VIEWP

OINT

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 turn to providers and platforms use it properly. Research shows
potential to cause a revolution to get a leg up on their that AI can increase enterprise
in business. The companies journey, but this too requires profit by 38 percentage points,
that use it best can re- invent cognizance of the right use while delivering
themselves and their services, case application, and often $14 trillion of additional gross
disrupting competition and results simply in “AI for AI’s added value to corporations by
grabbing market share. sake.” 2035.2
AI increases
But it’s hard to do. With the challenges, profit by 38
however, are resolutions
There’s a lot of noise and hype
that, if carried out
percentage
around the technology, a cloak There are many forms of AI,
effectively, improve efficiencies, points
that veils the challenge of including image recognition
create innovative operating
actually implementing and capabilities that
models, and — for those brave
enterprise AI. Only 8% of firms understand natural language.
enough — discover new These algorithms process data,
engage in core AI practices that
revenue streams. understand its context and
support widespread adoption.1
then take actions based on
A good business case is often What is intelligence.
lacking, processing power for
running complex algorithms is enterprise AI AI can create meaningful
insufficient, today? solutions to business problems.
Live Enterprise, an Infosys
AI terminology keeps
AI is running Amazon data initiative, uses its ‘Nia’ AI
executives in the dark and
centers and reading thousands capability to optimize employee
data scientists are in short
of legal contracts in minutes. It and partner requests and
supply. Many laggard firms
is finding cancerous tumors and resolutions. The platform
don’t have the working guides users to take actions
culture to predicting when customers will
get bored and leave a website. that save time, freeing them to
compete against titans like do more challenging work.
Google and Facebook, nor do It
they have the data to create can identify fraudulent credit
External Document © 2019 Infosys
sophisticated AI models. Many card behavior much better Limited
than other methods. It can
 To get the flywheel spinning, Arvind Krishna, IBM’s senior This bouquet of algorithms
executives must put a strong vice president of cloud and needs to function as a unit and
business case in place and cognitive software, points out scale up easily to work on
decide which use cases to that collecting and cleansing enterprise systems. They will
pursue. They must then this data is 80% of the work on vary among industries and will
understand the implications of an AI project, and smaller firms be dependent on data type (a
this road map on people, can take over a year to complex document of text and
process and technology — the complete this process.4 Many images is very different from a
challenges they will face, and lose faith and run out of text document of five lines).
possible resolutions they can patience.
turn to.
Challenge No. 3: Complex
Challenge No. 2: data types
AI challenges Complex algorithms Many algorithms exist for
faced by Algorithms exist for processing
data, deriving insights and
processing complex data
types such as video and
enterprises finally taking action on that audio. Their accuracy is often
intelligence. questionable, however, and
Challenge No. 1: The there is the further challenge
need for lots of clean A retail use case offers food for
of scaling the solutions.
data thought. A given purchase
order is fed into the system. Challenge No. 4: The
Google, Amazon, Facebook and The AI needs to need for faster
Apple (collectively known as uncover whether it is in fact a
GAFA) are good at AI. They
processing
purchase order as well as its
have data — lots of data quantity, price, Complex algorithms that work
to cleanse for processing. data and other variables. The on huge data sets, such as
Google can train AI algorithms order number can be those found in Amazon Alexa,
on the 3.5 billion searches that triangulated with the vendor can take days to train.
are made every 24 hours, supplying the product, According to AI consultant
while the 6 trillion likes posted providing insights into whether Andrew Burgess, any
to Facebook in the same period any discrepancies between improvement in processor
is ordering and speed will ensure that AI
a potential gold mine.3 GAFA delivery were made. This insight systems remain useful while
didn’t set out to be AI might then result in vendor using
companies at their genesis, blacklisting if too many up-to-date models.5 Cloud
but the tree of knowledge discrepancies were found in a computing
offers very lucrative fruit that six-month period.
is too tempting not to bite.

External Document © 2019 Infosys

Limited
is a boon here; AI-as-a-service used in day-to-day operations,” involved in the service design
uses specialized hardware that says Dr. Raghavan. need to pay particular
carries out processing on attention to the
demand — a sure-fire catalyst
Challenge No. 7: interfaces and user experience,
for the democratization of AI. Difficulty of knowledge and for many designers, AI is an
sharing and code reuse area where they will not have
Challenge No. 5: Real- expertise.
In larger firms, different
time processing
divisions may develop their Embedding complex algorithms
Real-time — or “stream” — own algorithms, leading to into existing operations and
processing and generation of duplication if code is not shared automation is also important.
insights is challenging without between teams. These This calls for creating
disrupting transaction systems. inefficiencies increase costs and production grade AI
“Most of the algorithms today are ultimately lead to inconsistency applications
built on offline insights,” says Dr. across the enterprise. that will work in legacy systems.
Raghavan, chief data scientist at This can even affect how A large telco such as AT&T
Infosys. “The algorithms are regulators view a firm’s use might be dealing with 130 off-
scored against a model built on of AI, particularly if it seems the-shelf software products, with
historical data. A new area of that there is no single data coming from 60% of
research that many experts are methodology or framework in those systems, which in turn
place. may come from 10,000
looking at is ‘How do we build
models that are valid for data Challenge No. 8: databases — not an easy task
that was generated in the last Dependence on niche, for any production shop.
half hour?’” expert talent Challenge No. 10:
Challenge No. 6: Multiple There is often a big talent gap Stakeholder involvement
providers with varying between firms already moving “Getting business and IT
strengths on their AI path and those still stakeholders around a table is
evaluating the technology. a significant challenge at most
The sheer variety of AI
Those without momentum often firms,” says consultant and
platforms and vendors can be
give up due to lack of dedicated author Andrew Burgess.8
daunting for any business
resources.6 One reason is the
executive to contend with. There So too is managing expectations
mammoth gulf between the
are tech giants such as GAFA of business executives once
number of AI practitioners
that have cloud-based they do collaborate with IT. For
(300,000 worldwide) and the
platforms to build models, many, AI is seen as a magic pill
number of job postings (in
understand granular data and The 300,000 AI
the millions). The best that will reduce the costs of
deploy models into production.
practitioners
candidates go to companies operations while improving
Then there are companies like production quality.
likeworldwide
DeepMind, Google, canAirbnb
RapidMiner, DataRobot and
and Intel — and others that
Dataiku that have built their choose from “Many executives don’t realize
tout exciting use cases and that it’s complex to create AI
own frameworks, with features Even more crucial
millionsmonetary
of AI is finding
job
significant rewards. 7
applications,” says Dr. Raghavan.
evolving every six months or talent with industry-specific
so. Finally, there are startups
postings
knowledge. If found, these “Often, executives have
exploring just one aspect of the workers become indispensable unrealistic estimates of
entire ecosystem. to a project until its completion. production, hoping that a new
system will be ready in the next
A large enterprise such as a Challenge No. 9: two to three weeks.”
big bank or consumer goods Developing the products
manufacturer, will run proofs
of concept to decide the
Once AI algorithms are in Resolutions
place, the business needs to A robust ‘workbench’ approach
functional areas in which to
commercialize the opportunity can be used to circumvent
run these solutions, or
through the creation of many of these challenges.
whether to use an enterprise
products and services. This Firms should also build up
wide solution at all. For this,
requires significant input from capabilities across people,
firms must employ in- house
product designers, business process and technologies. Pre-
talent to understand vendor
analysts and business built solutions are a way
nuances, the algorithms in use,
development teams. Those forward for those challenged to Infosys
ease of implementation, data External Document © 2019
Limited
sources, ease of training and so secure top talent or that have an
 process, with automation and • Multicloud API enabled the forest for the trees.
DevOps embedded in the Building up a pyramid of skills
Even within one enterprise, a
process. dedicated to AI- specific
good workbench must have the
systems is crucial.
“The AI workbench flexibility to support multiple
encapsulates the data layer for cloud environments, in order to For instance, data scientists
a data scientist,” says Dr. foster innovation and preserve classically use structured data.
Raghavan. “All they need to business unit autonomy (the However, with the proliferation
worry about is where to pull finance division in the U.K. might of text, speech, image and
the data from, giving them use Microsoft Azure, for example, video, an organization
freedom to build and validate while procurement in Germany will need to build up capabilities
models across geographies uses Amazon Web Services). in those specific areas too.
where hundreds of scientists This in turn solves the challenge Once the skills pyramid is
will be working on the same of getting business and IT created (a bank might need
problem.” stakeholders around the table to connoisseurs of text data while
push the AI strategy through a media firm will need video
This means that models can A good AI
quickly. and image analytics expertise),
be standardized, and workbench must firms will be more confident
algorithms can be reused
on a global scale. It also have multi-cloud that skilled workers won’t leave
flexibility to foster them partway through a
provides a single access
project.
point for innovation and They will also find that their
validated models to be • Guided experience
deployed into production. The
preserve autonomy reputation as an AI-driven
The guided experience is a organization will be enhanced,
workbench should consist of
“playbook” that increases giving them the motivation to
the following elements:
momentum in any AI project meet any one of the other 10
• API-driven development during the design stage. The challenges we listed.
A good workbench offers point is to showcase how the AI
For firms that are not highly
Application Programming system will work once deployed,
technical or those that are less
Interface-driven development, which can be used to attract
mature,
which forgoes dependence on data scientists to the project and
we advocate hiring
ultra-niche talent. ensure the right partner comes
experienced professionals
It also enables code reuse and along for the journey. A retail
who can evaluate external
knowledge sharing. Here, banking chatbot can be
talent that will ultimately
interactions with a program simulated,
build up the pyramid.
are standardized for instance, offering guidance
to key stakeholders on its look • Playbooks for specific
and essentially static. For
and feel, knowledge
example, IBM’s Watson (the
heavily lauded AI capability) is implications for back-end data Here, documentation and “wiki
just a series of APIs that each sources and ultimately the user pages” for a particular project
carry out a specific function, such experience. are developed to detail certain
as speech recognition or Q&A, The guided experience is aspects of use cases that an
which can be called by another crucial for AI projects with enterprise wants to build out.
program with the right access. complex algorithms For instance, a telco or retail
APIs also democratize AI, since and architectures. If solution company might develop a
the value is no longer in architects overlook this part of recommendation engine. The
the algorithm but in the the workbench, project “next best action” of the
amount of training that is development degenerates into engine will be built on
carried out on a given “spaghetti,” with too many what- playbooks that detail what
dataset. ifs occurring in design. specific components need to be
• Containers created and what data sources
Capability buildup need to be included. This
Containers — encapsulated • Data scientist-led talent reduces
applications with their own the complexity of building the
operating system and memory Very few firms make the
right code and simplifies
— mean that most of AI capabilities,
dialogue between business and
an AI application can be which, with some
IT CXOs.
External Document © 2019 Infosys
published as a service and thought, can be integrated
Limited
through effective partnering • Provider-specific training and
can be maintained and
a particular style and in a • A pre-built box of algorithms starting with speech detection
particular industry. If predictive and moving on to sentence
In-house development involves
AI for a mining company is the detection, sentence bonding
a lot of cost; identifying the
use case, having someone and looking at correlations
right resources to create AI
familiar with the mining across sentences. Each of
algorithms is expensive both in
industry will ensure that they these steps use standard
planning and development.
solve the right problems, since elements available in open
Having a pre-built library is a
they talk the same language source.
much more convenient way of
and understand the domain.
developing an algorithm that “You can then wire step one
Industry-specific experience
works for any given use case. to 10 as part of the
also helps ease adoption. The
sentiment analysis and
relevant person will know “If, for example, I’m doing
create a pre-built API,” says
exactly where to procure a clustering or time-series
Dr.
platform or pre- built solution, regression, having a library is
Raghavan. “Anyone who wants
and will know how always a blessing,” says Dr.
to use that API might have to
to influence key stakeholders, Raghavan.
do some minor customization,
while enabling easier adoption If artifacts are available at the but they won’t build anything
of AI into the overall system. library level, then a developer from scratch.”
Pre-built solutions can custom build any
So when it comes to
application very quickly.
Access to AI software libraries sentiment analysis,
speeds up development and Sentiment analysis is a good building and deploying
deployment, while proof-of- example of a pre-built box of solutions becomes much
concept (POC) solutions ensure algorithms to solve problems. easier if that pre-built layer
that the product is built Sentiment analysis can be already exists.
cheaply and quickly. Pre- embedded as part of a corporate • POC solutions
trained models working on communications application, or
it can be part of a product POC solutions are part of a
huge data sets help
development application. Here, natural agile cycle for building
democratize AI and counter the
pre-built components parse and deploying AI solutions,
talent challenge. External Document © 2019 Infosys
text. There are at least 10 allowing firmsLimited
to get critical
insight into AI platforms,
 To build an AI solution, questions or answers, you can challenges must be resolved,
underlying challenges must first ask your own questions and it including scarcity of talent, risk-
be understood. Very rarely is the will scan the entire document in averse cultures, and an inability
first solution similar to the final real-time and give you a good to imagine what the solution will
solution. Different algorithms answer nine times out of 10. In look like on completion.
must be tested, along with the future, insights will be
We advocate starting with
different software vendors and derived from even more
a workbench that bridges
software frameworks, and the complex data types such as
the gulf between
solution must images and video.”
development and
be tested in the user
This pre-built solution effectively production. Building
environment. The POC solution
reduces the dependence on capabilities in-house must
created builds confidence and,
expert coders and solves the also be a priority,
more importantly, ensures that
problem of processing complex accelerating changes across
the project runs on time and
data types in real- time — a people, process and
isn’t a failure. Once complete,
truly disruptive technology if technology. Finally, if these
the solution is scaled up.
ever there was one. skills are hard to come by, pre-
• AI-off-the-shelf built solutions trained on big
Infosys has built a pre- The future data can be employed to
generate momentum toward
trained functionality known
as Document enterprise AI maturity.
Comprehension. AI is guiding decisions on Before thinking about
“When you upload a everything from bank loans to implementation, executives
document, however complex, crop harvests. And it’s big must put a strong business case
the pre-trained model is so business. Harvard Business in place, and decide which use
sophisticated that it can Review estimates that AI will cases to pursue. They must then
derive questions and answers add understand the implications of
from the image and $13 trillion to the global this road map on their particular
document by itself,” says economy over the next industry. Only then will they
Harinder Cour, a consultant decade.9 However, as of this have clarity to make it back to
at the Infosys Center for writing, most firms have only shore and turn data into cash.
Emerging Technologies. “If applied AI in a single business
you don’t like those process. To improve efficiencies
and identify 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 Isaac LaBauve
Hughes Senior Consultant
Senior Consultant Infosys Knowledge
Infosys Knowledge Institute
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 : Stay

Connected
INFY