1

WHAT IS CB INSIGHTS?

CB Insights helps the world’s leading companies make

smarter technology decisions with data, not opinion.

Our Technology Insights Platform provides companies

with comprehensive data, expert insights and work
management tools to drive growth and improve
operations with technology.

CLICK HERE TO SIGN UP FOR A FREE TRIAL

1
CB Insights helps us compress our
time-to-decision when gathering and
analyzing data and getting an external
view on what's happening in the market
so we can quickly take action.

Meraj Mohammad
Vice President, Ventures Group, ADP
ENTERPRISE AI

We break down no-code AI, stream processing, data

governance, and other top trends in the space.
AI companies raised a record $33B in equity funding in 2020.
As commercial applications of AI scale rapidly, enterprises are
looking to overhaul existing data management practices to
become “AI-first.” Organizations are seeking best practices for
sourcing and storing big data in different formats, deploying AI
models, monitoring their performance, and developing ethical
solutions that are compliant with new regulations.
In this report, we look at trends in enterprise data management
in the context of AI, automation in IT and DevOps, no-code
solutions to enable non-experts to develop AI models, and more.

4
Track the latest
private enterprise
AI companies and
deals using the
CB Insights
AI Collection.

5
Contents
No-code AI platforms take off 7 Analytics vendors increase support for 40
unstructured data types
AIOps: IT and DevOps automation gains 15
Transformers, multilingual models 48
traction
improve enterprise NLP

Graph neural nets find mainstream 23 Data governance and explainable AI 56

enterprise applications

Stream processing: Capturing real-time 31

IoT data for AI applications

6
No-code AI platforms take off

7
What is Low-code and no-code solutions allow users without coding
expertise to build applications. While these have been around
no-code AI? for decades, no-code AI platforms are relatively nascent.
Companies that offer no-code AI solutions allow enterprise
users to build and deploy AI models through a “drag-and-drop”
interface.
No-code AI enables teams without IT or data engineering
experience to integrate machine learning (ML) applications
into enterprise workflows, automate data pre-processing,
reduce time-to-deployment, and narrow the skills gap in
machine learning.

8
 WHAT EXECUTIVES ARE SAYING

“Covid-19 has accelerated the urgent need for every business

to create no-code, low-code apps and workflows in hours or
days, not weeks or months. Power Platform [Microsoft’s AI-
powered BI platform] is already used by more than 3.4M
citizen developers and business decision-makers.”

Satya Nadella, Microsoft CEO,

Q3’20 earnings call

9
The emerging no-code AI ecosystem
Corporates VC-backed companies

CreateML AutoML,
Teachable Machine Stage: Series F Stage: Series B Stage: Series A Stage: Series B Stage: Series A
Total funding: $750M Total funding: $37M Total funding: $16M Total funding: $12M Total funding: $8.5M
Last raised: Dec ‘20 Last raised: Nov ‘20 Last raised: June ’20 Last raised: Dec ’20 Last raised: Dec ‘20

SageMaker Lobe.ai,
Autopilot AI Builder

ShareInsights Ex Machina Stage: Seed Stage: N/A Stage: Seed Stage: Seed Stage: Seed
Total funding: $7M Total funding: $4.6M Total funding: $4.2M Total funding: $3M Total funding: $2.3M
Last raised: Mar ‘20 Last raised: Nov ‘20 Last raised: Dec ‘20 Last raised: June ’20 Last raised: July ’20

Other developer tools

Stage: Seed Stage: Seed Stage: Seed Stage: Series A Stage: Series C
Total funding: $1.7M Total funding: $0.12M Total funding: NA Total funding: 18M Total funding: 37M
Y Combinator Last raised: Jan ‘21 Last raised: Apr ‘17 Last raised: Apr ‘20 Last raised: Apr ‘20 Last raised: Dec ’15
June ‘20
Source: cbinsights.com

10
What big tech companies are doing in no-code AI
Key highlights and product launches from Google, Microsoft, Apple, and Amazon

Google AutoML Vision,

Teachable Machine
For computer vision Microsoft’s Lobe.ai M&A
applications and in-browser ML Lobe.ai is a no-code tool for
experiments image classification
Microsoft AI Builder
Apple CreateML No-code tools for business
Apple introduces no-code model intelligence applications on Amazon SageMaker
development Microsoft's Power Platform
Autopilot
Automatically chooses the right
algorithm and fine-tunes the AI
model

Google AppSheet
Google acquires no-code
solution for app developers

2017 2018 2019 2020

11
C3.ai launches Ex Machina
C3.ai went public with a $4B valuation in
December 2020. Around the same time,
it launched Ex Machina, a no-code
platform with support for data
preparation, model development, and
integration with C3.ai’s business
intelligence suite.

C3.ai’s initial client list spans energy &

utility – where the company has an
established presence – as well as
banking, manufacturing, and
government agencies.

12
No-code computer vision lowers barrier to entry for
app developers

Lobe.ai: Computer vision Superannotate: Image Fritz AI: Augmented Teachable Machine:
tasks including gesture, annotation platform for reality applications and Pose, sound, and image
pose, color, and emotion training AI models ML model development recognition
recognition for Snapchat Lens and
mobile app developers
13
Mastercard, Lux Capital, and others are backing
industry-specific solutions

BANKING MEDIA INSURANCE

Signzy is developing no-code AI for RunwayML is building an AI-based Ushur creates end-to-end automation
customer verification and onboarding photo and video editing toolkit for solutions for customer engagement,
solutions for banks and financial content creators, including options to such as virtual assistant tech and
institutions. These include document create synthetic media using automated email processing. Ushur
and ID verification, as well as risk generating adversarial networks works with companies including
intelligence. (GANs). Cigna, Aetna, HealthSpire, and others.

Select investors: Mastercard, Vertex Select investors: Lux Capital Select investors: Plug and Play
Ventures; previously incubated in the Accelerator, 8VC, Plug and Play
Facebook India Innovation Hub and Ventures, Third Point Ventures, Iron
Google for Startups Accelerator Pillar, Aflac Corporate Ventures

14
AIOps: IT and DevOps automation
gains traction

15
What is AIOps, Enterprise IT infrastructure is becoming more complex with
hybrid cloud technology, on-prem, distributed databases,
or AI for IT containerization, and microservices architecture. As a result,
AIOps — using machine learning to automate IT and DevOps
operations? functions — is gaining traction.
AIOps can help enterprises detect anomalies in traffic based
on historic data, monitor logs to pinpoint the source of
performance issues, monitor applications across multi-cloud
and on-prem environments, and find security vulnerabilities in
code.
Cloud and software services vendors are adding the tech to
their offerings as enterprises face increasing costs incurred
from IT outages.

16
 WHAT EXECUTIVES ARE SAYING

"Our real-time analytics and monitoring offerings are mission-

critical for our customers as they face a growing complex
application environment. To simplify and automate their IT
operations, we recently launched AIOps, leveraging AI
machine learning and automation...”

Chuck Robbins, Cisco CEO,

marking the first-ever mention of AIOps on an earnings call

17
PRIVATE MARKET TRENDS

AIOps companies attracted record funding in 2020

Deals and equity funding ($M), 2016 – 2020
Funding amount ($M) Deal count
$750 35
31
30
$600
23 25
21
$450 20
20
15
15
$300

10
$150
5
$102 $109 $205 $433 $620
$0 0
2016 2017 2018 2019 2020
Note: We include startups providing AI-based IT automation and DevOps solutions. 18
New unicorns: DevOps vendors raise significant funding

Harness uses machine learning to automate CI/CD Snyk is an applications security company that uses AI
(continuous integration and continuous delivery, or to find and fix vulnerabilities in code. It supports open-
the handling of frequent code changes to source, container, and infrastructure-as-a-code
applications). SoulCycle reportedly used Harness to security. In Q3’20, Snyk acquired DeepCode, an ETH
reduce its deployment times from 60 minutes to 10- Zurich spin-off that helps software developers with
15 minutes. real-time analysis of their code as they write it.

Latest round: $115M Series C in Q1’20 at $1.7B Latest round: $175M Series E in Q1’20 at $4.7B
valuation valuation

Select investors: Norwest Venture Partners, Battery Select investors: Google Ventures, Accel, Salesforce
Ventures, Citi Ventures Ventures, Canaan Partners
Note: Unicorns are private companies valued at $1B+. 19
Notable deals in testing and anomaly detection

In Q4’20, Carbon Relay, which automates Kubernetes Logz.io detects anomalies in logs data. By cross-
app deployment, acquired automated performance referencing with crowdsourced data from forums like
testing platform StormForger and rebranded as Stack Overflow and GitHub, it surfaces relevant logs
StormForge to offer AI-based container application associated with a production issue.
testing and performance optimization.
Latest round: $23M Series E in Q4’20
Latest round: $63M Series B in Q1’20
Select investors: Giza Venture Capital, 83North,
Select investors: Foxconn Technology Ventures, OpenView Venture Partners, Vintage Investment
Insight Partners Partners, General Catalyst, next47

20
 ScienceLogic raises $105M from Intel, Goldman Sachs

Total funding: Latest round:

214M $105M Series E
in Q1’21

ScienceLogic develops automated

troubleshooting, hybrid cloud monitoring,
network monitoring, and other IT
automation solutions. The company was
granted 2 US patents in 2020, including
one for a “self-configuring network
management system.”

Select investors: Intel Capital, Goldman

Sachs, Silver Lake, NewView Capital, New
Enterprise Associates

Select clients: HCL, Kellogg, BlackBerry,

Cisco, JPL

21
Corporates are acquiring for AIOps

APPLICATION MONITORING LOG ANALYTICS OPERATIONAL INSIGHTS

IBM acquired Instana in Q4’20 ServiceNow acquired Israel- In its earnings call in Q3’20, HP
to bolster Watson AIOps (a based log analytics startup announced the launch of Aruba
product announced earlier in Loom Systems in Q1’20. Edge Services Platform (Aruba
2020). Instana builds ServiceNow made 4 AI ESP), combining AIOps and
application monitoring tools acquisitions last year, security features to “unify,
with AI-based 3D performance including Sweagle, an AI-based automate, and secure the
visualization and automated configuration management edge.” Additionally, to
notifications for DevOps tool. It also struck a strengthen its AIOps platform
teams. partnership with IBM for AIOps HPE Infosight, HP acquired IT
and IT management. infrastructure monitoring
software CloudPhysics in
Q1’21.

22
Graph neural nets find mainstream
enterprise applications

23
What are graph Most machine learning techniques are designed to work on
tabular data or relational databases. But the rise of graph
neural nets? databases such as Amazon Neptune, Neo4j, and TigerGraph
has created a need for machine learning techniques tailor-
made for graphs.
Graph databases consist of nodes (individuals/entities) and
edges (the relations between them). A graph-based approach
works well for applications like advanced material discovery,
drug R&D (where atoms are nodes, and the interactions
between them are the edges), anti-money laundering, anti-
fraud, and enterprise recommendation systems.
This has given rise to interest in graph neural networks
(GNNs), or applying machine learning and neural nets to a
graph database.

24
 WHAT EXECUTIVES ARE SAYING

“Applications of graph-based ML at Twitter are abundant and

exciting. Graphs in different forms are the core data assets
generated by people using our platform… Developing ML
methods capable of taking advantage of these data assets is
important for making Twitter a better place for
public conversation.”
Michael Bronstein, head of graph learning research, Twitter

25
PATENT ANALYTICS

Graph database patents trend up over last decade

Patent activity related to graph databases, 2001 – 2020

Note: By date of filing; patterned column(s) may show decline due to publishing delay. 26
Amazon launches Neptune ML for customer retention,
fraud detection, and more
Amazon launched its own graph
database called Neptune in
2017. Three years later, the tech
giant launched Neptune ML,
machine learning specifically
built for graphs. The solution
provides the database and
analytics support for enterprise
customers.

Neptune ML is built on the Deep

Graph Library, an open-source
graph neural network library with
contributions from Intel, Nvidia,
and others.

27
Graph neural nets power recommendation systems

SOCIAL NETWORKS E-COMMERCE ON-DEMAND DELIVERY

In 2019, Twitter acquired Fabula AI, a In 2019, Alibaba published a paper on Uber Eats uses GraphSAGE, a
company building GNNs to detect AliGraph, a graph neural net system framework developed by Stanford, to
social network manipulation. already deployed internally to power recommend dishes and restaurants to
In 2021, Twitter published research on personalized search and e-commerce users. The company reports
applications of deep learning on recommendations. “significant improvements in
dynamic graphs which “evolve over recommendation quality and
The tech was also used during 2020
time, with prominent examples relevancy.”
Singles Day in China to generate 3D
including social networks, financial models of items like furniture on
transactions, and recommender Alibaba’s e-commerce site Taobao.
systems.”

28
Google improves Maps,
advanced materials research
The Google Maps team has partnered
with Alphabet subsidiary DeepMind to
use GNNs to improve the service’s ETA
predictions.

In Q2’20, DeepMind published research

on using GNNs to understand the
formation of glass, with implications for
manufacturing and other sectors. In the
project, the nodes and edges were
particles and the interactions between
them, respectively. AI was used to
predict the movement of molecules as
they transitioned between solid and
liquid states.

29
Integrating ML analysis with
graph database-as-a-service
offerings
The graph databases market is
expected to grow at a 17.7% CAGR to
reach $4.6B by 2027, fueling a need for
machine learning techniques tailor-
made for graphs.
Neo4j, a popular graph database
vendor, announced graph machine
learning tools for enterprises in Q4’20. It
raised $30M that quarter, reaching a
valuation of $532M.

30
Stream processing: Capturing real-time IoT
data for AI applications

31
What is stream As the number of real-time data sources grows with the
proliferation of IoT, traditional batch processing methods –
processing? which store data and retrospectively analyze it in batches –
can result in missed opportunities for enterprises.
Organizations increasingly want instantaneous analysis and
decision-making capabilities. This has raised interest in
stream processing technologies, where data is viewed as a
“stream of events” that is constantly generated.
Stream processing powers AI apps that are responsive in real
time. Meanwhile, the streaming process itself can benefit
from the use of machine learning.

32
 WHAT EXECUTIVES ARE SAYING

“We're seeing a lot more interest in terms of real-time data

analytics, especially with 5G and IoT use cases… We have
streaming capabilities in the product just now, but we're also
investing to develop more streaming capabilities as we go
through 2021.”
Mark Culhane, CFO, Teradata
Q4’20 earnings call

33
MARKET SIZE

The streaming analytics market is a $52B opportunity

Note: Based on CB Insights market sizing aggregate data. 34

Technologies enabling the streaming ecosystem

OPEN-SOURCE TOOLS IN-MEMORY COMPUTING

In-memory computing pools together parallel
Stream processing technology has benefited from the
distributed random access memory (RAM) from
availability of open-source tools. Apache has several
multiple computers to store and process data at high
distributed streaming platforms under its umbrella,
speeds, making analysis thousands of magnitudes
lowering the barrier to entry for developers who want to
faster than traditional memory.
integrate streaming analytics into their enterprise
workflows. The advantages of in-memory computing make it a
natural enabler of stream processing.
Kafka, which LinkedIn contributed to developing, is one
of the most popular services currently used today.

35
Cloudera acquires streaming analytics vendor Eventador

Cloudera acquired Texas-based seed-

stage stream processing company
Eventador in Q4’20.
Eventador has introduced several
features to improve machine learning
pipelines for data science teams,
including introducing SQL queries for
use on streaming data.
Cloudera CEO Rob Bearden said
during the company’s Q3’21 earnings
call that it will be increasingly
focused on solutions like real-time
analytics for streaming data.

36
Vendors are expanding to offer solutions for the entire streaming
ecosystem
Hazelcast is one of the most popular
vendors in the in-memory computing
space, with $88M in funding from Bain
Capital Ventures, Earlybird Venture
Capital, and others.
Hazelcast has been iteratively building
more stream processing capabilities by
introducing Hazelcast Jet, which works
in tandem with its in-memory tech
Hazelcast IMDG (in-memory data grid).
In 2020, Hazelcast announced support
for machine learning inference in Jet.

37
Striim positions itself as an
end-to-end solutions provider
Striim, which raised its first round of
funding in 2013, is positioning itself as a
“one-stop shop” for building and
deploying streaming infrastructure,
including:
• Log-based change data capture, with
support for multiple source and
target types

• In-memory stream processing

• Streaming SQL (a SQL-like

programming language for
interacting with streaming data)

• Edge processing and cloud

integration capabilities

38
AI can help parse data from various sources automatically
Data streams can originate
from IoT sensors, social
media, or real-time changes
to relational databases. In
the image to the right,
leading data integration
vendor Informatica’s CLAIRE
AI engine automatically
recognizes the structure of
incoming data and parses it.
To strengthen CLAIRE’s
capabilities, Informatica
acquired data management
company GreenBay
Technologies in Q3’20.

39
Analytics vendors increase support for
unstructured data types

40
What is Around 80% of big data today is unstructured, meaning it is
without a predefined format and is not searchable by
unstructured organizations.

data analysis? The commercialization of AI in recent years has enabled

enterprises to mine unstructured data and unlock
information that was previously inaccessible. The number
of alternative data sources has also increased and
includes IoT sensors, images, social media posts, and
surveillance videos.
Support for unstructured data formats and pipeline
management solutions has become a necessity for cloud
data management and business intelligence vendors.

41
 WHAT EXECUTIVES ARE SAYING

“…it's why we're implementing Project Ascent, for us to bring

on more than 100 alternative data sources over the next
couple years... being able to bring new and unstructured
sources of data in and connecting it to a D-U-N-S Number to a
specific client...”
Anthony Jabbour, Dun & Bradstreet CEO,
Q2’20 earnings call

42
Alt and unstructured data become commonplace in business intel
ENVIRONMENT & ECONOMIC FORECASTING CUSTOMER & BRAND DATA

OUTBREAK MONITORING SATELLITE IMAGE ANALYTICS MEDIA AND CUSTOMER SENTIMENT ANALYSIS

OCEAN DATA

EMOTIONAL INTELLIGENCE GEO-LOCATION DATA

WEATHER INTELLIGENCE SUPPLY CHAIN RISK PREDICTION

ALT CREDIT SCORING/LENDING

CUSTOMER FEEDBACK AND VOICE DATA

(SALES & CRM)

SMART CITY AND TRAFFIC ALT DATA FOR PROPERTY DAMAGE AIR QUALITY DATA
DATA ASSESSMENT
CUSTOMER BEHAVIORAL BIOMETRICS

Note: This is a selection of alternate dataset-based product developers and analytics vendors. It is not a comprehensive list of all companies in a space. 43
PATENT ANALYTICS

Audio transcription patents trend up

Patent activity related to unstructured data types, 1999 – 2020

Note: By date of filing; patterned column(s) may show decline due to publishing delay. 44
Venture capital firms back unstructured document
analysis vendors

Latest round: $22M Latest round: 80M Latest round: $13M Latest round: $12M Latest round:
Series B in Q4’20 Series D in Q4’20 Series B in Q3’20 Series B in Q3’20 Series C in Q1’21 for
undisclosed amount
Select investors: Select investors: Select investors: Select investors:
.406 Ventures, Jump Bessemer Venture QED Investors, Grazia Equity, Plug Select investors:
Capital, Osage Partners, Battery Bullpen Capital, and Play Ventures, Insight Partners, Oak
Venture Partners, Ventures, FirstMark FinTech Collective, BlackFin Capital HC/FT Partners
Sandbox Insurtech Capital, Tiger Global RiverPark Ventures Partners
Ventures Management

45
Corporates mine audio data, climate risk scores, and more

AUDIO TRANSCRIPTION VIDEO, AUDIO ANALYTICS FINANCIAL RISK MANAGEMENT

S&P Global acquired analytics In a Q2’21 earnings call, Microsoft CEO Moody’s launched DataHub in Q1’21,
company Kensho in 2018, which Satya Nadella reported strong growth combining structured and
launched a transcription feature called in the company’s analytics business. unstructured datasets for financial
Scribe to extract unstructured audio He added that FedEx, Grab, P&G, and risk management with billions of data
data. Kensho has created other AI others use Microsoft’s Synapse “to points, including climate risk scores
solutions including ProSpread, which generate immediate insights from and ESG (environmental, social, and
supports data extraction in 9 massive amounts of structured and corporate governance) assessments.
languages using optical character unstructured data.” In June 2020
recognition and natural language alone, over 9M hours of speech were
processing. transcribed using Azure Cognitive
Services.
In 2018, S&P also acquired Panjiva, a
company that analyzes unstructured In Q3’20, Microsoft acquired Orions
shipping and trade data. Systems to tag and manage
unstructured data in video feeds.
46
Rubrik and Snowflake add unstructured data support

CLOUD DATA MANAGEMENT DATA WAREHOUSING

Rubrik expanded its unstructured data management Snowflake added support for unstructured data
capabilities with the acquisition of Igneous in Q4’20. management – including for images, PDFs, and video
Igneous initially built hardware for on-prem storage of files – in addition to structured and semi-structured
unstructured data and later expanded to cloud data data types. The feature was released in private preview
management. in Q4’20.

Total funding: $553M Exit date: Sep ’20 (IPO)

Valuation: $3.3B Exit valuation: $33.3B valuation

Select investors: Lightspeed Venture Partners, Khosla Select clients: Siemens, Comcast, Instacart, Logitech
Ventures, Bain Capital Ventures, Greylock Partners

47
Transformers, multilingual models improve
enterprise NLP

48
What are Google introduced a language model called Transformer in
2017. The following year, it launched BERT, another model
Transformer based on Transformer. With these models, Google
delivered breakthrough improvements in natural language
models? processing (NLP) and understanding. Around the same
time, OpenAI launched its now-popular Generative Pre-
trained Transformer (GPT) AI series.
Transformer-based models, or Transformers, are “pre-
trained” without the need for labeled datasets, removing a
huge bottleneck in NLP progress. In pre-training, an AI
model is trained on an enormous amount of text readily
available on the internet. This way, the model understands
context of words and relations between sentences.
Transformers are leading to breakthroughs in sentiment
analysis, translation, reading comprehension, gaming, and
more.
49
 WHAT EXECUTIVES ARE SAYING

“And just as AlexNet 7 years ago kind of was the watershed

event for a lot of computer vision-oriented AI work, now the
Transformer-based natural language understanding model
and the work that Google did with BERT really is a watershed
event also for natural language understanding… that
breakthrough is really quite significant.”
Jensen Huang, Nvidia CEO,
Q3’20 earnings call

50
Recent breakthroughs in natural language processing
Key research highlights from Google and OpenAI

Google’s Transformers
New neural net architecture for OpenAI’s GPT
language understanding Transformer-based
outperforms older approaches unidirectional contextual AI
model, where a word is taken in
the context of words preceding OpenAI’s GPT-2
it in a sentence
AI pre-trained with 8M pages of
internet text
GPT-3 produces
Google’s BERT human-like text
Transformer-based bi- Larger and more compute-
directional AI pre-trained with intensive than GPT-2; OpenAI
Wikipedia text, where a word is releases an API integration
taken in the context of
preceding & succeeding words

2017 2018 2019 2020

51
NLP models get bigger and better
Published in Q1’20, The number of parameters in AI language models over time, Q1’18 – Q1’20
Microsoft claimed its
Turing Natural Language
Generation (T-NLG) model,
with 17B parameters,
outperformed others in
tasks like question
answering and
summarization.
In Q2’20, OpenAI beat this
record with its GPT-3, with
100B+ parameters. Google
took the lead in Q1’21,
releasing a model with 1T+
parameters.

Note: Image published in Q1’20 before the release of OpenAI’s GPT-3 and Google’s 1T+-parameter model. Image source: Microsoft, DistilBERT 52
APIs make advanced NLP tech
accessible to enterprises
In Q2’20, OpenAI released GPT-3, larger
and even more compute-intensive than
its predecessor GPT-2.
Due to potential for misuse, OpenAI
initially didn’t release the entire source
code of GPT-3, but later licensed the
tech to Microsoft and made it available
via a limited beta API.

“Unlike most AI systems which are

designed for one use-case, OpenAI’s
API today provides a general-purpose
‘text in, text out’ interface, allowing
users to try it on virtually any English
language task.” – OpenAI blog

53
Early applications of GPT-3’s language generation

GAMING CUSTOMER SERVICE VR/GAMING

Utah-based AI Dungeon is Sapling uses GPT-3 to compose In Q1’21, Modbox, a sandbox for
developing a text-based personalized responses to assist PC/AR/VR multiplayer games, released
adventure game where an AI sales and customer support teams a demo of an AI-driven NPC (non-player
model generates open-ended with customer response. Features character) using GPT-3 and Replica
storylines based on GPT-3. The include autocomplete and spelling software for natural language
game reportedly attracts 1.5M and grammar checks. understanding and speech synthesis.
active users per month.

54
Facebook AI’s milestone in language translation
In Q4’20, Facebook open-sourced a multilingual
machine translation model, M2M-100, that can
translate between 100 languages “without
relying on only English-centric data.” The model
relies on a dataset with 7.5B sentences.
Earlier, in Q3’19, Facebook developed its own
version of BERT, called RoBERTa, to moderate
hate speech on its platform.

That same quarter, FB open-sourced a speech

recognition model that bypasses the need for
manually annotated transcripts. Called
wav2vec, it is particularly useful for speech
recognition in non-English languages, where
annotated training data is sparse.

55
Data governance and explainable AI

56
What are data Establishing protocols for sourcing, handling, and using
data is crucial for developing ethical AI solutions and
governance and preventing algorithmic bias in outcomes.

explainable AI? Strong governance is the foundation of AI explainability.

Enterprises may not know what an algorithm is “seeing” in
the input data or why it arrives at a certain conclusion. For
example, in lending, where consumers are protected
against discrimination, banks need visibility into the inner
workings of an algorithm.
To comply with regulations like GDPR and CCPA, vendors
are increasingly focusing on explainability, bias monitoring,
and strong governance.

57
 WHAT EXECUTIVES ARE SAYING

“Data governance is top of mind for every business leader

and will grow into an important category on its own as critical
as any AI or analytics category today. We are investing to
participate in this growth…Azure Purview provides an end-to-
end view of an organization's data state across on-premise,
multi-cloud, and SaaS apps that previously was impossible.”
Satya Nadella, Microsoft CEO,
Q2’21 earnings calls

58
AI regulation and ethics in focus
Recent news mentions

Ethical concerns mount as AI takes Why companies are thinking twice

bigger decision-making role in more about using artificial intelligence
industries
Oct ’20 Jan ‘21

[Global] alliance aims to accelerate Why this is the year for regulation that
the adoption of inclusive, trusted and finally reins in AI
transparent AI worldwide
Jan ‘21 Feb ‘21

59
Companies are using AI for data governance, which is
key for ethical AI

Dathena is an AI-enabled data privacy Privacera develops software-as-a- Securiti builds an AI-powered
and security vendor that monitors on- service for cataloging sensitive data PrivacyOps platform, RPA solutions,
prem and cloud data. Dathena has a co- across multi-cloud environments. and tech to automatically link
sell partnership with Microsoft. personal data to users.
Latest round: $50M Series B in Q1’21
Latest round: $12M Series A in Q2’20 Latest round: $50M Series B in Q1’20
Select investors: Accel, Point72
Select investors: Jungle Ventures, Ventures, Cervin Ventures, Alchemist Select investors: Mayfield Fund,
CapHorn Invest, CerraCap Ventures; Accelerator General Catalyst
previously participated in Microsoft AI
Factory and Nvidia Inception
60
NEWS TRENDS

Media mentions of explainable AI peak in 2020

• “Fiddler Secures Strategic
Investment from Amazon Alexa
Fund” – FinSMEs
• “Appen Partners with World
Economic Forum to Create
Responsible AI Standards” – Unite.AI “Explainable AI: A Must For
• “NIST Issues Draft Publication on Nuclear Nonproliferation,
‘Explainable’ AI” – ExecutiveGov National Security”
– Pacific Northwest
National Laboratory

“Temenos Acquires a
SaaS-based, Patented,
Explainable AI (XAI)
Platform” – AiThority

“FICO Announces xML

[explainable ML] Challenge”
“DARPA sees future wars won – PR Newswire
with hypersonic weapons and
artificial intelligence” – Business
Insider (mentioning DARPA’s
explainable AI research program)

Note: Based on CB Insights news trends tool. 61

Alexa Fund, Lockheed Martin back explainable AI company

Total funding: Valuation:

13M $31M

Fiddler develops a platform for

enterprises to build explainable AI
models. Lockheed Martin recently
partnered with Fiddler, calling
explainable AI tech key to building
successful AI solutions.

Select investors: Amazon Alexa

Fund, Lockheed Martin Ventures,
Lux Capital, Lightspeed Venture
Partners, Haystack Ventures,
Bloomberg Beta

62
Risk management: Monitoring AI performance and bias

ML EXPERIMENT TRACKING MODEL MONITORING

Weights & Biases develops tools to track ArthurAI’s platform tackles model monitoring
model performance and ensure AI experiments and performance optimization, bias detection,
are reproducible. Its tech is used by companies and explainability.
including OpenAI and John Deere.
Latest round: $15M Series A in Q4’20
Latest round: $45M Series B in Q1’20
Select investors: Index Ventures, Homebrew,
Select investors: Insight Partners, Trinity Work-Bench, AME Ventures, Plexo Capital,
Ventures, Bloomberg Beta, Coatue Acrew Capital
Management
63