VIEWPOINT

IMPLEMENTING AI IN BUSINESS —
CHALLENGES AND RESOLUTIONS

Large enterprises need to implement AI solutions quickly and efficiently,

but many don’t have the right people, process and technology. Firms can
accelerate progress to AI maturity through use of a robust workbench, an
operating model built on a data science-led pyramid of skills, and pre-built
solutions when data complexity and talent scarcity are too big a challenge.
Artificial intelligence has the potential turn to providers and platforms to get use it properly. Research shows that
to cause a revolution in business. The a leg up on their journey, but this too AI can increase enterprise profit by 38
companies that use it best can re- requires cognizance of the right use percentage points, while delivering
invent themselves and their services, case application, and often results $14 trillion of additional gross added
disrupting competition and grabbing simply in “AI for AI’s sake.” value to corporations by 2035.2
market share.
With the challenges, however,
But it’s hard to do. are resolutions that, if carried out AI increases profit by 38
effectively, improve efficiencies, create percentage points
There’s a lot of noise and hype around
innovative operating models, and —
the technology, a cloak that veils the
for those brave enough — discover
challenge of actually implementing There are many forms of AI, including
new revenue streams.
enterprise AI. Only 8% of firms engage image recognition and capabilities
in core AI practices that support
widespread adoption.1
What is enterprise AI that understand natural language.
These algorithms process data,
A good business case is often lacking, today? understand its context and then take
actions based on intelligence.
processing power for running AI is running Amazon data centers and
complex algorithms is insufficient, reading thousands of legal contracts in AI can create meaningful solutions to
AI terminology keeps executives in minutes. It is finding cancerous tumors business problems. Live Enterprise,
the dark and data scientists are in an Infosys initiative, uses its ‘Nia’ AI
and predicting when customers will
capability to optimize employee and
short supply. Many laggard firms get bored and leave a website. It
partner requests and resolutions. The
don’t have the working culture to can identify fraudulent credit card
platform guides users to take actions
compete against titans like Google and behavior much better than other that save time, freeing them to do
Facebook, nor do they have the data to methods. It can bestow extreme more challenging work.
create sophisticated AI models. Many competitive advantage on those that

External Document © 2019 Infosys Limited

 To get the flywheel spinning, Arvind Krishna, IBM’s senior vice This bouquet of algorithms needs to
executives must put a strong business president of cloud and cognitive function as a unit and scale up easily
case in place and decide which use software, points out that collecting to work on enterprise systems. They
cases to pursue. They must then and cleansing this data is 80% of the will vary among industries and will be
understand the implications of this work on an AI project, and smaller dependent on data type (a complex
road map on people, process and firms can take over a year to complete document of text and images is very
technology — the challenges they will this process.4 Many lose faith and run different from a text document of five
face, and possible resolutions they can out of patience. lines).
turn to.
Challenge No. 2: Complex Challenge No. 3: Complex data
AI challenges faced by algorithms types

enterprises Algorithms exist for processing data,

deriving insights and finally taking
Many algorithms exist for processing
complex data types such as video
Challenge No. 1: The need for action on that intelligence. and audio. Their accuracy is often
lots of clean data A retail use case offers food for
questionable, however, and there is
the further challenge of scaling the
Google, Amazon, Facebook and Apple thought. A given purchase order is
solutions.
(collectively known as GAFA) are good fed into the system. The AI needs to
at AI. They have data — lots of data uncover whether it is in fact a purchase Challenge No. 4: The need for
to cleanse for processing. Google can order as well as its quantity, price, faster processing
train AI algorithms on the 3.5 billion data and other variables. The order
Complex algorithms that work on
searches that are made every 24 number can be triangulated with
huge data sets, such as those found
hours, while the 6 trillion likes posted the vendor supplying the product,
in Amazon Alexa, can take days to
to Facebook in the same period is providing insights into whether any
train. According to AI consultant
a potential gold mine.3 GAFA didn’t discrepancies between ordering and
Andrew Burgess, any improvement
set out to be AI companies at their delivery were made. This insight might
in processor speed will ensure that
genesis, but the tree of knowledge then result in vendor blacklisting if too
AI systems remain useful while using
offers very lucrative fruit that is too many discrepancies were found in a
up-to-date models.5 Cloud computing
tempting not to bite. six-month period.

External Document © 2019 Infosys Limited

is a boon here; AI-as-a-service uses used in day-to-day operations,” says Dr. involved in the service design need
specialized hardware that carries out Raghavan. to pay particular attention to the
processing on demand — a sure-fire interfaces and user experience, and for
Challenge No. 7: Difficulty of
catalyst for the democratization of AI. many designers, AI is an area where
knowledge sharing and code they will not have expertise.
Challenge No. 5: Real-time reuse
processing Embedding complex algorithms into
In larger firms, different divisions may existing operations and automation is
Real-time — or “stream” — processing develop their own algorithms, leading also important. This calls for creating
and generation of insights is to duplication if code is not shared production grade AI applications
challenging without disrupting between teams. These inefficiencies that will work in legacy systems. A
transaction systems. increase costs and ultimately lead to large telco such as AT&T might be
“Most of the algorithms today are built inconsistency across the enterprise. dealing with 130 off-the-shelf software
on offline insights,” says Dr. Raghavan, This can even affect how regulators products, with data coming from 60%
chief data scientist at Infosys. “The view a firm’s use of AI, particularly of those systems, which in turn may
algorithms are scored against a model if it seems that there is no single come from 10,000 databases — not an
built on historical data. A new area of methodology or framework in place. easy task for any production shop.
research that many experts are looking Challenge No. 8: Dependence Challenge No. 10: Stakeholder
at is ‘How do we build models that are on niche, expert talent
valid for data that was generated in
involvement
the last half hour?’” There is often a big talent gap “Getting business and IT stakeholders
between firms already moving on their around a table is a significant
Challenge No. 6: Multiple AI path and those still evaluating the challenge at most firms,” says
providers with varying strengths technology. Those without momentum consultant and author Andrew
The sheer variety of AI platforms and often give up due to lack of dedicated Burgess.8
vendors can be daunting for any resources.6 One reason is the
mammoth gulf between the number So too is managing expectations of
business executive to contend with.
of AI practitioners (300,000 worldwide) business executives once they do
There are tech giants such as GAFA
and the number of job postings (in collaborate with IT. For many, AI is seen
that have cloud-based platforms to
the millions). The best candidates go as a magic pill that will reduce the
build models, understand granular
to companies like DeepMind, Google, costs of operations while improving
data and deploy models into
Airbnb and Intel — and others that production quality.
production. Then there are companies
like RapidMiner, DataRobot and tout exciting use cases and significant “Many executives don’t realize that it’s
Dataiku that have built their own monetary rewards.7 complex to create AI applications,” says
frameworks, with features evolving Dr. Raghavan. “Often, executives have
every six months or so. Finally, there The 300,000 AI practitioners unrealistic estimates of production,
are startups exploring just one aspect hoping that a new system will be ready
worldwide can choose from in the next two to three weeks.”
of the entire ecosystem.
millions of AI job postings
A large enterprise such as a big bank
or consumer goods manufacturer,
Resolutions
Even more crucial is finding talent with A robust ‘workbench’ approach can
will run proofs of concept to decide
industry-specific knowledge. If found, be used to circumvent many of these
the functional areas in which to
these workers become indispensable challenges. Firms should also build up
run these solutions, or whether to
to a project until its completion. capabilities across people, process and
use an enterprise wide solution at
all. For this, firms must employ in- Challenge No. 9: Developing the technologies. Pre-built solutions are a
house talent to understand vendor products way forward for those challenged to
nuances, the algorithms in use, ease of secure top talent or that have an AI-
Once AI algorithms are in place, the averse culture.
implementation, data sources, ease of
business needs to commercialize the
training and so on. The robust workbench
opportunity through the creation of
“Evaluating platforms is not easy work. products and services. This requires A workbench brings together
You must test them out, understand significant input from product development, deployment and
their core production and capability, designers, business analysts and ongoing maintenance under one
and determine whether they can be business development teams. Those roof. It offers users a factory-like

External Document © 2019 Infosys Limited

 process, with automation and DevOps • Multicloud API enabled the forest for the trees. Building up a
embedded in the process. pyramid of skills dedicated to AI-
Even within one enterprise, a good
specific systems is crucial.
“The AI workbench encapsulates the workbench must have the flexibility to
data layer for a data scientist,” says support multiple cloud environments, For instance, data scientists classically
Dr. Raghavan. “All they need to worry in order to foster innovation and use structured data. However, with
about is where to pull the data from, preserve business unit autonomy (the the proliferation of text, speech,
giving them freedom to build and finance division in the U.K. might use image and video, an organization
validate models across geographies Microsoft Azure, for example, while will need to build up capabilities in
where hundreds of scientists will be procurement in Germany uses Amazon those specific areas too. Once the
working on the same problem.” Web Services). This in turn solves the skills pyramid is created (a bank might
challenge of getting business and IT need connoisseurs of text data while a
This means that models can be
stakeholders around the table to push media firm will need video and image
standardized, and algorithms can
the AI strategy through quickly. analytics expertise), firms will be more
be reused on a global scale. It also
confident that skilled workers won’t
provides a single access point for
A good AI workbench must leave them partway through a project.
validated models to be deployed into
They will also find that their reputation
production. The workbench should have multi-cloud flexibility to as an AI-driven organization will be
consist of the following elements:
foster innovation and preserve enhanced, giving them the motivation
• API-driven development to meet any one of the other 10
autonomy
A good workbench offers Application challenges we listed.
Programming Interface-driven • Guided experience For firms that are not highly technical
development, which forgoes or those that are less mature,
dependence on ultra-niche talent. The guided experience is a “playbook”
that increases momentum in any AI we advocate hiring experienced
It also enables code reuse and professionals who can evaluate
knowledge sharing. Here, interactions project during the design stage. The
point is to showcase how the AI system external talent that will ultimately
with a program are standardized build up the pyramid.
and essentially static. For example, will work once deployed, which can be
used to attract data scientists to the • Playbooks for specific knowledge
IBM’s Watson (the heavily lauded AI
capability) is just a series of APIs that project and ensure the right partner Here, documentation and “wiki pages”
each carry out a specific function, such comes along for the journey. A retail for a particular project are developed
as speech recognition or Q&A, which banking chatbot can be simulated, to detail certain aspects of use cases
can be called by another program with for instance, offering guidance to that an enterprise wants to build out.
the right access. APIs also democratize key stakeholders on its look and feel, For instance, a telco or retail company
AI, since the value is no longer in implications for back-end data sources might develop a recommendation
the algorithm but in the amount and ultimately the user experience. engine. The “next best action” of the
of training that is carried out on a The guided experience is crucial for engine will be built on playbooks that
given dataset. AI projects with complex algorithms detail what specific components need
• Containers and architectures. If solution architects to be created and what data sources
overlook this part of the workbench, need to be included. This reduces
Containers — encapsulated the complexity of building the right
project development degenerates into
applications with their own operating code and simplifies dialogue between
“spaghetti,” with too many what-ifs
system and memory — mean that business and IT CXOs.
occurring in design.
an AI application can be published
• Provider-specific training and hands-
as a service and can be maintained Capability buildup
on experience
and monitored very easily. From a • Data scientist-led talent
deployment perspective, containers If an enterprise wants to build an Azure
ensure that downstream applications Very few firms make the most of stack, an Azure developer is needed
don’t worry about specific aspects of AI capabilities, which, with some who knows the Azure data lake and is
modules, libraries and functions or thought, can be integrated through familiar with Azure machine learning,
the AI application itself. Containers effective partnering and ready-made Azure Cortana and so on. This might
help with code reuse and sharing (like solutions. Most often this is due to seem obvious, but is often overlooked.
API calls) and reduce the number of a lack of in-house data science skills So too is ensuring that workers
complex algorithms in use. and AI expertise that prevents seeing have used a particular platform, in

External Document © 2019 Infosys Limited

a particular style and in a particular • A pre-built box of algorithms starting with speech detection and
industry. If predictive AI for a mining moving on to sentence detection,
In-house development involves a lot
company is the use case, having sentence bonding and looking at
of cost; identifying the right resources
someone familiar with the mining correlations across sentences. Each
to create AI algorithms is expensive
industry will ensure that they solve of these steps use standard elements
both in planning and development.
the right problems, since they talk available in open source.
Having a pre-built library is a much
the same language and understand
more convenient way of developing “You can then wire step one to 10
the domain. Industry-specific
an algorithm that works for any given as part of the sentiment analysis
experience also helps ease adoption.
use case. and create a pre-built API,” says Dr.
The relevant person will know exactly
Raghavan. “Anyone who wants to use
where to procure a platform or pre- “If, for example, I’m doing clustering or
that API might have to do some minor
built solution, and will know how time-series regression, having a library
customization, but they won’t build
to influence key stakeholders, while is always a blessing,” says Dr. Raghavan.
anything from scratch.”
enabling easier adoption of AI into the If artifacts are available at the library
overall system. So when it comes to sentiment
level, then a developer can custom
analysis, building and deploying
Pre-built solutions build any application very quickly.
solutions becomes much easier if
Access to AI software libraries speeds Sentiment analysis is a good example that pre-built layer already exists.
up development and deployment, of a pre-built box of algorithms to • POC solutions
while proof-of-concept (POC) solve problems. Sentiment analysis can
be embedded as part of a corporate POC solutions are part of a natural
solutions ensure that the product is
communications application, or it can agile cycle for building and deploying
built cheaply and quickly. Pre-trained
be part of a product development AI solutions, allowing firms to get
models working on huge data sets
application. Here, pre-built critical insight into AI platforms,
help democratize AI and counter the
components parse text. There are at potential providers and ensuring
talent challenge.
least 10 different steps in the analysis, that the product is built cheaply
and quickly.

External Document © 2019 Infosys Limited

 To build an AI solution, underlying questions or answers, you can ask challenges must be resolved, including
challenges must first be understood. your own questions and it will scan scarcity of talent, risk-averse cultures,
Very rarely is the first solution similar to the entire document in real-time and and an inability to imagine what the
the final solution. Different algorithms give you a good answer nine times solution will look like on completion.
must be tested, along with different out of 10. In the future, insights will be
We advocate starting with a
software vendors and software derived from even more complex data
workbench that bridges the
frameworks, and the solution must types such as images and video.”
gulf between development and
be tested in the user environment.
This pre-built solution effectively production. Building capabilities
The POC solution created builds
reduces the dependence on expert in-house must also be a priority,
confidence and, more importantly,
coders and solves the problem of accelerating changes across people,
ensures that the project runs on time
processing complex data types in real- process and technology. Finally, if
and isn’t a failure. Once complete, the
time — a truly disruptive technology if these skills are hard to come by, pre-
solution is scaled up.
ever there was one. built solutions trained on big data can
• AI-off-the-shelf be employed to generate momentum
Infosys has built a pre-trained The future enterprise toward AI maturity.
functionality known as Document AI is guiding decisions on everything Before thinking about implementation,
Comprehension. from bank loans to crop harvests. And executives must put a strong business
“When you upload a document, it’s big business. Harvard Business case in place, and decide which use
however complex, the pre-trained Review estimates that AI will add cases to pursue. They must then
model is so sophisticated that it can $13 trillion to the global economy understand the implications of this
derive questions and answers from over the next decade.9 However, as road map on their particular industry.
the image and document by itself,” of this writing, most firms have only Only then will they have clarity to
says Harinder Cour, a consultant applied AI in a single business process. make it back to shore and turn data
at the Infosys Center for Emerging To improve efficiencies and identify into cash.
Technologies. “If you don’t like those new revenue opportunities, many

External Document © 2019 Infosys Limited

External Document © 2019 Infosys Limited
 References
1.
Building the AI-Powered Organization, HBR, 2019
2.
How AI Boosts Industry Profits and Innovation, Accenture
3.
Andrew Burgess, The Executive Guide to Artificial Intelligence, Palgrave Macmillan
4.
Data challenges are Halting AI Projects, IBM Executive Says
5.
Andrew Burgess, The Executive Guide to Artificial Intelligence, Palgrave Macmillan
6.
AI Adoption Held Back by Company Culture, Talent Shortage, Data Issues”, WSJ, 2019
7.
It’s Recruiting Season for AI’s Top Talent, and Things Are Getting a Little Zany”, MIT Technology Review
8.
Andrew Burgess, The Executive Guide to Artificial Intelligence, Palgrave Macmillan
9.
Building the AI-Powered Organization, HBR, 2019

SME
Dr N.R. Srinivasa Raghavan
Global Chief Data Scientist
Infosys

Authors
Harry Keir Hughes Isaac LaBauve
Senior Consultant Senior Consultant
Infosys Knowledge Institute Infosys Knowledge Institute

External Document © 2019 Infosys Limited

About Infosys Knowledge Institute
The Infosys Knowledge Institute helps industry leaders develop a deeper understanding of business and technology trends
through compelling thought leadership. Our researchers and subject matter experts provide a fact base that aids decision
making on critical business and technology issues.
To view our research, visit Infosys Knowledge Institute at infosys.com/IKI

For more information, contact [email protected]

© 2019 Infosys Limited, Bengaluru, India. All Rights Reserved. Infosys believes the information in this document is accurate as of its publication date; such information is subject to change without notice. Infosys
acknowledges the proprietary rights of other companies to the trademarks, product names and such other intellectual property rights mentioned in this document. Except as expressly permitted, neither this
documentation nor any part of it may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, printing, photocopying, recording or otherwise, without the
prior permission of Infosys Limited and/ or any named intellectual property rights holders under this document.

Infosys.com | NYSE : INFY Stay Connected