features used in model
continuous_features (list) : list of continuous
features
target (str) : target column name
model_type (str) : classification or regression
Returns:
explainer (object) : counterfactual explainer object
feature_range (dict) : dictionary containing
permissible feature ranges for continuous features
"""
df_model = train_df[feature_list + [target]]
####Round the decimal upto 4 digits

df_model[continuous_features] = (
df_model[continuous_features].astype(float).round(4)
) # .astype(str)
print("continuous features", continuous_features)
#### dice data initialisations

data_dice = dice_ml.Data(
161
dataframe=df_model, # For perturbation strategy

continuous_features=continuous_features,
outcome_name=target,
)
## dice model initialisation

if model_type == "regression" or model_type == "time
series":
model_dice = dice_ml.Model(model=model,
backend="sklearn", model_type="regressor")
else:
model_dice = dice_ml.Model(model=model,
backend="sklearn")
# getting model and data together

explainer = dice_ml.Dice(data_dice, model_dice,
method="random")
df_model.drop(target, axis=1, inplace=True)
feature_range = {}
for i in continuous_features:
feature_range[i] = [
df_model[i].astype(float).min(),
df_model[i].astype(float).max(),
]
return explainer, feature_range
def preprocess_encode(df,mappings,feature_list):
if mappings != None:
encode_df = df[feature_list]
cat_cols = (encode_df.dtypes == object)
cat_cols = list(cat_cols[cat_cols==True].index)
162
encode_df = df.copy()
if len(cat_cols)>0:
encode_df.update(encode_df[cat_cols].apply(lambda
col: col.map(mappings[col.name])).astype(int))
encode_df[cat_cols] = encode_df[cat_cols].
astype(int)
else:
encode_df = df.copy()
return encode_df
def postprocess_decode(df,mappings):
decoded_output = df.copy()
if mappings != None:
inv_mapping_dict = {cat: {v: k for k, v in map_dict.
items()} for cat, map_dict in mappings.items()}
decoded_output[list(mappings.keys())] = decoded_
output[list(mappings.keys())].astype(int)
decoded_output = decoded_output.replace(inv_
mapping_dict)
return decoded_output
def generate_output_counterfactuals(explainer,feature_
range,query_instance,model_type,model,feature_list,desired_
class_or_range,continuous_features,features_vary='all',num_
cf=3,mappings={}):
"""
Generating the counterfactuals output for a particular
query_instance (record), given various other inputs like
features to vary, etc.
Args:
explainer (object) : Counterfactuals explainer object
163
feature_range (dict) : dictionary containing

permissible feature ranges for continuous features
query_instance (dataframe) : selected record to explain
(user input)
model_type (str) : classification or regression
model (object) : model object, required only in the
case of classification
feature_list (list) : list of features used in model
desired_class_or_range (list) : if classification then
desired output class; if regression then desired output
range (user input)
continuous_features (list) : list of continuous
features
features_vary (list) : list of features that can be
varied (user input)
num_cf (int) : number of counterfactuals to generate
mappings (dict) : encoder mapping dictionary
Returns:
cf_output (dataframe) : counterfactuals output
dataframe
"""
query_instance[continuous_features] = query_
instance[continuous_features].astype(float).round(4)
query_instance_processed = preprocess_encode(query_
instance,mappings,list(query_instance.columns))
display(query_instance_processed)
if model_type == 'classification':
cf_exp = explainer.generate_counterfactuals(query_
instance_processed, total_CFs=20,desired_class=desired_
class_or_range,permitted_range=feature_range,features_
to_vary=features_vary)
164
else:
cf_exp = explainer.generate_counterfactuals(query_
instance_processed, total_CFs=num_cf,desired_
range=desired_class_or_range,permitted_range=feature_
range,features_to_vary=features_vary)
cf_exp_df = cf_exp.cf_examples_list[0].final_cfs_df
if model_type == "classification":

cf_exp_df["Probability"] = model.predict_proba(cf_exp_
df[feature_list])[:, desired_class_or_range]
cf_exp_df = (cf_exp_df.sort_values("Probability",
ascending=False).reset_index(drop=True).head(num_cf))
cf_exp_df = postprocess_decode(cf_exp_df,mappings)
cf_output = cf_ouput_df_fn(cf_exp_df,query_instance)
return cf_output
def cf_ouput_df_fn(df,org_df):
newdf = df.values.tolist()
org = org_df.values.tolist()[0]
for ix in range(df.shape[0]):
for jx in range(len(org)):
if str(newdf[ix][jx]) == str(org[jx]):
newdf[ix][jx] = '-'
else:
newdf[ix][jx] = str(newdf[ix][jx])
return pd.DataFrame(newdf, columns=df.columns,
index=df.index)
features = [x for x in X_train.columns]
train_df = X_train.copy()
train_df["Churn"] = y_train
165
model_type='classification'
target = 'Churn'
explainer, feature_range = initialize_counterfactuals(
train_df, churn_classifier, features, numeric_columns,
target, model_type
)
[Out]:
continuous features ['tenure', 'MonthlyCharges',
'TotalCharges']
Stage 3: What-If Analysis

Step 7: Using the explainer object to generate possible strategies to
retain customers
Now we have the explainer counterfactual object that can be used
to produce counterfactual explanations, i.e., what changes in the inputs
can be made to get the desired output. In our case, it would be: “For the
customers who are likely to churn, what can be changed (services, account
related, demographics) so that he/she is retained.” Let’s understand this by
looking at an example.
The following are the inputs DiCEML expects:
• The trained machine learning model.
• The input instance for which we want to generate a

counterfactual. We will take a sample user from a pool
of users who are predicted to be churning.
• The features to vary. For this example, we will see if

providing services such as paperless billing, online
security, online backup, device protection, tech
support, Internet services for streaming TV or movies,
one- or two-year contract, or if changing the monthly
plan or a combination of features can change the
customer’s mind and they might plan to stay with the
telecom operator.
166
• The desired output or the target class for which we

want to generate a counterfactual. In this case, it will be
nonchurn, i.e., 0 (from 1).
Here’s how it works:
[In]:
test_df = X_test.copy()
test_df["Churn"] = y_test
test_df_churners = pred_df[pred_df['Churn'] == 1].sort_

values(by=['prob_1'], ascending=False)
test_df_churners_input = test_df_churners.drop(['Churn',
'pred', 'prob_0', 'prob_1'], inplace=False, axis=1)
test_df_churners_input_record = test_df_churners_input[61:62]
features_vary = ['PaperlessBilling', 'MonthlyCharges',

'OnlineSecurity_Yes', 'OnlineBackup_Yes',
'DeviceProtection_Yes', 'TechSupport_Yes',
'StreamingTV_Yes', 'StreamingMovies_Yes',
'Contract_One year', 'Contract_Two year']
dice_exp = explainer.generate_counterfactuals(test_df_churners_
input_record, total_CFs=4, desired_class="opposite", features_
to_vary=features_vary)
dice_exp.visualize_as_dataframe()
[Out]:
100%|██████████| 1/1 [00:31<00:00, 31.02s/it]
[In]:
test_df_churners_input_record[features_vary]
[Out]:
167
168
Chapter 7
Paperless Monthly Online Online Device Tech Streaming Streaming Contract_ Contract_
Billing Charges Security_ Backup_Yes Protection_Yes Support_Yes TV_Yes Movies_ One year Two year
Yes Yes
0 101.15 0 0 1 0 1 1 0 0
Case Studies
[In]:
dice_exp.cf_examples_list[0].final_cfs_df[features_vary]
Paperless Monthly Online Online Device Tech Streaming Streaming Contract_ Contract_
Billing Charges Security_Yes Backup_Yes Protection_ Support_ TV_Yes Movies_ One year Two year
Yes Yes Yes
0 55.36 1 0 1 0 1 1 0 0
0 35.86 1 0 1 0 1 1 0 0
0 22.32 1 0 1 0 1 1 0 0
0 67.89 1 0 1 0 1 1 0 0
From the previous output, we see that the explainer model has
suggested four possible ways to retain the customer.
• Change monthly charges from 101.15 to 55.36 and

provide online security service.



However, there might be certain constraints for the business to take

into consideration, and one cannot rely only on the suggestions from
the counterfactual analysis. One might want to take the inputs from
the counterfactual analysis, tweak them a bit based on the business
constraints, and then check if those tweaked inputs generate the desired
output. This is possible through what-if analysis.
Step 8: Using what-if analysis for getting the right set of strategies to
retain customers (based on constraints)
What-if analysis is a technique used to understand the behavior of a
machine learning model under different scenarios, or what-if, situations.
It involves testing the model with different input values or scenarios to
understand how the model behaves in those situations.
Now suppose for the given counterfactual recommendations, we have
a constraint that the customer monthly charges cannot be reduced by
more than 10 percent, as it can result in large revenue losses. So, for the
given customer, we cannot bring the monthly charges below 91. Also, we
do not have any constraint on providing services for free, so let’s see with
the given constrained monthly charges and a combination of services if we
can retain the customer or not.
169
Options:
• Change monthly charges from 101.7 to 91; provide

online security and online backup services.
[In]:
custom_input = test_df_churners_input_record.copy()
custom_input['MonthlyCharges'] = custom_
input['MonthlyCharges'].replace(101.15, 91)
custom_input['OnlineBackup_Yes'] = custom_
input['OnlineBackup_Yes'].replace(0, 1)
custom_input['OnlineSecurity_Yes'] = custom_
input['OnlineSecurity_Yes'].replace(0, 1)
if churn_classifier.predict(custom_input)==0:
print("Successfully retained the customer")
else:
print("Sorry, customer could not be retained")
print("\nProbability of Churn:",churn_classifier.
predict_proba(custom_input)[0][1].round(2))
[Out]:
Sorry, customer could not be retained
Probability of Churn: 0.57
We see that with the given inputs, we cannot retain

the customer. Let’s try another set of inputs,
• Change monthly charges from 101.7 to 91; provide

online security and online backup services with a
2-year contract.
[In]:
custom_input2 = test_df_churners_input_record.copy()
170
custom_input2['MonthlyCharges'] = custom_
input2['MonthlyCharges'].replace(101.15, 91)
custom_input2['OnlineBackup_Yes'] = custom_
input2['OnlineBackup_Yes'].replace(0, 1)
custom_input2['OnlineSecurity_Yes'] = custom_
input2['OnlineSecurity_Yes'].replace(0, 1)
custom_input2['Contract_Two year'] = custom_
input2['Contract_Two year'].replace(0, 1)
if churn_classifier.predict(custom_input2)==0:
print("Successfully retained the customer")
else:
print("Sorry, customer could not be retained")
print("\nProbability of Churn:",churn_classifier.
predict_proba(custom_input2)[0][1].round(2))
[Out]:
Successfully retained the customer
Probability of Churn: 0.47
We can see that with the given set of inputs in this option that we
were successful in retaining the customer with a churn probability of the
customer coming down to 47 percent. So, this option can be converted to a
personalized offering to the customer.
Counterfactuals and “what-if” analysis can be used to detect bias in
machine learning models. It involves analyzing counterfactual outputs
or testing the model’s sensitivity to changes in the input data or model
parameters to identify areas where the model may be biased. Let’s see how
bias can be detected using counterfactual analysis + what-if for the same
previously used record, adding demographics SeniorCitizen and gender
into the features to vary the matrix.
171
[In]:
test_df_churners_bias_inputs = test_df_churners_input[(test_
df_churners_input['SeniorCitizen'] == 1) | (test_df_churners_
input['gender'] == 1)]
features_vary_bias = ['SeniorCitizen', 'gender']
dice_exp_bias = explainer.generate_counterfactuals(test_
df_churners_bias_inputs[100:130], total_CFs=4, desired_
class="opposite", features_to_vary=features_vary_bias)
dice_exp_bias.visualize_as_dataframe()
The previous counterfactual output, in a few instances, provides

recommendations to change either the gender or age (senior citizen to
nonsenior citizen) for favorable outcomes. This indicates there is a need
to check for bias in the model. It is important to note that these methods
are not foolproof for detecting and mitigating bias in AI systems, and it is
crucial to use a combination of techniques, data validation, auditing, and
diversity and inclusion practices to ensure AI systems produce equitable
outcomes. Fairness metrics such as disparate impact, equal opportunity
difference, false positive rate parity, equalized odds, and calibration
are used to assess whether an AI model is treating different groups of
individuals fairly, and the choice of metrics depends on the specific
context and goals. Disparate impact is one such measure of fairness that
examines whether the AI model’s outcomes have a disproportionate
impact on particular groups based on protected characteristics, and its
formula is used to calculate the disparate impact ratio for two groups being
compared. Let’s calculate the disparate impact value for the Gender and
SeniorCitizen columns in our dataset.
[In]:
protected_columns = ["gender", "SeniorCitizen"]
target="pred"
for pro_col in protected_columns:
categories = list(pred_df[pro_col].unique())
172
selection_rate_list = []
for cat in categories:
selection_rate = pred_df[(pred_df[target]==1)&(pred_
df[pro_col]==cat)].shape[0]/pred_df[pred_df[pro_
col]==cat].shape[0]
selection_rate_list.append(np.round(selection_rate,3))
disparate_impact = np.round(min(selection_rate_list)/
max(selection_rate_list),3)
print("Disparate Impact for",pro_col,":",disparate_impact)
[Out]:
Disparate Impact for gender : 0.843
Disparate Impact for SeniorCitizen : 0.438
The disparate impact score for the SeniorCitizen column is quite low,
suggesting a bias toward this feature, while the Gender column has a score
of about 84%, indicating a very fair treatment, which exceeds the threshold
of 80 percent. However, the SeniorCitizen score of about 44 percent is
significantly lower than the threshold value, indicating an unfair bias. It is
important to use multiple fairness metrics in conjunction with each other
to arrive at a conclusion.
ase Study 2: Mobile Phone Pricing/

C
Configuration Strategy
Problem Statement: A mobile phone manufacturing company wants
to improve its pricing strategy by utilizing the power of AI and machine
learning algorithms. The company has a large database of mobile phone
models with various features such as screen size, camera quality, processor
speed, battery life, and others. However, it currently lacks an accurate
and efficient method for predicting the prices of its mobile phones based
on these features. The company is seeking a solution that can help them
173
forecast prices for new phone models with a high degree of accuracy,
allowing them to optimize their pricing strategy and stay competitive in
the market.
Solution:
The AI team proposes a hybrid solution to solve the given problem.
The plan is to first build a machine learning model that takes historical
mobile details such as screen size, camera quality, processor speed,
battery life, and others along with the price. The model output will be
used to predict the potential price of the handset. The team also proposes
a counterfactual generation engine that can provide various scenarios
wherein the predicted outcomes change as per the expected price range.
For example, if the price of a handset is predicted to be X and if the
manufacturer wants to sell it at a price 10 percent to 15 percent more
than the predicted one, the engine will suggest changes in the phone’s
configuration to the expected price range. This change in the input
parameters can be used as part of product strategy. A third component
of a what-if analysis solution can help the decision-makers to tweak
the counterfactual suggestions based on certain constraints and check
whether those changes can still help them sell the phone at the expected
price range.
The AI team suggests the following benefits of using the proposed
hybrid solution for mobile pricing prediction:
• Accurate price predictions: AI algorithms can analyze

vast amounts of data and identify patterns that humans
may miss. This enables the tool to accurately predict
mobile handset prices based on various features with
greater accuracy.
• Real-time pricing: With an AI-driven tool, pricing

decisions can be made in real time based on changes
in market demand and supply. This allows the mobile
174
phone manufacturer to respond quickly to changing

market conditions and adjust its pricing strategies
accordingly.
• Improved profit margins: By accurately predicting

prices, the manufacturer can optimize its pricing
strategies to maximize profit margins. The tool
can identify the optimal price point for a product,
taking into account factors such as production costs,
competition, and market demand.
• Increased customer satisfaction: By setting the right

price for a product, the manufacturer can increase
customer satisfaction. Customers are more likely to
purchase a product if they feel that it is priced fairly and
accurately reflects its value.
• Better decision-making: An AI-driven tool can provide

the manufacturer with valuable insights into customer
preferences and buying behavior. This can help make
informed decisions about product development,
marketing, and pricing strategies.
Dataset Details:
These are the dataset details:
• Product_id: ID of each cell phone
• Price: Price of each cell phone
• Sale: Sales number
• weight: Weight of each cell phone
• resolution: Resolution of each cell phone
• ppi: Phone pixel density
• cpu core: Type of CPU core in each cell phone
175
• cpu freq: CPU frequency in each cell phone
• internal mem: Internal memory of each cell phone
• ram: RAM of each cell phone
• RearCam: Resolution of rear camera
• Front_Cam: Resolution of front camera
• battery: Battery capacity (in mA)
• thickness: Thickness of the phone
Source: https://www.kaggle.com/datasets/mohannapd/mobile-
price-prediction
Stage 1: AI Model Creation
The goal of this activity is to build a machine learning model that
can fairly predict mobile prices. Although advanced data processing
techniques and ML algorithms are available, we will use the ones that give
satisfactory results to ensure we are spending more efforts toward the goal,
i.e., building a decision intelligence system.
[In]:
import pandas as pd
import dice_ml
from dice_ml import Dice
from sklearn.metrics import r2_score
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split
Step 2: Loading the dataset
[In]:
all_data = pd.read_csv('cellphone_price_data.csv')
all_data.head()
[Out]:
176
Product_ Price Sale weight resolution ppi cpu cpu internal ram RearCam Front_ battery thickness
id core freq mem Cam
203 2357 10 135 5.2 424 8 1.35 16 3 13 8 2610 7.4

880 1749 10 125 4 233 2 1.3 4 1 3.15 0 1700 9.9
40 1916 10 110 4.7 312 4 1.2 8 1.5 13 5 2000 7.6
99 1315 11 118.5 4 233 2 1.3 4 0.512 3.15 0 1400 11
880 1749 11 125 4 233 2 1.3 4 1 3.15 0 1700 9.9
Chapter 7
177
Case Studies
Step 3: Checking the data characteristics to see whether there

are any discrepancies in the data (e.g., missing values, wrong data
types, etc.)
[In]:
all_data.info()
[Out]:
RangeIndex: 161 entries, 0 to 160
--- ------ -------------- -----
0 Product_id 161 non-null int64
1 Price 161 non-null int64
2 Sale 161 non-null int64
3 weight 161 non-null float64
4 resolution 161 non-null float64
5 ppi 161 non-null int64
6 cpu core 161 non-null int64
7 cpu freq 161 non-null float64
8 internal mem 161 non-null float64
9 ram 161 non-null float64
10 RearCam 161 non-null float64
11 Front_Cam 161 non-null float64
12 battery 161 non-null int64
13 thickness 161 non-null float64
dtypes: float64(8), int64(6)
memory usage: 17.7 KB
As we see from the previous details, the dataset does not have any
missing values or any issues related to the data. So, no data preprocessing
is required here, and we can directly move on to the modeling.
178
Step 4: Modeling
We split the dataset for training and inference.
[In]:
x=all_data.drop(['Price', 'Product_id'],axis=1)
y=all_data['Price']
x_train,x_test,y_train,y_test=train_test_split(x,y,test_
size=0.2,random_state=0)
Now, we build a random forest regressor model and use it to make

predictions on the test data, saving and evaluating them.
[In]:
rf = RandomForestRegressor()
rf.fit(x_train,y_train)
[Out]:
RandomForestRegressor()
[In]:
predictions = rf.predict(x_test)
print('R2 score for Training Data: ', rf.score(x_train,
y_train))
print('R2 score for Inference Data: ', r2_score(y_
test,predictions))
[Out]:
R2 score for Training Data: 0.9945696702046988
R2 score for Inference Data: 0.9678002631345382
[In]:
x_test['predicted_price'] = predictions
x_train.info()
[Out]:
Int64Index: 128 entries, 80 to 47
179
--- ------ -------------- -----
0 Sale 128 non-null int64
1 weight 128 non-null float64
2 resolution 128 non-null float64
3 ppi 128 non-null int64
4 cpu core 128 non-null int64
5 cpu freq 128 non-null float64
6 internal mem 128 non-null float64
7 ram 128 non-null float64
8 RearCam 128 non-null float64
9 Front_Cam 128 non-null float64
10 battery 128 non-null int64
11 thickness 128 non-null float64
dtypes: float64(8), int64(4)
memory usage: 13.0 KB
[In]:
x_train['Price'] = all_data['Price']
Stage 2: Generating Counterfactuals

We will be using a combination of counterfactual and what-if analysis
to drive decision intelligence. Unlike in the classification use case, we will
be using the expected price range of the mobile as the desired outcome as
this is a regression use case.
[In]:
d_mobile = dice_ml.Data(dataframe=x_train, continuous_
features=['Sale', 'weight', 'resolution', 'ppi', 'cpu core',
'cpu freq', 'internal mem', 'ram', 'RearCam', 'Front_Cam',
'battery', 'thickness'], outcome_name='Price')
m_mobile = dice_ml.Model(model=rf, backend="sklearn", model_
type='regressor')
explainer_mobile = Dice(d_mobile, m_mobile, method="random")

180
Stage 3: What-If Analysis

Step 5: Using the explainer object to generate possible pricing
strategies
Now we have the explainer counterfactual object that can be used to
produce counterfactual explanations, i.e., what changes in the inputs can
be made to get the desired output. In our case, it would be: “For the given
predicted price, what can be changed (mobile configurations like screen
size, resolution, etc.) so that the mobile can be sold at a higher price.” Let’s
understand this through an example.
[In]:
input_data = x_test.drop(['predicted_price'], axis=1)
input_record = input_data[0:1]
input_record.head()
[Out]:
181
182
Chapter 7
Sale weight resolution ppi cpu cpu internal ram RearCam Front_Cam battery thickness
core freq mem
Case Studies
302 149 5.5 534 8 1.6 32 3 16 8 3000 7
[In]:
x_test[0:1].head()
[Out]:
Sale weight resolution ppi cpu cpu internal mem ram RearCam Front_ battery thickness predicted_
core freq Cam price
302 149 5.5 534 8 1.6 32 3 16 8 3000 7 2941.57

For the given record, we see that the predicted price is about
2.9K. Suppose if the manufacturer wants to sell the mobile at a price 10
percent to 15 percent higher than the predicted one, the explainer can be
used to see recommendations in the mobile configuration that can achieve
the objective.
[In]:
predicted_price = x_test[0:1].predicted_price
expected_price_range = [round(float(predicted_price*1.10), 0),
round(float(predicted_price*1.15), 0)]
random_mobile = explainer_mobile.generate_
counterfactuals(input_record, total_CFs=2,
desired_range=expected_price_range)
random_mobile.visualize_as_dataframe(show_only_changes=True)
[Out]:
100%|██████████| 1/1 [00:00<00:00, 1.40it/s]
Query instance (original outcome : 2942)
183
184
Chapter 7
Sale weight resolution ppi cpu cpu internal ram RearCam Front_ battery thickness Price
Case Studies
core freq mem Cam
302 149 5.5 534 8 1.6 32 3 16 8 3000 7 2942
Diverse Counterfactual set (new outcome: [3236.0, 3383.0])
Sale weight resolution ppi cpu cpu internal ram RearCam Front_ battery thickness Price
core freq mem Cam
- - - - - 1.6 84.8 5.7 - - - - 3278.169922

- - - - - 1.6 - 5.2 - - 5116 - 3305.25
From the previous output, we see that the explainer model has suggested
two possible ways to sell the mobile at the expected price range.
1. Change the internal memory from 32 to 84.8 GB,

and increase the RAM from 3 to 5.7 GB.
2. Change the RAM from 3 to 5.2 GB, and increase

battery capacity from 3000 to 5116 mA.
However, there might be certain constraints for the business to take

into consideration, and one cannot rely only on the suggestions from
the counterfactual analysis. One might want to take the inputs from
the counterfactual analysis, tweak them a bit based on the business
constraints, and then check if those tweaked inputs generate the desired
output. This is possible through what-If analysis.
Step 6: Using What-If analysis for getting the right set of pricing
strategies for the given handset
Now suppose for the given counterfactual recommendations, we have
a constraint that expensive parts such as RAM cannot be increased, but
cheaper components such as internal memory can be increased by two to
four times the present one, and battery can be increased by a maximum
of 50 percent. So, for the given mobile, we cannot increase the RAM. Let’s
check which configuration can help us get the desired price.
Here are the options:
1. Increase the internal memory by two times and the

battery by 50 percent.
[In]:
expected_internal_memory = int(input_
record['internal mem']*2)
expected_battery = int(input_
record['battery']*1.5)
custom_inputs = input_record.copy()
185
custom_inputs['internal mem'] = custom_

inputs['internal mem'].replace(int(input_
record['internal mem']), expected_
internal_memory)
custom_inputs['battery'] = custom_
inputs['battery'].replace(int(input_
record['battery']), expected_battery)
print('New price based on updated
configuration: ', rf.predict(custom_inputs))
if rf.predict(custom_inputs)>=expected_price_
range[0]:
print("The given configuration helps in
getting the minimum expected price")
else:
print("The given configuration does not
help in getting the minimum expected price
of",expected_price_range[0])
[Out]:
New price based on updated
configuration: [3043.53]
The given configuration does not help in getting
the minimum expected price of 3236.0
We see that increasing the internal memory by 2x

and battery size by 50% does not help in getting the
expected price range. Let us see if we can achieve
the objective with a different combination.
2. Increase the internal memory by four times and the

battery by 50 percent.
186
[In]:
expected_internal_memory2 = int(input_
record['internal mem']*4)
expected_battery2 = int(input_
record['battery']*1.5)
custom_inputs2 = input_record.copy()
custom_inputs2['internal mem'] = custom_
inputs2['internal mem'].replace(int(input_
record['internal mem']), expected_internal_
memory2)
custom_inputs2['battery'] = custom_
inputs2['battery'].replace(int(input_
record['battery']), expected_battery2)
print('New price based on updated
configuration: ', rf.predict(custom_inputs2))
if rf.predict(custom_inputs2)>=expected_price_
range[0]:
print("The given configuration helps in
getting the minimum expected price")
else:
print("The given configuration does not
help in getting the minimum expected price
of",expected_price_range[0])
[Out]:
New price based on updated
configuration: [3247.54]
The given configuration helps in getting the
minimum expected price of 3236.0
We can see that the second option gives us the right configuration to
achieve the objective of increasing the handset price by at least 10 percent
of the predicted one.
187
Conclusion
In this chapter, we saw how decision-making can be supported by applying
AI/ML and other advanced techniques (counterfactuals and what-if
analysis). We saw how organizations not only can predict which customers
are going to churn but also can take proactive measures to avoid the
churn, with the help of counterfactuals and what-if analysis. We also saw
how a combination of machine learning predictions, counterfactuals,
and what-if analysis can help mobile manufacturers design their product
configurations (screen size, RAM, etc.) and get the desired price over the
predicted ones. Through these use cases we covered decision intelligence
for classification and regression problems.
188
Index
A Authority-based decision-
making, 63, 65
Accountability, 126
Autonomous systems, 81
AI-driven tool, 175
AI/ML models, 103, 108
AI-powered decision-making, 12
B
Artificial intelligence (AI), 109
Bayesian inference, 19
advanced technologies, 1
Bayesian networks, 77
challenges, 7, 8
Bias detection
customization, 9
data analysis, 2 auditing, 133
goal, 1 data analysis, 133
integration, 9 expert review, 134
key ethical considerations, 125 testing, 133
accountability, 126 Bias detection tools, 134
bias and fairness, 126 data analysis techniques, 134
human values and fairness metrics, 135
rights, 126 human review, 135
privacy and security, 126 machine learning
transparency and algorithms, 135
explainability, 126 statistical analysis
machine learning techniques, 135
algorithms, 1 Biased decision-making, 132
risk management, 9 Biases, 127
transparency, 9 algorithmic bias, 131
types, 1–5 automation bias, 131
variety of applications, 1 cognitive biases, 129
Attribution model package, 98, 99 confirmation bias, 130

https://doi.org/10.1007/978-1-4842-9568-7
INDEX
Biases (cont.) adoption barriers, 24

group attribution bias, 131 agility and flexibility, 11
impact on organization, AI life cycle, 16, 17
132, 133 AI model to business tools,
sampling bias, 130 120, 121
selection bias, 130 mapping data, 121, 122
Business analytics, 6 AI predictions to business tools,
Business tools, 118, 120 118, 119
applications, 14–16
biases, 23, 127
C build/integrate apps, 114, 115
Churning, 141 with business workflows, 110
Churn prediction, 140, 141 challenges, 13, 14
Cognitive bias effects, 73 integrating with business
Cognitive biases, 128–130 workflows, 110, 111
Collaborative decision-making, 12 integrating with existing
Computer-based system, 75 systems, 111, 112
Consensus-based decision- cognitive biases, 10, 128, 129
making, 65, 66 competitive advantage, 10, 22
Counterfactual analysis, 144, 145, complex and uncertain
159, 160, 169, 171, 185 environments, 10
Counterfactuals, 144, 145, 159, 160 complexity, 23
CRM systems, 118, 120–122 continuous improvement, 23
cost-effective, 23
customization, 22
D data quality issues, 23
Data analytics, 109 decision-making, 11–13
Data collection, 16 dependence on data, 24
Data-driven decision-making, 12 development, apps, 115, 116
Data preparation, 16 enhanced accuracy, 22
Decision analysis, 18, 76, 78 ethical concerns, 23
Decision intelligence (DI), 109 history, 6
accountability, 11 human expertise, 5
actions, 127 human-in-the-loop, 127
190
INDEX
improved decision-making, 22 define clear metrics, 38–40

improved decision quality, 10 framework, 30, 31
improved risk management, 22 guidelines/paths, 29, 30
increased efficiency, 22 myths, 28
industry, 24 organizational challenges, 27
information overload, 10 planning, 31–33
integration challenges, 24 return on investment (ROI),
machine intelligence, 5 41, 42
methodologies, 18–21 value, 40, 41
organization, 24 value per decision, 42, 43
overreliance on technology, 23 Decision intelligence
personalization, 11 methodologies, 47
privacy concerns, 23 conditions, 86
Procter & Gamble, 25 linear models, 86
real-time decision-making, 22 nonlinear models, 90
risk management, 11 nonlinear problem, 91
scalability, 22 nonlinear solutions, 94
steps, business workflows, production optimization, 87
112, 113 PuLP library, 88
technical expertise, 23 Decision-making, 17, 47, 53
techniques, 5 alternatives, 48
transparency, 11 authorized service centers, 56
UPS, 25 components, 47
user-friendly interfaces, 116, 117 constraints, 53–55
Decision intelligence (DI) definition, 47
requirements evaluation, 55
AI predictions, 43–45 methodologies
AI projects, 28 categories, 57
approach, 33–35 humans and machines, 58
approval mechanism/ overview, 50
organization problem definition, 52
alignment, 36, 37 problem statement, 52
artificial intelligence (AI), 28 strategic, tactical, and
decision-making, 28 operational, 50
191
INDEX
Decision-making (cont.) First acceptable match, 69

structured approach, 52 Fuzzy logic, 19
types, 48
Decision theory, 6
Deontology, 61 G
Descriptive analytics, 2 Game theory, 6, 19
Diagnostic analytics, 3 Group decision-making, 48, 49, 65
Disparate impact, 172
H
E Human decision-making
Ethical AI, 125 history, 59
Ethics-based decision-making Human-in-the-loop (HITL)
types, 60 systems, 127, 136–141
Ethics-based/moral-based Human-machine
decision-making, 60 decision-making, 81
advantages, 61 Humans-only decision-making, 58
disadvantages, 61 Hyperparameter tuning, 155
Experience-based Hypothetical telecom business, 146
decision-making, 62
advantages, 62
disadvantages, 63
I, J, K
evidence-based, 62 Individual decision-making, 48
intuition-based, 62 Instagram, 103
rule-based, 62 Instruction/rule-based systems, 74
Explainer counterfactual object, Intuition-based
160, 166, 181 decision-making, 12
F L
Fairness metrics, 135, 172 Linear mathematical models, 90
Feedback through human Linear models, 86, 90
intervention, 136–138 Linear programming, 76
192
INDEX
M Optimal solution function, 92

Optimizing or maximizing
Machine learning algorithms, 103
decision-making, 71
Machine learning model, 144, 174
Marketing channels, 96
Markov chain models, 95, 96 P, Q
Mathematical and probabilistic
Personal fulfillment, 61
models, 103
Predictive analytics, 3
Mathematical decision system, 75
Prescriptive analytics, 3
Mathematical models, 76, 85, 87
Probabilistic
Mobile pricing prediction, 174
decision-making, 77
Model deployment, 17
Probabilistic models, 94
Model training, 17
Probability-based decision
Monitoring and optimization, 17
systems, 78, 79
Monte Carlo simulation, 78
Production optimization, 87
Multicriteria decision
Python libraries, 88
analysis (MCDA), 19
Python package, 96
Multinational corporation, 103
Python programming
Multiple-criterion decision-making
language, 85
(MCDM), 49
advantages, 50
SCDM, 50 R
Random decision-making, 59
advantages, 59
N
disadvantages, 60
Nonlinear mathematical
Random forest model, 157
models, 86
Return on investment (ROI),
Nonlinear programming, 76
41, 42
Risk index (RI), 91
O Rule-based decision-making, 12
Rule-based human-machine
Operational decision-making, 51
decision-making, 74
Operations research (OR), 6
193
INDEX
S U
Satisfactory model Unsupervised learning, 79
performance, 158 User-friendly interfaces, 112, 116, 117
Single-criterion decision-making User interface, 115
(SCDM), 49 Utilitarianism, 60
Stratified K-fold cross
validation, 106
V
Supervised learning, 79
Virtue ethics, 61
Voting-based decision-making, 66
T
Tactical decision-making, 51 W, X, Y, Z
Threshold-based What-if analysis, 144, 145, 159, 169,
decision-making, 68 171, 174, 180, 185, 188
TotalCharges, 150, 152 Workforce constraints, 87
194

Introduction To Prescriptive AI: A Primer For Decision Intelligence Solutioning With Python

Uploaded by

Copyright:

Available Formats

Introduction To Prescriptive AI: A Primer For Decision Intelligence Solutioning With Python

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introduction To Prescriptive AI: A Primer For Decision Intelligence Solutioning With Python

Uploaded by

Copyright:

Available Formats

Introduction to

ISBN-13 (pbk): 978-1-4842-9567-0 ISBN-13 (electronic): 978-1-4842-9568-7

Copyright © 2023 by Akshay Kulkarni, Adarsha Shivananda, Avinash Manure

About the Technical Reviewer�������������������������������������������������������������xi

Chapter 1: Decision Intelligence Overview�������������������������������������������1

Chapter 2: Decision Intelligence Requirements����������������������������������27

Chapter 3: Decision Intelligence Methodologies��������������������������������47

Chapter 4: Interpreting Results from Different Methodologies�����������85

Decision Intelligence Methodology: Probabilistic Models�����������������������������������94

Chapter 5: Augmenting Decision Intelligence Results into the

Chapter 6: Actions, Biases, and Human-in-the-­Loop������������������������125

Example: Customer Churn���������������������������������������������������������������������������������140

Chapter 7: Case Studies��������������������������������������������������������������������143

Adarsha Shivananda is a data science and

Avinash Manure is a seasoned machine

By the end of the book, you will have a solid understanding of

© Akshay Kulkarni, Adarsha Shivananda, Avinash Manure 2023 1

As mentioned, prescriptive AI systems differ from descriptive and

• Diagnostic AI: Diagnostic AI, also known as diagnostic

For example, a hospital may use diagnostic AI to

• Predictive AI: Predictive AI, also known as predictive

• Prescriptive AI: Prescriptive AI, also known

using optimization algorithms to identify the best

For example, a logistics company may use

In summary, as shown in Figure 1-1, descriptive AI is used to describe

Figure 1-1. Types of AI

Prescriptive AI typically involves analyzing large amounts of data

Decision Intelligence History

• Operations research: Operations research (OR)

• Game theory: Game theory is a branch of mathematics

involves using data to inform decision-making and has

• Decision intelligence: DI emerged in the early 2000s

• Data quality and availability: AI systems rely heavily

• Technical expertise: Implementing/maintaining AI

• Cost: AI systems can be expensive to implement and

• Resistance to change: There may be resistance to

• Ethical considerations: AI systems can raise ethical

• Integration with existing systems: Integrating AI with

• Regulation and compliance: AI systems may need

• Lack of understanding: There may be a lack of

To overcome these challenges, businesses need to carefully evaluate

 ow Can DI Help Bridge the Gap Between AI

• Integration: DI can help integrate AI technology into

• Customization: DI can be customized to fit the specific

• Risk management: DI can help identify and mitigate

• Transparency: DI can help improve transparency

The Need for Decision Intelligence

• Information overload: With the explosion of data and

• Complex and uncertain environments: Many

• Cognitive biases: Human decision-making is often

• Competitive advantage: In today’s competitive

• Improved decision quality: DI can help decision-­

• Risk management: Many decisions involve some level

• Agility and flexibility: In today’s fast-paced and

• Transparency and accountability: DI can help ensure

• Personalization: DI can help personalize decisions

Overall, DI can help organizations and individuals make better

The Evolution of Decision-Making

• Intuition-based decision-making: In the early days

About the Technical Reviewer��xi

Chapter 1: Decision Intelligence Overview��1

Chapter 2: Decision Intelligence Requirements��27

Chapter 3: Decision Intelligence Methodologies��47

Chapter 4: Interpreting Results from Different Methodologies��85

Decision Intelligence Methodology: Probabilistic Models��94

Chapter 6: Actions, Biases, and Human-in-the-Loop��125

Example: Customer Churn��140

Chapter 7: Case Studies��143

Decision Intelligence History

ow Can DI Help Bridge the Gap Between AI

The Need for Decision Intelligence

• Improved decision quality: DI can help decision-

The Evolution of Decision-Making

nderstanding Where Decision Intelligence

Decision Intelligence Methodologies

Some Potential Pros and Cons of DI

E xamples of How Companies Are

Why Do AI Projects Fail?