Co

m
pl
im
en
ts
of
The Future
of Analytics
The New Landscape of Artificial
Intelligence & Machine Learning
Applications

Dan Darnell, Rafael Coss

& Patrick Hall

REPORT
 The Future of Analytics
The New Landscape of Artificial
Intelligence and Machine
Learning Applications

Dan Darnell, Rafael Coss, and Patrick Hall

Beijing Boston Farnham Sebastopol Tokyo

The Future of Analytics
by Dan Darnell, Rafael Coss, and Patrick Hall
Copyright © 2021 O’Reilly Media, Inc. All rights reserved.
Printed in the United States of America.
Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA
95472.
O’Reilly books may be purchased for educational, business, or sales promotional use.
Online editions are also available for most titles (http://oreilly.com). For more infor‐
mation, contact our corporate/institutional sales department: 800-998-9938 or
[email protected].

Acquisitions Editor: Rebecca Novack Proofreader: Abby Wheeler

Development Editor: Angela Rufino Interior Designer: David Futato
Production Editor: Katherine Tozer Cover Designer: Randy Comer
Copyeditor: nSight, Inc. Illustrator: Kate Dullea

October 2020: First Edition

Revision History for the First Edition

2020-10-16: First Release

The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. The Future of
Analytics, the cover image, and related trade dress are trademarks of O’Reilly Media,
Inc.
The views expressed in this work are those of the authors, and do not represent the
publisher’s views. While the publisher and the authors have used good faith efforts
to ensure that the information and instructions contained in this work are accurate,
the publisher and the authors disclaim all responsibility for errors or omissions,
including without limitation responsibility for damages resulting from the use of or
reliance on this work. Use of the information and instructions contained in this
work is at your own risk. If any code samples or other technology this work contains
or describes is subject to open source licenses or the intellectual property rights of
others, it is your responsibility to ensure that your use thereof complies with such
licenses and/or rights.
This work is part of a collaboration between O’Reilly and H2O.ai. See our statement
of editorial independence.

978-1-492-09173-8
[LSI]
 Table of Contents

Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

1. The Converging World of Analytics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Challenges in Current Analytics Projects 1
From Business Intelligence to Augmented Analytics 4
The Role of Automation in the Future of Analytics at Scale 5
The AI App Revolution 7
Current AI App Development Challenges 11

2. Modern AI Applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
The Anatomy of a Modern AI Application 15
Detailed Application Examples for Key Industries and
Functions 19
CPG Sales Forecasting with COVID-19 Data 21
Hospital Staffing Optimization (Healthcare Industry AI
Apps) 22
Marketing Lead Optimization (Line of Business AI
Applications) 23
Data Augmentation (AI Apps for Data Teams) 26

3. Case Studies: Real Impacts of AI Application in Business. . . . . . . . . . 29

ArmadaHealth 29
Hortifrut 31
Jewelers Mutual 32

iii
4. Adoption Challenges for Next-Generation Analytics. . . . . . . . . . . . . 35
Ineffective Data and AI Principles 35
Lax Security Practices 36
Inadequate Human Review 36
Downplaying Traditional Domain Expertise 37

Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

iv | Table of Contents
 Introduction

In 2015, when I started writing The Evolution of Analytics with my

colleagues Wen Phan and Katie Whitson, we made the case for
machine learning in business. Five years later, it’s time to make the
case to use machine learning the right way in business. While I cer‐
tainly don’t know all the answers, a few themes stand out to me,
looking back over the past five years. On the negative side, there’s
the endless hype about artificial intelligence (AI) and the tendency
to deploy it in creepy and discriminatory ways. On the positive side,
I see the growing government and public awareness of AI. I hope
this awareness translates into the regulation of AI, improved interac‐
tion design in AI apps, and more corporate responsibility and gov‐
ernance for AI.
In this report, we’ll introduce AI-driven applications that boost tra‐
ditional data analytics with machine learning. While these apps may
beat the odds, provide useful insights, and drive organizational
value, such success stories don’t serve as guarantees. In fact, every‐
one involved in organizational AI projects would be wise to take an
inventory of AI’s impacts on businesses, consumers, and the general
public.
Another bright spot in AI over the last five years has been the devel‐
opment of technologies that increase human trust and
understanding in machine learning.1 These inventions have trans‐

1 See the following examples: “This Looks Like That: Deep Learning for Interpretable
Image Recognition”, “Intelligible Models for HealthCare”, “A Unified Approach to
Interpreting Model Predictions”, “Introducing AI Fairness 360”, “Introducing Tensor‐
Flow Privacy”, and the What-If Tool.

v
formed machine learning from a field of black-box algorithms to a
field that is now capable of fierce debate around the concepts of
algorithmic transparency, accountability, and fairness. This techno‐
logical progress not only enables the nuts and bolts of regulatory
oversight, but it also gives companies the power to govern their AI
systems like the enterprise software assets they are. If you can block
out the hype, you’ll see that AI is really just software. And like all
other enterprise IT resources, AI systems should be documented,
managed, monitored, and governed.
Looking forward, I see successful AI deployments being aware of the
risks of AI, taking on the associated governance burdens, and ena‐
bling humans to work together with computers to solve big prob‐
lems. For businesses climbing to the next plateau in digital
transformation, don’t settle for any AI system. You’ll need AI sys‐
tems that are documented, transparent, managed, monitored, and
minimally discriminatory. Moreover, these AI systems must support
AI apps that are flexible, explainable, and, when appropriate, auto‐
matic. That’s why Dan, Rafael, and I have written this new report,
The Future of Analytics. It’s a necessary update to the original Evolu‐
tion of Analytics report, and we hope you find it to be a timely and
useful guide through the new world of AI-powered analytics apps.

— Patrick Hall

vi | Introduction
 CHAPTER 1
The Converging World of Analytics

In the broadest sense, analytics is the systematic analysis of data.

This analysis makes the data consumable by people and systems,
with the goal of understanding past outcomes and helping to predict
future events. The adoption of analytics has driven a wave of digital
transformation across industries where companies use data to
power decision-making processes. Analytics projects, however, have
not been without their drawbacks.

Challenges in Current Analytics Projects

Like many changes in business thinking, the first forays into data-
driven decisions led down accessible but less useful paths. One such
path was using dashboards to view historical trends to drive human
insights from data, as shown in Figure 1-1. We now know that these
traditional analytics dashboards alone can be insufficient to make
better decisions as they provide only a historical summary.

Figure 1-1. Typical dashboard with historical information

1
While historical trends are useful and can be predictive, they can
also provide a false sense of confidence, when the future does not
mirror the past. In the end, historical information and descriptive
analytics alone leave business leaders to use their best judgment
about trends in order to make decisions based on their own experi‐
ence and limited view of the data. The result of this process is then
highly dependent on the individual decision-maker’s expertise,
which yields highly variable outputs.
Using machines to find patterns in data and make predictions is
another area of great promise for decision support. Machine learn‐
ing models, trained on historical data, can look at new data and pre‐
dict what is likely to happen. For example, credit card companies
use machine learning models to determine who has access to credit
and is likely to carry a balance on their credit card bill each month.
Such models can even prescribe actions for users to take and recom‐
mend products or content of interest. This ability to predict future
outcomes and prescribe actions has made machine learning a hot
technology—and data science a hot profession.
For all its promise, machine learning has not reached widespread
usage in production or within business applications where it can
provide value and support business decisions. The challenges vary
by organization and use case. However, the common themes in AI
and machine learning adoption revolve around a few key areas,
including a lack of resources, lack of business trust in models and
their outputs, difficulty putting models into production and keeping
them running, lack of consistent business involvement, and bottle‐
necks in putting predictive results into business applications. Let’s
discuss some of these dilemmas below:
Talent gap
“Data scientist” has become the hottest title and the hardest to
fill position for many organizations. While many data analysts
changed their titles to capture the new wave of interest, they
lack the skills in languages like Python and R and an under‐
standing of modeling algorithms and techniques needed for
predictive model development. The lack of qualified data scien‐
tists in the market has put many companies on the slow track in
AI development. Even with qualified data scientists on staff,
companies find their output is limited due to the handcrafted
nature of their work and the high maintenance costs of such
models in production.

2 | Chapter 1: The Converging World of Analytics

Model trust
Trust and the responsible use of machine learning models are
evolving fields and the topics of entire books. What is clear is
that for a business to use insights from predictive models, that
organization must first trust the models’ outputs. Trust can
develop over time as users make decisions and begin to see that
the model provides valuable insights. However, even before this
can happen, data science teams must explain the models in
business terms using techniques that show what factors contrib‐
ute to each prediction, how the model makes such predictions,
and how a model will behave on new data. Without a level of
trust in a model, business users will not use predictive models to
automate business decision processes.
Model operations
Another barrier to machine learning adoption is operational.
Building an experiment to make predictions is one thing, but
deploying a model into production is quite another. Again, this
is a broad topic discussed in other books. At the summary level,
models will need to undergo an extensive process of testing and
validation, model packaging, and then production deployment
and ongoing management for which many organizations are
not prepared. Once these issues are exposed, models that once
looked promising in the lab never find their way into produc‐
tion use.
Domain involvement
Probably the most overlooked barrier to AI adoption is the
involvement and access by domain experts and end users. Too
often, machine learning and AI projects take on lives of their
own during development. While the business user may define a
need or use case and even provide the budget for technology
and data science teams, that is often the end of their involve‐
ment. Without a feedback loop between the domain experts and
the development team, integrating AI into existing applications
—or projects to create new applications—fails to get off the
starting line.
Application development
All too often, even successfully deployed predictive models end
up only as services. For predictive models to provide value to
the business, these services must be incorporated into new or
existing applications by application development teams.

Challenges in Current Analytics Projects | 3

 Without this application development step, the domain experts
and end users can’t provide feedback on the model because it is
not in their context. This step also involves a new team of appli‐
cation developers in the project, and this team has a backlog of
projects that, once again, can send your AI project to the
dustbin.
To navigate the future of analytics, organizations must deal with
each of these challenges. In this report, we will examine the solution
to domain involvement and application development, as other
reports and books cover trust, operations issues, and solutions.

From Business Intelligence to Augmented

Analytics
Analytics means many things and covers a host of technologies that
can sometimes seem quite different from one another. The two lead‐
ing areas of analytics are business intelligence and artificial intelli‐
gence. Each of these fills a need for the systematic analysis of data,
but they use entirely different means and typically support different
user groups within the organization.
Business intelligence (BI) is the ability of an organization to use his‐
torical data analysis to understand business performance and ideally
to improve business decisions and outcomes. Traditional business
intelligence provides reporting technology for business users and
analysts to view summarized historical information. Recent innova‐
tions in BI have added more real-time information to these reports,
but the context remains to analyze available historical data. Com‐
mon BI applications include key performance indicator (KPI) met‐
rics reporting, executive dashboarding, and ad hoc data analysis.
Artificial intelligence (AI) is the ability for machines to perform tasks
that people generally do, such as making inferences and decisions
based on data, identifying trends, sorting images, responding to
speech or text, and more. Traditionally, AI in the form of machine
learning is used to create models that predict what will happen in
the future based on past information. These models are used in
batch processes to score or categorize records offline, or in real time
to provide predictions in decision-making processes.
AI and BI are converging. Gartner, Inc., a leading IT analyst firm,
views these parallel tracks converging and consolidating in support

4 | Chapter 1: The Converging World of Analytics

of a more unified analytics platform.1 This convergence will drive
better overall insights, understanding, and decision making, and
overcome many of the challenges caused by using different analytics
tools. In doing so, this consolidation will dramatically improve busi‐
ness agility and productivity, and create many opportunities for new
and enhanced AI-driven business applications.
With the need to provide predictive and prescriptive insights, ven‐
dors across the analytics spectrum are changing their stripes. Com‐
panies that formerly focused on dashboards are incorporating
predictive intelligence into their products. Similarly, companies that
used to focus on data science and machine learning are offering ana‐
lytic dashboards, including predictive insights. This convergence of
business intelligence and data science with machine learning is
called augmented analytics. Again, entire books exist on this topic.
Many companies want to be the one-stop shop for descriptive and
predictive analytics across the enterprise. However, simply combin‐
ing different analytics styles may be exciting, but prove insufficient
to meet business needs for actionable insights in a business context.
Similarly, data science tools that help teams collaborate, manage, or
automate some parts of the machine learning life cycle may also
miss the mark. Indeed, the future of augmented analytics may prove
to be something much different in the form of AI applications.

The Role of Automation in the Future of

Analytics at Scale
If the future of analytics is one where AI reaches into every corner of
business to enhance applications and processes, then automation
plays a crucial role in achieving that vision. The traditional data sci‐
ence process involves weeks or even months of painstaking work to
hand-code models using languages like Python or R and test various
machine learning algorithms. This process can produce outstanding
results, depending on the skill of the data scientist. Unfortunately, if
AI is to scale, most organizations can afford neither the talent
required for such artisanal work, or the time to wait for hundreds or
thousands of such projects to complete.

1 “2020 Planning Guide for Business Analytics and Artificial Intelligence”, Gartner
Research, October 7, 2019.

The Role of Automation in the Future of Analytics at Scale | 5

The solution is the robust automation of the data science and
machine learning process, known as automated machine learning
(AutoML). AutoML does not replace the data scientist, but instead
helps to automate the steps typically done by advanced data scien‐
tists. With AutoML, novice users can achieve robust modeling
results, and expert data scientists can improve productivity by auto‐
mating data science functions, including:
Feature engineering
AutoML will identify and transform data into numeric forms
that machine learning algorithms can understand. Advanced
data scientists will also tease signal out of data by looking for the
ways to combine input data columns together. AutoML should
perform both functions, using the power of computing infra‐
structure to try different permutations of features to determine
the ones that provide the best signal.
Algorithm selection
For any given problem, there are many mathematical
approaches that the data scientist could take. AutoML will typi‐
cally look at the data and the type of problem, and then decide
which algorithms are best suited. The AutoML will then run
through these algorithms to see which ones provide the highest
accuracy, most interpretable results, or the fastest scoring. The
goals of the project will determine which model provides the
best fit.
Parameter tuning
For each algorithm, there are many different settings that will
impact how the model is built and determine the accuracy and
speed of predictions. AutoML can easily run through different
tuning options to determine which parameter settings yield the
best results, based on goals for the model.
AutoML should also include testing and validation of models to
ensure high quality under various conditions and to avoid common
pitfalls. The output of the AutoML process is a model that can be
used for production scoring that includes all transformations to take
in raw data and produce a score or response in a production envi‐
ronment. The automation of the data science and machine learning
process allows more people to engage in the data science process
from a variety of technical backgrounds, thus increasing the number
of possible projects. Automation also dramatically decreases the

6 | Chapter 1: The Converging World of Analytics

time to complete projects and produce a production-ready model.
These two factors lead to a dramatic increase in the number of AI
projects that reach production.
Automation in production operations is also critical. Once deployed
into a production environment, a predictive model will need to be
monitored and maintained over time. When data patterns change,
and the model is no longer able to predict as it once did, that model
will need to be updated. Automation can play a crucial role in opera‐
tions, allowing operations teams to monitor models without detailed
knowledge of model techniques, algorithms, or modeling languages.
Monitoring by exception, where guardrails are established around
models to alert the operations team to issues, allows a relatively
small group to manage a large number of models. This operations
paradigm also removes the need to continually involve the data sci‐
ence team in production issues, which takes them away from build‐
ing new models.
Automation of the end-to-end modeling and production process
also lessens the dependence on specific individuals and makes the
process more repeatable and robust. That is not to say that outstand‐
ing individuals cannot make outsized contributions, but rather that
these contributions, like those of other users, must be captured in a
repeatable framework. With this approach, when an individual is
reassigned or leaves the organization, the process continues. Taking
a systematic approach using automation across all data science and
production projects leads to a kind of corporate memory, rather
than individual memory, that allows the organization to scale up
with thousands of AI-powered processes.
The result of such robust automation across the machine learning
process is that models are not only easier and faster to develop, but
once built, facilitates the automation of ongoing maintenance such
as periodic retraining. This automation of updates is the key to
building and maintaining hundreds or even thousands of AI-driven
applications in an organization.

The AI App Revolution

Many in the analytics community believe that the last mile in AI is
the production operations of models and endpoint services that
deliver model results. You can see this in the way many vendors talk
about consuming model results in batch processes or as an online

The AI App Revolution | 7

service at the end of AI development. However, the truth is that no
model can provide value—no matter how robustly implemented or
how many computing nodes it’s deployed to for scale and availabil‐
ity—if end users can’t access it at the critical decision point, in con‐
text and in a way that they deem useful. This interface between the
business user and AI happens in an AI-driven application.
Using AI in applications is not a new concept. Thinking about use
cases and driving AI adoption from a business need should be the
norm; however, as we have already seen, many such processes break
down due to issues like trust, operations, and a lack of business
context.
AI applications are the natural end product of a mature AI program,
as outlined in Figure 1-2.

Figure 1-2. The AI application development life cycle

Each AI application starts as an idea from a domain expert or busi‐

ness unit with a problem to solve, or where AI is likely to improve
efficiency or customer experiences. Although the terms are often
used interchangeably, AI applications should not be confused with
the business use cases related to them. The AI application is used by
the business user to help aid or solve the problem from the use case.
The process for developing AI applications can be broken down into
the following steps:
Data
The first question in developing an AI application always
revolves around data. What data do we have and what data do
we need for this use case? If internal data is insufficient, then

8 | Chapter 1: The Converging World of Analytics

 external data sources, either public or private, will be required
to augment existing datasets. This set of tasks is often handled
by data engineers who access data sources and create datasets
using data management tools for use by the data science team.
Modeling
With sufficient data, a data scientist can begin experimenting to
develop a working model that can provide a useful prediction or
prescriptive information for the use case. Without automation,
this is an iterative process that can take weeks, or even months,
to find the most accurate predictions that meet the business
requirements and create a model that can be deployed in
production.
Validation
Once this model is developed and trained, it must be tested and
validated, typically by a group outside of the data science team
that developed it. These teams will use historical data to “back
test” the model to prove that it would have generated acceptable
results under a variety of conditions. Validation may also
explore machine learning explanations to look for bias, to better
understand how it makes the predictions, and to see any poten‐
tial limits.
Operations
If the model proves robust from a data science perspective, it
will undergo further testing by the operations team. In these
tests, the model will endure production loads on production
systems to ensure that it can meet service level agreements for
downstream applications. Once deployed, the operations team
will manage and maintain the model.
Applications
With a robust and reliable production model, application devel‐
opers can now integrate those predictive insights into existing
or new AI-driven applications. This process is best done with
intense interaction with the domain experts to ensure that the
application meets the business need.
This entire process is a cycle because as business needs change, new
data becomes available and new algorithms or techniques emerge.
Applications will need to be updated along with the machine-
learning models.

The AI App Revolution | 9

AI-driven applications can take several forms, from enhancements
to existing business applications to new, purpose-built decision sup‐
port tools.
Enhancing existing app
Currently, the most common AI-driven application augments
an existing business application with AI insights. For example,
showing a predicted trend line on an existing dashboard,
adding a customer churn score (the probability the customer
will churn), or recommending offers to an existing call center
application. These projects provide value but can take
application-specific development resources to integrate AI
insights, making projects expensive, and taking many months to
complete. Also, when adding AI insights into applications, those
applications may need to be refactored or wholly redone to take
full advantage of predictions or put the insights into context.
Purpose-built AI apps
AI applications built from the ground up to support a specific
decision or role are another way to deliver AI-driven insights.
These applications include descriptive, predictive, and prescrip‐
tive insights focused on a particular problem. These purpose-
built applications benefit from modern frameworks and focus
on meeting immediate user needs. On the downside, applica‐
tion development resources are notoriously tricky to find, limit‐
ing the ability to develop these focused applications.
Prebuilt AI apps
One solution to the time-consuming nature of application
development is to purchase AI apps developed by specialists.
These prebuilt apps can act as a springboard for the use of AI
within an organization. The downside of such apps is that they
have to be somewhat generic by design to accommodate any
number of users. An example of a prebuilt AI app would be an
application for automated clustering analysis. This application
takes data as input and then outputs information and plots on
clusters identified within the data. Such clusters of customers,
for example, are useful in marketing to create segments and tar‐
geted campaigns. Prebuilt AI apps may also need to be custom‐
ized to fit a particular organization’s needs or data. When
looking at prebuilt applications, business leaders should factor
in the cost of such customization in their analysis.

10 | Chapter 1: The Converging World of Analytics

AI applications are the value endpoint that many business leaders
have been looking for to make sense of investments in AI and
machine learning technologies and investments in data science
teams. Incorporating AI or building new AI apps is how business
users can finally see the value of smart models in their everyday
work to improve customer experiences, increase employee produc‐
tivity, and optimize business operations.

Current AI App Development Challenges

Incorporating AI into new and existing applications is perhaps more
complicated than it seems. AI and machine learning predictions
might seem like just another data source or another API service. To
transform a business with AI, however, will take more than just
putting trend lines on dashboards. Domain experts will need to
partner with data scientists to build apps for their specific business
needs quickly. These data scientists will also have to partner with
application developers to create production apps. The apps they cre‐
ate will be interactive and help business users make smarter deci‐
sions. Understanding why existing tools and teams don’t fulfill needs
around AI apps is critical to understanding what organizations need
to build the right team and adopt the right technology.

The Impedance Mismatch Between AI and Existing Web

Frameworks and Teams
Software web development frameworks for building business appli‐
cations are well established. When you want to build a web applica‐
tion, developers will use web frameworks and languages like Django,
Flask, Express, Ruby on Rails, HTML, JavaScript, Spring, and more.
For business applications, developers may default to classic BI
reporting and dashboard tools or others. To engage these teams in
new AI application development creates new issues. First, applica‐
tion developers and web developers are already swamped with
projects. Bringing them a new project that involves AI will likely put
the project at the end of the queue in favor of projects that can be
delivered with existing tools. Then there is the impedance mismatch
between the DevOps and MLOps life cycles. The lack of knowledge
about AI and machine learning also means that developers might
not understand what the application is doing; why, when, and how

Current AI App Development Challenges | 11

the model should be applied; how to interpret the results; and when
to retrain a model, which can lead to delays and expensive rework.

Web development is not AI-friendly

While applications using AI may seem like any other web applica‐
tion, there are key differences that will doom any project conceived
and executed using AI and web frameworks. First, many web frame‐
works are designed for simple data interaction patterns. Typical web
data applications tend to be transaction-focused with clicks and sim‐
ple inputs driving subsequent changes in pages to provide new expe‐
riences. These applications tend to work with web frameworks that
know how to interact with transactional or operational databases.
The analytics in these applications often only work with the opera‐
tional database and not the analytics databases (enterprise data
warehouses), and therefore have minimal analytics.

Generic AI Apps
One approach to AI apps is to build generic apps that allow users to
input values and see results from predictive models. These apps have
names like What-if, Optimizer, and Predictor. These applications are
little more than a front end to the model API. While such applica‐
tions can help business users interact with predictions, these appli‐
cations are not tailored to a business need or user. The result is that
such generic applications become little more than a way for business
users to test models and become comfortable with their outputs. In
addition, generic applications are unlikely to be designed for specific
needs like peak loads in a call center or for online shopping. These
applications also have static workflows, so there is minimal addi‐
tional exploration, simulation, and evaluation.

Noninteractive Experience
Many analytics applications are used to pull data from a data ware‐
house to generate reports or dashboards. Interactivity with such data
is limited to filtering and slicing the available data. In this frame‐
work, additional requests will pull a new set of data. As the user
looks at dashboards, they expect to see the latest data and to have the
data visualization dynamically update as new data arrives, but this is
not the case. Even for IoT applications that process data in real-time
streams, the user’s dashboard will not update unless the user specifi‐
cally requests it. For AI app development, systems should have both

12 | Chapter 1: The Converging World of Analytics

push and pull capabilities to provide an interactive experience for
the user, which should also include real-time predictions in context.

Scarce Development Resources

In many organizations, application developers are as scarce and val‐
uable a commodity as data scientists. As mentioned earlier, getting a
project into the queue of IT app developers can postpone a project
by months. When thinking about AI application development,
organizations should consider a two-stage approach that includes
rapid prototyping and final production development. This rapid
prototyping phase should be managed by analytics or data science
resources who can work closely with the domain experts in the busi‐
ness. Once an AI application design is completed and has shown
business potential, only then does the application development team
become involved.
The world of analytics is changing. With business intelligence and
artificial intelligence merging into augmented analytics, the land‐
scape is ready for a new set of technologies and applications for
building AI into applications. Automation of AI model develop‐
ment, training, and operations play a critical role, but focusing on
the traditional data science life cycle or leveraging automation in
AutoML alone is insufficient to provide better customer experiences
and smarter decisions for business users. The next step for analytics
is the creation of AI applications. These new AI applications will
drive new levels of business value by bringing descriptive, predic‐
tive, and prescriptive intelligence to users across the enterprise.

Current AI App Development Challenges | 13

 CHAPTER 2
Modern AI Applications

For AI to have a serious impact on organizations, users of all types—

from the manufacturing floor to the executive suite—must have
access to AI to make better decisions, improve customer experien‐
ces, and optimize business operations. This ubiquitous access to AI
will come in the form of applications used to make everyday deci‐
sions and drive processes. These AI-driven applications will change
the way employees do their jobs and how companies interact with
their customers and partners. In this chapter, we’ll explore the
makeup of such applications and look at detailed examples of appli‐
cations focused on specific business problems.

The Anatomy of a Modern AI Application

The modern AI application is the product of augmented analytics,
automation, and application development. To better understand
how the AI application differs from traditional applications and
dashboards, let’s dive into the various components and key features
of a modern AI application (see Figure 2-1):
Specific
Users of AI applications typically have specific goals and deci‐
sions in mind. The interface of the application is designed to
support those specific decisions, such as making hospital staff‐
ing assignments, as shown in Figure 2-1. Unlike a dashboard,
however, the interface is interactive; the user inputs questions or
parameters, and then interacts with the analysis to further refine
the results or to ask additional questions.

15
 Figure 2-1. Example of a modern interactive AI application

Interactive
Modern AI applications are also push-driven versus pull-driven.
Many legacy BI dashboards pull data on request or periodically
from a database. The information displayed is correct as of the
last refresh. Modern AI apps are push-driven because the inter‐
face is continuously being updated as data arrives and is upda‐
ted along with the latest predictions, creating a more dynamic
experience for the user and ensuring that they have the latest
information.
Automated
Modern AI applications include automation to maintain their
performance over time. A critical function in modern AI appli‐
cations is the ability to build and rebuild AI and ML models to
support new data or changes in request parameters. An open
and extensible AutoML engine is used by the developer or data
scientist involved in the AI app project to build models used in
the app. Using AutoML, versus a hand-coded approach, is desir‐
able for the business and the users as it allows the application to
be updated without needing resources from the data science
team, which are typically busy with other projects. The use of
AutoML also creates a well-documented and repeatable process
so that applications don’t break down as they are updated.

16 | Chapter 2: Modern AI Applications

Key Components of an AI Application Platform
Modern AI applications have a variety of components that facilitate
interactivity and allow for easy updates and management:
End user application interface
The business user will access AI applications either within exist‐
ing business applications or as standalone applications. End
users will access applications through a browser with a specific
URL or through an application catalog that showcases available
apps. The interface consists of a series of steps to collect data
and processing directions. This workflow guides the user
through the steps required to create the outputs they are look‐
ing for.
Development environment
Application developers and data scientists will have their own
interface to develop apps with a software development kit
(SDK). Ideally, this development environment can be deployed
to their development environment or laptop where they can
iterate through application changes. Developers can use this
environment to create rapid prototypes and to share these pro‐
totypes with domain experts for review. Once the application is
ready to be shared with the user community, the developer can
publish the app to the server environment for sharing.
Application server
The application server provides the needed services and scala‐
bility for production applications. Once applications are pub‐
lished and access controls are set, then end users can see
applications that are available to them.
Application catalog
If the user has access to more than one application, then an
application catalog capability is required so that the user can
access, load, and run multiple applications, as shown in
Figure 2-2. Using this functionality, the administrator can also
load new applications for the users. When each user logs into
the system, they will see the applications available to them.

The Anatomy of a Modern AI Application | 17

 Figure 2-2. Application catalog gives users in an organization
access to available apps

Security
Among the many functions of the application environment, the
security of the data and application is highly important. In this
system, AI applications should only be accessed by designated
users and updated by administrators. As data travels between
the application server and the user’s computer, the information
should be encrypted to ensure that it cannot be manipulated in
transit. This is especially true when personal or medical infor‐
mation is present.
Provisioning and management
AI applications are deployed from the server for the use of each
user. The model is similar to “app stores” that mobile phone
users are used to. This is different from shared software, where
multiple users share the same application. This separation
allows the users to customize the apps to meet their needs and
keep those settings for future use. The application management
system then has to provision the application for the user to con‐
sume and track who is using the app and what version they have
downloaded. Application updates are then made available for
those users to update their app to gain access to new functional‐
ity. Administrators may also choose to update applications
directly in the event of a security concern or other serious issue.
The framework and key components of modern AI applications
allow for usage in a variety of use cases to create specific, interactive
applications across industries.

18 | Chapter 2: Modern AI Applications

Detailed Application Examples for Key
Industries and Functions
AI applications may seem abstract to business leaders. The following
examples show how AI applications are used across industries and
lines of business functions, especially with business changes due to
COVID-19.

Mortgage Lending (Financial Services AI Applications)

In early 2020, COVID-19 shutdowns caused massive unemployment
and other societal disruptions. The models that banks had previ‐
ously used to determine mortgage eligibility were suddenly ineffec‐
tive due to a dramatic change in conditions.
To better understand potential borrowers in these new conditions, a
leading US bank partnered with an AI technology company to
develop a new mortgage lending application that would take into
account COVID-19 related data, including infections and unem‐
ployment. The new application centers around scenario planning
and gives the user the ability to input various factors and review
possible scenarios for risk (Figure 2-3).
The output is a series of scenarios that are specifically designed for
the mortgage lender to understand what happens with changes in
data. By reviewing the scenarios, the bank now has a better under‐
standing of the factors that will lead to defaults under changing
conditions.

Detailed Application Examples for Key Industries and Functions | 19

Figure 2-3. Mortgage lending AI application using localized unemploy‐
ment and other data

20 | Chapter 2: Modern AI Applications

CPG Sales Forecasting with COVID-19 Data
When the shutdowns started, many businesses found that years of
experience and data were unable to predict future demand. For con‐
sumer packaged goods companies, this was especially true as
demand for commodities like cleaning supplies and toilet paper
changed overnight. Forecasting models built on typical seasonal
demand and regional patterns could not predict demand under a
new set of conditions based on COVID-19.
A leading European CPG company with a large number of products
was quickly put under pressure as their products went out of stock
across geographies worldwide. To replenish supplies, the manufac‐
turer needed to know which products would be most in demand
and where. Working with a leading AI technology company, they
developed an AI application to predict demand for each SKU at
each retailer and region in a matter of days (Figure 2-4).
This new application used an updated forecasting model that
enriched historical patterns with new data sources to predict the
new demand patterns. A new predictive model alone would have
been insufficient. What was also needed was an interactive applica‐
tion that business leaders could use for scenario planning.

Figure 2-4. CPG sales forecasting AI application using COVID-19 data

The demand-forecasting AI app allowed the business leader to input

the products, regions, and retailer partners of interest to see the

CPG Sales Forecasting with COVID-19 Data | 21

predicted demand over the forecast period to plan for production
and delivery of products across geographies.

Hospital Staffing Optimization (Healthcare

Industry AI Apps)
Faced with COVID-19, US hospitals suddenly experienced potential
shortages in resources, including ventilators, personal protective
equipment, and nursing staff to fight the epidemic. Under normal
circumstances, forecasting models were not really required because
needs for equipment and staff were relative to the size of the hospital
and service area. With a new and deadly disease, however, resources
would be needed based on a new set of factors including infection
rates and demographics.
A leading US integrated healthcare company, serving over twelve
million people, realized in early March that they did not know
where to place resources across their hundreds of hospitals and
medical offices. They partnered with a leading AI technology com‐
pany to build a new AI application to forecast needs for ventilators
and nursing staff across locations.
The new AI application allowed users to input geographic areas,
hospitals or clinics of interest and then see how infection rates and
demographics for their patient community would impact resource
needs (Figure 2-5).
Using the new app, the administrative team was able to allocate
resources to the sites with the most need and get people and equip‐
ment there ahead of outbreaks in cases, which helped save lives for
both patients and healthcare workers.
Another critical area for AI apps is at the line of business level,
including key business functions like marketing, sales, services, and
support.

22 | Chapter 2: Modern AI Applications

Figure 2-5. Hospital resource forecasting AI application using
COVID-19 data

Marketing Lead Optimization (Line of

Business AI Applications)
Marketing departments in B2B companies generate leads from mar‐
keting programs that are followed-up by the sales department. Not
all leads, however, are of equal importance. For a given company, or
even at a given time of year, the factors that make a good lead could
change. Traditional lead scoring is done by rules based on the items
of content that a prospective customer views or downloads. A cus‐
tomer who comes to a live webinar might get more points than
someone who just downloads a solution brief, and so on. Once a
lead is passed to sales, each salesperson uses their own experience
and bias to determine which leads to follow up with. The problem
with this method is that it relies on human intuition to determine
what makes a good lead and is often “one size,” leading to missed
opportunities.
The lead scoring application allows the business development, mar‐
keting, and sales team to predict which prospects or accounts are
likely to make a purchase in a given period. A machine learning

Marketing Lead Optimization (Line of Business AI Applications) | 23

model looks at the history of previous prospects and the interactions
with the company to build a predictive model. This model is then
used to score new prospects. The higher the score, the more likely a
prospect will be to engage with sales and spend money on products
in a given time period.
A more granular ML lead scoring approach can impact many areas
of the business, from marketing to sales. Knowing which leads have
the highest potential helps everyone prioritize where to spend their
time. Marketing knows exactly what leads to pass along to sales and
which leads they need to nurture further. Sales knows which meet‐
ings to prioritize based on the relative score.
The lead scoring application uses AutoML to produce a model to
predict which leads in the funnel are good or bad. AutoML trains
and tunes several models based on the training data provided. To
run the lead optimization app, the user first selects the settings for
the geographic regions and data sources they would like to use
(Figure 2-6).
The output of the application depends on the settings the user
selected. Figure 2-7 shows a collection of reports that the user has to
explore the data about the leads, both good and bad. This analysis
allows the user to understand what factors are important to create
good leads and what contributes to poor leads. Marketing users can
then use this to tune marketing campaigns to those areas that pro‐
duce better results. Sales teams can use this information to find
other potential customers with similar profiles. Depending on the
needs of the users, the application can also produce a list of scored
leads that can be used directly for lead follow-up.

24 | Chapter 2: Modern AI Applications

Figure 2-6. Users adjust the settings to produce their desired analysis

Marketing Lead Optimization (Line of Business AI Applications) | 25

Figure 2-7. Lead dashboard provides detailed information about what
makes a good lead

Data Augmentation (AI Apps for Data Teams)

Another key step in the development of any analytic process is
access to data. When an organization’s data is insufficient or when
additional information could improve model accuracy or insight,
then external data sources will need to be acquired and integrated
with existing data. With the onset of COVID-19, many companies
found that existing predictive models no longer provided accurate

26 | Chapter 2: Modern AI Applications

results. Rather than look for more of the same data, i.e., a longer
dataset, some organizations looked to create wide datasets by adding
new data from external sources. With traditional processes, finding
datasets and discovering how to merge them, and then teasing these
signals out of the data can take months of painstaking development
and testing.
The data augmentation AI application uses fuzzy matching techni‐
ques to find relationships across datasets. This matching algorithm
saves time by finding new related data quickly using information
like time, location, or names in the data. Included in the application
are public datasets, like US census data, crime data, housing data,
and so on. The application also includes samples of private datasets
for users to try. With access to these datasets and an easy way to
match them to private data, organizations can quickly determine if
public or private data can provide additional signals to improve
model accuracy or provide insights to business users.
Let’s explore the data augmentation workflow as part of the lending
app (Figure 2-8). The process begins with selecting the initial dataset
for data augmentation. Users then select what datasets they would
like to consider for augmentations and the method for augmenta‐
tion (automatic or guided). Next, users are presented with a variety
of datasets that have matching columns that can be merged with the
first dataset.
The original datasets are then automatically blended, and more fea‐
tures are added. This newly enriched data can then be used to do
further analysis and potentially build better machine learning mod‐
els. This automated enrichment process allows data analysts and
data scientists to quickly and easily explore public and private data‐
sets to see if they provide new insights for their use case.
When data augmentation is combined with AutoML, data scientists
can very quickly see if new data provide valuable signals by creating
new models and comparing them to existing models. Without this
ability to rapidly and cheaply build models, exploring new datasets
for small increases in signal is not cost-effective.

Data Augmentation (AI Apps for Data Teams) | 27

Figure 2-8. Workflow guides the user to select data augmentation level
and augmentation datasets

28 | Chapter 2: Modern AI Applications

 CHAPTER 3
Case Studies: Real Impacts of
AI Application in Business

AI and machine learning can have significant impacts on business

transformation. Many companies and platforms use AI to create sol‐
utions that have a countless degree of impact on businesses. In this
chapter, we are going to focus on just a few of these solutions, and
how H2O.ai has worked with companies to solve their problems.

ArmadaHealth
ArmadaHealth is a health data science and services company
founded to help people access the right physician or expert for them.
Their unique solution, QualityCare ConnectSM, combines big data
and expert clinical insights which points straight to the root cause of
healthcare access problems. ArmadaHealth does this by applying
sentiment analysis on customer reviews and advanced analysis of
experts’ wisdom to understand the consumers, objectively finding
providers that meet their needs and preferences, preparing them,
and delivering timely access to a choice of the most appropriate
physicians for their condition (see Figure 3-1).

29
Figure 3-1. The QualityCare Connect app on the ArmadaHealth
website

Challenges
Finding the right specialist is the first step to receiving the right care.
However, consumers are often not equipped to navigate the complex
and confusing healthcare system. It can be challenging for patients
to discover which specialist they should approach for different
health situations and, even with a referral from a primary physician,
it can still be a long process to find the right specialist who can accu‐
rately treat them while also providing a satisfactory patient experi‐
ence. Finding the right match between patient and doctor quickly
can solve major problems and save lives.

Solution
AutoML is an essential part of reaching ArmadaHealth’s goal of
delivering accurate patient-expert matches through their online
application. Using the H2O.ai platform, including automatic
machine learning, the company is able to build and train a natural
language processing (NLP) model to identify the sentiment (posi‐
tive, negative, neutral) in each customer review. The company looks
at three main aspects in each review: treatment outcome, communi‐
cation, and attitude. These three aspects are critical to finding the
expert that best matches customer preferences.

30 | Chapter 3: Case Studies: Real Impacts of AI Application in Business

Results
ArmadaHealth has achieved its goal of helping people access the
right physician or expert for them using a purpose-built AI app.
Among the various positive results of using AI, ArmadaHealth saw
increased net promoter scores and faster model building and
deployment.

Hortifrut
Hortifrut, based in Chile, is the largest producer of blueberries in the
world, and operates farms in Peru, Chile, Mexico, Argentina, the
United States, Spain, Morocco, and China, with distribution across
37 countries. Hortifrut holds 25% of the world’s blueberry market
and uses AI to make distribution decisions across their expansive
operations. They are able to predict the quality of the blueberries
from origin to final destination, improving the consumer experience
with higher quality products, and increasing revenue throughout
the supply chain.

Challenges
Transporting fruit from the farm can take weeks, so Hortifrut has to
predict the quality of produce upon arrival. Not being able to do this
accurately can impact customer experience and revenue loss. But
getting such predictions accurately can be a difficult task, given the
complexity of the distribution channel, weather data, variety of data‐
sets, shipping times, and more. Traditional machine learning meth‐
ods and toolkits took months to build accurate predictions and
production-ready models. To scale the use of AI under these condi‐
tions would require hiring additional data science talent and
increasing the budget.

Solution
Hortifrut leveraged the H2O.ai platform to have better predictive
insights into the quality of their blueberries. They used capabilities
such as feature engineering, NLP, explainability, time-series analysis,
visualization, and scoring pipelines. Hortifrut is now able to scale
their data science efforts in order to deliver use cases such as pre‐
dicting the quality of blueberries based on features like variety, farm
origin, shipping time, vessel, and packaging, without hiring

Hortifrut | 31
additional data science talent in the team. These results are delivered
directly to business users making decisions, so they can take the cor‐
rect actions for each shipment.

Results
Hortifrut achieved the following key benefits using AI apps:

• Hortifrut has saved a significant amount of money by reducing

perishable claims. If the berries are spoiled at their destination,
there can be a loss of revenue from customer claims and also
cost valuable customer satisfaction.
• Hortifrut has been able to deliver real business results with a
small data science team by improving the productivity of the
team instead of hiring more people.
• Hortifrut is able to reduce the model development time from
three to five months down to three to five weeks.

Jewelers Mutual
Jewelers Mutual is one of the United States’ and Canada’s most
established and trusted providers of affordable and comprehensive
insurance for jewelers and consumers. As a leader in driving
customer-focused innovation and providing the latest technology to
a long-standing industry, Jewelers Mutual uses AI to deliver excep‐
tional customer experiences, prevent losses, and provide better pro‐
tection and policies for both jewelers and consumers.

Challenges
The leadership at Jewelers Mutual recognized the need to invest in
analytics, AI, and machine learning for improving overall customer
experiences. Their business relies on both being able to effectively
protect their customers’ businesses, and providing personal insur‐
ance directly to consumers—both with innovative customer experi‐
ences. Jewelers Mutual has been at the bleeding edge in adopting AI.
They collected data already available from losses, customers, and
multiple other sources, which weren’t tapped into before.

32 | Chapter 3: Case Studies: Real Impacts of AI Application in Business

Solution
Jewelers Mutual standardized on the H2O.ai platform to develop
predictive models. Their first deployed AI application helped com‐
mercial underwriters understand their customers better and pro‐
vided the reason codes as to why decisions were made by machine
learning models. These insights were then made available through
an app to the underwriters. Having interpretability and explainabil‐
ity as part of AI model development and deployment was also
instrumental in convincing business stakeholders to use AI to make
business decisions.

Results
The success that Jewelers Mutual has seen in adopting AI in their
business is a testament to the fact that regulated industries can ach‐
ieve real competitive advantage using AI. For example:

• Jewelers Mutual has been able to offer more competitive jewelry

protection insurance rates to its customers.
• AI insights are leading to some interesting outcomes; during the
2019 California wildfires and power outages, the Jewelers
Mutual team was able to identify jewelers that would need addi‐
tional physical security personnel to protect their properties and
inventory.

Jewelers Mutual | 33
 CHAPTER 4
Adoption Challenges for
Next-Generation Analytics

The new world of AI apps promises to make AI available to every‐

one across organizations for every function, from frontline employ‐
ees to executives. Before jumping into this AI-powered revolution,
there are a few critical issues that organizations should consider as
they invest. In addition to widely-discussed staffing and technology
issues, AI is presenting several less well-known challenges that will
be the focus of this chapter.

Ineffective Data and AI Principles

According to AlgorithmWatch, dozens of organizations, including
governments, have published data and AI usage principles.1 These
principles attempt to set broad guidelines for the use of data and AI
within an organization and also signal to the organization’s peers,
competitors, employees, or customers that they are considering cer‐
tain data and AI risks. Unfortunately, AlgorithmWatch also recently
published a report stating that only 10 of 160 reviewed sets of prin‐
ciples were enforceable.2 So, if your organization does create data
and AI principles, keep in mind that a major pitfall to avoid is

1 See AlgorithmWatch’s “AI Ethics Guidelines Global Inventory”.

2 See AlgorithmWatch’s “In the Realm of Paper Tigers: Exploring the Failings of AI Eth‐
ics Guidelines” by Leonard Haas and Sebastian Gießler, with additional research by
Veronika Thiel, April 28, 2020.

35
ineffectiveness. Getting a technological perspective, along with ethi‐
cal, legal, oversight, and leadership perspectives into organizational
AI and data principles are perhaps the best way to avoid such issues.

Lax Security Practices

In the hype around machine learning and data science, and in a rush
to build end-to-end data and AI products (the kind of technology
that reaches into data centers, out to the public, and back into data
centers), data scientists can be intentionally or accidentally given too
many privileges in an IT system. This is an ethics and security prob‐
lem. If the same person can manipulate a database, create a predic‐
tive model, and make it operational, they can make a predictive
model do what they want it to do, and in very subtle ways. (Maybe
it’s to give their girlfriend’s mother a giant loan, or to deny loans to
people in a political or socially discriminatory way.) Regardless,
these kinds of insider attacks against AI can cost your organization
money and are another avenue by which discrimination can enter
into AI. More standard concerns about data privacy must also be
recognized by AI practitioners to ensure solid security. Hence, AI
systems and the teams working on them should be under the same,
if not stronger, security constraints as other employees.

Inadequate Human Review

Related to the practice of model risk management, the concept of
effective challenge is used to improve AI implementation at large
financial services organizations in the US. An interpretation of an
effective challenge is that, when building AI systems, one of the best
ways to guarantee good results is to actively challenge and review
each step of the development process. Of course, a culture of effec‐
tive challenge must apply to everyone developing an AI system, even
so-called “rock-star” engineers and data scientists. For instance, the
Federal Reserve System’s famous SR 11-7 guidance on model risk
management makes no exceptions for rock-star data scientists, and
there’s probably a good reason for that: a rigorous human review of
AI systems is one of the best known methods for mitigating risks
associated with AI. One easy way to start to build a culture of effec‐
tive challenge is to hold mandatory weekly meetings where alterna‐
tive design and implementation choices for AI systems are put
forward, questioned, and discussed.

36 | Chapter 4: Adoption Challenges for Next-Generation Analytics

Downplaying Traditional Domain Expertise
Real-world success in AI often requires input from humans with a
deep understanding of the given problem domain. Of course, such
experts can help with feature selection and engineering, and inter‐
pretation of AI system outputs. But they can also serve as a basic
sanity and usefulness check mechanism. For instance, if you’re
developing a medical ML system, you should consult with physi‐
cians and other medical professionals. How will generalist data sci‐
entists be able to understand the subtlety and complexity inherent in
medical data and the results of AI systems trained on such data?
They might not be able to, and this can lead to AI incidents when
the system is deployed. The social sciences deserve a special callout
in this regard as well. Sometimes called “tech’s quiet colonization of
the social sciences,” technology companies are pursuing AI projects
that either replace decisions that trained social scientists should
make, or use practices, such as facial recognition, for criminal risk
assessments that have been highly criticized by social scientists.3

AI Security and Privacy

Like nearly every other powerful commercial technology, AI sys‐
tems are subject to failures and attacks. These can include the kind
of hacks that plague other public-facing IT systems, wherein attack‐
ers block services with massive amounts of incoming web traffic or
insert themselves between an AI service and a consumer. These can
also include specialized concerns regarding training and output data
privacy and security, or even highly specialized attacks on underly‐
ing machine learning algorithms.
In terms of data privacy and security, there are traditional data secu‐
rity concerns related to the confidentiality, integrity, and availability
of input training data, intermediate data generated by the AI system,
and output response data from the AI system, but there are also
increasing legal and regulatory obligations around data privacy.
These can include everything from the legal basis for data collection,
to anonymization requirements, data retention limitations, and

3 See, for example, “To Really ‘Disrupt,’ Tech Needs to Listen to Actual Researchers”,
Wired, June 26, 2019; Rumman Chowdhury’s post on Twitter; “AI Researchers Say Sci‐
entific Publishers Help Perpetuate Racist Algorithms”, MIT Technology Review, June 23,
2020.

Downplaying Traditional Domain Expertise | 37

alignment with organizational privacy policies. Moreover, security
and privacy breaches can also trigger breach reporting require‐
ments. Because AI is so hungry for data, all of these can indirectly,
or even directly, impact an organization’s use of AI.
For specialized attacks against machine learning algorithms that
underpin most of today’s AI systems, organizations should have sev‐
eral known attack vectors on their radar, including:

• Insider manipulation of training data (i.e., “data poisoning”)

• Manipulation of model outcomes by external adversaries
• The theft of intellectual property, like models and data, by exter‐
nal adversaries
• Trojan horse code or manipulations buried in complex machine
learning software and related artifacts, like model weights that
give a favorable outcome under certain conditions only known
to hackers or external adversaries

While basic security practices are an effective shield against some

attacks on machine learning, it’s important to consider these attacks
as part of updated model risk management or information security
policies. Organizations can also leverage security audits, bug boun‐
ties, and red-teaming to help understand their vulnerabilities and to
fortify their defensive measures.
The future of analytics in the enterprise is bright, but before you
begin or scale up on your journey to build AI applications across the
enterprise, there are clearly some organizational and security issues
to consider and mitigate before taking the plunge.

38 | Chapter 4: Adoption Challenges for Next-Generation Analytics

 Conclusion

The future of analytics is sure to take many forms. With the conver‐
gence of traditional business intelligence with artificial intelligence
and machine learning, the possibilities are endless and exciting.
With AI integrated into business and customer experiences, the
hope is that every business user will be more productive and
empowered and that every customer experience will be exciting,
resulting in new levels of customer satisfaction and engagement and
new growth opportunities for companies.
AI applications present a compelling way to implement AI in the
enterprise. AI apps are different from traditional dashboards and
business applications in that AI apps are designed together with
domain experts to meet their specific needs for descriptive, predic‐
tive, and prescriptive insights. AI app development is accelerated
using AutoML and rapid prototyping frameworks so that organiza‐
tions can scale AI access across the business. The results of AI appli‐
cations are tangible, such as helping patients find the right
physician, reducing waste in fruit shipments, or optimizing insur‐
ance underwriting.
The future, however, is far from certain. Barriers remain, both tech‐
nological and organizational. To reach this transformational future
will require a new wave of innovators and leaders with the vision to
find the right technology partners and create the AI applications
that will drive their business and even their industries for years to
come.

39
About the Authors
Dan Darnell is a seasoned product marketer with over 20 years of
experience in leading technology companies. For the past nine
years, he has been working on AI platforms and applications,
including senior roles at H2O.ai, DataRobot, ParallelM, Talend, and
Baynote. Before that, Dan was focused on analytics and optimiza‐
tion technologies at Adchemy, Interwoven, Oracle, and Siebel Sys‐
tems. He holds an MBA from Carnegie Mellon University and a BS
in engineering from The University of Colorado at Boulder.
Rafael Coss is Director of Technical Marketing and Developer Rela‐
tions at H2O.ai in Mountain View, CA. Before joining H2O.ai, he
was Director of Technical Marketing, Community, and Data Evan‐
gelist at Hortonworks. He also has served as the DataWorks Summit
program cochair for three years. Before Hortonworks, he was a
senior solution architect and manager of IBM’s World Wide Big
Data Enablement team and coauthored Hadoop for Dummies. At
IBM, he also had roles in technical product enablement, quality
engineering, and software development across multiple products
and initiatives: XML database tools, federated database, and object-
relational database. He holds an MS in computer science and a BS in
civil engineering, both from California Polytechnic State University.
Patrick Hall is principal scientist at bnh.ai in Washington, DC, a
boutique law firm focused on data analytics and AI. Patrick also
serves as a visiting professor in the Department of Decision Sciences
at George Washington University and as an advisor to H2O.ai.