AI-ENABLED

SOFTWARE-AS-A-SERVICE (SaaS)
The Next Frontier for Global SaaS Start-Ups from India

IN COLLABORATION WITH
About IFC
IFC—a member of the World Bank Group—is the largest global development institution focused on the private sector in emerging
markets. We work in more than 100 countries, using our capital, expertise, and influence to create markets and opportunities in
developing countries. In fiscal year 2021, IFC committed a record $31.5 billion to private companies and financial institutions in
developing countries, leveraging the power of the private sector to end extreme poverty and boost shared prosperity as economies
grapple with the impacts of the COVID-19 pandemic. For more information, visit www.ifc.org.

About Stellaris Venture Partners

Stellaris Venture Partners is an early-stage tech-focused VC firm that partners with fearless founders who dream big, think differently,
and have an unstoppable desire to challenge the status quo. We work with these teams at nascent stages to help transform their ideas into
resilient businesses. We do so with a team of seasoned, diverse operators and entrepreneurs who bring significant experience in building
businesses, and by providing access to some of the most successful founders and professionals in India and the United States. Stellaris
has financed a number of successful and promising software as a service (SaaS) companies such as Whatfix, Signzy, and Slintel. For more
information visit www.stellarisvp.com

© International Finance Corporation. First printing, February 2022. All rights reserved.
2121 Pennsylvania Avenue, N.W.
Washington, D.C. 20433
www.ifc.org/thoughtleadership

The material in this work is copyrighted. Copying and/or transmitting portions or all of this work without permission may be a
violation of applicable law. IFC encourages dissemination of its work and will normally grant permission to reproduce portions of the
work promptly, and when the reproduction is for educational and non-commercial purposes, without a fee, subject to such attributions
and notices as we may reasonably require.
IFC does not guarantee the accuracy, reliability or completeness of the content included in this work, or for the conclusions or
judgments described herein, and accepts no responsibility or liability for any omissions or errors (including, without limitation,
typographical errors and technical errors) in the content whatsoever or for reliance thereon. The boundaries, colors, denominations,
and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal
status of any territory or the endorsement or acceptance of such boundaries. The findings, interpretations, and conclusions expressed in
this volume do not necessarily reflect the views of the Executive Directors of The World Bank or the governments they represent.
The contents of this work are intended for general informational purposes only and are not intended to constitute legal, securities,
or investment advice, an opinion regarding the appropriateness of any investment, or a solicitation of any type. IFC or its affiliates
may have an investment in, provide other advice or services to, or otherwise have a financial interest in, certain of the companies and
parties named herein.

Rights and Permissions

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 IGO License.
All other queries on rights and licenses, including subsidiary rights, should be addressed to
IFC’s Corporate Relations Department, 2121 Pennsylvania Avenue, N.W., Washington, D.C. 20433.

Permissions for Redistributing Content

Translations—If you create a translation of this work, please add the following disclaimer along with the attribution: This
translation was not created by International Finance Corporation and should not be considered an official International Finance
Corporation translation. International Finance Corporation shall not be liable for any content or error in this translation.
Non-Commercial—You may not use the material for commercial purposes.
No Derivatives—If you remix, transform, or build upon the material, you may not distribute the modified material.
No additional restrictions—You may not apply legal terms or technological measures that legally restrict others from doing anything
the license permits.
Third-party content—IFC does not necessarily own each component of the content contained within the work. IFC therefore does
not warrant that the use of any third-party-owned individual component or part contained in the work will not infringe on the rights
of those third parties. The risk of claims resulting from such infringement rests solely with you. If you wish to re-use a component
of the work, it is your responsibility to determine whether permission is needed for that re-use and to obtain permission from the
copyright owner. Examples of components can include, but are not limited to, tables, figures, or images.
All other queries on rights and licenses, including subsidiary rights, should be addressed to IFC’s Corporate Relations Department,
2121 Pennsylvania Avenue, N.W., Washington, D.C. 20433.

Image credit for cover and pages 8, 16, 26, 34, 63, 68, and 72: Shutterstock
AI-ENABLED
SOFTWARE-AS-A-SERVICE
(SaaS)
The Next Frontier for Global SaaS Start-Ups
from India
About the Authors
ALOK GOYAL, Stellaris Venture Partners
RUCHIRA SHUKLA, Regional Lead, South Asia, Disruptive Technologies—Direct Equity and VC Funds, IFC.
ARPIT MAHESHWARI, Stellaris Venture Partners

Content Advisors
Disruptive Technologies and Funds | William Sonneborn, Ruchira Shukla
Office of the Chief Operating Officer | Stephanie von Friedeburg
Partnerships, Communications and Outreach | Nadine Shamounki Ghannam, Kelly Alderson

Acknowledgments
This report was developed under the direction of Alok Goyal, Stellaris Venture Partners, and Ruchira Shukla,
Regional Lead, South Asia, Disruptive Technologies – Direct Equity and VC Funds, International Finance
Corporation (IFC), with support from IFC’s William Sonneborn, Senior Director of Disruptive Technologies
and Funds.
The core working group was led by Arpit Maheshwari, Stellaris Venture Partners. In addition to Arpit, the
team was comprised of Anagh Prasad and Naman Lahoty from Stellaris; and Sayan Ghosh, and Rachit
Kuchhal from IFC. Additionally, Lana Graf, Principal Industry Specialist, Artificial Intelligence/Machine
Learning, IFC, was an invaluable sounding board during the project.
The AI4Biz initiative was run in collaboration with Infosys, Cisco, Freshworks, Amazon, and Github, which
generously supported companies in the AI4Biz cohort with credits, mentoring, and business development.
The working group is grateful to all of the researchers, advisors, and experts (many of whom are quoted in
the report) who assisted in this report’s creation, and we hope that its content will support the emerging AI
software as a service start-ups in India. In particular, we would like to thank C. N. Raghupathi, Senior Vice
President & Head of India Business for Infosys, who connected us to some outstanding people during the
development of this report. Feedback from these experts was invaluable in the creation of this report.

Note to the Reader

This report has been prepared solely for the purpose of information to provide a perspective on the global
opportunity for Indian AI software start-ups. Projected market and financial information, analyses, and
conclusions contained in this report should not be construed as definitive forecasts or guarantees of future
performance or results. Also, IFC’s internal data were not used in preparing this report.

Project Team
Project Manager | Ruchira Shukla
Publishing | Thomas Rehermann
Editors | Ann Bishop, Matt Benjamin
Research Assistants | Alexander Christopher Vanezis, Adyasha Mohanty, Wen Kang Chow
Composition and Design | Rikki Campbell Ogden
CONTENTS
9 | INTRODUCTION

11 | VOICES FROM THE ECOSYSTEM

14 | EXECUTIVE SUMMARY

17 | CHAPTER 1 Enterprise AI—A Massive Disruption

23 | CHAPTER 2 Drivers of AI Adoption in Enterprises

27 | CHAPTER 3 India—Primed to Create Value in an AI-Driven Future

35 | CHAPTER 4 Opportunities for Indian AI SaaS Start-Ups

36 | 4.1 Infrastructure Layer—Limited Opportunity
37 | 4.2 Data Layer—Limited Opportunity
37 | 4.3 AI Platforms Layer—Limited Opportunity
38 | 4.4 MLOps Layer—Potential Opportunity
41 | 4.5 Application Layer—Significant Opportunity
49 | 4.6 Services Layer—Significant Opportunity
53 | 4.7 The Global AI Opportunity
55 | 4.8 The Domestic AI Opportunity

57 | CHAPTER 5 Risks and Considerations

57 | 5.1 Challenges Ahead
59 | 5.2 Policy—Learnings From the World
61 | 5.3 India’s AI Policy Approach—Laying a Robust Foundation

64 | APPENDIX The AI4Biz Challenge

69 | ENDNOTES

73 | FURTHER READING

75 | REFERENCES
Figures

FIGURE 1.1 Simultaneous Confluence of Multiple AI Enablers 18

FIGURE 1.2 Applications of AI Are Limited Only by Our Imagination 19
FIGURE 1.3 Data and Models Are Key to Developing AI Software 19
FIGURE 1.4 The AI Stack is Very Different from the Software Stack 21
FIGURE 1.5 AI is Not a Feature But a New Wave, Leading to Creation of a New Category 22
FIGURE 1.6 AI Will Drive > $800 Billion in Annual Revenue Over the Next Decade 22

FIGURE 2.1 A Wide Range of Problem Types Are Addressed by AI 23

FIGURE 2.2 Readiness and Risk Will Define the Extent of AI Adoption 24
FIGURE 2.3 Readiness and Risk Levels for Healthcare Use Cases 25

FIGURE 3.1 Information Technology BPM Revenue and Employment Growth in India 28
FIGURE 3.2 India is Already a Notable Participant in the Global SaaS Wave 30
FIGURE 3.3 Indian Companies Can Sell Products Across the ACV Spectrum 31
FIGURE 3.4 India-Based Sales Teams Lead to Highly Capital Efficient GTM 31
FIGURE 3.5 Strong Second-Order Effects of the SaaS Wave in India 33

FIGURE 4.1 Within the AI Stack, India Has an Advantage Across Three Layers 35
FIGURE 4.2 Worldwide Cloud Infrastructure Services Spending 36
FIGURE 4.3 DevOps is a Large Software Category on Its Own 38
FIGURE 4.4 Two Break-Out Winners in DevOps: Postman and BrowserStack 39
FIGURE 4.5 MLOps Surpasses the Complexity of DevOps 40
FIGURE 4.6 MLOps is a Fast-Growing Software Category 40
FIGURE 4.7 India Has a Greater Advantage Where Service Intensity is Higher 42
FIGURE 4.8 Micro-Verticalization Will Play to India’s Advantage 44
FIGURE 4.9 Venture Capital Investment in Indian SaaS Companies 46
FIGURE 4.10 AI-SaaS Will Need More Capital Than Traditional SaaS Companies 47
FIGURE 4.11 Methodology: Sizing the AI Applications Opportunity for India 47
FIGURE 4.12 AI Application Companies Can Create ~$400 Billion Market Cap by 2030 48
FIGURE 4.13 India Could be the Natural Destination for AI Services 50
FIGURE 4.14 Methodology: Sizing the AI Services Opportunity for India 53
FIGURE 4.15 AI Services Companies Can Create ~$200 Billion Market Cap by 2030 54
FIGURE 4.16 Impact of India’s Play in the Global AI Market 54
FIGURE 4.17 AI Could Add $500 Billion to India’s GDP by 2025 55
FIGURE 4.18 AI Start-Ups Have Started to Show Impact Across Sectors 56

4
FIGURE 5.1 Top 10 Countries: Number of AI Authors on arXiv 58
FIGURE 5.2 AI Policy Initiatives by Frontrunners 60

FIGURE A.1 108 Diverse, High Quality Applications 64

FIGURE A.2 Generous Support for Cohort Members from Industry Leaders 65
FIGURE A.3 AI4Biz: Top 10 Winners 66
FIGURE A.4 AI4Biz: Additional Finalists 67

Case Studies

CASE STUDY 4.1 Signzy: Examples of AI Applications That Do Not Require Customer-Specific Training 42
CASE STUDY 4.2 Synapsica: Marketplace and AI-Assistant for Efficient Radiology 43
CASE STUDY 4.3 Inspektlabs: Computer Vision for Asset Inspection 45
CASE STUDY 4.4 Taskmonk: Making Data Labeling More Efficient 50
CASE STUDY 4.5 Playment: Data Prep Entails Unique Business Opportunities 51
CASE STUDY 4.6 Scry Analytics: Turnkey AI Solutions Companies—Solution as a Service 52

5
Abbreviations and Acronyms

ACV
Annual Contract Value GDP
Gross Domestic Product
AI
Artificial Intelligence GPT
Generative Pre-Trained Transformer
AI4BIZ
Artificial Intelligence for Business GTM
Go-To-Market
API
Application Programming Interface HITL
Human-In-The-Loop
ARR
Annual Recurring Revenue HPC
High-Performance Computing
AWS
Amazon Web Services ICTAI
International Centre of
Transformational AI
B2B
Business-to-Business
ID
Identification
B2C
Business-to-Consumer
IFC
International Finance Corporation
BD2K
Big Data to Knowledge
INR
Indian Rupee
BPM
Business Process Management
IOT
Internet of Things
BPO
Business Process Outsourcing
IP
Intellectual Property
CAC
Customer Acquisition Cost
IPO
Initial Public Offering
CADM
Custom Application Development
and Maintenance iSPIRT Indian Software Product Industry
RoundTable
CAGR
Compounded Annual Growth Rate
IT
Information Technology
CEO
Chief Executive Officer
ITeS Information Technology Enabled
CMM
Capability Maturity Model
Services
CNN
Convolutional Neural Network
KYC
Know-Your-Customer
COE-DSAI Centre of Excellence for Data
LiDAR Light Detection and Ranging
Science and Artificial Intelligence
ML
Machine Learning
CORE
Centre of Research Excellence
MLOPs Machine Learning Operations
CRM
Customer Relationship Management
MRI
Magnetic Resonance Imaging
CT
Computerized Tomography
NASA
National Aeronautics and Space
DARPA
The Defense Advanced Research
Agency
Projects Agency
NASSCOM National Association of Software
DEVOPs Development Operations
and Service Companies
DOE
Department of Energy
NEP
National Education Policy
EHR
Electronic Health Record
NIH
National Institutes of Health
ERP
Enterprise Resource Planning
NITI
National Institution for Transforming
EV
Enterprise Value India
FAQs Frequently Asked Questions NLP
Natural Language Processing
FMCG
Fast-Moving Consumer Goods NSF
National Science Foundation
FOS
Feet-On-Street OCR
Optical Character Recognition
GCP
Google Cloud Platform OTT
Over-The-Top

6
 POC
Proof of Concept
PPP
Public-Private Partnership
R&D
Research and Development
RNN
Recurrent Neural Network
RPA
Robotic Process Automation
SAAS Software as a Service
SDR
Sales Development Representative
SKU
Stock-Keeping Unit
SMB
Small and Medium Business
SME
Small and Medium Enterprise
SQL
Structured Query Language
STEM
Science, Technology, Engineering,
and Mathematics
SVP
Senior Vice President
TCS
Tata Consultancy Services
TFLOPs Teraflops
TIFR
Tata Institute of Fundamental
Research
UI
User Interface
UK
United Kingdom
US
United States
US$
United States Dollar
UX
User Experience
VC
Venture Capital
WB
World Bank
WBG World Bank Group
YOY
Year-on-Year
Note: All dollar amounts are U.S. dollars unless otherwise indicated

7
INTRODUCTION

Historically, technological megatrends have had a transformational impact on businesses,

industries, and society at large. From electricity and steam power to personal computing and
the Internet, these innovations have completely changed how we work, communicate, and live.
While it is hard to predict the timing and full extent of megatrends accurately, their early signals
provide opportunities for businesses and countries alike to leverage technological innovation to
transform entire sectors, while also driving economic growth and improving human welfare.

We at IFC have over six decades of experience in In early 2020, IFC and Stellaris started working
investing in innovative companies in emerging markets. together to study the role that AI-led software as a
We work with a wide range of industry stakeholders service (SaaS) start-ups from India could play in the
across the entire technology ecosystem, including global market. Looking at the increasing adoption
early-stage start-ups,1 and seed, venture capital, and of AI across a wide range of enterprises, we posed
growth equity funds. In doing so, we focus on spotting two questions:
disruptive technology trends at an early stage, and we
a. What opportunities does this megatrend present,
support such innovation through our investment and
and what challenges and bottlenecks must be
advisory work.
overcome to take advantage of these opportunities?
Stellaris Venture Partners, an IFC investee fund in the
b. What role can India, which is already a leader
start-ups ecosystem in India, collaborated with us in
in the global software ecosystem, play in an
preparing this report. Stellaris, a leading tech-focused
AI-led future?
VC fund, invests in early-stage start-ups that leverage
technology to provide innovative solutions across three We began formulating our views by making the
areas—the emerging 500 million transacting Internet following hypotheses:
consumers in India, the digitization of more than 50 1. AI has arrived and it is ready for prime time:
million Indian small and medium-sized businesses, and
A multitude of AI-enabling technologies are reaching
global enterprise software companies.
inflection points simultaneously and are setting the
Like the many innovations IFC has identified and stage for AI to start delivering on its many promises.
supported over the years, the increasing adoption These include the availability of enormous amounts
of artificial intelligence (AI) in enterprise processes of data, storage capacity, and computing power, as
could not only promote business growth, but also well as highly sophisticated algorithms.
offer innovative solutions to complex challenges in
2. The disruption of enterprise processes
emerging markets and beyond. AI shows early promise
is inevitable:
in addressing a breadth of challenges, ranging from
improving agricultural yields and providing credit AI can improve judgment, which leads to smarter,
to people previously excluded from formal financial faster, and more consistent decision making,
services, to increasing doctors’ capacity to make and the resultant gains in both efficiency and
rapid and accurate diagnoses, and contributing to productivity are likely to be much greater than the
some of humanity’s most ambitious projects such gains previously achieved with computer software.
as autonomous transport, space travel, and ground- Consequently, most if not all business processes will
breaking medical treatments. be rethought and not just automated.

9
3. India can be a significant contributor to described in the annex, allowed us to interact with over
AI disruption: 100 start-ups and have in-depth conversations with the
India is already recognized as a global leader in 20 finalists. These interactions gave us valuable insights
software development. The country’s software into the opportunities and challenges that AI SaaS
services industry has driven massive value creation start-ups face in India.
through companies such as Tata Consultancy This report combines IFC’s and Stellaris’s extensive
Services (TCS), Infosys, Wipro, and HCL, which investment experience with the data and lessons
have a combined market cap of almost $400 learned from our market assessment and the AI4Biz
billion.3 India is now becoming the third SaaS global Challenge. Before proceeding further, we would like to
powerhouse, after the United States and Israel. point out some methodological limitations, as well as
Following the success of Freshworks (and previously clarify the scope of this report.
Zoho), within a short timeframe India has created
We used primary and secondary research, along with
thirteen unicorns4 and at least 50 SaaS companies
our judgment to forecast up to 2030 the evolution of
that are now worth more than $l00 million. India
enterprise AI adoption and the associated software
also has an enormous trained workforce that can
market. In producing this report, we focused on
enable the provisioning of human and data support
identifying trends in AI applications, as well as the
in developing algorithms, and this gives India an
challenges and opportunities that Indian start-ups face
advantage in the creation of AI-led applications5 and
in targeting global markets, and what they need in
the tailoring of these for different markets.
order to succeed.
4. In addition to economic impact, AI presents
There are several important aspects related to enterprise
an unprecedented opportunity to achieve
AI that we did not include in this document.
social impact:
AI is likely to cause significant social change—for
Emerging economies often lack the resources
example, AI could lead to job losses due to automation,
required to scale critical services such as education,
but it could also generate substantial employment
health, finance, logistics, agriculture, and
in both the private and public sectors. Furthermore,
transportation. AI-led software has the potential
although AI can be used to invade people’s privacy, it
to augment both physical and human resources
can also prevent identity theft and subsequent financial
to expand reach, affordability, and quality that
losses. While concerns about both job losses and
otherwise would be difficult to achieve.
privacy are extremely important and require extensive
Much of the work behind this report was completed analysis and attention, they were not within the scope
in 2020 and in the first half of 2021, when IFC and of this project.
Stellaris conducted a comprehensive market assessment
Finally, in this report we presented current policies
to validate these hypotheses. As we spoke with
related to AI from the vantage point of private
participants across the software ecosystem—company
enterprise. As we recognize that policymakers operate
founders, investors, academics, and government and
in a complex environment, we did not include any
industry leaders—we grew even more bullish about
specific policy recommendations. 2
the AI-led software opportunity, both globally and in
India. In 2020, we also organized the AI4Biz Challenge It is our hope that this document will prove useful
to identify and recognize groundbreaking, early-stage for the start-up ecosystem as we approach what is a
Indian AI SaaS companies. This contest, which is seminal era for businesses.

10
VOICES FROM THE ECOSYSTEM
The future of software is not SaaS (software as a service), but SaaS (solution
as a service). Products are context-agnostic, solutions have context; India can
build context-aware solutions by combining machines and people. The Indian
information technology (IT) industry today faces the innovator’s dilemma in
the wake of the AI adoption wave.
—DR. HARRICK VIN, TCS Fellow and Chief Services Innovation Officer,
Tata Consultancy Services (TCS)

The hype about creating a General Purpose AI system started in the 195Os and
died by 1975. By 2010, following Moore’s law, hardware had become significantly
cheaper and companies started focusing on Specific Purpose AI systems. By 2035,
AI will be as embedded in our digital world as motors and engines are in our
physical world today. In doing so, these breakthrough technologies of the past
will become commonplace and standardized.
—ALOK AGGARWAL, Chairman, Chief Executive Officer (CEO), and Chief Data Scientist, Scry Analytics

If I could change one thing about all the AI projects I have done in the past, I
would choose to be more realistic in my goal setting. AI start-ups should begin by
targeting simple problems rather than solving the problems that “sound interesting.”
—KAPIL TANDON, VP Product Growth, Tricentis; ex-Global Product Lead for AI-powered Fraud
Protection solutions, Microsoft

There are AI-first and AI-second companies; the latter mostly have machine
learning (ML)/AI added to the existing functionality, whereas the former will
not exist without ML/AI. AI-first software companies may take longer to reach
the first million dollars when compared to regular software companies, but the
scale-up to the next $99 million may be significantly faster as they tend to disrupt
the space they are in. In such companies, the importance of models is often
underrated, and that of data is overrated.
—MANISH SINGHAL, Founding Partner, pi Ventures

Research results in AI are focused on controlled conditions and often betray the
true complexity of launching the model in a real business environment.
—MANISH GUPTA, Director, Google Research India

11
Democratizing AI is a massive opportunity, and the market is up for grabs.
The India advantage in AI lies in the vast amount of consumer data available
for training. Progress in AI will happen once companies start seeing research
and development (R&D) through an investment lens, and not a “tax-break”
lens. Research funding has to wean itself off the tax incentives and stand on its
own merits.
—PRASAD JOSHI, Senior Vice President, Emerging Technology Solutions, Infosys

You are not building an AI company, just as you are not building a Cloud
company. These are technologies that help you to solve a business problem
and should not be confused with the business itself; AI is a means to an end.
Instead, make yourself known as a product/solution for “X” problem.
—DILIP KHANDELWAL, Global CIO and India MD, Deutsche Bank

As an Indian company, you are always tempted to use people rather than
automation to solve a repetitive problem, but the world does not want to interact
with humans when using their software. There is no faster way to kill your
product than bandaging the rough edges with people. If humans already have
a 95 percent accuracy in achieving a task, AI systems have to shoot for near-
perfect 99 percent to have a reasonable chance of replacing humans.
—NISHITH RASTOGI, Founder and CEO, Locus.sh

Contrary to popular belief, AI innovation in call centers, chat-bots, next-best

action, and so on, are far from done and dusted. There is a lot of value yet to
be claimed.
—RAJESH RADHAKRISHNAN, Executive Vice President and General Manager, Front Office
Automation, Automation Anywhere

AI will become a commodity. In the trade-off between working on building the

best AI model versus solving data collection and workflow problems, an AI
applications start-up should always optimize for the latter.
—MAYANK KUMAR, CEO, Tangent.ai

In AI, data ownership leads to more adoption, thus leading to even more data
accumulation that results in ever larger companies that already own user data;
India should architect public data to achieve public good.
—SUCCESSFUL SILICON VALLEY AI ENTREPRENEUR (ANONYMOUS BY REQUEST)

12
Factors of production have evolved, e.g., deep leverage data is the new
opportunity frontier. That implies there is a disproportionate advantage
accruing to companies that are already bang in the middle of data pipes and
flows as they are sitting on a potential AI goldmine. In certain domains, it is
the incumbents who have the ability to deliver the full potential of AI with their
ready access to the key pools of data—a huge advantage for an already good
customer relationship management (CRM) solution to transform itself into an
amazing offering with new AI layered on.
—RAVI GURURAJ, Founder and CEO, QikPod

Increasingly, companies are realizing that AI is not a stand-alone piece of

software; it is something that has to work seamlessly with existing systems and
their complexities. The risk of the unknown is another significant roadblock
in AI buyers’ minds—it is not about what AI can handle; it is about what has
never been seen before.
Data is an overrated factor when it comes to winning in AI software, whereas
in an industry where AI is trying to solve problems, domain expertise is
immensely underrated.
—RAGHUNATH R, Chief Revenue Officer, Platform & Technology Group at SAP

There are some things that AI can do better than humans, and there are some
things that humans can do better than AI. A lot of my time goes into figuring
out where to put machines, and where to have humans in the loop, and advising
customers on where to use AI, and where not to.
—ASHWINI ASOKAN, Founder and CEO, Mad Street Den

AI, as machine intelligence, will elevate humans to focus on what really matters.
And yet, paradoxically, the idea of intelligent machines will be a pipe dream
without humans helping train and retrain the models that will be the underbelly
of these machines. This is true symbiosis.
—DHEERAJ PANDEY, Founder, Investor, and then CEO, Nutanix; Board Member, Adobe

13
EXECUTIVE SUMMARY

AI will drive the next wave of business for AI, whereas we are not ready yet to engage in
transformation unconstrained conversations with machines. Risk is
a measure of what can go wrong if a machine makes
Artificial intelligence (AI)—the use of technology to
decisions that have always been made by humans,
aid human judgment, as opposed to simply automating
repetitive tasks—is ready for prime time. Several or through a rule-based system. For example, the
simultaneous inflection points have created the recommendation of a search engine is a low-risk
perfect opportunity for AI adoption: the availability use case, whereas a machine serving as a judge in a
of enormous amounts of data, storage capacity, and criminal court would be a high-risk use case. As one
computing power, as well as highly sophisticated would expect, AI use cases with high readiness and
algorithms. The efficiency and effectiveness gains from low underlying risk will be the first AI use cases we
the proper use of AI can create value that far exceeds adopt. Conversely, some use cases with low readiness
traditional software as a service (SaaS) solutions. and high risk such as autonomous surgery may never
Globally, AI applications are mushrooming at a rapid become mainstream, or they will only be used in limited,
pace, and their use is limited only by our imagination. controlled circumstances.

AI solutions are fundamentally different from SaaS in Across industries and functions, there are already a
their data intensity. What code is to SaaS, data is to AI. large number of use cases that have been disrupted
In addition to coding skills, which are the primary skills through the use of AI, or that are ripe for disruption.
needed for traditional software development, building AI As both computing and storage costs continue to
solutions requires deep data management, data science, decrease exponentially, these improvements, coupled
and modeling skills. There is an inherent separation of with advances in AI algorithms, mean that we will likely
data and software in traditional software, whereas with be pleasantly surprised by how quickly we are ready to
AI that distinction disappears because it is the data that accept new applications of AI.
“write” the software. The AI stack and the skill sets
India is well positioned to be at the
required to build AI solutions are very different from
forefront of the AI revolution
those required for traditional SaaS applications. Thus,
AI will not just be a “feature” in existing applications, it India has come a long way in the last decade with
will help produce a new class of applications and create regard to building globally successful SaaS companies.
the next enterprise disruption, which will be akin to Consequently, Indian companies have gained experience
the impact of the Cloud at the turn of the 21st Century. in solving complex problems through SaaS, and in
Estimates suggest that by 2030, AI infrastructure, executing a variety of go-to-market strategies, from
applications, and services will become a market worth inside sales6 to feet-on-the-street, and for annual
more than $800 billion. contract values (ACVs) that range from a few thousand
dollars to more than a million. India has a large talent
Adoption of AI will vary across use cases, pool of developers and strong process expertise,
depending on readiness and risk along with a fast-growing pool of niche talent such as
For a particular use case, readiness refers to the level of designers and data scientists.
impact that AI can have, the availability of data, and The overall ecosystem for early-stage companies is quite
the sophistication of algorithms that justify the case mature in India. Also, the country has no shortage of
for change. For example, a chatbot used for customer risk capital, with investors increasingly focusing on
support in banking is an illustration of high readiness software investments, and several industry-specific

14
associations and accelerators to support companies in the same time, AI solutions are likely to be “stickier”
the early stages of their journey. In addition, a growing and they can command better prices given the higher-
number of successful business founders are looking to value unlock potential. Thus, we expect AI SaaS
“give back” by helping younger entrepreneurs. companies to command attractive valuations at scale.
These, in combination with the increased service By 2030, AI applications and services together have
intensity inherent to AI, plus the democratization of the potential to create additional market capitalization
AI infrastructure through big Cloud companies and of more than $500 billion for Indian companies (see
open-source software, mean that in comparison to the Chapter 4, sections 4.7 and 4.8 for more detail).
SaaS wave, India is in a much better position to drive
the AI wave. Unleashing the opportunity will require
concerted industry and policy interventions
Through AI disruption, India could create
While the opportunity ahead of us is massive, several
more than $500 billion in value by 2030
challenges must be overcome. Currently, the biggest
With regard to AI, the immediate opportunity for India roadblocks facing India are insufficient senior talent,
lies in leveraging the capabilities developed during the knowledgeable capital, and data and research capabilities.
information technology (IT) and SaaS waves in order to
build global AI applications and services businesses. The size of the AI prize is too large to ignore, and
both the government and industry need to act now to
AI applications will come in many variations; however,
mitigate the challenges to AI adoption. Global examples
India’s advantage will lie in the ones that have a higher
suggest that well-designed policy measures can have a
intensity of service, those that are more vertical specific,
positive impact on a country’s competitiveness in AI,
or those that are built on India’s unique data sets. India
which will be measured by metrics such as the number
could also be the natural choice as a global destination
of research papers published, and the start-ups that are
for a set of services related to data collection, cleaning,
and labeling, as well as for implementing services and created and funded. In this context, national bodies
monitoring results on an ongoing basis. However, we such as the Indian government think tank, NITI Aayog,
expect AI SaaS companies to be somewhat less capital- the National Association of Software and Services
efficient than has been the case with the traditional SaaS Companies (NASSCOM), and the Indian Software
companies. This is due to the additional costs involved Products Industry Round Table (iSPIRT) have proposed
when building the underlying AI models. These costs sound measures for India, and the government has taken
include data acquisition, processing, and the team of some significant steps. However, speedy execution is
data scientists and machine learning (ML) engineers. At crucial for building on this momentum.

15
CHAPTER 1

Enterprise AI—A Massive Disruption

AI: A Brief History7 • AI had its first era of excitement in the 1960s
and 1970s, when researchers demonstrated the
AI is a misused and often misunderstood term.
One definition of AI is “the ability of a computer ability of machines to perform specific tasks,
or computer-controlled robot to perform tasks albeit in controlled environments—for example,
commonly associated with intelligent beings.”8 Ever proving mathematical theorems and solving
since computers were invented, machines with general specific types of algebra and calculus problems.
intelligence—common sense, and the ability to perform However, performing these tasks in an uncontrolled
tasks that involve learning, reasoning, and processing environment led to a combinatorial explosion of
complex information—have been enthusiastically possibilities, which the computers of that time could
anticipated. In 1965, I.J. Good,9 a mathematician, not handle, and this led to the first “AI winter.”
when defining the potential for an “intelligence • A second era of excitement about AI occurred
explosion” argued as follows: in the 1980s with the Japanese Fifth Generation
Let an ultraintelligent machine be defined as a Computer Systems Project—a public/private
machine that can far surpass all the intellectual partnership set up to create massive parallel
activities of any man, however clever. Since the computing infrastructure to serve as a platform
design of machines is one of these intellectual for AI projects. This led to a large number of
activities, an ultraintelligent machine could expert systems and computers that could perform
design even better machines; there would then only one task. Although the systems worked, the
unquestionably be an ‘intelligence explosion’, utility of acquiring a computer for solving just
and the intelligence of man would be left far one problem was suspect, at best, and led to the
behind. Thus, the first ultraintelligent machine next “AI winter.”
is the last invention that man need ever make, • A new wave of optimism began in the early
provided that the machine is docile enough to 1990s with the introduction of newer types of
tell us how to keep it under control. algorithms that mimic natural systems. One
Since the mid-1950s, AI has gone through several popular type of algorithm here is a neural
boom and bust cycles, with periods of intense hope and network, which is inspired by how neurons
excitement followed by “AI winters.” function in the human brain.

• In 1956, 10 scientists gathered at Dartmouth College AI: The Status Quo

for a two-month research project funded by the
Rockefeller Foundation. The project proposal stated, AI is consistently achieving human-like or superhuman
“An attempt will be made to find [out] how to make performance levels across a range of tasks.
machines that use language, form abstractions and • Convolutional neural networks (CNNs) have led
concepts, solve the kinds of problems now reserved to a dramatic increase in the accuracy of object
for humans, and improve themselves. We think that recognition within images. Error rates for computers
a significant advance can be made in one or more have been below human levels since 2016,10 and
of these problems if a carefully chosen group of computers’ image recognition performance is
scientists work on it together for a summer.” consistently improving.

17
• Recurrent neural networks (RNNs) have led context-related data points such as physical location
to a similar improvement in computer speech and type of location (home versus shopping mall versus
recognition, with error rates for computers school) in addition to time and type of action taken.
falling below human levels since 201711 and the Additionally, there has been a veritable explosion in
performance of computers consistently improving. the number of interconnected devices (popularly called
• Over time, computers have surpassed humans in the Internet of Things—IOT), which are constantly
playing a number of games12—for example, checkers sending data points to a central repository—for
(1994), Othello (1992), chess (1997), Scrabble example, temperature and pressure sensors in a power
(2002), Jeopardy (2010), and Go (2016). plant. As a result, there is no lack of data available to
train AI’s increasingly more sophisticated and data-
Simultaneous inflection points in several underlying
efficient algorithms. Add to these the continuing effect
technologies have set the stage for the rapid
of Moore’s law on storage and computing power and
proliferation of AI. These inflection points are
ever-increasing Internet bandwidth, and the problem
exponential growth in the availability of data, data
of “combinatorial explosion” seems a lot less daunting
storage, computing power, and AI algorithmic
than it did a few decades ago.
efficiency. Based purely on its computing power and
storage capacity, a typical 2020 smartphone would The coming of age of AI has led to a combinatorial
have been considered a supercomputer in 1990. explosion of a different kind—AI use cases themselves.
Content consumption has shifted dramatically to AI is being applied to a large number of use cases
mobile devices, which generate significantly more across horizontals (functions) and verticals (industries).

FIGURE 1.1 Simultaneous Confluence of Multiple AI Enablers

Data Availability Data Storage Cost

Data: Over 70 billion IoT devices
worldwide in 2020; to triple in 5
2020: 44 ZB 1960: $3.5 B
years. Supervised data is the next
challenge, which is on track to be
ZetaBytes (ZB)

solved 50% through automation.

$ / TB

Computing: Next-gen GPUs, TPUs

and AI specific chips which can have
1,000s of cores.
2013: 4.4 ZB 2015: $27

2000 2010 2015 2020 1960 1980 2000 2020 Storage: The least limiting factor for
AI development; storage costs are
manageable because the training
Compute Power Compute Cost
data need not be stored indefinitely.
2017: 93,000 TFLOPS 1961: $1.1 T
Algorithm: Improvement in
TFLOPS / Supercomputer

algorithmic ability to parallelize

$ / GFLOPS

training, and allow for training using

exponentially larger data sets.

1964: 0.0000005 TFLOPS 2015: $0.08

1960 1980 2000 2020 1960 1980 2000 2020

Sources: World Economic Forum, OpenAI, Economist, Seagate

18
FIGURE 1.2 Applications of AI Are Limited Only by Our Imagination

Industries Functions

Insurance Human Resources

Assessment of size, location, Automated chat-based interviews of
constructed area, proximity to blue collar workers
fire station, roof condition, etc., of
individual homes, using satellite images

Manufacturing Sales
Identification of potential breakdown Prediction of best prospects for a
on a shop floor through analysis of the B2B company based on hundreds of
sound signatures of equipment structured/unstructured parameters
and past successes

Education Supply Chain

Use of chatbot as a teaching assistant Stock capacity planning for new SKUs
in school/college education based on proximity to historical SKUs
and past sales data

Financial Services Quality

Fraud detection through keystroke Video-based continuous monitoring
fingerprinting of every item produced, instead of
random sampling

A Georgia Tech professor uses AI as a teaching Hang on, isn’t AI the same as traditional
assistant; healthcare start-ups are using AI for online software?
consultations with patients; manufacturing companies
While there are several similarities between AI and
are using AI to predict future failures and prevent
traditional software, developing AI software is very
downtimes; and financial institutions and retailers are
different from developing traditional software—this
using AI to prevent fraud. The breadth of AI use cases
difference is primarily because AI software relies on
is enormous and growing.
data and models, in addition to code.

FIGURE 1.3 Data and Models Are Key to Developing AI Software

AI Specific Processes

Data Model Traditional Software Processes

Collection Development
• Architecture design
• Input collection
• Input processing (ML models
are used here)
Data Model
Processing Deployment • Output

19
A typical AI software development cycle involves the
deployment of an enterprise-grade, machine learning
model before “software” kicks in, which broadly SaaS and AI-SaaS are poles apart;
involves the following four steps. the services-to-software ratio for
• Data Collection: This involves putting in place AI companies is significantly higher
a process for collecting relevant raw data to train than regular SaaS. One of the
the underlying AI models—for example, a credit-
strongest competitive moats a SaaS-
scoring algorithm may reasonably require historical
data on borrowers’ profiles, loan parameters, and oriented AI company can build is
repayment performance. Underlying data sets [to] own a workflow that enables a
continuously change as additional data become single model across customers, with
available. These data sets can come from internal
the ability to regenerate training
sources such as other enterprise software, or from
external sources such as Dun & Bradstreet. data, continuously.
• Data Processing: This involves a range of steps —KRISH MANTRIPRAGADA,
Chief Product Officer, Seismic
to make the data usable for machine learning.
Data sets need to be cleaned (removing outliers
and extraneous variables) and data need to be
made richer and easier for machine consumption
through annotation. Often, two or more additional set of specialized skills—chief among which
underlying data sets are combined, which results are data engineering and machine learning—the AI
in a larger and more contextual data set. stack also looks notably different from the traditional
• Model Development: This involves both model software stack (see Figure 1.4). In particular, AI
selection, where different types of ML models development is significantly more service intensive—
can lead to varying levels of accuracy for the up to three times more intensive—when compared to
same problem statement and the same data set,13 traditional software, with service requirements across
as well as training the actual model with the data preparation, model development, deployment, and
processed data set. ongoing maintenance and improvement of the model.
• Model Deployment: The trained model needs to What this means is that AI will be a different category
be deployed in a production environment, which of software altogether, as opposed to a feature within
requires integration with an enterprise’s existing the existing software workflow. AI is expected to
software workflows. This also requires building trigger massive disruption because of its ability to
feedback loops wherein both correct and incorrect fundamentally change the structure, functionality, and
outputs of the model can be used to reinforce or efficacy of technology solutions by orders of magnitude
modify the model’s parameters. that are far beyond what traditional software can do.
Owing to the presence of both data and models, AI In our opinion, the last such disruption was the Cloud.
software development is much more dynamic than While many other trends such as analytics and mobility
traditional software development. The same version have occurred since then, they did not necessarily
of an application can lead to differing results due to create a massive disruption.
changes in the underlying training data, even though The implications of a disruptive AI wave are
no change has occurred to the underlying source significant for the start-up ecosystem. During such
code. Not only does AI software development need an waves, the incumbents are incremental in their

20
FIGURE 1.4 The AI Stack is Very Different from the Software Stack

Description Example

Data cleaning/labeling during

Services: Data Plumbing,
dev; Ongoing maintenance;
Data Prep, Ongoing Tuning
Ongoing model updates

Applications or packaged APIs

Applications: End User
that rely on underlying AI for a
Applications, AI APIs
business use case

Tools for deployment of AI/ML

Dev: DevTools, MLOps code, versioning, collaboration,
CI/CD, testing

AI Platform: AI/ML Algorithms that describe how a

Algorithms, Model Training, machine should process & learn
Feature Extraction from data

Data: Warehouse, Software layer that collects &

Ingestion, Cleaning, stores data from different
Visualization sources in a scalable format

Infra: Data Centers, Computing and storage

Distributed Computing hardware

approach to adopting new technology, whereas of the current incumbents—a phenomenon we are all
insurgents take higher risks, as they start without closely watching for.
baggage from the past. Also, they bring different
Regardless of who wins this race, we believe that
skills and approaches and, in the process, are able to
over the next decade significant opportunities
innovate and scale much faster than the incumbents.
will be created for all stakeholders—businesses,
While SAP had every opportunity to build the first
Cloud enterprise resource planning (ERP) software, investors, academics, and governments. By 2030, AI
and Siebel had the same opportunity for Cloud is expected to create an $800 billion market across
customer relationship management (CRM) software, applications, infrastructure, and services. To put
the winners in these categories were Netsuite and this into perspective, the global cloud infrastructure
Salesforce. Similarly, the disruption caused by AI services market was estimated to be at a revenue run
could bring forth new winners that could leap ahead rate of $200 billion as of Q3 2021.14

21
FIGURE 1.5 AI is Not a Feature But a New Wave, Leading to Creation of a New Category

Major Disruptions of Enterprise Software AI

Cloud

Web

Client/
Server

Mini

Mainframes

1940’s 50’s 60’s 70’s 80’s 90’s 00’s 10’s 2020’s

IBM DEC SAP i2 Salesforce

Unisys HP Oracle Siebel SuccessFactors
Burroughs Honeywell Sybase ePiphany Workday
DB2 IBM SAP/Oracle ServiceNow

FIGURE 1.6 AI Will Drive > $800 Billion in Annual Revenue Over the Next Decade

875 • Overall market to clock ~40% annual

AI Services growth till 2025 and continue to grow
at ~30% for the next 5 years.
AI Applications

AI Hardware • Implementing AI tools will be service-

intensive, and this segment will capture
more than a third of overall value.
Global AI Market Size
(US $ Billions) • Software to be dominated by enterprise
applications, with less than 20% of the
software spend on underlying
developer platforms.
256
• Growth in the AI hardware market will
be dominated by large incumbents;
51 competition for the next generation of
AI processors and memory is on.

2020 2025 2030

Sources: Gartner, IDC, Grandview, SEC filings

22
CHAPTER 2

Drivers of AI Adoption in Enterprises

As was the case with previous megatrends, the types—big or small, manufacturing or services, new
adoption of a new technology, once it begins, is or old—will lead to the unprecedented availability of
followed by a series of catalytic factors that accelerate the underlying data needed to make the adoption of AI
the trend through the interplay of several dependent a megatrend. In such circumstances, a different lens is
variables. For example, the increasing adoption of AI necessary to understand AI adoption by businesses—
creates virtuous cycles as the greater availability of one that focuses on the underlying use case as opposed
data leads to models that are more accurate, which in to the industry or function.
turn encourage more enterprises to adopt AI. It is also Use cases within an enterprise can be thought of as
likely that when the benefits of adopting AI are seen, a combination of problem type and data type. AI is
other enterprises will follow. However, it is difficult to employed to solve a variety of underlying problems—for
predict when specific AI use cases will be adopted. example, classification, clustering, and recommendations
When forecasting the adoption of AI, conventional that utilize different types of data as input such as text,
wisdom dictates considering AI’s impact across audio, images, video, and time-series data that are in
different industries (or verticals) and functions (or either a structured or unstructured form.
horizontals). For example, how would AI be adopted For example, tumor detection in healthcare is a
in chemical manufacturing versus education, or across classification problem that uses images as inputs. Use
a finance function versus a recruitment function? In cases can also involve multiple problem types. This is
our view, the inevitable digitization of businesses of all the situation with voice assistants that take in audio

FIGURE 2.1 A Wide Range of Problem Types Are Addressed by AI

Classification Clustering Continuous Estimation Ranking

Categorize new input as Create a set of categories Estimate next numeric Order the results of a
belonging to an existing set in which items have value in a sequence; query/request using
of categories similar features prediction pre-set criteria

Recommendation Anomaly Detection Data Generation All Other Optimizations

Provide recommendations Given a set of data, Generate appropriately Create a set of outputs
based on a specific set determine if specific novel set of data based on that optimize for an
of data; a special type of inputs are abnormal; training data objective function
optimization subset of Classification

23
as input and apply techniques related to classification, legal action arising from an AI malfunction. An
clustering, and data generation in order to convert example of legal risk is a class-action lawsuit that
speech to text, make sense of the statements made results when personally identifiable information is
through voice, and then deliver the system’s response leaked through a security breach. A malfunctioning
back in a spoken fashion. AI system can lead to loss of life as well—for
example, if a nuclear power plant goes out of
Readiness and Risk are the primary factors governing
control. Thus, as is the case with Readiness, there
whether an enterprise will adopt AI for a particular
are Risks when using AI, and these need to be
use case.
considered for each use case.
• Readiness is essentially driven by performance. To
Based on the interplay of readiness and risk, use cases
judge readiness, we need to check whether there
fall into four categories when it comes to the extent of
is tangible evidence that AI will improve factors
AI adoption (see Figure 2.2).
such as accuracy or costs. These indicate whether
the deployment of AI will have a reliably beneficial • Established—High Readiness, Low Risk
impact on a business. For sufficient readiness, there These are use cases where the suitability of AI is
must be a large enough pool of underlying data obvious. Examples include document processing
(available continuously and in a manageable format), through robotic process automation (RPA), and
algorithms that are advanced enough to take care of recommendation engines such as those that suggest
use-case specific edge cases15 and other challenges, the next-best purchase option on an e-commerce
and a manageable level of process complexity. website or the next movie to watch on a video
In general, the higher the number of processes streaming service. Given the early stage of AI
impacted, the lower the level of readiness will be. adoption, globally, it is worth pointing out that even
Each use case has a different readiness level, which with established use cases, penetration levels within
depends both on the type of problem involved and an industry are likely to be low, which indicates
the available data. For example, image recognition, significant room for growth.
which requires the classification of visual inputs
(still images or video), has reached accuracy levels
that surpass human performance, which means FIGURE 2.2 Readiness and Risk Will Define the
Extent of AI Adoption
that image recognition has reached a high level
of readiness. Conversely, using AI for long-term HIGH

equity investing is much more challenging, and

consequently has a low level of readiness.
• Risk refers to the consequences if the AI model fails Emerging Extreme
to do its job properly, and there are primarily three
types of risk: financial, reputational, and legal.
Financial risk is the amount of money at stake if
AI malfunctions. For example, a hedge fund that Risk
relies on algorithmic trading could lose hundreds
of millions of dollars if a misfiring algorithm
goes unchecked. Reputational risk is often the
intangible cost of damage to a brand. For example, Established Early
if a chatbot’s perceived rudeness or insensitivity
when dealing with an irate customer is then widely
criticized on social media, this will have negative LOW
consequences for the brand of the business using the HIGH LOW
chatbot. Legal risk is the potential for an adverse Readiness

24
• Emerging—High Readiness, High Risk
FIGURE 2.3 Readiness and Risk Levels for Healthcare
These use cases rank second with regard to their Use Cases
degree of AI adoption. Although readiness is high,
HIGH
risk is significant too. For example, an automated
know-your-customer (KYC) process used for Drug Development

financial transactions cuts staff costs, but at the Radiology Diagnosis

potential risk of approving customers without
Gene Analytics & Editing
the necessary documentation, or with fraudulent
documentation. In these use cases, despite the risk, Personalized Medication

the benefits of AI are large enough for businesses

to adopt AI solutions. Also, the risk can often Risk
be mitigated by a human-in-the-loop (HITL)
Blood Sample Analysis
mechanism that augments the AI system with
Psychology Consultation
constant human monitoring to deal with a high-risk
Patient Data Analytics
edge case if it occurs.
• Early—Low Readiness, Low Risk
Pricing & Risk
In low-risk circumstances, unproven technologies
LOW
are often adopted by businesses that face workforce
HIGH LOW
problems due to cost or lack of personnel. Examples Readiness
of such AI applications include advanced chatbots
that help shoppers when they are considering online AI must meet, which in turn leads to challenges
purchases. Most of the chatbots used by enterprises with regard to adoption. An example of this is fully
today provide responses to simple questions, much automated robotic surgery. It is worth pointing out,
like the kind of assistance provided through a however, that failing to meet goals should not deter
list of frequently asked questions (FAQs). These researchers from undertaking moonshot projects, as
chatbots offer very limited value if a customer is these often develop extremely beneficial technology.
making an expensive and complex purchase such For example, light detection and ranging (LiDAR)
as a home theatre system. In these situations, what performance has improved enormously in recent
matters is the contextual awareness of the chatbot, times. While the relationship is hard to quantify,
and its ability to personalize responses based on much of this improvement in LiDAR is a result of
the conversation’s history. A chatbot is likely to existing firms and start-ups conducting research and
be deployed when there are not enough staff to experiments to create autonomous vehicles. While
deal with each customer. Therefore, as long as the widespread use of autonomous vehicles is still
the chatbot can deliver some positive results (the some years away, LiDAR improvements have been
customer makes a purchase) and some intermediate immediately applicable in other use cases such as
results (the chatbot recognizes that failure is better estimation of ore deposits in mining.
imminent, and a human is alerted to take over),
With this framework in mind, adoption of AI will
there is relatively little risk of a bad outcome such as
happen in a non-uniform fashion across use cases
an unsatisfactory conversation.
within industries. For example, while analytics of
• Extreme—Low Readiness, High Risk patient data is a fairly well-established use case
These scenarios, which are best described as within healthcare, using AI to develop drugs has yet
“moonshots,” often imply that there is no clear to become mainstream. With regard to the latter, the
prospect of commercial adoption. Often, the technology is not mature enough yet, and negative
extremely high underlying risk of failure leads to consequences such as unintended side effects could be
exceptionally high-performance benchmarks that extremely severe.

25
CHAPTER 3

India—Primed to Create Value in an

AI-Driven Future
To understand how Indian companies can capture operated primarily as data entry providers for domestic
value in the AI megatrend, it is important to companies that had just started to digitize their books
understand the developments that have shaped the and records. Toward the latter half of the 1970s, Indian
talent pool and business ecosystem in India. Two IT services companies began to participate frequently
major enterprise software trends in India serve as the in international onsite projects, most of which were
precursors to the AI software wave we are seeing today. billed based on time and materials, and which involved
The first is the IT service boom of the late 1960s that low-level code development and the maintenance of
eventually matured in the first five years after 2000, with custom applications.
India’s dominance in business process outsourcing (BPO). A small company, Tata Consultancy Services (TCS),
Much of India’s technical talent ecosystem and software was established in these early days without much
product legacy can be attributed to this multi-decade fanfare, but it had large ambitions. As a part of a much
wave. It is worth noting that significant talent has also larger conglomerate, the Tata Group, TCS stood out
come from the Indian offices of large tech companies among Tata’s largely traditional industries such as steel
such as Microsoft, SAP, Oracle, and Symantec. and motor vehicle production.
The second is the software as a service (SaaS) wave that
started in the early-2000s and has created significant
business value since then. This wave, which is the Many years ago, there was an
more recent phenomenon, has significantly shaped
industrial revolution; we missed
India’s preparedness to participate in the next wave of
it for reasons beyond our control.
enterprise software development—artificial intelligence.
Today, there is a new revolution—a
The Making of an IT Service Juggernaut revolution in information
India’s journey in the world of computing began in technology, which requires neither
1955 with the installation of HEC-2M (a computer mechanical bias nor mechanical
designed in England) at the Indian Statistical Institute temperament. Primarily, it requires
in Calcutta (now Kolkata).16 Also in 1955, a team at the ability to think clearly. This we
the Tata Institute of Fundamental Research (TIFR) have in abundance. We have an
in Bombay (now Mumbai) started designing and opportunity to participate in this
fabricating a computer. From a handful of people in
revolution on an equal basis; we
India who knew about computers in the 1950s, India’s
information technology workforce grew to over 4.7
have an opportunity even to assume
million by 2020.17 As indicated in Figure 3.1, India has leadership in this revolution.
come a long way over the past few decades. —F C KOHLI, TCS Deputy Chairman, 1975
Speech to the Computer Society of India
The first wave of Indian IT services companies emerged
in the late 1960s and early 1970s. These companies

27
FIGURE 3.1 Information Technology BPM Revenue and Employment Growth in India

200 4.5M

180 4M

160
3.5M

140
3M
Revenue (US $ Billions)

120

No. of jobs
2.5M
100
2M
80

1.5M
60

40 1M

20 .5M

0 0
FY 1980 FY 1990 FY 2000 FY 2010 FY 2019

Domestic Export Employment

Sources: Nasscom, McKinsey, Gartner, Fujitsu

In 1980, the Indian IT services industry had only 25 phenomenal growth was a direct result of the shift
firms and employed just 2,000 people. By then, early toward the offshoring model, and India’s preparedness
success in IT services exports had begun and the to capture the lion’s share of this trend. The IT
industry’s combined export value was $30 million.18 industry itself did exceedingly well to master the
Based on high-quality and consistently reliable evolving standards of offshoring, as measured by its
service delivery, toward the end of the 1980s, Indian capability maturity model (CMM)20 assessments.
IT services firms started enjoying higher project During this period, offshoring moved beyond just
ownership and accountability. As a result, a great deal traditional custom application development and
of onshore work started coming back to India. By maintenance (CADM) into higher value addition
1990, the percentage of revenue earned from on-site services such as systems integration and consulting.
coding had dropped from 90 percent to 80 percent, For example, in 1993, Motorola’s software
but it remained high. By this time, software services development center in Bangalore was the first
exports had quadrupled to reach $128 million, but commercial software development center in the world
this still accounted for only 0.05 percent of India’s to achieve CMM level 5. By 2003, 75 percent of the
gross domestic product (GDP). world’s CMM level 5 software centers were in India. 21
The steady growth of India’s IT services in the 1970s This excellence in process standardization and
and 1980s reached hyper speed in the 1990s, and quality was in no small measure attributable to the
this has continued into the 2000s. In 2007, the IT early establishment of captive development centers
services’ and BPO industry’s revenue was $23.5 in India by some of the world’s leading software
billion (including $18 billion from exports).19 This companies. Notable among the other reasons were: a)

28
the significant depreciation of the Indian rupee against well-defined playbooks for product development,
the dollar (depreciating roughly 17 percent in real customer success, or inside sales.6 Also, these early
terms22 [adjusted for inflation from 1991–1999]); b) SaaS companies’ revenues were skewed toward services
relatively liberal foreign investment regulations in the revenue more than recurring product revenue.
software industry; c) strengthened telecommunications As the first batch of SaaS companies grew in size
infrastructure; and d) establishment of software between 2010 and 2015, their capabilities evolved. A
technology parks equipped with satellite better understanding of product user interface (UI)
communication and tax incentives for exports. design, user experience (UX) design, and inside sales
Besides the policy impetus and favorable macro- practices was established. Indian SaaS revenue started
factors, another development around this time that to come increasingly from global customers and
supported the industry’s growth was the establishment revenue became more recurring.
of the National Association of Software and Services In 2020, 75 percent of India’s SaaS revenue came from
Companies (NASSCOM) in 1988. Over the years, the global markets, and we have seen this trend continue
role of NASSCOM in putting the industry’s concerns to accelerate in 2021. 25 This also led to a fresh crop of
in front of policymakers and in pushing for reforms entrepreneurs who had firsthand experience of scaling
cannot be overstated—NASSCOM has been a great product teams, as well as engineering and sales, due
champion of the IT industry. to their work in the previous generation of successful
The Indian IT industry achieved a seminal milestone on SaaS companies. These entrepreneurs achieved faster
October 8, 202023 when TCS became the most valuable product development and go-to-market results, as they
IT services company globally, which made India’s had learned from the mistakes of their predecessors.
dominance in IT services official. The success of India’s In parallel, India’s fast-growing business-to-consumer
IT services sector and its emergence as the preferred (B2C) start-ups provided the impetus for the
software partner across the world has paved the way development of the SaaS ecosystem. By 2015, B2C
for the next wave of SaaS products from India. start-ups such as Flipkart, Paytm, Myntra, Zomato,
and others had grown rapidly; they served as fertile
Dawn of the SaaS Era ground for new product ideas; and they were also early
While the SaaS wave started in the United States in the adopters of the SaaS products that were built to solve
late 1990s, India did not have its first SaaS companies their problems.
until the mid-2000s. Most early SaaS companies were As of 2020, Indian SaaS companies had firmly
in fact new avatars (incarnations) of legacy software established themselves on the global map, and India’s
businesses that were sold as on-premises installations. SaaS revenue was $3.5 billion, 26 which represented 3.4
Zoho, a company that offers online productivity tools percent of global SaaS revenue, up from 1.5 percent in
for businesses, is a classic example of this. Zoho, 2015. At this rate, India’s SaaS revenue is expected to
a current poster child of India’s SaaS success story, reach $13 to $15 billion by 2025, when it will represent
started in 1996 as AdventNet, a network management 6 to 7 percent of global SaaS revenue.27 Other estimates
company, and only pivoted to become a SaaS product suggest Indian SaaS companies are set to generate
suite in 2005. 24 $30 billion revenue by 2025.28 This revenue will place
Between 2005 and 2010, India’s first SaaS companies India firmly in the top five SaaS countries in the world.
were launched and they had several characteristics However, given the massive size of the global SaaS
in common. Many of the companies were started market, India’s share is still small, which presents a
by engineers who had worked in the technology significant opportunity for further growth.
teams of global companies in the United States, or As of January 2022, India has thirteen SaaS unicorns
in the Indian software development centers of these and the list is growing rapidly, with six entrants in
global companies. Most of these companies sold their 2021 alone. In addition, more than 50 companies have
services to the Indian market, and they did not have valuations of over $l00 million.

29
FIGURE 3.2 India is Already a Notable Participant in the Global SaaS Wave

India’s SaaS Leaders

Unicorn club ($1B+ valuation)

$100M+ valuation club*

*representative, non-exhaustive list

Making Way for AI-First SaaS Companies persona, buying behavior, cultural nuances, ability
to customize for buyers’ unique requirements, and
While Indian SaaS companies have grown faster than
effective sales methods that can predictably create
their global counterparts since 2015, it could be argued
that, overall, India has been a late entrant in the SaaS revenue in the target market. Over the decade since
wave. Most new SaaS categories have been created in 2010, Indian SaaS companies have made successful
the United States first, and their Indian SaaS companies forays into both the European and U.S. markets.
have been launched three to five years later. We believe Through several trials and errors, a playbook has
that the time lag has narrowed since 2015, and this emerged for different go-to-market approaches
narrowing is more evident in AI SaaS companies. that include inside sales, feet-on-street enterprise
Several AI-first SaaS companies have launched in India sales, product-led growth, and open-source. There
at the same time as their global counterparts, and they have been encouraging success stories in companies
are competing for the same global business. They are such as Mindtickle, Whatfix, Freshworks,
also significantly more prepared to compete with their Moengage, Vymo, and Chargebee that have created
global counterparts than was the case with the first differentiated positioning and built strong demand
crop of Indian SaaS companies that launched around generation functions.
2005. This change can largely be attributed to the The expertise gained is further distributed across
groundwork laid by India’s IT services companies and, the start-up ecosystem when employees from
more recently, to the SaaS companies, as well as key successful organizations join younger start-ups or
market developments: launch new businesses and deploy tried-and-tested
• Global Go-to-market (GTM) Expertise: Go-to- playbooks that succeed in global markets.
market expertise is a combination of sales, demand • Services Expertise and Advantage: As
generation, and product marketing skills that discussed earlier, firms’ development of AI
enable a business to sell in a particular market. SaaS solutions requires much higher human
This requires a deep understanding of the buyer involvement than was the case with traditional

30
FIGURE 3.3 Indian Companies Can Sell Products Across the ACV Spectrum

Distinct advantage in remote go-to-market

for small and medium businesses

• Proven playbook with demand generation,

lead qualification, and inside sales capabilities

• ACVs up to $50K sold remotely

• Can onboard sub-$5000 ACV accounts with

human support

Significant experience in enterprise sales

• $100K and above contracts sold with

feet-on-street (FoS) sales in USA and Europe

• FoS enterprise sales team supported by lead

geneneration and qualification team in India

• Cost-effective services contract a

differentiator in large SaaS deals

FIGURE 3.4 India-Based Sales Teams Lead to Highly Capital Efficient GTM

Cost of salespeople ($/Year, 2020) US India

58,000
Account
Executive
(Inside Sales)
11,300

Sales 48,000
Development
Representative
(SDR) 6,600

0 10,000 20,000 30,000 40,000 50,000 60,000 70,000

Sources: Salary.com, Indeed.com, Glassdoor.com

31
 SaaS. This is primarily due to requirements such of start-ups. Over 20 venture capital (VC) funds
as cleaning, collecting, and preparing data, as were active investors in SaaS in 2021, as opposed
well as the HITL requirements for handling edge to only five to seven funds in 2010. Good start-ups
cases in high-risk situations. These are also the are now “spoiled for choice,” with more than two
reasons why India is better prepared to capture dozen high-quality seed funds and accelerators
value in the AI wave. In addition to having a offering financing. Lastly, industry bodies such
large base of trained personnel, Indian companies as SaaSBOOMi and NASSCOM have played
have their proven ability to manage large-scale an integral role in identifying and showcasing
services projects. IT services companies such as promising AI start-ups, as well as helping them
TCS and Infosys solve some of the largest and to gain much needed credibility during the early
most complex technology process challenges for stages of their journey.
Fortune 1000 companies. As a result, India has • Talent Pool: According to LinkedIn, India had
a whole generation of managers and technical over one million software developers at the end
leaders who are uniquely trained and skilled to of 2021, and by 2023 India is set to overtake
manage large-scale IT projects for multinational the United States as the country with the highest
corporations; they thus have skills that are highly number of developers. 29 While India’s developer
useful in service-heavy AI software development talent pool has been one of the more obvious
and deployment. results of the IT services boom during the late
• Maturing Ecosystem: The success of India’s SaaS 1990s and early 2000s, the SaaS boom since
start-ups has resulted in advantageous second-order 2010 has resulted in the development of the niche
effects for the next generation of AI-first software talent that is required to successfully launch
businesses. These businesses can learn from the global product businesses, and especially highly
product building, go-to-market, and fundraising specialized AI software businesses. Most notable
expertise of the successful entrepreneurs who are among these skills is data science, followed by
actively supporting the start-up ecosystem. Given user interface/user experience (UI/UX) design.
their appetite for trying new things and supporting Analytics services companies such as MuSigma,
start-ups early in their journey, these successful Fractal, Axtria, and many others have contributed
later-stage start-ups have acted as early adopters of significantly to creating a vast pool of data-related
AI-first software. This creates a virtuous cycle and talent in India. India’s advantage lies not only in
enables early-stage start-ups to find initial traction. the availability of talent, but also in the cost of
Additionally, a lot more risk capital is available talent. Sales, engineering, and services talent in
for AI software start-ups in 2022 than was the India cost 75–90 percent less than the equivalent
case in 2010, and this capital is available both at talent in the United States, which makes Indian
the early/seed stage as well as in the growth stage start-ups significantly more capital efficient.

In the AI value chain, most lower-level processes such as data management are
dominated by open source, and this trend is slowly making its way upstream in the
creation of AI models. Creating an open-source AI company is a tricky process—
it requires a skillful combination of horizontal open-source technologies with
proprietary knowledge and understanding of the business problems being addressed.
—JISHNU BHATACHARJEE, Managing Director, Nexus Venture Partners

32
FIGURE 3.5 Strong Second-Order Effects of the SaaS Wave in India

Founders participating in
Founder
coaching and early-stage
Network
investments

20+ VC funds actively invested

Venture
in SaaS in 2021; against 5–7
Capital
in 2010

Multiple SaaS focused seed

Seed Funds/
funds and accelerators have
Accelerators
come up in the last 5 years

Dedicated forums and industry

Industry
bodies have given much
Forums
needed visibility

Mature SaaS organizations

Early
provide an early market for
Adopters
upcoming SaaS start-ups

• Democratization of AI Infrastructure: In projects are related to AI. This includes TensorFlow

the 1960s and 1970s, only IBM could develop from Google, Neo-AI by AWS, H2O, and the
applications on top of IBM mainframes, as the Microsoft Cognitive Toolkit, among others.
underlying stack was proprietary, which gave IBM
Even though not open-source, the one project that
an unfair advantage. Two things have happened
has captured the most attention recently is Generative
over the last few decades: (1) architectures have
Pre-trained Transformer 3 (GPT-3). This is a language
become open—for example, anyone can develop an
application on Microsoft SQL or Oracle Database; prediction model created by OpenAI, an AI research
and (2) the core intellectual property (IP) of lab based in San Francisco, in the United States.
infrastructure layers has become open-source. The GPT-3’s full version has been trained on 175 billion
concept of open source is not new—it has existed parameters.31 Introduced in May 2020, GPT-3 is
since Unix/Linux was developed, but the number part of a trend in natural language processing (NLP)
of different open-source projects and their adoption systems of pretrained language representations. Before
has dramatically changed. Open source is a norm the release of GPT-3, the largest language model was
for today’s software developers. Microsoft’s Turing NLG, which was introduced in
Github, the world’s de facto developer collaboration February 2020. The capabilities of such systems are
platform with over 56 million developers, recorded a available to developers globally, including developers
40 percent year-on-year increase in open-source project in India, which reduces the competitive advantage of a
creation in 2020.30 A large number of open-source few large players.

33
CHAPTER 4

Opportunities for Indian AI SaaS

Start-Ups
India has a unique opportunity to capitalize on the are areas where we expect that India will have distinct
growing global AI market. As discussed in the previous advantages, and areas where it will not.
chapter, India’s vast experience and its advantages In this chapter, we explore where the advantages
in services, the skills gained in building global SaaS lie for Indian companies and the reasons for these,
companies since 2010, and a level playing field created including some early evidence of success with attractive
through Cloud infrastructure players and open-source opportunities. It is important to note, however, that
software, have created the right enablers for India’s the AI landscape will rapidly evolve, and so will India’s
success. However, as we de-layer the AI stack, there capabilities. Therefore, we would not be surprised

FIGURE 4.1 Within the AI Stack, India Has an Advantage Across Three Layers

India’s Right to Win (or Not)

Large pool of technical talent;

Services: Data Plumbing,
Existing know-how for creating global
Data Prep, Ongoing Tuning
services businesses

Success factors similar to SaaS for apps;

Applications: End User Increased advantage for apps & APIs
Applications, AI APIs where service intensity is higher or data
collection has an India advantage

Largest number of developers globally

Dev: DevTools, MLOps by 2024. MLOps can be a significant
opportunity

AI Platform: AI/ML Lack of cutting-edge research

Algorithms, Model Training, on AI fundamentals—either
Feature Extraction in industry or academia

Data: Warehouse,
Deep tech tools, largely from
Ingestion, Cleaning,
academic research or Big Tech
Visualization

Infra: Data Centers, Massive scale & scope requirements;

Distributed Computing Hard to catch up

35
if our predictions are either too conservative or too requirements for computing and storage for Amazon’s
aggressive, even in the medium term. e-commerce business. Similarly, GCP has emerged from
We believe that India’s AI opportunities will broadly Google’s internal requirements. Azure, even though
be in the upper part of the AI stack—that is, closer it launched later than the other two, has grown very
to applications and services. As is explained in rapidly on the back of Microsoft’s deep enterprise
the sections to follow, India has fewer competitive relationships, and its own Cloud application portfolio,
advantages in the lower parts of the stack—the such as what it provides for Office 365. In a relatively
infrastructure, data, and platform layers.  rapid time, the criteria for success have shifted from
providing just servers to providing the application
4.1 Infrastructure Layer—Limited Opportunity components that developers need for front end,
The business of data centers and distributed computing back end, and analytics. The pace of innovation and
infrastructure is highly capital intensive and has scale has made it very hard to penetrate this market.
significant scale advantages. As a result, over the last Although other large companies including IBM,
decade, the infrastructure layer has largely become Oracle, and SAP have attempted to break into this
a three-horse race between Amazon Web Services market, they have not had much success. The three top
(AWS), Microsoft Azure, and the Google Cloud players controlled 61 percent of market share in 2021,
Platform (GCP). AWS is the pioneer and market leader with a year-on-year increase of three percentage points,
in this category, and it has its roots in the internal as seen in Figure 4.2.32,33

FIGURE 4.2 Worldwide Cloud Infrastructure Services Spending

US $27.5 Billion US $36.5 Billion US $19.4 Billion

Growth: 36.4% Growth: 32.5% Growth: 35%

39%
45% 42%

8%
6% 7%

17% 21%
19%

33% 32% 32%

Q3 2019 Q3 2020 Q3 2021

AWS Microsoft Azure Google Cloud Others

36
We believe that as data privacy and data localization happened since 2010, and several such as Hadoop have
policies become more stringent, pressure will increase even come into prominence, and then faded away.
for service providers to keep the data, and therefore Pushing the boundaries of the data stack requires
the computing resources, in India. However, given the
fundamental technical innovation, which has been
growing size of the Indian market, we also expect that
dominated by top-tier doctoral degree (PhD) scientists
the leading global companies will set up local data
and software engineers in Silicon Valley in the United
centers in India and, in fact, AWS has already done so.
States. Currently, such talent is not abundantly available
The chances are slim that Indian data center start-ups
in India, and, consequently, Indian success stories in
will achieve scale before the major global companies set
this category have been rare and sporadic. The best-
up their own competing data centers in India.
known Indian company in the data layer has been
Along with the issue of achieving scale, the technical Qubole, but it has not been able to become a global
depth required to build such platforms is substantial, leader. Recent exceptions are Hevodata, which has
and catching up on these technical capabilities will be developed a leading data ingestion and transformation
hard for Indian players. As Figure 4.2 shows, despite tool, Atlan, a data governance platform, and Acceldata,
their considerable resources and their distribution which has seen early success with its data observability
reach, even Microsoft and Google have had a tough platform. These companies stand out not just because
time making a dent in AWS’s dominance in this of their tech-differentiation, but also because they are
market. In Jeff Bezos’s words: exceptions rather than the rule in the Indian landscape
AWS had the unusual advantage of a seven- of innovations within the data layer.
year head start before facing like-minded However, we do expect the availability of talent in
competition. As a result, the AWS services are by India to consistently improve, and it will be driven by
far the most evolved and most functionality-rich. Indian engineers’ exposure to the latest technologies,
It is important to note that infrastructure is a critical and the greater proportion of research and development
and foundational component for AI to take off, and (R&D) for global technology companies that is being
we believe that infrastructure provision is a large and carried out in India.
growing market. However, for the reasons we just
discussed, the opportunities for Indian start-ups in this 4.3 AI Platforms Layer—Limited Opportunity
market are likely to be limited.
The complexity of building AI applications today
has been significantly lowered by the underlying AI
4.2 Data Layer—Limited Opportunity
platforms, which make various AI algorithms and model
Data infrastructure such as data warehouses, ingestion training infrastructure available. This infrastructure
technologies, data cleansing platforms, and data leverages the research undertaken by universities as well
visualization software have matured significantly in the as companies, and it is generating newer algorithms
United States and Europe since 2010. The development at an unimaginable pace. Most of the AI algorithms
of these technologies has correlated highly with the data are open-source and therefore also offered by Cloud
complexity and requirements of industry and academia in infrastructure players. For these players, the primary
these countries. The requirements of advertising technology monetization mechanism lies in their computing and
(adtech), IoT, e-commerce, and government/defense storage infrastructure, and they are able to offer the
applications have driven innovation in the data stack. platform at negligible costs. It is no surprise, therefore,
Therefore, the West naturally innovated ahead of India. that this market is dominated by the big three Cloud
Examples of these technologies include NoSQL infrastructure providers—AWS, Azure, and GCP. IBM,
databases such as MongoDB, Cloud data warehousing with its Watson initiative, and many others have also
technologies such as Snowflake, data science platforms built sophisticated AI platforms, but they have found it
such as Databricks, and data cleansing software such hard to compete, given the massive reach achieved by the
as Tamr and Trifecta. A host of such innovations have big three Cloud infrastructure players.

37
The development of many of these platforms has been 4.4 MLOps Layer—Potential Opportunity
driven by the needs of large technology companies
In the era of on-premises software, the codes were
such as Amazon, Microsoft, and Google. Each of
clunky and took forever to integrate, test, and
these has hundreds of millions of customers, and
deploy. The update frequency was very low and
customers’ increasing expectations, as well as the need
new releases could take as long as six months.
for differentiation, has led them to use sophisticated
Low update frequency was primarily due to on-
AI algorithms for their core businesses. As an example,
premises deployment of software, as any change
consider how much AI is used to autocomplete
in code had to be upgraded across every instance
sentences in email and in search engines. Much
of every customer. This was particularly difficult,
like AWS was spun out of Amazon’s eCommerce
as customers often had different versions and
requirements, the AI infrastructure initiatives of these
customized code on top of their instances. However,
giants have also taken advantage of internal needs. The
situation in China is very similar, with Alibaba Cloud this changed with the arrival of Cloud applications
being a leading AI platform. where all customers are served with the same
instance and, therefore, they have a single base code.
Despite the presence of giants, some start-ups in the
Cloud-based SaaS companies made it easier and
United States have shown signs of success. Notable
faster to do updates, and engineering teams have
among these is H20, an open-source AI platform that
moved faster than ever before.
provides application programming interfaces (APIs)
and application components (for example, fraud This has led to different development paradigms—
detection) across a variety of industries and functions. for example, the Waterfall34 methodology has been
Similarly, Clarif.ai has created an AI platform replaced by the Agile35 methodology. Multiple tools
specifically for computer vision applications. However, that automated one or more steps of committing a
for the most part, much like data and infrastructure, code to production emerged, and eventually this led
this layer is dominated by the large incumbents, which to the advent of the DevOps category that has enabled
are unlikely to be easily dethroned. teams to release updates much more frequently.

FIGURE 4.3 DevOps is a Large Software Category on Its Own

Code Monitor Operate

Build
Deploy

Test Release

38
FIGURE 4.4 Two Break-Out Winners in DevOps: Postman and BrowserStack

• A collaborative platform for API development, • A cloud-based testing platform that enables
automated testing, version control, and developers to test applications across platforms
continuous monitoring and documentation (browsers, operating systems, and mobiles)
without the need of installation
• Valued at $5.6 billion+ in 2021—only six years
after inception • Raised $200M at a valuation of $4 billion from
Accel, BOND, and Insight Partners
• Raised $430M so far from Nexus, Insight
Partners, and CRV • Powers 2 million tests every day and provides
instant access to 2000+ real mobile devices and
• Counts global giants like Microsoft, PayPal,
browsers
Cisco, and Shopify as its clients
• Counts global giants like Microsoft, Twitter, RBS,
and Expedia as its clients

DevOps as a separate category began emerging, complexities associated with developing ML models
globally in 2008. However, until 2010 there was have led to the emergence of a new set of practices
little notable activity in this space from Indian start- called MLOps (machine learning operations).
ups. But as the software ecosystem matured in India, An MLOps framework must cover everything that a
the country’s developers started following the best DevOps framework does because it is essentially still
engineering practices, and these developers became based on collaborative code development. However,
available as a fertile testing ground for locally created MLOps extends far beyond DevOps because a typical
developer tools. Despite being a late entrant in the AI system contains two more dynamic components:
space, India has created two winners in DevOps since data and ML models. The interplay of data and models
2015—POSTMAN and BrowserStack. on top of the code in AI systems leads to additional
requirements such as statistically validating data,
4.4.1 MLOps—An Emerging Category to Support
AI-First Start-Ups handling continuous data updates, building a complex
pipeline of individual ML models that feed into each
While the popular DevOps tools of today enable other, experimenting with multiple slices of datasets/
developers to deploy production-grade code at scale, model parameters, deploying resource-heavy models
they cannot cater to the added complexities of data and with reliability built in, constantly monitoring the
model management, both of which are essential for AI model’s performance with live data, and so on.
systems. As AI rapidly moves from being a scientific
discipline to an engineering one, and from research/
experimentation to deploying AI in businesses, the
challenges of implementing commercial AI systems have The space of MLOps is in the early
become clear. As of December 2019, only 22 percent of stage of infancy, but there is plenty
companies using AI for commercial use had been able of scope; just as it was with SaaS
to successfully deploy machine learning models, which
products 7–8 years back.
actually constitute the “intelligence” of AI systems.36
Moreover, in companies with dedicated data science —MANEESH SHARMA, Country Manager,
India, Github
teams, data scientists were spending at least 25 percent
of their time just on deployment efforts.37 The unique

39
FIGURE 4.5 MLOps Surpasses the Complexity of DevOps

Practice DevOps MLOps

Code version control + Data versioning +
Version Control Code version control
Model version control
Data Validation N/A Statistical validation
Model Pipeline N/A Training ML pipeline + Serving ML pipeline
Validation Unit tests Unit tests + Model validations
Continuous Integration/ Deploy code to Deploy code to production + Deploy trained
Continuous Deployment production models
Service level Service level objectives-based + Differential
Ongoing Monitoring
objectives-based monitoring + Statistical sliced monitoring

A complete MLOps infrastructure that can help (data preparation/model development/experimentation/

with the tasks just listed will comprise parts that are deployment/monitoring), there has also been a recent
considerably more complex when stacked against surge in end-to-end MLOps platforms. Given the
analogues from the DevOps world (Figure 4.5). complex and customized nature of AI use cases, the wide
In 2019, more than 200 MLOps tools were available, applicability of such generic end-to-end MLOps platforms
globally and earning over $350 million in revenue. is still uncertain. A chart depicting the popular categories,
MLOps continues to grow rapidly and is poised to reach a as well as the end-to-end MLOps tools, is shown in Figure
global market size of $4 billion by 2025.38 While most of 4.6. Going forward, by 2025 we expect more MLOps
today’s MLOps tools are provided by specialized players tools that target a specific use case (for example, NLP or
that offer solutions for one or two parts of the value chain an entity classification) to come to market.

FIGURE 4.6 MLOps is a Fast-Growing Software Category

Data Development Tuning/ Deployment/

Experimentation Monitoring

40
4.4.2 MLOps from India—Opportunities and Challenges39 targeting the North American SME market. Zoho and
Freshworks are the two best-known companies.
While MLOps is certainly a promising new category, it is
still unclear if India will be able to capture a meaningful We believe that India’s success in AI-led SaaS will also
market share. India’s strategic advantages in producing have some dominant themes. In Chapter 1 we discussed
winners in MLOps comprise (a) having a large talent how “AI software” is different from traditional SaaS
pool of developers and data scientists that can act as beta in terms of service intensity, which is the case not
users/early adopters; and (b) having a rapidly evolving just during development but also during and after
tech start-up ecosystem wherein large tech start-ups that implementation. India’s edge in supplying talent,
aspire to follow globally benchmarked engineering best as well as its cost arbitrage with regard to services,
practices could be codevelopment partners. However, provide a unique competitive advantage for India’s AI
Indian developers have traditionally lagged behind start-ups. The higher the services intensity, the greater
their global counterparts in adopting engineering best India’s advantage will be.
practices. This is due to a combination of India’s deep AI-applications can be divided into four categories,
services “DNA” and the fact that Indian product start- based on the increasing level of their training and
ups have only recently become large enough to see the service complexity, and therefore their increasing
benefits of following best practices. For India’s large pool attractiveness for Indian AI start-ups (see Figure 4.7).
of developers to become a material advantage for Indian
1. No/Little Training: Think of an application that
MLOps companies, that trend will need a quick reversal.
uses Google’s speech-to-text API as part of a video
It is a bit early to predict how successful India will be conferencing application for the transcription of notes.
as a source of global MLOps companies. However, Technically, this is an AI application; however, the
there are a few promising start-ups. For example, application provider is not investing in any training in
Segmind is a start-up focused on a unified MLOps this scenario—instead it is relying on Google to train
platform for computer vision applications; Waterdip. the API, and it simply uses the API as it is, with no
ai is creating an observability solution for assessing modifications. Such software is no different from any
model drift; and True Foundry is creating a platform traditional SaaS application using an external API.
for end-to-end flow for deployment and monitoring of We classify such applications as AI-enabled, and we
ML applications. believe that India has no special advantage above and
beyond what it has with traditional SaaS.
4.5 Application Layer—Significant Opportunity
2. Train Once, Deploy Everywhere: Many AI
While AI applications could be sold either through applications do not require separate training for
the traditional licensing-based model or through the each customer, but instead utilize the same model
software as a service (SaaS) model, in this and the (and therefore the same underlying training) for
following sections we use the terms AI applications each customer. Signzy, which provides onboarding
and AI-led SaaS interchangeably, since we believe that solutions for financial services institutions, is
over the next decade SaaS will become the preeminent one such example. One of the steps in customer
business model for Indian AI companies. onboarding is know-your-customer (KYC), which
Since 2010, India’s SaaS industry has seen breakout involves matching people’s faces across multiple IDs/
companies across different industries, functions, and documents, and ensuring the consistency of text that
customer types. However, India has had the most is extracted from multiple documents using optical
success within the small and medium enterprise (SME) character recognition (OCR). While underlying
space, where annual contract values (ACVs) are low models for face matching and OCR need to be
and unit economics are not easy to make work with developed for the KYC use case, they do not need to
U.S.-based sales costs. Remote inside sales from India be trained differently for different customers. Case
dramatically lowers the customer acquisition cost (CAC), Study 4.1 shows two examples of Signzy’s capabilities
and Indian SaaS companies have been very successful in that do not require customer-specific training.

41
FIGURE 4.7 India Has a Greater Advantage Where Service Intensity is Higher

Type What it means India’s advantage

• Uses 3rd party APIs inside other

AI-Enabled
I. No/Little workflows with little/no additional
training • Similar to regular SaaS
Training
• Use of AI is just a feature

• Application uses 3rd-party or open- • Access to the right

II. Train Once, source algorithms/partially trained training data is important
Deploy algorithms, but need to supplement • Companies like Skit and
Everywhere with own data training. Does not Increasing
Inspektlabs have solved it advantage
require any customer specific training in creative ways for India

• Longer implementation
AI-First III. Customer- • AI needs training with every customer cycles
Specific as data and underlying patterns are
Training unique to every customer • Higher data-related
services

• Customer needs full solution to a • Can be combined with

IV. Human in
problem and inaccuracies of AI are not BPOs at a later stage, but
the Loop
acceptable—e.g., medical diagnostics, AI app providers will need
(HITL) AI
where human backup is required to do this in early stages.

CASE STUDY 4.1 Signzy: Examples of AI Applications That Do Not Require Customer-Specific Training
Industry: Financial Services
Function: Customer Onboarding
Location: India/US/Middle East

• Real-time ID verification
against established
databases
• Matching face on ID with
face in video (with %
confidence score)

42
3. Customer-Specific Training: A large number of AI application’s provider delivers this service. In such
applications will need customer-specific training— cases, the boundary between product and services
for example, demand forecasting. Machine learning begins to blur. For example, Synapsica’s technology
models for demand forecasting are dependent on can read through MRI/CT scans to identify and
historical demand data for each specific company. measure relevant biomarkers; however, since a
Thus, while the technique used for forecasting mistake could be fatal, the company has a remote
may be the same across different fast-moving radiologist who checks the machine’s diagnosis.
consumer goods (FMCG) companies, the actual 4.5.1 Micro-Verticalization—India’s Edge in
model that is deployed is likely to be very different, AI-Applications
and even more so across customers from different
Memorial Sloan Kettering Hospital in New York is
industries. Limechat.ai is a conversational AI start-
regarded as one of the best cancer hospitals in the
up that is targeting this category of problem. The
world. One of the reasons for this is the hospital’s
training complexity for these kinds of applications
level of specialization. Even within onco-surgery, there
is naturally higher than that of the previous two
are different surgeons for ovarian cancer and breast
categories, described above. cancer. This is important because a surgeon who only
4. Human-in-the-Loop (HITL) AI: Unlike the performs ovarian cancer surgeries is able to identify
demand forecasting example just discussed, for patterns and nuances that a general onco-surgeon
many AI-applications achieving 100 percent might not be able to. AI models are no different.
accuracy is necessary. As discussed in Chapter 2, The more specific you make them (while, of course,
examples of high-risk applications include robotic providing sufficient data), the better they become.
surgery and AI for diagnosis of complex diseases. Google’s Dialogflow was designed to be a booking
Since AI has deficiencies, and in many cases it assistant that could be used by a variety of small
cannot provide sufficiently accurate output, human businesses such as spas, beauty salons, and dental
backup is required. Sometimes this can be provided practices. However, making Dialogflow work for
by the customers themselves, and in other cases the a specific application such as a spa is not easy.

CASE STUDY 4.2 Synapsica: Marketplace and AI-Assistant for Efficient Radiology

Industry: Healthcare
Function: Diagnostics
Location: India/US

Problem Solution Impact

• Margin pressures, as 85% of • Automatically detecting case • Improve margins through
the radio-diagnostic market information, e.g., body part, workflow automation and
is fragmented, and 24x7 orientation, etc., to determine accessing a flexible pool of
radiologist cover is costly reporting needs radiologists via an Uber-like model
• Wasted time and earning • Automated detection and • Save time with automated
opportunities due to the tedious characterization of the biomarkers reports and diagnostic assistance.
work of adding objective of pathologies in patients’ scans Create access to geographically
evidence to a report distributed earning opportunities
• Combining Al findings with NLP to
• Opaque reports without auto-generate relevant text/table/ • Generate better quality reports
objective evidence lead to images in radiology report and visibility in medical care
subpar medical assistance

43
Dialogflow does not understand the nuances of a spa In working with AI companies, we believe that greater
appointment, and the buyer, in this case a spa, needs specialization is often necessary for both the accuracy
technology that is trained for its specific use case. of AI as well as the overall value proposition of the
Interpreting the vocabulary, intent, and grammar software because specific workflows and integrations
for dentists is very different from that required for are required across different verticals (industries).
spas. If someone wants to come in for a hairstyling
It is not surprising, therefore, that AI start-ups are
appointment, perhaps the booking assistant should ask
specializing aggressively and targeting micro-verticals
if the customer also needs a shampoo, hair coloring,
across industries. Since 2010, a great many start-ups
and other related services. The assistant must also
have launched within the industrial IoT space. At the
know the amount of time required for each of a small
business’s services, understand the availability and core of all of these is a common value proposition—
skills sets of different service team members, and they collect data that is continuously emitted from a
be able to schedule appointments accordingly. If the vast number of machines and sensors on shop floors,
assistant is integrated into the customer relationship and they utilize the data to predict faults and prevent
management (CRM) software, the system should know downtime. In theory, if this works for one business,
that the customer has a teenage daughter, and ask if the it should work for all. However, in practice, it does
daughter also needs an appointment at the same time. not. An automotive plant and its equipment are very
These specificities of spas will make the application different from a specialty chemical plant, and two
very different from what is required for a similar front specialty chemical plants are very different from each
office solution for dentists. True Lark is one example of other. Two brands of pumps have different technical
a product for the front office of a spa (see Figure 4.8). specifications and, often, two pumps from the same

FIGURE 4.8 Micro-Verticalization Will Play to India’s Advantage

• Front-desk assistant that handles inbound • Dialogflow provides an end-to-end “build

customer queries over text once, deploy anywhere” development suite
for chatbots
• Sharp focus on beauty and wellness, fitness
centers, and dental clinics • Requires a developer to build the chatbot
using Dialogflow, which is an expensive
• Understands the service portfolio of these
proposition for most SMEs
micro-verticals as well as the inbound
customer’s context for product enquiries, and • Will require training on the salon-specific
booking data to understand the customer’s booking
intent, service catalogue, etc.
• Integrates with standard booking software
that SMBs use—Acuity, Booker, etc. • No out-of-the-box integrations

Can I come AI understands that ‘head Shows the available slots and
for a head massage’ is a service, and masseuse details by pulling it
massage the customer is asking to from its integration with the
next week? book this service next week. booking management system.

44
manufacturer have different normal/abnormal signals We believe India could be a provider of many such
in different situations. APIs, particularly in areas where India is data rich.
Start-ups in general, and those in India in particular, Other examples of APIs could lie in healthcare (for
have a unique advantage as big software companies example, assessing x-rays and MRI scans), financial
such as Google and SAP will find it hard to justify services (for example, OCR of structured financial
investing in specialized use cases; they would rather data) and identification (ID) matching (for example, of
create platforms, on top of which others can build driving licenses).
specific applications. 4.5.3 Unit Economics and Capital Requirements for
4.5.2 The API Opportunity Building AI-Application Companies
Given the sizable global SaaS opportunity for
Much like Twilio, many SaaS companies do not
companies from India, VC funding in Indian SaaS
build full-fledged applications. Instead, they provide
companies has grown rapidly since 2010, as shown
APIs that are repeatedly used across many different
in Figure 4.9.40 In 2019, VC funding of $1 billion was
applications. Such examples are abundant within the
raised by Indian SaaS companies, which was more
AI stack as well; the Google speech-to-text API is a
than twice the amount raised in 2018. In 2021, Indian
very popular API service. In addition to applications,
Indian companies will also find opportunities to create SaaS companies attracted $4.5 billion of fresh capital,
AI-APIs where they have a data collection advantage a 170 percent increase over 2020.41 Freshworks alone
over global alternatives. had raised over $300 million. Almost every major VC
fund in India invests in SaaS today, and many specialist
We are already seeing examples of such companies in SaaS funds have been created—for example, Exfinity,
India. Inspektlabs has built APIs that assess damage Avataar, and Together.
to cars based on uploaded photos of the cars. Given
India’s bad traffic conditions, cars have a much higher The success of SaaS investing will drive capital toward
incidence of dents, scratches, and other damage, which AI-application companies, as well. However, we believe
provide rich data for training. Within a few months, that in comparison to the traditional SaaS companies,
Inspektlabs collected millions of photographs at a very creating AI-application companies will require more
low cost. Today, the company’s APIs are being used by capital for the following reasons:
insurance providers to assess vehicle damage with no • Higher initial costs in data acquisition and training
need for a visit by a company inspector. to reach a minimum viable product;

CASE STUDY 4.3 Inspektlabs: Computer Vision for Asset Inspection

Industry: Automotive, Insurance, Mobility

Function: Asset Inspection
Location: India/Global

Problem Solution Impact

• Every time a vehicle changes hands, • Customer takes a 360° video of • Automates car
there is a need to perform inspection the asset which is shared with inspections for motor
on the asset for any damages. Inspektlabs’ pretrained damage insurance and sharing
detection models via APIs for real- economy (e.g., car
• Current process for inspections is
time quality scores. rental players).
completely manual and slow—taking
up to 10 days, expensive—up to INR • The model has been trained on 5Mn+
1000 per inspection, and rife with videos/images for accurate detection
fraud (5–15%). of dents, scratches, etc.

45
FIGURE 4.9 Venture Capital Investment in Indian SaaS Companies

Capital Invested

$0.1B $0.1B $0.2B $0.2B $0.4B $0.4B $0.3B $0.4B $1.0B $1.7B $4.5B
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021

• Longer sales cycles—many deployments will require • Given the higher value of solutions, higher ACVs
custom proofs of concept (POCs); can be expected;
• Longer implementation cycles where customer- • In many businesses, there will be data network
specific training is required; effects. More customers ==> more data ==> higher
• Higher Cloud infrastructure costs for data storage accuracy ==> higher value ==> higher number of
as well as training; customers;

• Higher research and development (R&D) team costs • The increasing availability of open source
because, in addition to those with regular software components and APIs can reduce development costs.
skills, data architects, data scientists, and machine Our research and experience suggest that AI-
learning engineers will be needed. application companies will take longer to create a
Some of these costs will be offset by some advantages minimum viable product, they will typically have lower
that AI-applications will have over regular SaaS: gross margins in their initial years, and they will take
longer to scale due to more complex sales cycles. Gross
• Given the deeper integration and learning of
margins for AI-application companies will typically
models over time, lower customer churn and higher
start at 50–65 percent, compared to 70–80 percent
customer stickiness is likely;
for traditional SaaS companies. In the long run, these
margins should move toward those of traditional SaaS
companies owing to lower churn (loss of customers),
When selling AI SaaS, keep your and better pricing once value has been demonstrated.
POCs quick; longer POCs can be a In conclusion, to achieve an annual recurring revenue
death knell for an AI SaaS start-up. (ARR) of $100 million, we conservatively estimate that
an AI-application company will require ~$120 million
—RAJESH RADHAKRISHNAN, EVP and
General Manager, Front Office Automation, in capital compared to ~$75 million for a traditional
Automation Anywhere SaaS company, and could also take a year longer to
achieve ARR of $100 million.

46
FIGURE 4.10 AI-SaaS Will Need More Capital Than Traditional SaaS Companies

Gross margins of SaaS companies Capital requirement to reach $100M ARR

70–80% $115–125M

50–65%

$75–80M

SaaS Companies AI-Led SaaS SaaS Companies AI-Led SaaS

However, this does not mean that AI-led applications 4.5.4 Sizing the Opportunity for AI Applications
are not an attractive opportunity. In fact, quite the To estimate the size of the opportunity available for AI
contrary. We believe that many cost components will applications from India, we estimated the aggregate
decline over time. Also, in the early years, valuation revenue for Indian AI applications companies, and applied
premiums will be high, as leaders will be created in a valuation multiple in order to arrive at the market
different categories, and the scarcity of similar assets capitalization that these companies could create. We used
will drive investors to pay high valuations for the a decade-long forecast interval that allows sufficient time
emerging winners. for the category to display signs of maturity.

FIGURE 4.11 Methodology: Sizing the AI Applications Opportunity for India

Market Size for

AI Applications
(2020)

Market Cap
Global AI Market Share
Valuation of Indian AI
Applications for Indian
Multiple Application
Revenue Companies
Companies

Growth Rate for

AI Applications

Notes:
1. Market growth rate benchmarked against growth rate for Cloud applications in early stage of maturity.
2. Market share for AI applications benchmarked against market share for SaaS players from India.
3. Valuation multiple estimates based on late-stage private and public market data.

47
Between 2010 and 2020, the global revenue of SaaS technology and data management skills; experience
companies—the last major wave of significance in the in remote sales of software products to international
global software industry—grew by more than l0 times, SaaS customers; significantly lower labor costs; and
from $9.2 billion in 201042 to $104 billion in 2020,43 the development of a vibrant start-up ecosystem, with
and this revenue is expected to reach $479 billion by diverse sources of capital and know-how available to
2030.44 As of 2020, AI applications revenue, including young start-ups.
both SaaS and licensed models, was estimated to be Since several of these strategic advantages are either
around $20 billion.45 However, this segment is expected the same as or similar to those for traditional SaaS, we
to grow rapidly, with healthy growth rates of between expect that the trajectory of India’s market share in AI
25 percent and 40 percent between now and 2030, and applications is likely to closely mirror that of India’s
it is likely to constitute a significant proportion (over traditional SaaS market share. Moreover, we assume
$300 billion) of an estimated $1.3 trillion in global that SaaS will be the predominant business model for
software spending by 2030.46 the new AI applications companies coming out of India
As discussed in earlier chapters, India is likely to enjoy and, therefore, that Indian AI applications companies
some unique advantages in developing AI applications will, for all practical purposes, be AI SaaS businesses.
due to: the country’s decades of global leadership in Within the traditional SaaS market, India’s revenue
traditional software services; a large workforce with share has risen from practically nothing in 2010 to

FIGURE 4.12 AI Application Companies Can Create ~$400 Billion Market Cap by 2030

Cumulative Market Cap of Indian AI Application Companies

63 382

16.6
Revenue (US $ Billion)

21.6

2.1
0.4
0.3 4.2

2020 2025 2030

Potential Market Cap Base estimate Upside as per optimistic estimate

1. Base market growth rate of 35% until 2025 and 25% for 2025–30, benchmarked against cloud growth since 2008.
2. Optimistic market growth rate of 40% until 2025 and 30% for 2025–30, consensus estimates.
3. India’s share in global SaaS ~6% and 10% in 2025 and 2030, respectively; our scenarios assume that AI apps
will either command a similar share from 2025 onwards, or will take two years longer to reach similar market shares.
4. At maturity, assumed EV/ARR multiple to be 10.

Sources: Gartner, NASSCOM, IDC, Grandview, SEC filings

48
3.4 percent47 in 2020, and it is expected to continue Unlike applications, where significant value will be
to outperform the market and reach 9.5 percent48 by captured by start-ups, we believe that a significant
2030. We have assumed that by 2030, India’s AI-SaaS part of the value in services will be captured by the
market share will lag behind India’s SaaS market share incumbents. With regard to expertise, we believe
by zero to two years. that skills will transfer more easily across services
In the previous section on unit economics, we or, in many cases, that AI services will not be as
compared the economics of AI-SaaS businesses to difficult to build as applications. Incumbents’ existing
that of traditional SaaS businesses. In our view, AI client relationships will also make it easier for them
companies are likely to be less capital efficient than to capture opportunities, as opposed to start-ups,
traditional SaaS companies, with a slightly slower which will take a long time to build their client bases.
growth trajectory to achieve similar outcomes. However, some opportunities are being captured
However, for the foreseeable future, these businesses by start-ups, several of which are discussed in the
are likely to have “stickier” customers and revenues, following section.
command higher pricing over time, and present Data-Related Services
opportunities for category leadership. On balance, the
A large quantity of data must go through multiple
longer-term average valuation multiples for traditional
different steps before it can be used to train AI models
SaaS businesses—an Enterprise Value (EV)/ARR of
with reasonable accuracy. The first step is data
~10 for a representative sample of listed companies—
extraction, which involves accurately picking data
are the best proxy for the future valuation multiples of
from structured or unstructured sources that can be
AI-SaaS companies.49
either human or machine generated. Examples include
Based on these assumptions, we believe that Indian AI invoices, emails, documents, images, and videos. The
application start-ups could create a market capitalization next step is data cleansing, which ensures that the data
of anywhere from $200–$400 billion by 2030. adhere to underlying norms such as anonymization
of information, removing erroneous information, and
4.6 Services Layer—Significant Opportunity standardizing formats—for example, across locations.
Since 1980, India has developed significant capacity Once data have been cleaned, they are often sent for
to operate services-oriented businesses, and the labeling to add richness to the data; this can involve
country is now a preferred destination for outsourcing the selection of pre-specified objects within the data, as
various services such as technical support, technical well as the addition of specific metadata—for example,
implementation, customer support, lead generation, identifying different elements in a traffic image for
data entry, and many other back-office business training self-driving car algorithms. Finally, training
processes. What makes India the number one choice is data can also be augmented to add robustness such
its large pool of skilled English-speaking, tech savvy, as through the generation of synthetic data, which
affordable professionals, and a proven track record for are used when real-world training data are either
managing large and complex projects. unavailable or too costly to generate—for example, in
As discussed earlier in this report, the primacy the case of Waymo One, 50 a self-driving taxi service
of data and models results in significantly higher launched in China by Alphabet. Augmentation can
service intensity in the AI software development also occur through combining multiple data sets into a
and deployment process. Service requirements are larger data set that has greater context.
unavoidable, and are found at all the key stages of the Depending on the requirements of the model being
process—from data preparation, model development generated, and the nature of the underlying data, data
and deployment, to ongoing support. This has created an preparation services can vary substantially. Taskmonk
entirely new area of opportunity for AI-focused services. and Playment are two such examples. Both provide
Figure 4.13 lists the different kinds of AI-related tools and services for data labeling, and they also
services where Indian companies have opportunities. automate parts of the labeling process.

49
FIGURE 4.13 India Could be the Natural Destination for AI Services

Retrieval of data from structured/unstructured sources;

Data Extraction machine generated data, e.g., invoices, error logs; human
generated data, e.g., emails, documents

Ensure data adheres to norms/standards—remove erroneous

Data Cleansing information, standardize formats, anonymize information, fill in
missing values

Data annotation (selection of pre-specified objects), and

Data Labeling
addition of metadata

Synthetic data, e.g., blurred images for autonomous vehicles;

Data Augmentation
combining data sets, e.g., adding location data to FMCG spend

Implementation Services Model development, testing and deployment

Rule changes within enterprises or their workflows can throw

Exception Handling models off kilter; back-up teams are always needed for mission
critical tasks

Model Correction ‘Drift’ is inevitable and needs constant course correction

Data related services Implementation services Ongoing support

CASE STUDY 4.4 Taskmonk: Making Data Labeling More Efficient

Industry: Horizontal
Function: Data Ops
Location: India/US/SEA

Problem Solution Impact

• Enterprise-grade AI solutions • Taskmonk’s unified data labeling • Effective auto-AI decreases
need massive amounts of platform contains active-AI human labeling time by up
labeled data for training. (pre-labeling with existing to 60%.
standard algorithms) built in.
• Up to 25% time in a typical • Unified data labeling platform
enterprise AI project is spent on upgrades employment-intensive
data labeling. small BPOs, enabling them to
win large enterprise projects.

50
CASE STUDY 4.5 Playment: Data Prep Entails Unique Business Opportunities

• Provides both tools and services for

data labeling.
• Has built tools to automate parts of
the labeling process.
• Crowdsources data labelers to create
a flexible capacity model.
• Started with a focus on autonomous
driving that requires high accuracy
and needs massive amounts of data
to be labeled.
• Data labeler costs in India can be as
low as $200 per month.

Implementation Services: Model Development Ongoing Support

and Deployment
Most AI applications require regular, ongoing
As enterprises adopt AI, they often require a tailored support. In general, support requirements are for
approach to make an AI system work with their unique two types of use cases. The first are the high-risk
data and existing processes. Model development requires use cases, with very high costs if they fail—for
dedicated intervention by highly skilled machine example, medical diagnosis, which requires a
learning engineers and data scientists. Additionally, human-in-the-loop (HITL). The second case is
deployment often needs customized efforts—for “model drift,” which essentially is a change in the
example, integrating with legacy IT systems that have relationship between a model’s input and output
been custom developed and therefore do not have variables. Model drift is a natural consequence of
standard APIs. Most traditional enterprises do not have the dynamic nature of AI, and it can occur over
highly skilled in-house AI talent and prefer to outsource time as a result of regime changes in a model’s
the model development, testing, and deployment process. underlying data set(s). Consider the case of a
This is creating an opportunity for companies like SCRY manufacturing plant that collects data from a
Analytics (see Figure 4.15), which works closely with variety of sensors and uses predictive modelling to
each customer and implements a custom-fit AI solution forecast machine failures before they occur. When
from its wide suite of use case-based product lines. equipment in the plant is upgraded to collect data

There is a trade-off between cost and returns in getting from 80 percent to 99

percent accuracy for AI; more often than not, it makes sense to stay at 80 percent,
and use human aid to make up the difference. India’s advantage in AI services
stems from its proven expertise in handling large-scale software services projects.
—SIDDHARTH MALL, Co-founder and CEO, Playment

51
CASE STUDY 4.6 Scry Analytics: Turnkey AI Solutions Companies—Solution as a Service

What: Collection of Collatio Risc

AI-enabled enterprise
applications across different Ingesting, harmonizing, and Predicting financial and
industries and functions reconciling data business risks for financial,
pharma, and service industries
How: Work closely with each
customer, using a service heavy
approach, and implement a
custom fit AI solution from
their wide suite of use-case
Anomalia Medsocial
based product lines
Identifying potential fraud risk, Community and platform for
Why: As enterprises adopt
conflict, and noncompliance in insights from global medical
AI, they require a tailored
financial and legal transactions conversations
approach to making the system
work with their unique data,
and existing processes Data Collection Data Processing Integrations Decision Workflows

Location: India/US
Human-in-the-loop ‘Drift’ Improvement User-training

from more points and with higher accuracy, the 4.6.1 Sizing the Opportunity for AI Services
plant’s previous predictive model may not be as To estimate the market size for AI services from India,
effective and the model will need to be retrained we used a similar approach to the one we used for AI
with the new sets of data coming from the upgraded applications. We divided AI services into two broad
equipment. As a consequence, machine learning categories: (1) IT Services (services such as model
models need regular maintenance services so that testing and deployment, which are already being
they continue functioning as optimally as possible. performed by Indian IT services companies), and (2)
It is worth pointing out that the market for AI services business services (such as data labeling and HITL
is one where skilled employees would be needed to services), which are being provided by Indian business
succeed, and as a corollary, labor cost arbitrage, process outsourcing (BPO) companies. We then forecast
while useful, is an insufficient competitive advantage. the aggregate revenue of India’s AI services by 2030,
Companies that provide AI services—be this data and subsequently applied a valuation multiple to arrive
preparation or model development/deployment/ at the estimated market capitalization while using a
maintenance—use sophisticated technology and conservative estimate for long-term valuation multiples.
employ highly qualified and trained personnel to run
their operations. The most important skills needed for
success are (a) technical capabilities and (b) managing
AI Services will generate more revenue
large delivery projects at scale. India is well placed on
than AI Applications themselves.
both counts, as it has a large population of software
developers, an increasing number of data scientists, —SUMIT GUPTA, Head of Product
Management, Google
and decades of experience managing highly complex
IT delivery projects.

52
FIGURE 4.14 Methodology: Sizing the AI Services Opportunity for India

Market Size for

AI Services
(2020)

Market Cap
Global AI Market Share
Valuation of Indian AI
Services for Indian
Multiple Services
Revenue Companies
Companies

Growth Rate for

AI Services

Notes:
1. Market growth rate for services assumed to be same as AI applications.
2. Market share for AI services benchmarked against market share for ITeS players from India.
3. Valuation multiples estimated based on data from top 4 publicly listed Indian players.

To estimate the revenue accruing to Indian AI services around 4.5.56 While the talent cost of servicing AI
companies, we estimated the global size of the AI requirements would be higher than for traditional IT/
services market, as well as India’s market share. BPO services, we believe that automation will reduce
The total global AI services market in 2020 was operations and process costs. Furthermore, AI service
about $18.4 billion, 51 of which AI IT services vendors can command higher pricing, and would be
accounted for an 80 percent share, and business harder to replace due to sensitivities regarding data
services accounted for the rest. Similar to AI sharing. Therefore, on balance, we have assumed a
applications, AI services are also expected to grow similar EV/revenue multiple for AI services as well.
by 25 to 40 percent between now and 2030. Based on these assumptions, over the next decade,
We expect that India’s market share for AI IT and we estimate that for AI services, India could create
business services will follow a similar trajectory as market capitalization of somewhere between $130
that of traditional IT services and BPO industries, billion and $200 billion.
respectively. India is a leading destination, globally,
for both IT services (~$1 trillion52 global market) and 4.7 The Global AI Opportunity
BPO services ($166 billion53 global market), with a
To conclude, in terms of financial impact, India’s
market share in 2020 of 10 percent, 54 and 25 percent, 55
participation in the global AI opportunity could
respectively. With a conservative assumption of a 6
percent compound annual growth rate (CAGR) for create $500 billion worth of market value by 2030, 57
Indian services until 2030, versus a ~3 percent global which could add an estimated 900,000 new white-
growth rate until 2030, India’s market share in IT and collar jobs, and more than 3.6 million indirect jobs.58
BPO services would be 13 percent and 33 percent, To put things in perspective, in 2020, the total
respectively. However, the traditional IT/BPO services market capitalization of IT services companies in the
industry is fairly mature in India, so with regard to “Nifty 50” (India’s 50 largest traded companies), was
global market share, it is likely that AI services from approximately $420 billion; there were only about
India will lag behind by up to two years. 12.5 million white-collar employees in India;59 and
The enterprise value (EV)/revenue multiple for the overall net employment growth rate was well
India’s IT and BPO services businesses averages below one million jobs per year.

53
FIGURE 4.15 AI Services Companies Can Create ~$200 Billion Market Cap by 2030

Cumulative Market Cap of Indian AI Services Companies

46 198

15.5
Revenue (US $ Billion)

28.4
2.2
0.1 8.1

2020 2025 2030

Potential Market Cap Base estimate Upside as per optimistic estimate

1. Base market growth rate of 35% until 2025 and 25% from 2025–30, assumed to be at same pace as AI Apps & benchmarked
against 2000–05 growth in Indian IT companies.
2. Optimistic market growth rate of 40% until 2025 and 30% from 2025–30, same as AI Apps.
3. India's share of global ITeS ~10% and 12% in 2025 and 2030 respectively. Our scenarios assume that AI services will either
command similar share from 2025 onwards, or will take two years longer to reach similiar market shares.
4. At maturity, assumed EV/revenue multiple to be 4.5.

Sources: Gartner, NASSCOM, IDC, Grandview, SEC filings

FIGURE 4.16 Impact of India’s Play in the Global AI Market

$580 900 3.6

Impact of Billion Thousand Million
AI by 2030 worth of potential new white- new indirect
market value to collar jobs to jobs to be
be created be created generated

54
4.8 The Domestic AI Opportunity

India’s AI Potential AI start-ups will be pleasantly

So far in this report, we have discussed the massive surprised by the vibrancy and size
opportunity for Indian companies to create value as a of the domestic opportunity.
direct result of global AI adoption. While that has been —C. N. RAGHUPATHI, SVP and Head of India
the primary focus of this report, we must also offer our Business, Infosys
views on the potential for AI adoption in India.
Within India, immense value could be created and
captured through AI, not just by for-profit enterprises,
pushing AI adoption in India is not the explicit cost of
but also by the government and society at large.
getting the job done by humans, but the implicit cost of
However, so far, the adoption of AI in large sectors
delays and inefficiencies.
in India such as education, healthcare, agriculture,
financial services, and retail has been quite limited. AI has already made inroads into the everyday lives
The scale of operations required in India cannot be of millions of Indians without much fanfare or notice.
solved solely through physical and human resources. Banks have embraced digital KYC processes to reduce
Indian highways are often choked at tollbooths, the customer onboarding time from a few days to a
hospitals are overwhelmed with incoming patients, few hours. Tax collection and assessment is now being
and banks are overloaded with know-your-customer aided by ML algorithms to identify tax evasion and
(KYC) processes that lead to significant delays. As a payout errors.60 These are not just isolated examples,
result, the economic cost of delays is greater than the but are part of a larger push toward a more efficient
cost of investment in AI technology. Therefore, what is system of work across India.

FIGURE 4.17 AI Could Add $500 Billion to India’s GDP by 2025

Impact of AI

Retail Agriculture
• Predictive analytics • Precision farming for higher yields
• Targeted marketing 90B • Crop failure prediction 60B
• Quality assessment • Improved crop insurance

Finance Logistics
• AI-enabled underwriting of • Road network optimization with
loans for SMEs and individuals 60B geospatial data 50B
• Financial risk modeling and • Traffic pattern analysis
customized pricing

Energy Healthcare
• Energy demand prediction and • Remote diagnostics and treatment
usage optimization 50B • Optimized health resource 25B
• Rationalization of asset allocation
utilization • Cheaper and faster diagnostics

GDP increase

Sources: NASSCOM. 2020. “Unlocking Value from Data & AI: The India Opportunity.”

55
On the domestic front, by 2025, $500 billion could example, Niramai is able to provide highly accurate
be added to India’s economy through leveraging breast cancer detection in a noninvasive manner, which
AI,61 spread across key sectors such as agriculture, has enabled many more women to come forth for
healthcare, logistics, energy, and finance. For screening. This in turn enables early diagnosis, which
clarification, this is the impact on GDP, as opposed to significantly improves a patient’s chances of survival.
market capitalization. For example, agricultural yield Another company, Plantix, helps farmers by providing
in India is among the lowest in the world. India’s paddy them with a highly accurate diagnosis of plant diseases
rice yield in 2018 was 3.8 tons per hectare compared to within a few hours, and is already used by more than
8.6 tons per hectare in the United States and 7 tons per one million Indian farmers.
hectare in China,62 and this productivity gap cannot Other application areas of AI are equally promising.
be bridged by infrastructure development alone. AI For example, companies today use AI-driven models
can play a big role by helping farmers directly through to assess the creditworthiness of individuals who do
precision agriculture, crop failure prediction, pest not have a financial history, and this is increasing
control, and so on, and indirectly through improved financial inclusion. On the healthcare front, mFine, a
crop insurance and financing. AI-led improvements in medical consultation start-up, is using AI to increase
sectors such as insurance and financing will also have doctors’ patient screening capacity by three to four
incremental second-order effects on industries that times. This productivity improvement has immense
heavily depend on them such as agriculture, retail, value in a country where the doctor/patient ratio is 1
and pharmaceuticals. to 1,445.63 Accessing medical specialists is one of the
In addition to AI’s contribution to India’s gross biggest bottlenecks in healthcare in India. We are also
domestic product (GDP), there is massive potential optimistic and confident about AI’s role in helping to
for India to create social impact through AI. We are improve the quality of life of Indian citizens, and it is
already seeing signs of AI start-ups creating social this potential that excites us the most when we think
impact by solving problems unique to India. For about what AI could achieve by 2030.

FIGURE 4.18 AI Start-Ups Have Started to Show Impact Across Sectors

• Breast cancer is the most common form • Farmers need to deal with a variety of problems
of cancer among women and kills over such as pests and disease during the growing
600,000 women each year. cycle. Problems are often new, with the potential
to spread rapidly. Given small landholdings, delays
• Indian women with breast cancer have only
in solving the problem often lead to massive
a 50% chance of survival.
income shocks for Indian farmers.
• Niramai’s AI-based diagnostics software
detects cancer earlier than traditional • The current solution involves visiting local
methods in a noninvasive, radiation-free, retailers/large farmers, which takes significant
and painless manner. time. These inputs are prone to error.

• The technique has shown >90% • Plantix takes images of infected crops—which
farmers upload through the Plantix app—and,
sensitivity and is 27% more accurate than
mammography. Now 32% more women within a few hours, provides a diagnosis along
walk into hospitals for breast screening. with recommended solutions. It is currently used
by >1 million farmers in the country.

56
CHAPTER 5

Risks and Considerations

5.1 Challenges Ahead on the Internet, only 313 were in India. It has been
estimated66 that the computer science departments
India’s path to capitalizing on the AI opportunity,
of the top 20 Indian engineering universities produce
both globally and domestically, is not without
approximately 500–600 PhDs per year, and of these,
challenges. A good way to understand these challenges
only about 100 PhDs are in the AI/ML field. However,
is to look at the fundamental building blocks of AI
several leading academic institutions are taking
software businesses:
measures to increase the strength and quality of AI
1. Research: Research is at the core of designing researchers in India, and much of this may be done in
advanced algorithms and AI models that can push partnership with government and industry.
the boundaries of what is commercially viable and
create a competitive edge. Moreover, research also
improves the viability of AI solutions with respect
to their accuracy, cost of training, and the data There is a sore need to fund research
appetite of models. projects in cutting edge AI. A thousand
2. Data: Data is the lifeblood of AI. The continuous such AI projects, challenged with at
and economical availability of usable data is critical least a million dollars each in funding,
for all good AI models.
will surely deliver a vastly greater and
3. Capital/Infrastructure: AI software businesses
longer-lasting impact for those at the
are more capital intensive than regular software
businesses due to the cost of collection and
bottom of the pyramid than the same
processing of large datasets, model development, billion dollars handed out as a subsidy.
and continuous training. As a result, the risk capital —RAVI GURURAJ, Founder and CEO, QikPod
required to properly fund AI software start-ups
must be both large and patient.
4. Talent: Creating AI software solutions requires
more than just coding skills. It requires skillful
application of data science and mathematics to India could have done better as an IT
create, iterate, and improve AI models. industry powerhouse to establish its
Today, India faces systemic challenges across these leadership in AI; we now need to catch
four building blocks. According to the Global AI up with heavy government/industry
Talent Report 2020,64 India ranks ninth, globally, with investments in basic AI research, and
1,008 AI authors on the open access research-sharing
the creation of open datasets.
platform, arXiv, compared to 26,818 from the United
States and 6,401 from China. Also, according to the —MANISH GUPTA, Director, Google Research
India
same 2020 report,65 out of 4,149 self-reported AI
researchers whose contact information was available

57
FIGURE 5.1 Top 10 Countries: Number of AI Authors on arXiv
57,654 total authors as of 2019
53% average growth since 2007

US

China

UK

France
Country

Germany

Canada

Japan

Australia

India

Italy

0K 2K 4K 6K 8K 10K 12K 14K 16K 18K 20K 22K 24K 26K 28K

Authors on arXiv

India does not fare too well with regard to data

availability. Large public data sets are usually the
There is encouraging mindshare
domain of government, and a great deal of public data
are still offline in unconnected “silos.” For the data that
on AI development from the
are online, a coherent effort is required to aggregate government today, and it only bodes
the data—for example, across state governments— well for the industry going forward.
with a plan to make data accessible to AI practitioners Data availability is a better long-term
in a regulated fashion, with the requisite safeguards
incentive than cash to help AI
to ensure data privacy. For example, while India has
start-ups succeed.
adopted an electronic health record (EHR) policy, the
sharing of data among hospitals is still at an early stage —PRAVIN RAO, President and Board Member,
since different hospital chains have interpreted records’ Infosys Ltd

digitization differently.

The availability of infrastructure and capital

requirements for AI start-ups go hand in hand. The
For AI start-ups to flourish, the
computing costs required to train world-class AI
government should make more data
models are extremely high—some estimates predict
public, and public data more accessible. that it took approximately $10–12 million to run each
—A WELL-KNOWN INDIAN AI epoch for GPT-3 (in ML, an epoch is one complete
ENTREPRENEUR, identity protected pass of the training data through the neural network).
Experts believe that a well-trained model such as GPT-

58
3 will require between 10 and 100 epochs—which Indian skill testing company, revealed that 80 percent
is, at minimum, an infrastructure bill of over $l00 of Indian engineers are not trained for a job in the
million. This is a significant cost for start-ups, as well knowledge economy, and only 2.5 percent possess the
as for larger companies. Also, most of India’s high- AI skills that industry requires.
performance computing (HPC) infrastructure is not Although this is a challenge, it also an opportunity
built for the AI workload. In a list of the world’s top across multiple areas. We believe that the growing
500 supercomputers, which was released in November demand for AI/data professionals across the spectrum
2020, only three were in India: Param Siddhi, Mihir, will create opportunities for new education companies.
and Pratyush. Of these, the latter two are purpose- In the last few years, we have seen significant growth
built for the specific, non-AI tasks of meteorology in the companies teaching coding to children. We
and weather forecasting. Moreover, most of India’s believe that many more Indian companies will be
supercomputers are only available to a few premium created to target AI education, and some will also
educational institutions and government agencies and, target upskilling/reskilling of workers displaced by
hence, unavailable to businesses, including start-ups. automation. There will be second-order opportunity
On the talent front, India fares reasonably well, with effects as well—for example, lending for tech-enabled
most estimates of AI talent supply placing India second education aimed at people who want to develop the
only to the United States. However, it is possible that skills required in the AI age. Similarly, we believe that
China may have a bigger talent pool, as most reports on there will be targeted recruitment portals/companies
the topic use LinkedIn for their research, and LinkedIn catering to the demand and supply pools for AI skills.
is not widely used in China. The Global AI Talent
Report divides AI roles67 (barring research) into the 5.2 Policy—Learnings From the World
following categories: data engineering and architecture, The challenges mentioned in the previous section
AI/ML engineering, and AI/data productization. The are not unique to India. To varying degrees, every
report estimates that of over 470,000 professionals country in the world faces similar challenges.
in these roles, globally, India has 75,900, placing the While private enterprises are expected to step in
country second only to the United States, which has and capture the value from AI adoption, they need
186,700 professionals. China’s numbers are not reliable policymakers to remove or blunt the challenges under
because of the underreporting issue mentioned above. each of the four pillars of data, talent, research,
On their own, these numbers paint a good picture. and capital/infrastructure. Policymakers must
However, India’s higher education curriculum has ensure that: the requisite supporting infrastructure/
not traditionally encouraged the mingling of arts and capital is in place for promising businesses; data are
science subjects, which means that engineers do not readily and affordably available; training capacity
learn about the softer aspects of decision making, and is expanded to ensure a ready supply of quality AI
arts students typically do not take courses focused talent; and there is a flourishing research ecosystem
on computational skills. To put this in perspective, for developing local expertise and transferring
a 2019 employability survey by Aspiring Minds, an technology from research to industry.

To leapfrog into the future, we will have to re-skill every engineer to be an

ML engineer and every employee to be an AI-native. The best thing that the
government can do is train the future generation of kids in machine learning, the
way they are trained in math or science, making it part of basic literacy.
—ASHWINI ASOKAN, Founder & CEO, Mad Street Den

59
FIGURE 5.2 AI Policy Initiatives by Frontrunners

China UK Singapore US

Govt. selected ‘AI national 16 new centers for

Eight strategic priorities
Champions’, e.g., Baidu doctoral studies, Expedited granting
areas for federal R&D
for autonomous driving; graduating 1,000 PhDs of patents for AI
Research investment; NSF invests
and Tencent for computer in 5 years, and govt. skills inventions to as fast
$0.5 billion/year on
vision and medical and talent package such as six months
fundamental research
diagnosis as Turing AI Fellowships

Department of
Local cities sponsoring ‘Industrial Strategy
‘100 Experiments’ Medicine, Department
AI funds for integrating Challenge Fund’ allocates
program supports of Defence, Food & Drug
AI solutions in city £725 million to tackle
companies in Administration, and
Capital infrastructure, e.g., City grand AI challenges;
deploying AI in National Institutes of
of Tianjin to establish British Business Bank
a co-investment Health develop own AI
$16 billion fund for AI allocates £2.5 billion
model research and investment
development ‘patient capital’
policy

400 new university Increasing the number of AI Apprenticeship 2018, five-year strategy
majors related to AI Tier 1 (Exceptional Talent) Program, TechSkills for STEM with AI focus;
and big data; lucrative Visa up to 2000/year; Accelerator, and agency fellowships; and
Talent
research grants to bring and industry sponsors for joint curriculum training funding from
back foreign PhD scholars new AI master’s degrees with IBM to help DARPA, NASA, DOE, NIH,
of Chinese origin in UK universities citizens USDA

Three pilot projects ‘Trusted data City-level regulations on

Public data made available under the ‘data-trust’ sharing framework’ the use of citizens’ data,
to businesses that partner framework for safe, and ‘model AI and defining privacy
Data with the government secure, and equitable governance statutes; Big Data to
in facial recognition, data transfer; and framework’ for data Knowledge (BD2K) to
healthcare, etc. making high quality partnerships and AI support data accessibility
public data available deployment in biomedical research

Sources: Secondary research, news

Across the globe, some countries are already reaping those in the United States.68 China’s researchers have
the benefits of coherent design and implementation gained critical insights from their exchanges with
of AI-enabling policies. Among these countries, foreign universities, and they have brought back
China has been the most notable. With a clear and knowledge to share with other academics in China.
ambitious target from the central government, and This initiative, which started in the early 2000s, has
its development of a “new generation AI plan 2030” now begun to yield results. As shown in a study69 of
to attain global leadership in AI, local governments, academic AI research papers published up to the end of
departments, and universities in China have taken 2018, by 2025 China should produce the top 1 percent
further initiatives to ensure effective implementation. of the most-cited AI research papers. China has created
Leading Chinese universities such as Tsinghua several generous research grants to bring back AI talent
University have been forging AI research collaborations that has migrated to other countries. There has also
with leading foreign universities, and especially with been a tremendous push to develop AI education in

60
China. Four hundred new university majors in the field • Privacy and security concerns, including the lack
of AI and big data were launched in 2019.70 of formal regulations about the anonymization of
City administrations in China are also offering funds data; and
for the development of private AI solutions that can be • The absence of a collaborative approach to the
integrated into the city’s administrative infrastructure. adoption and application of AI.
For example, the City of Tianjin has pledged to establish NITI Aayog has also highlighted the long-term
a $16 billion fund for AI, alone.71 Meanwhile, the challenges concerning privacy, security, and
central government has created a specific program for collaboration that affect the sustainability of the AI
“AI national champions,” which are companies playing technology ecosystem in India.
a key role in developing cutting-edge AI technology that
Second, on the research front, NITI Aayog proposes
can solve local problems, but also be adopted globally
setting up organizations that draw inspiration from
and establish China’s leadership in AI development. A
global best practices:
few prominent “AI national champions”72 are Baidu
(for autonomous driving) and Tencent (for computer • The Centre of Research Excellence (CORE) in
vision and medical diagnosis). Besides regulatory AI, which would focus on developing a better
support, these companies get preferential access to large understanding of the current research on AI. This
government-owned data sets, which they can use to would also emphasize the development of infrastructure
create increasingly better AI models. tools for the application of basic research.
Among the other countries that have done • The International Centre of Transformational
exceptionally well in providing support to AI AI (ICTAI), which would focus on providing
businesses are the United Kingdom and Canada. the ecosystem for application-based technology
Both have focused on attracting top research talent development and deployment. This would be an
through a mix of research grants, tax incentives, industry-led initiative that would take up high-level
and favorable visa policies. AI challenges, or inter-ministerial projects focusing
on AI-based solutions. The initiative would also
5.3 India’s AI Policy Approach—Laying a focus on the commercialization of early-stage AI
Robust Foundation technologies and prototypes.
Despite the challenges, it is encouraging to see that In addition to NITI Aayog, prominent industry
India’s concerted push toward AI aligns with the rest bodies such as NASSCOM and iSPIRT have their
of the world’s timeline. Most countries announced own recommendations, and have contributed their
dedicated AI policies between 2015 and 2020. In 2018, expertise in laying the groundwork for AI-capability
NITI Aayog (India’s policy think tank) came up with development in India.
a comprehensive policy report on AI.73 This report is Some of NASSCOM’s key recommendations include:
a first step in acknowledging India’s challenges and
• Creating a nodal agency within government
suggesting several policy measures to address them.
for development and diffusion of AI in order to
First, among the key challenges that NITI Aayog nudge government departments to develop their
has identified facing the development of India’s AI capabilities to adopt AI;
capabilities are:
• Creating a collaborative framework for engagement
• The lack of broad-based expertise in research and between government, industry, and academia
the applications of AI; (public-private partnerships—PPPs) to support start-
• The absence of enabling data ecosystems, and lack ups and promote innovative localized solutions;
of access to intelligent data; • Developing an all-encompassing data strategy,
• The high cost of resources and low awareness of the integrating public data, and evaluating alternative
benefits of AI; data-sharing models;

61
• Addressing skill gaps, revising curricula, creating a by making the development of future-ready talent,
marketplace for skilled AI professionals, attracting and especially talent in the field of AI, a key feature
skilled foreign talent, and including private players of its recent revision of the National Education
in STEM (science, technology, engineering and Policy (NEP). However, the results of such changes,
mathematics) training for citizens; and especially when it comes to developing human capital,
• Addressing governance challenges in AI, adopting will take time, and the NEP’s benefits are only likely
elements of algorithmic impact assessment, and after 2025. The groundwork for creating the national
recommending AI specifications for explainability, digital health blueprint, which will create a national
robustness, and safety. public database of citizens’ health records, is also
underway. This should be a boon for healthcare
In addition to these, NASSCOM has undertaken
companies that aim to use machine learning to
several other initiatives. For example, NASSCOM has
understand and identify disease patterns and early
partnered with the Government of Karnataka to set
up a center of excellence for data science and AI (CoE- signals, and enhance drug efficacy.
DSAI), based on a PPP model. The goal is to catalyze Similarly, several local governments and
innovation by providing technology and infrastructure administrations have adopted AI-based solutions to
know-how to innovators and start-ups, in partnership solve local needs. The traffic police of Goa is evaluating
with private players such as IBM, Intel, Nvidia, a computer vision-based street monitoring system, and
Microsoft, and Digital Ocean. Chennai police have experimented with a somewhat
Similarly, iSPIRT has made recommendations about controversial facial recognition technology system to
data availability and privacy, including a framework identify public offenders.
for “data laws” with principles based on the merits Several efforts of public and private institutions and
of data democracy, the creation of a level playing government bodies are praiseworthy for their focus on
field for Indian companies, and a design to enable building long-term AI capacity. Foremost among these
compliance and convenience. is the Indian Institute of Technology Delhi’s School of
Implementation of these policy recommendations Artificial Intelligence. Started in 2021, this institution
is still a “work in progress.” The government has offers PhDs in the field of ML/ AI, and is rolling out
made commendable efforts to bridge the talent gap postgraduate courses in 2022.74

62
APPENDIX

The Al4Biz Challenge

As part of our larger goal to understand the external funding. Two-thirds of the companies,
opportunities for Indian AI SaaS start-ups, we which comprised a broad range of industries and
conducted a competition to identify the best early- applications, had annualized recurring revenue (ARR)
stage AI SaaS start-ups in the country. We undertook of less than $250,000. In reviewing the applications,
this competition, which began in September 2020, we were positively surprised to see some high-quality
in collaboration with Infosys, Cisco, AWS, Github, MLOps and AI platform companies.
and Freshworks, and we announced the winners in In total, 20 short-listed start-ups pitched to a jury
December 2020. comprised of seasoned investors, academics, start-up
We received a total of 108 applications from a range founders, and industry leaders from the United States
of start-ups that, by design, were mostly early- and India. Each start-up had 30 minutes to present
stage. More than 60 percent of the companies had and field questions from the jury, and then the top 10
either raised no capital or less than $250,000 in finalists were chosen.

FIGURE A.1 108 Diverse, High Quality Applications

Entries by Cumulative Funding

37 24 27 20

Not raised so far Less than $250K $250K–$1M $1M–$5M

Entries by Current ARR

16 26 32 23 11

Pre-revenue Less than $50K $50K–$250K $250K–$1M >$1M

Entries by Target Industry

33 20 14 12 7 22

Horizontal AI Tools Retail Financial Services Healthcare Others

64
A Highly Qualified Jury Supported the PRASAD JOSHI, Senior Vice President,
Selection of Competitors Emerging Technology Solutions, Infosys

ALOK GOYAL, Partner, Stellaris Venture

RAHUL GARG, Professor of Computer
Partners
Science, IIT Delhi

C.N RAGHUPATHI, Sr. Vice-President &

RAHUL PANGAM, Vice President,
Head of the India Business, Infosys
Paypal

DILIP KHANDELWAL, Global CIO and

RUCHIRA SHUKLA, Regional Lead, South Asia,
India MD, Deutsche Bank
Disruptive Technologies—Direct Equity & VC
Funds, Global HealthTech Sector Lead, IFC
KRISH MANTRIPRAGADA, Chief Product
VISHAL GOENKA, Founder, CEO of Stealth
Officer, Seismic
Mode AI Startup

LANA GRAF, Principal Industry Specialist,

Each of the finalists was eligible for generous support,
Artificial Intelligence/Machine Learning, IFC
which was made possible through the industry leaders
who supported us in this initiative. It is worth pointing
MANEESH SHARMA, Country Manager,
out that Stellaris made an investment in one of the
India, Github
10 winners in the final cohort. Also, IFC has chosen
to pilot the AI-powered human resources assistant,
PANKAJ MITRA, Director & India Head, Amber, which is offered by inFeedo, another start-up
Cisco Investments & Acquisitions among the top 10 winners.
A brief overview of each of the final 20 start-ups,
several of which have been discussed in this report, is
presented in Figures A.3 and A.4.

FIGURE A.2 Generous Support for Cohort Members from Industry Leaders

Generous credits Mentorship sessions Introductions to

from global experts potential customers
and practitioners

Introduction to business Potential funding

units and accelerators

65
FIGURE A.3 AI4Biz: Top 10 Winners

Company Description

Data intelligence start-up leveraging satellite data+AI to generate environmental

datasets across a range of air, land and water parameters.

Empowers HR leaders to identify gaps in culture and employee experience through

human-like chat bot.

Computer vision platform focused on automation of physical asset inspections (such

as cars) using photos/videos.

Shopify for chat; provides a personalized shopping experience for D2C companies on
chat platforms using L3 conversational AI.

Uses AI on video and voice to enhance resolution on edge devices with low bandwidth;
target use cases include media consumption among rural India and moving vehicles.

Helps pulmonologists reduce misdiagnosis and improve follow-up care through

accurate analysis of chest CT imaging by using AI.

SaaS platform, applies AI in solar power plants to identify causes for

underperformance and suggests work orders to plant crew.

Enables balanced decision-making through techniques that identify and remove

human biases in AI training datasets.75

MLOps platform to help teams collaborate on their machine learning products.

Improves efficiency and decision making in radiology. Offerings include a cloud

marketplace of radiologists for diagnostic centers as well as an AI-driven PACS
radiology workflow solution.

66
FIGURE A.4 AI4Biz: Additional Finalists

Company Description

Leverages insights from geospatial data to help outdoor service professionals—

lawn care, pest control, landscaping, etc.—better manage their businesses.

Provides contextual intelligence for target accounts for salespeople and integrates
deeply with underlying CRM systems.

Leverages AI to create the equivalent of a credit score for talent. Automates and
crowdsources candidates’ technical skills (acquired by InfoEdge).

Analytics platform that enables B2B marketers to make better decisions through
goal-driven, automatically ‘surfaced’ insights from data.

Demand forecasting solution for supply chains in the CPG/Retail sector; Reduces
waste and stockouts.

Computer vision and NLU solution that understands, analyzes, and auto-grades
free-form student responses (digital and handwritten).

Enterprise data labeling platform for data science teams developing

production-level AI.

Operationalizes AI for factories, enabling them to improve equipment reliability

and process efficiency.

By automating document processing with AI, helps mortgage lenders and servicers
reduce the time and cost to close mortgages.

AI-powered video analytics platform that helps businesses monitor internal and
external process compliance.

67
ENDNOTES
1 For the purpose of this report, a start-up is defined as any company or project that is created by an entrepreneur to develop
and validate a business model. Please note that there are several definitions for this term. The definition used by the Indian
government is available at the following link: https://www.startupindia.gov.in/content/sih/en/startup-scheme.html
2 It is worth noting that the government of India has already implemented several policies with a view to developing an Indian AI
ecosystem, e.g., the India Semiconductor Mission and the Hyperscale Data Centre Policy.
3 As of January 15, 2022.

4 In the financial world, a unicorn is a privately owned start-up with a valuation greater than $1 billion.

5 Throughout this document, wording such as “AI-led applications,” “AI applications,” “AI software,” and “AI SaaS” have been
used interchangeably.
6 Inside sales is the process of selling an organization’s products or services through digital channels such as phone and email.

7 Bostrom, Nick. 2017. “Superintelligence: Paths, Dangers, Strategies.” London, UK: Oxford University Press.

8 Copeland, B. J. “Artificial Intelligence.” In The Encyclopedia Britannica. August 11, 2020. https://www.britannica.com/
technology/artificial-intelligence
9 Good, I.J. 1965. “Speculations Concerning the First Ultraintelligent Machine.” Advances in Computers, Vol 6 (In Bostrom,
Nick. 2017. “Superintelligence: Paths, Dangers, Strategies.” London, UK: Oxford University Press).
10 Electronic Frontier Foundation. 2017. “AI Progress Measurement.” Electronic Frontier Foundation project. https://www.eff.org/
ai/metrics
11 Electronic Frontier Foundation. 2017.

12 Bostrom, Nick. 2017.

13 Choosing the right model is a critical decision while building AI solutions. For example, a neural network will generally be more
accurate than a linear regression model, especially when there are underlying non-linearities, but it will need more data points
as input. As a result, the model is likely to be slower and possibly more expensive.
14 Canalys. 2021. “Global Cloud Services Spend Hits Record US$49.4 billion in Q3 2021.” Press Release. Canalys.com. October
28, 2021. https://www.canalys.com/newsroom/global-cloud-services-q3-2021
15 Edge cases are highly unlikely events that occur under extreme operating conditions.

16 Rajaraman, V. 2012. “History of Computing in India: 1925–2010.” Bangalore: Supercomputer Education and Research Centre,
Indian Institute of Science. https://ethw.org/File:Rajaraman,_V._History_of_Computing_in_India,_1955-2010.pdf
17 NASSCOM (National Association of Software and Services Companies). 2018. “NASSCOM Artificial Intelligence Primer 2018.”
https://nasscom.in/knowledge-center/publications/nasscom-artificial-intelligence-primer-2018
18 Ghemawat, P. and S.A. Altman. 2007. “The Indian IT Services Industry in 2007.” Industry Case Study. August 2007.
https://www.ghemawat.com/wordpress/wp-content/uploads/2016/l0/Indian-lT-lndustry.pdf
19 NASSCOM and BCG (National Association of Software and Services Companies and Boston Consulting Group). 2007.
“Innovation Report 2007: Unleashing the Innovative Power of Indian IT-ITeS Industry.” p. 54. Available from NASSCOM.
20 Capability maturity model (CMM) is a methodology used to refine software development methodologies within organizations.
The model classifies organizations across five levels of maturity, with level 5 being the highest. A distinguishing feature of
CMM level 5 systems is their continuous improvement.
21 Dataquest. 2003. “Why ‘India Inside’ Spells Quality.” Dataquest, October 27, 2003.

22 Ghemawat, P. and S. A. Altman 2007.

23 ET Bureau. 2020. “TCS Surpasses Accenture in M-Cap to Become the World’s Most Valuable IT Company.” The Economic
Times, October 8, 2020. https://economictimes.indiatimes.corn/markets/sloeks/news/tcs-surpasses-accenlure-in-market-cap/
articIeshow/78558794.ems
24 Zoho.com. 2020. About us. Zoho Corporation website. https://www.zoho.com/aboutus.html

25 NASSCOM (National Association of Software and Services Companies). 2020a. “Indian SaaS Landscape Growing at ~1.5x
Faster Than Global SaaS: NASSCOM Report in Partnership with SaaSBOOMi and Zinnov.” NASSCOM Press Release. July
23, 2020. http://Press-Release-NASSCOM-SaaS-Report-2020-23-Jul-2020.pdf
26 NASSCOM (National Association of Software and Services Companies) and SaaSBOOMi. 2020. “Riding the Storm: Towards
the Giant India SaaS Opportunity.” July 2020. https://nasscom.in/knowledge-center/publications/nasscom-riding-storm-
towards-giant-india-saas-opportunity
27 NASSCOM (National Association of Software and Services Companies) and SaaSBOOMi. 2020.

28 Sheth, Arpan, et al. 2021. “India SaaS Report 2021.” Bain & Company. https://www.bain.com/insights/india-saas-report-2021/

69
29 Business Wire. 2018. “Global Developer Population and Demographic Report 2018: Developer Population Growth Shifts
Toward China, India and Emerging Countries.” Businesswire.com. October 9, 2018. https://www.businesswire.com/news/
home/20181009005840/en/
30 GitHub. 2020. “The 2020 State of the Octoverse: Empowering Healthy Communities.” https://octoverse.github.com/static/
github-octoverse-2020-community-report.pdf
31 Larochelle, Hugo, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, Hsuan-Tien Lin, editors. 2020. “Advances in
Neural Information Processing Systems 33.” Annual Conference on Neural Information Processing Systems (NeurIPS 2020),
December 6–12, 2020, Virtual Conference. https://proceedings.neurips.cc/paper/2020
32 Canalys. 2020. “Global Cloud Infrastructure Market Q3 2020.” Canalys Newsroom. October 29, 2020. https://www.canalys.
com/newsroom/worldwide-cloud-market-q320; Canalys. 2021. “Global Cloud Services Spend Hits Record US$49.4 billion in
Q3 2021.” Press Release. Canalys.com. October 28, 2021. https://www.canalys.com/newsroom/global-cloud-services-q3-2021
33 Canalys. 2021.

34 The waterfall model of software development involves breaking down a project into sequential phases, where every phase
requires completion of the immediately preceding phase before work on that phase can begin.
35 The Agile approach to software development involves continuous iteration of both development and testing throughout a
project’s lifecycle. Unlike Waterfall, the Agile method involves simultaneous development and testing.
36 Algorithmia. 2019. “2020 State of Enterprise Machine Learning.” Algorithmia.com. https://www.coriniumintelligence.
com/2020-state-of-enterprise-machine-learning-algorithmia-whitepaper-download
37 Algorithmia. 2019.

38 Cognilytica. 2020. “ML Model Management and Operations 2020 (MLOps).” Abstract. March 3. 2020. https://www.
cognilytica.com/document/ml-model-management-and-operations-2020-mlops/
39 Please note that since this work was originally done, Stellaris Venture Partners has done significant work in this space. A more
current and detailed view on this topic can be accessed at https://www.stellarisvp.com/mlops-2021/ from Anagh Prasad’s
article, MLOps – Opportunities in powering 2020s’ AI-Revolution.
40 Rajan Priya. 2020. “The Indian SaaS Landscape.” Silicon Valley Bank Blog. July 10, 2020. https://www.svb.com/blogs/priya-
rajan/the-indian-saas-landscape
41 Sheth, Arpan, et al. 2021.

42 CBR Staff Writer. 2010. “SaaS Revenue with EAS Market to Reach $9.2bn in 2010: Gartner.” TECHMONITOR. December 4,
2020. https://techmonitor.ai/technology/software/saas-revenue-within-eas-market-to-reach-92bn-in-2010-gartner-151210
43 Gartner. 2020b. “Gartner Forecasts Worldwide Public Cloud Revenue to Grow 6.3% in 2020.” Press Release. Gartner.com.
July 23, 2020. https://www.gartner.com/en/newsroom/press-releases/2020-07-23-gartner-forecasts-worldwide-public-cloud-
revenue-to-grow-6point3-percent-in-2020
44 Gartner. 2019. “Gartner Forecasts Worldwide Public Cloud Revenue to Grow 17% in 2020.” Press Release. Gartner.com.
November 13, 2019. https://www.gartner.com/en/newsroom/press-releases/2019-11-13-gartner-forecasts-worldwide-public-
cloud-revenue-to-grow-17-percent-in-2020
45 Dunay, Neil. 2020. “Artificial Intelligence Software Market Forecasts 2020 Report.” PowerPoint Presentation. June 29,
2020. Omedia.com. https://omdia.tech.informa.com/-/media/tech/omdia/brochures/ai/artificial-intelligence-software-market-
forecasts-report-2020.aspx
46 Mines, Chris. 2019. “The Next Decade of Software Spending: A Paradox.” Forrester.com Blog. September 5, 2019. https://
go.forrester.com/blogs/the-next-decade-of-software-spending-a-paradox/
47 NASSCOM (National Association of Software and Services Companies) and SaaSBOOMi. 2020. “Riding the Storm: Towards
the Giant India SaaS Opportunity.” July 2020. https://nasscom.in/knowledge-center/publications/nasscom-riding-storm-
towards-giant-india-saas-opportunity
48 NASSCOM (National Association of Software and Services Companies) and SaaSBOOMi. 2020.

49 Of late, there are signs that valuation multiples in public markets may have undergone a more permanent reset. We have
continued with a conservative EV/ARR multiple of 10.
50 Wiggers, Kyle. 2020. “Waymo Is Using AI to Simulate Autonomous Vehicle Camera Data.” Venture Beat. May 20, 2020.
https://venturebeat.com/2020/05/20/waymo-is-using-ai-to-simulate-autonomous-vehicle-camera-data/
51 IDC (International Data Corporation) 2020. “IDC Forecasts Strong 12.3% Growth for AT Market in 2020 Amidst Challenging
Circumstances.” Press Release. IDC.com. August 4, 2020. https://www.idc.com/getdoc.jsp?containerId=prUS46757920
52 Gartner. 2020a. “Gartner Says Global IT Spending to Reach $3.9 Trillion in 2020.” Press Release. January 15, 2020.
gartner.com https://www.gartner.com/en/newsroom/press-releases/2020-01-15-gartner-says-global-it-spending-to-reach-
3point9-trillion-in-2020
53 Gartner Research. 2019. “Forecast Analysis: Business Process Outsourcing, Worldwide.” https://www.gartner.com/en/
documents/3956441/forecast-analysis-business-process-outsourcing-worldwide
54 IBEF (Indian Brand Equity Foundation). 2019. “Indian IT and ITeS.” IBEF Presentation. July 2019. https://www.ibef.org/
download/IT-ITeS-July-2019.pdf

70
55 IBEF (Indian Brand Equity Foundation). 2019.
56 The average EV/Revenue multiple for the top four Indian IT services companies (TCS, Infosys, Wipro, and HCL Tech). Each
company’s multiple was taken as the five-year average of their public market reports from 2016 to 2020.
57 The report provides an overall global opportunity and is not an operational guide to building an AI SaaS company. As such,
the report does not go into detail to map out the opportunities in different geographies such as Europe, Asia, North America,
or Latin America.
58 Based on revenue per employee for existing Indian 1T services and SaaS companies.

59 Vyas Mahesh. 2020. “Job Losses in White and Blue Collar Workers.” Centre for Monitoring Indian Economy. https://www.
cmie.com/kommon/bin/sr.php?kall=warticle&dt=2020-09-14%2021%3A47%3A53&msec=416
60 Moharkan, Furquan. 2019. “India First Country to Use AI/ML in Tax Assessment.” Deccan Herald, September 2, 2019.
https://www.deccanherald.com/business/business-news/india-first-country-to-use-ai/ml-in-tax-assessment-758580.html
61 NASSCOM (National Association of Software and Services Companies). 2020b. “Unlocking Value From Data And AI– The
India Opportunity.” Presentation. August 2020. https://nasscom.in/knowledge-center/publications/unlocking-value-data-and-
ai-india-opportunity
62 FAO (Food and Agriculture Organization of the United Nations). 2020. FAO Statistics. http://www.fao.org/faostat/
en/#compare.
63 Goel, Samiksha. 2020. “The Doctor-Population Ratio in India 1:1456 against WHO Recommendations.” Deccan Herald.
January 31, 2020. https://www.deccanherald.com/business/budget-2020/the-doctor-population-ratio-in-india-is-11456-against-
who-recommendation-800034.htm
64 Gane, J.F. 2020. “5 Highlights from the 2020 Global AI Talent Report.” jfgane Blog. https://jfgagne.ai/blog/5-highlights-from-
the-2020-global-ai-talent-report/
65 Gane, J.F. 2020.

66 itihassa Research and Digital. 2018. “Landscape of Artificial Intelligence/Machine Learning Research India.” itihassa.com.
https://www.itihaasa.com/public/pdf/LandscapeofAI-MLResearch.pdf
67 Gane, J.F. 2020.

68 O’Meara, Sarah. 2019. “AI Researchers in China Want to Keep the Global-Sharing Culture Alive.” Nature. May 29, 2019.
https://www.nature.com/articles/d41586-0l9-0l681-x
69 Schoenick, Carissa. 2019. “China May Overtake US in AI Research.” AI2 Blog. Allen Institute for AI Research. https://
medium.com/ai2-blog/china-to-overtake-us-in-ai-research-8b6blfe30595
70 Xi, Chen. 2019. “China to Open 400 Big Data, AI Majors in Universities for Global Competition.” Global Times. February 27,
2019. http://en.people.cn/n3/2019/0227/c202936-9550508.html
71 Reuters Staff. 2018. “China’s City of Tianjin to Set Up $16 Billion Artificial Intelligence Fund.” Reuters Technology News.
May 16, 2018. https://www.reuters.com/article/us-china-ai-tianjin-idUSKCN1II0DD
72 Zhang, Jane. 2020. “China’s AI Champions Are Powering an Array of Processes That Will Rise in 2020.” Tech in Asia.
January 1, 2020. https://www.techinasia.com/chinas-ai-champions-powering-processes
73 NITI Aayog. 2018 “National Strategy for Artificial Intelligence #AIforall.” NITI Aayog Discussion Paper. June 2018.
https://niti.gov.in/writereaddata/files/document_publication/NationalStrategy-for-AI-Discussion-Paper.pdf
74 Education Desk. 2020. “IIT-Delhi Sets up School of Artificial Intelligence, to Offer Degree, Professional Courses.” Indian
Express. September 2, 2020. https://indianexpress.com/article/education/iit-delhi-to-set-up-school-of-artificial-intelligence-to-
offer-degree-professional-courses
75 Received strong support from the jury given the high social relevance and complexity of the problem being solved; however, the
company failed to find product market fit.

71
FURTHER READING
Additional reports about investing in challenging markets and the role of
technology in emerging markets, as well as a list of EM Compass Notes
published by IFC Thought Leadership. Visit http://ifc.org/thoughtleadership.

Artificial Intelligence
in Emerging Markets:
Opportunities, Trends, and
Emerging Business Models
March 2021, Second and Expanded
Edition—148 pages

Artificial intelligence (AI)—the science

of making machines act in rational,
intelligent ways—is rapidly making inroads
into business operations and society. AI
is already being applied in many areas
of our lives, with high penetration in
financial services followed by e-commerce,
healthcare, post-secondary learning,
agriculture, and manufacturing.
Emerging markets can benefit significantly
from AI: Its applications are providing new
ways to leapfrog infrastructure gaps and
solve pressing development challenges in
critical sectors.
This report explores the latest AI
applications and trends in emerging
markets and includes several examples
of how AI is expanding opportunities and
contributing to the achievement of the
Sustainable Development Goals. It also
sheds light on how investors, clients, and
governments can harness its full potential
while minimizing its risks—when managed
effectively and with safeguards in place,
AI can facilitate private investment to
reduce poverty and improve lives at a pace
inconceivable only a decade ago.

73
Reinventing Business Through Disruptive
Technologies: Sector Trends and
Investment Opportunities for Firms in
Emerging Markets
March 2019—108 pages

Technology disrupts and transforms. And disruptive technologies

are critical to achieving the Sustainable Development Goals, many
of which can be advanced and accelerated through technological
innovations.
For a comprehensive examination of the ways these innovations
alter private sector business models in emerging markets, IFC
conducted a tour of the technology horizon in eight selected
sectors—power, transport, water and sanitation, digital
infrastructure, manufacturing, agribusiness, education, and financial
services—and six selected themes, from gender and climate-smart
cities to e-logistics and personal identification, among others.
This report examines each of these sectors and themes in terms of
what true disruption looks like, which technologies are most likely
to have a dramatic impact, and the specific opportunities they offer.
It also identifies the challenges and constraints that will need to
be surmounted if the private sector is to seize these opportunities.
Lastly, it presents how IFC supports companies and investors in
their efforts to enter into or expand in emerging markets.

Blockchain: Opportunities for Private

Enterprises in Emerging Markets
January 2019, Second and Expanded Edition—88 pages

Over the course of two years, IFC worked with key influencers and
experts in the worlds of distributed ledgers and digital finance to
create a series of nine papers examining the potential and perils
of blockchain. An initial report with six chapters was published
October 2017. Since then, three additional in-depth notes have been
added to broaden and deepen our understanding of this burgeoning
technology, its enormous potential, and its many challenges.
These documents collectively examine the general contours and
technology underlying blockchain and its implications for emerging
markets.
Specifically, this report provides an examination of blockchain
implementation in financial services and global supply chains; a
regional analysis of blockchain developments in emerging markets;
a new focus on blockchain’s ability to facilitate low-carbon energy
solutions; and a discussion of the legal and governance issues
associated with the technology’s adoption.

74
REFERENCES
Algorithmia. 2019. “2020 State of Enterprise Machine Learning.” Algorithmia.com. https://www.coriniumintelligence.
com/2020-state-of-enterprise-machine-learning-algorithmia-whitepaper-download
Bostrom, Nick. 2017. “Superintelligence: Paths, Dangers, Strategies.” London, UK: Oxford University Press.
Business Wire. 2018. “Global Developer Population and Demographic Report 2018: Developer Population Growth Shifts
Toward China, India and Emerging Countries.” Businesswire.com. October 9, 2018. https://www.businesswire.com/
news/home/20181009005840/en/
Canalys. 2020. “Global Cloud Infrastructure Market Q3 2020.” Canalys Newsroom. October 29, 2020. https://www.
canalys.com/newsroom/worldwide-cloud-market-q320
CBR Staff Writer. 2010. “SaaS Revenue with EAS Market to Reach $9.2bn in 2010: Gartner.” TECHMONITOR.
December 4, 2020. https://techmonitor.ai/technology/software/saas-revenue-within-eas-market-to-reach-92bn-in-2010-
gartner-151210
Cognilytica. 2020. “ML Model Management and Operations 2020 (MLOps).” Abstract. March 3. 2020. https://www.
cognilytica.com/document/ml-model-management-and-operations-2020-mlops/
Copeland, B. J. “Artificial Intelligence.” In The Encyclopedia Britannica. August 11, 2020. https://www.britannica.com/
technology/artificial-intelligence
Dataquest. 2003. “Why ‘India Inside’ Spells Quality.” Dataquest, October 27, 2003.
Dunay, Neil. 2020. “Artificial Intelligence Software Market Forecasts 2020 Report.” PowerPoint Presentation. June 29,
2020. Omedia.com. https://omdia.tech.informa.com/-/media/tech/omdia/brochures/ai/artificial-intelligence-software-
market-forecasts-report-2020.aspx
Education Desk. 2020. “IIT-Delhi Sets Up School of Artificial Intelligence, to Offer Degree, Professional Courses.”
Indian Express. September 2, 2020. https://indianexpress.com/article/education/iit-delhi-to-set-up-school-of-artificial-
intelligence-to-offer-degree-professional-courses
Electronic Frontier Foundation. 2017. “AI Progress Measurement.” Electronic Frontier Foundation project. https://www.
eff.org/ai/metrics#Vision
ET Bureau. 2020. “TCS Surpasses Accenture in M-Cap to Become the World’s Most Valuable IT Company.” The
Economic Times, October 8, 2020. https://economictimes.indiatimes.com/markets/stocks/news/tcs-surpasses-accenture-
in-market-cap/articleshow/78558794.cms
FAO (Food and Agriculture Organization of the United Nations). 2020. FAO Statistics. http://www.fao.org/faostat/
en/#compare
Gane, J.F. 2020. “5 Highlights from the 2020 Global AI Talent Report.” jfgane Blog. https://jfgagne.ai/blog/5-highlights-
from-the-2020-global-ai-talent-report/
Gartner. 2019. “Gartner Forecasts Worldwide Public Cloud Revenue to Grow 17% in 2020.” Press Release. Gartner.
com. November 13, 2019. https://www.gartner.com/en/newsroom/press-releases/2019-11-13-gartner-forecasts-
worldwide-public-cloud-revenue-to-grow-17-percent-in-2020
Gartner. 2020a. “Gartner Says Global IT Spending to Reach $3.9 Trillion in 2020.” Press Release. January 15, 2020.
gartner.com https://www.gartner.com/en/newsroom/press-releases/2020-01-15-gartner-says-global-it-spending-to-
reach-3point9-trillion-in-2020
Gartner. 2020b. “Gartner Forecasts Worldwide Public Cloud Revenue to Grow 6.3% in 2020.” Press Release. Gartner.
com. July 23, 2020. https://www.gartner.com/en/newsroom/press-releases/2020-07-23-gartner-forecasts-worldwide-
public-cloud-revenue-to-grow-6point3-percent-in-2020
Gartner Research. 2019. “Forecast Analysis: Business Process Outsourcing, Worldwide.” https://www.gartner.com/en/
documents/3956441/forecast-analysis-business-process-outsourcing-worldwide
Ghemawat, P. and S. A. Altman. 2007. “The Indian IT Services Industry in 2007.” Industry Case Study. August 2007.
https://www.ghemawat.com/wordpress/wp-content/uploads/2016/l0/Indian-lT-lndustry.pdf
GitHub. 2020. “The 2020 State of the Octoverse: Empowering Healthy Communities.” https://octoverse.github.com/
static/github-octoverse-2020-community-report.pdf
Goel, Samiksha. 2020. “The Doctor-Population Ratio in India Is 1:1456 against WHO Recommendations.” Deccan
Herald. January 31, 2020. https://www.deccanherald.com/business/budget-2020/the-doctor-population-ratio-in-india-
is-11456-against-who-recommendation-800034.html

75
Good, I.J. 1965. “Speculations Concerning the First Ultraintelligent Machine.” Advances in Computers, Vol 6 (In
Bostrom, Nick. 2017. Superintelligence: Paths, Dangers, Strategies. London, UK: Oxford University Press).
IBEF (Indian Brand Equity Foundation). 2019. “Indian IT and ITeS.” IBEF Presentation. July 2019. https://www.ibef.
org/download/IT-ITeS-July-2019.pdf
IDC (International Data Corporation) 2020. “IDC Forecasts Strong 12.3% Growth for AT Market in 2020
Amidst Challenging Circumstances.” Press Release. IDC.com. August 4, 2020. https://www.idc.com/getdoc.
jsp?containerId=prUS46757920
itihassa Research and Digital. 2018. “Landscape of Artificial Intelligence/Machine Learning Research India.” itihassa.
com. https://www.itihaasa.com/public/pdf/LandscapeofAI-MLResearch.pdf
Larochelle, Hugo, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, Hsuan-Tien Lin, editors, 2020.
“Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems
(NeurIPS 2020), December 6–12, 2020, Virtual Conference. https://proceedings.neurips.cc/paper/2020
Mines, Chris. 2019. “The Next Decade of Software Spending: A Paradox.” Forrester.com Blog. September 5, 2019.
https://go.forrester.com/blogs/the-next-decade-of-software-spending-a-paradox/
Moharkan, Furquan. 2019. “India First Country to Use AI/ML in Tax Assessment.” Deccan Herald, September 2, 2019.
https://www.deccanherald.com/business/business-news/india-first-country-to-use-ai/ml-in-tax-assessment-758580.html
NASSCOM (National Association of Software and Services Companies). 2018. “NASSCOM Artificial Intelligence
Primer 2018.” Key Findings. https://nasscom.in/knowledge-center/publications/nasscom-artificial-intelligence-
primer-2018. Report available from NASSCOM.
NASSCOM (National Association of Software and Services Companies). 2020a. “Indian SaaS Landscape Growing at
~1.5x Faster Than Global SaaS: NASSCOM Report in Partnership with SaaSBOOMi and Zinnov.” NASSCOM Press
Release. July 23, 2020. http://Press-Release-NASSCOM-SaaS-Report-2020-23-Jul-2020.pdf
NASSCOM (National Association of Software and Services Companies). 2020b. “Unlocking Value From Data And AI
– The India Opportunity.” Presentation. August 2020. https://nasscom.in/knowledge-center/publications/unlocking-
value-data-and-ai-india-opportunity
NASSCOM and BCG (National Association of Software and Services Companies and Boston Consulting Group). 2007.
“Innovation Report 2007: Unleashing the Innovative Power of Indian IT-ITeS Industry.” p. 54. Report available from
NASSCOM.
NASSCOM (National Association of Software and Services Companies) and SaaSBOOMi. 2020. “Riding the
Storm: Towards the Giant India SaaS Opportunity.” July 2020. Key Findings. https://nasscom.in/knowledge-center/
publications/nasscom-riding-storm-towards-giant-india-saas-opportunity. Report available from NASSCOM.
NITI Aayog. 2018 “National Strategy for Artificial Intelligence #AIforall.” NITI Aayog Discussion Paper. June 2018.
https://niti.gov.in/writereaddata/files/document_publication/NationalStrategy-for-AI-Discussion-Paper.pdf
O’Meara, Sarah. 2019. “AI Researchers in China Want to Keep the Global-Sharing Culture Alive.” Nature. May 29,
2019. https://www.nature.com/articles/d41586-019-01681-x
Rajan Priya. 2020. “The Indian SaaS Landscape.” Silicon Valley Bank Blog. July 10, 2020. https://www.svb.com/blogs/
priya-rajan/the-indian-saas-landscape
Rajaraman, V. 2012. “History of Computing in India: 1925–2010.” Bangalore: Supercomputer Education and
Research Centre, Indian Institute of Science. https://ethw.org/File:Rajaraman,_V._History_of_Computing_in_
India,_1955-2010.pdf
Reuters Staff. 2018. “China’s City of Tianjin to Set Up $16 Billion Artificial Intelligence Fund.” Reuters Technology
News. May 16, 2018. https://www.reuters.com/article/us-china-ai-tianjin-idUSKCN1II0DD
Schoenick, Carissa. 2019. “China May Overtake US in AI Research.” AI2 Blog. Allen Institute for AI Research. https://
medium.com/ai2-blog/china-to-overtake-us-in-ai-research-8b6b1fe30595
Vyas Mahesh. 2020. “Job Losses in White and Blue Collar Workers.” Centre for Monitoring Indian Economy. https://
www.cmie.com/kommon/bin/sr.php?kall=warticle&dt=2020-09-14%2021%3A47%3A53&msec=416
Wiggers, Kyle. 2020. “Waymo Is Using AI to Simulate Autonomous Vehicle Camera Data.” Venture Beat. May 20, 2020.
https://venturebeat.com/2020/05/20/waymo-is-using-ai-to-simulate-autonomous-vehicle-camera-data/
Xi, Chen. 2019. “China to Open 400 Big Data, AI Majors in Universities for Global Competition.” Global Times.
February 27, 2019. http://en.people.cn/n3/2019/0227/c202936-9550508.html
Zhang, Jane. 2020. “China’s AI Champions Are Powering an Array of Processes That Will Rise in 2020.” Tech in Asia.
January 1, 2020. https://www.techinasia.com/chinas-ai-champions-powering-processes
Zoho.com. 2020. About us. Zoho Corporation website. https://www.zoho.com/aboutus.html

76
IFC Disruptive Technologies and Venture Capital
IFC’s Disruptive Technologies and Funds (CDF) Department focuses on disruptive technologies and works
across the entire entrepreneur ecosystem, including venture capital (VC) direct investing, as well as seed,
VC, and growth equity funds. In 2020, IFC’s VC portfolio exceeded $1.5 billion in total commitments,
with over 100 direct and co-investments, and over 50 investments in VC funds, seed funds, and
accelerators. We invest in innovative business models and technologies that disrupt the conventional sectors
across IFC’s focus areas such as health, education, agribusiness, financial services, and climate. For more
information, please visit: http://ifc.org/vc
IFC
2121 Pennsylvania Avenue, N.W.
Washington, D.C. 20433 U.S.A.

ifc.org/ThoughtLeadership

FEB 2022