W31059

ZHUIYI TECHNOLOGY: DEVELOP OR DIVERSIFY?

Chongqiao Xiong, Haiyang Li, Yijia Tang, and Jianan Lu wrote this case solely to provide material for class discussion. The authors
do not intend to illustrate either effective or ineffective handling of a managerial situation. The authors may have disguised certain
names and other identifying information to protect confidentiality.

This publication may not be transmitted, photocopied, digitized, or otherwise reproduced in any form or by any means without the
permission of the copyright holder. Reproduction of this material is not covered under authorization by any reproduction rights
organization. To order copies or request permission to reproduce materials, contact Ivey Publishing, Ivey Business School, Western
University, London, Ontario, Canada, N6G 0N1; (t) 519.661.3208; (e) cases@ivey.ca; www.iveypublishing.ca. Our goal is to publish
materials of the highest quality; submit any errata to publishcases@ivey.ca.

Copyright © 2023, Ivey Business School Foundation Version: 2023-09-15

As Shenzhen Zhuiyi Technology Co., Ltd. (Zhuiyi), celebrated its third anniversary in February 2019, over
700 million consumers were chatting with the company’s artificial intelligence (AI) chatbot platform,
designed for customer service. The company, which augmented call centre agents with AI algorithms, was
a leading AI start-up and natural language processing (NLP) provider located in Shenzhen, China. It had
outperformed its competitors in underlying NLP algorithms by securing a technical advantage in accuracy
rates. In 2020, Zhuiyi had ranked second place, after Google LLC (Google, owned by Alphabet Inc.), in
SuperGLUE, one of the most difficult and authoritative competitions in the NLP field (see Exhibit 1).
Zhuiyi offered its AI chatbot services to corporate clients, who could take advantage of the reinforced
learning features of the AI system to reduce customer service costs,1 as chatbots could answer 80 per cent
of low-level questions faster than their human counterparts. Additionally, Zhuiyi’s AI response-correction
rate could also reach more than 92 per cent, and its chatbot could respond to clients’ inquiries automatically
in both text and voice formats.

In April 2019, Zhuiyi raised US$41 million in Series C funding, and its workforce grew by 30 per cent.2
The round was led by China Merchants Capital Management Co., Ltd., and followed by Sinovation
Ventures, Morningside Venture Capital, Gaorong Capital, and GGV Capital. In the months following the
expansion, the company’s revenue growth slowed, but costs steadily climbed. However, new-customer
acquisition became increasingly competitive as a few vendors started to take a loss leaderpricing approach.
In June 2021, Zhuiyi raised Series D financing. While continuing to acquire customers in the finance sector,
the company’s executives knew they had to grow both revenues and profit margins to justify the valuation
for additional rounds of fundraising (see Exhibit 2).

Two broad options were on the table. First, Zhuiyi could invest more to develop new AI products. Previous
efforts in this direction had been fruitful, but there were always risks involved in the new and evolving market.
New applications would demand higher investments in research and development (R&D), which already
represented a large portion of operating costs. The second option was to invest further in industry
customization efforts and diversify the client base into other sectors. Thus far, Zhuiyi had built its reputation
in the niche finance industry by servicing most household bank brands. This track record provided a
foundation for expansion into other industries. However, competitors such as Baidu, Inc. (Baidu), and Iflytek
Co., Ltd. (Iflytek), had already accumulated large data sets in the retail, travel, and education sectors, which
would require Zhuiyi to spend time and effort to catch up on industry solution customization. As 2022 began,
Joey Wu—one of the co-founders and the chief executive officer of Zhuiyi—met with the company’s top
executives in their Shenzhen office to mull over the options. Which one held the most promise?
Page 2 W31059

HISTORY OF ZHUIYI

Getting Started

In March 2016, Joey Wu, the general manager of the AI platform at Tencent Holdings, Ltd. (Tencent), and
three of his colleagues left the Chinese tech giant to start their own venture. Zhuiyi was envisioned as an
enterprise service AI platform, but Wu was still exploring various options for what the company would
actually do. Having technical backgrounds, Wu and his team built the Chinese version of Google’s
AlphaGo, a computer program designed to play a popular Chinese board game. However, the team believed
that the technology they had developed had application possibilities far beyond playing Go. Venture capital
investors were the first to identify the technology’s potential. Even before Zhuiyi had a working product,
angel investors had already fuelled the company with sufficient funds. While Wu and his team were
bootstrapping their idea, Didi Chuxing Technology Co. (Didi), the Chinese ride-sharing monopoly,
approached them for help. At that time, Didi was experiencing triple growth in its user base, and the
increasing amount of service requests necessitated an upgrade of the company’s technology infrastructure
and service staff. Didi’s customer service centre employed over 10,000 full-time and part-time agents and
operated on a real-time response basis. The service centre was at full capacity and could not handle the
influx of cases and calls.

Zhuiyi demonstrated significantly better results in the proof of concept than traditional players, whose
technologies were largely based on keywords and semantic rules. This helped Zhuiyi secure its first client:
Didi. Out of the 30,000 daily inquiries, more than 80 per cent were repetitive and could be answered by
standard template answers. A chatbot was programmed to answer frequently asked questions without the
help of a human operator. The other 20 per cent were non-standard and complicated questions, and they
were redirected to human agents. Six months after inception, Didi managed to reduce two-thirds of the
original staff volume in its call centre while also answering three times more customer inquiries (see Exhibit
3). Employees that were no longer needed in the call centre were transferred to sales for premium
membership and insurance products, which brought in significant amounts of additional revenue.

Accelerating the Business

Learning from Didi’s successful application, the Zhuiyi team quickly replicated the service with other
Internet service clients, including Hangzhou Beigou Tecnhology Co., Ltd., Xiaomi Inc., and Ctrip.com
International, Ltd. In 2018, Zhuiyi expanded its pool of clients to the finance sector when it entered into an
agreement with China Merchants Bank Company Limited China (China Merchants Bank), the largest
shareholding bank in China. “Financial institutions were the type of corporations that needed AI technology
the most,” said Wu. “They had a strong sense of innovation and complete [information technology]
infrastructure to support the implementation of interactive applications.” The China Merchants Bank case
allowed Zhuiyi to dive deep into new user case scenarios that were more connected to business than just
customer service. Some of those scenarios included credit card debt collection, installment sales, money-
laundering detection, and identity verification. Conversational AI that could navigate customers through
different tasks was referred to as a taskbot. A taskbot required far more data for training purposes than
regular chatbots. Both tagging and processing were a significant amount of work, as was connecting to
different information systems within the financial sector. The level of complexity meant that Zhuiyi was
building an industry moat that was harder for competitors to replicate.

Working with financial institutions brought Zhuiyi much more revenue than did its Internet service clients.
“In China, the biggest clients were always governments or state-owned companies,” said Peja Hu, Zhuiyi’s
chief marketing officer. “They were not price sensitive, and once you were in their vendor system, it was
Page 3 W31059

hard to get kicked out.” In the financial sector, Zhuiyi served 70 per cent of the top 20 domestic banks and
50 per cent of the top 10 insurance companies and AA-level securities companies. Benchmark clients
included China Merchants Bank, Shanghai Pudong Development Bank Company Limited, and six other
major state-owned banks (see Exhibit 4).

A BRIEF HISTORY OF AI AND NLP COMMERCIALIZATION

The AI era in China started around 2015, and research and business applications in the field developed
quickly from there.3 Capital investments in AI companies increased from $7.18 billion in 2015 to $25.35
billion in 2020.4 In the early stages, AI companies in China mainly focused on computer vision, using AI
algorithms in image processing. The market was dominated by four companies: SenseTime Group Inc.,
Megvii Technology Ltd., Cloudwalk Technology Company Ltd., and Yitu Network Technology Co., Ltd.5
The conversational AI industry, established on the development of NLP technologies, had gradually taken
centre stage and continued to rapidly progress.

The core task of conversational AI could be divided into three parts: speech recognition, speech
understanding, and speech synthesis—that was, creating machines with the ability to hear, understand, and
talk. 6 NLP technology processed language through corpus acquisition, corpus preprocessing, feature
engineering, feature selection, and model training to enhance the conversational interaction between
machines and humans. In recent years, the extensive applications and continuous breakthroughs of deep
learning technology had greatly promoted the landing of NLP, of which the development of a
conversational AI platform was an important commercialization direction. Conversational AI platforms
integrated NLP with other technologies, including automatic speech recognition, text to speech, machine
learning, and knowledge maps so that machines could identify and understand user needs as well as offer
feedback, thus enabling human-machine interactions in conversational service scenarios.7 Conversational
AI platforms could be loaded on intelligent hardware to meet the requirements of manufacturing operation
and control, enhancing the authenticity of human-computer interaction. Such platforms could also enable
service scenarios, such as customer service, marketing, voice robots, intelligent quality inspection, agent
assistance, multimodal digital human, and enterprise information services that applied text robots.

THE CONVERSATIONAL AI INDUSTRY IN CHINA

The conversational AI industry surged in China in just five years. The industry’s market size was $460
million in 2014, and it reached $2.39 billion in 2018.8 In 2019, conversational AI accounted for 22 per cent
of the market share of China’s AI industry, second only to computer vision, which had already been put
into large-scale commercial use.9

The industry was characterized by a three-tiered partitioning pattern: Internet giants such as Baidu, Alibaba
Group Holding Limited (Alibaba), and Tencent; leading specialists on conversational AI technologies
represented by Iflytek; and conversational AI start-ups that focused on narrower and more segmented
scenarios.10 In 2020, Iflytek accounted for 60 per cent of the market share in the conversational AI industry,
and 20 per cent of the market share was occupied by Internet giants such as Baidu and Alibaba. 11
Downstream, the industry was divided into consumer-level markets (direct to consumer and business to
consumer [B2C]) that targeted smart-life scenarios, such as home applications and vehicles, and enterprise-
level markets direct to-business(2B) that targeted medical care, government affairs, and education.

The Internet giants competed for key accounts based on abundant product-market experience and resources.
Access to large-scale user data allowed these companies to engage in user-oriented product design and
accurate targeting. Their product ecosystems and platforms were used for the launch of self-developed
Page 4 W31059

products as well as intra-industry cooperation through technologies that were embedded into partners’
products or business scenarios. Leading specialists like Iflytek occupied most of the market share because of
their strong and focused technological capabilities. However, as conversational AI technology matured and
as technological barriers were reduced, first-mover advantages were gradually weakened, and niche
competition became increasingly fierce. As such, the profitability of the leading specialists’ original business
model of technology licensing and outsourcing was limited. These specialists sought to transform by
establishing AI developer platforms to provide integrated technological services and modular customized
solutions, which created greater value for downstream clients. Conversational AI start-ups adopted a niche
strategy to develop conversational AI products or solutions focused on a single scenario or industry (e.g.,
voice input, automotive voice interaction, health-care assistance, smart conferences). These start-ups targeted
narrow, homogenous, and lower-end customers, and the niche strategy allowed them to specialize in specific
technologies and avoid direct competition with the other giants and leaders in the industry.

Baidu

Baidu entered the AI area in 2010 and established the Institute of Deep Learning, China’s first AI research
institute in 2012.12 Since 2015, Baidu had started to restructure its core business around mobile ecosystems
and AI technologies.13 The Understanding and Interaction Technology (UNIT) platform, Baidu’s key output
in the field of conversational AI, served as one of the main products competing with Zhuiyi’s Bot system. The
platform was designed to provide developers with services for intelligent dialogue development and
customization at low cost and entry barriers. By opening access to the underlying NLP technologies, such as
voice, interaction, and knowledge mapping, UNIT provided solutions and packages at different customization
levels. The platform had helped clients develop products of intelligent customer service, artificial intelligence
of things, and intelligent cockpit systems that were applied in scenarios such as financial services, government
affairs, logistics, education and training, smart vehicles, and home automation.

Baidu’s premium brand helped attract benchmark customers, such as utility companies and operators.14 For
instance, comprehensive solutions for the intelligent interactive voice response and customer service assistant
were developed based on UNIT for the China United Network Communications Group Co., Ltd., Tianjin branch,
which significantly enhanced the efficiency and quality of customer service.15 The platform also upgraded the
intelligent customer service system for the China Southern Power Grid Company Limited in Guangdong
Province. Xiaodu, Baidu’s voice assistant product, was released in 2017 and rapidly gained acknowledgment in
the market. The number of monthly voice interactions using Xiaodu reached 280,000, and the number of smart
devices equipped with the Xiaodu Assistant System grew rapidly, reaching 12.5 billion in 2020.16

Iflytek

Iflytek had long maintained a dominant position in developing and providing AI technologies of speech
and language in China.17 Its main business included research on AI voice and language, software and chip
development, voice and information services, and e-government system integration. Iflytek had over 80 per
cent of the market share in the telecommunications, utility, and social security industries.18

Iflytek had long-standing experience in AI-related R&D. It partnered with domestic universities and
institutions, including the University of Science and Technology of China, Tsinghua University, and South
China University of Technology to undertake a series of the State Key Laboratories Development
Program.19 From 2015 to 2019, the company’s average R&D expense rate was 21.8 per cent, far exceeding
the industry’s average investment. 20 The immense R&D investments secured Iflytek’s world-leading
position in speech recognition and synthesis technology.21 The extensive NLP capabilities that Iflytek
Page 5 W31059

accumulated supported its development of a complete product matrix and highly competitive platforms for
voice, semantics, and knowledge and included leading customers in the financial and insurance industries.22

Iflytek Open Platform specialized in providing standardized conversational technologies, including speech
recognition, synthesis and analysis, and multilingual technologies to developers. Currently the platform had
gathered more than 1.12 million AI developer teams, developed more than 750,000 applications, and
provided 287 different AI capabilities and solutions.23 For example, Beijing ByteDance Technology Co.
Ltd. accessed Iflytek’s real-time voice transcript service to enhance voice interaction for their Feishu
product. Iflytek’s platform was also applied in Huawei Technologies Co., Ltd.’s (Huawei), IdeaHub office
products, in real-time voice interaction for Meituan Inc.’s (Meituan) takeaway riders, and in the exhibition
hall of the Sino-Singapore Tianjin Eco-city.24 The Iflytek Yanzhi Platform served as a direct competitor to
Zhuiyi’s Bot/Call. This platform was designed to provide intelligent robots and customer service solutions
in the fields of finance, insurance, telecommunications, and governments.

Other Competitors

Other competitors included emergent AI start-ups such as Emotibot Technologies Limited (Emotibot).
Emotibot targeted the specialized technologies of brain-like dialogue robots, speech emotion, visual
understanding, and multimodal affective computing.25 Bot Factory, Emotibot’s conversational AI platform,
offered customized chatbots with few coding requirements and precise operating outcomes based on a
combination of its highly accurate speech and semantic recognition.26 Bot Factory also specialized in user
portrait through a conversation process that empowered personalized recommendation. Partnering with
Huawei Cloud, this platform included enterprise services for government affairs, manufacturing, medical
care, and Internet of things.27

Domestic Internet giants developed cloud platforms around their ecosystems to enable the
commercialization of speech technologies. For instance, Tencent Cloud and its affiliated AI Open Platform
served as the main output channels for Tencent’s voice technologies developed by Tencent AI Lab.28
Speech recognition had been applied to Tencent’s internal businesses, such as WeChat, games, and the
video platform. 29 Alibaba applied voice technologies developed by the DAMO Academy’s speech
laboratory to its internal scenarios in the value chain.30 Alibaba’s conversational AI assistant had been used
at Taobao and Tmall, Alipay hotline, and phone robots for Cainiao Smart Logistics Network Limited, and
it accounted for 11 per cent of China’s intelligent customer service market in 2020.31

ZHUIYI’S PRODUCTS

Bot

Zhuiyi’s customer service chatbot was packaged into all modular products. Bot answered customer
inquiries in text format and could be connected to customer service centre instant-messaging platforms
(software that connected callers with and agents). Corporations could integrate Bot into their systems
without disrupting the flow of customer service. The product incorporated re-enforced learning algorithms
that allowed computers to learn by themselves and continuously improve their rates of accuracy. Bot was
also equipped with emotional analysis tools that could detect customer emotions and that requested human
agent intervention when a situation escalated.

As of December 2019, over 85 per cent of the top 100 banks in China were using Zhuiyi’s Bot. Other clients
included insurance and security companies, retailers, and education services. Bot was also used to answer
consumer questions on client apps, WeChat public accounts, and other social media platforms.
Page 6 W31059

Call

In late 2018, Zhuiyi replicated its Bot success with Call, a voice-enabled product that spoke to customers
in humanized tones. Call was often mistaken as a real human agent, as it could have several rounds of
conversations with customers without interruption. Like Bot, Call was engineered to not only answer
customer inquiries but also to actively reach out to customers. China Merchants Bank used Call to collect
credit card debts and other receivables.

The Call market was bigger than the Bot market because consumers in China were more used to talking
over the phone, which made them feel more serviced. The rise of short-form user videos on social media
platforms also promoted the use of Call. According to an industry survey, outbound telemarketing had a
projected size of $2.1 billion, and it was still trailing at a high growth rate.32 More use case scenarios were
found to feature the Call product.

Face

Face took the form of a digital twin that could be a virtual representation of any person, cartoon character,
or even animal. As an interactive product, it integrated a number of AI technologies, including semantic,
voice, and video multimodal algorithms. Face could simulate the human voice, emotional expression, and
body movements. Based on industry needs, Face served in sectors such as finance, telecommunication,
government affairs, and media and entertainment (see Exhibit 5).

There were two types of Face. The broadcaster type self-generated content, which could significantly
reduce the costs of content production. It was mainly used in news anchoring and live streaming. The
interactor type could greatly improve the usage experience by engaging in conversations with customers. It
was often used in customer service, consultation, and shopping guide scenarios.

Efficiency Tools

While Bot and Call were the two products whose end users were mass consumers, Zhuiyi had also
developed applications for enterprise clients to improve productivity. For example, Learn was an AI-
assisted training program for new employee onboarding, and See oversaw human agents in call centres for
quality inspection. Pal assisted human call centre agents by prompting recommended answers in front of
their screens so they did not have to memorize anything. Those products all served as human-machine
interaction tools to enhance corporate objectives (see Exhibit 6).

Wu and his co-founders envisioned an open ecosystem. They had licensed all of Zhuiyi’s core algorithms
into application programming interfaces so their developer partners could use the AI capabilities in their
own products. The company did not intend to compete in the business-to-consumer (B2C) segment of the
NLP market, which was mostly driven by small to medium-sized companies and individual developers.

PRICING MODELS

Three main pricing models were commonly used in the conversational AI industry:

1. On premises was the most basic pricing model. It allowed for data storage on local servers with a one-
time licensing fee and additional costs depending on the customization specifics (e.g., customization
Page 7 W31059

costs typically depended on the number of staff and time involved, which varied according to the
degree of customization and the robustness of client data).
2. Public cloud pricing was introduced as corporations increasingly worked with cloud computing
providers for data storage. It allowed clients to use algorithms trained across multiple clients’ data sets
for a lower fee while maintaining its own data security. A public cloud could be purchased through a
subscription model that depended on the usage amount.
3. Cost per success was a revenue-sharing scheme that allowed corporations to only pay for what they
saved/made. Many start-ups experimented with this model by charging a 20 per cent upside of the
additional revenue that corporations gained through adopting the conversational AI services.

In 2019, more than 85 per cent of Zhuiyi’s clients used the on-premises model. Most clients chose the on-
premises model for purposes of data security. Corporations in China, financial institutions specifically, built
their physical firewalls in compliance with government regulations on data privacy.33 The on-premises
model granted a level of security that many corporations desired, but it striped Zhuiyi from collecting
reoccurring revenue from those clients on a single product, except for additional licenses. Of the two
models, the on-premises one could not be updated in real time, and all product delivery and maintenance
had to be done on site. The public cloud model offered more enhanced algorithms, which generally implied
a higher accuracy rate (see Exhibit 7). Zhuiyi had been continuously trying to convert on-premises clients
to public cloud clients, but it was an industry-wide challenge that showed little success.

COMMERCIALIZATION PROBLEMS

Many traditional companies had digitalization agendas, but only a few really knew how to pursue them.
Salespeople had a tendency to promise effects of AI that scientists and engineers had a hard time delivering.
When IBM’s Watson supercomputer beat Ken Jennings, the best human Jeopardy! player ever, many
assumed it would have a similar performance on large, complex initiatives. However, scientists and product
people understood that Watson was custom-built for the quiz show.34 Because of AI’s highly publicized
image, companies adopted it with higher expectations that might not have actually been possible. When
they learned about the complex nature of AI deployment and its true capabilities, they were often
dissatisfied with AI products. As a result, it was very common for AI enterprise software products to have
low customer satisfaction rates. This was also an issue for Zhuiyi.

Many factors contributed to low satisfaction rates. By 2019, Zhuiyi’s product bundle was sold to more than
300 key accounts. A key account was categorized by the amount of AI usage and size of the company,
which was typically a Fortune 500 company. Large-scale companies were used to outsourcing their non-
core businesses to service vendors and usually had very established procurement systems. Thus, these
companies tended to look for mature and sophisticated solutions that were plug and play. However, Zhuiyi’s
founding team believed strongly in the minimum viable product principle, which was typical in the start-
up space. Products were developed in the least amount of time and then tested in the market for customer
feedback. Improvements were made later after launches. The short product cycle enabled companies to
quickly capture market share and gain a competitive advantage over other players. This model had been
proven successful in China’s high-growth B2C market, as consumers had a high tolerance for product
misfunctions. Consumers sometimes did not even notice changes, since most bug fixes were updated within
six hours of release, as consumer-facing apps and software were hosted on clouds. However, it was an
entirely different case for enterprise clients. Business-to-business (B2B) software was expected to be
complete and perfect at the time of delivery, and bug fixes could be very costly. Customer trust was hard
to rebuild after even one failure event. It took Zhuiyi’s executives at least two years to realize this clash in
corporate cultures. Alan Wen, chief product officer, expressed his concerns over different skill sets of
Page 8 W31059

direct-to-consumer versus enterprise software: “The team was quite impetuous. I thought the biggest
problem was self-cognition. We were an all-engineer team previously from Tencent. We had worked in the
agile model for more than 10 years [and were] very used to the culture of small steps and fast iterations at
Tencent. Trunk-based development required relatively low product quality.”35

The challenge of AI platform vendors versus industry specialists also played a role in the low satisfaction
rate. None of Zhuiyi’s executives worked outside the technology industry before starting Zhuiyi, nor did
many of its employees (see Exhibit 8). Employees often felt that they did not speak the same language as
their clients. In addition, B2B software-as-a-service generally required more sophisticated industry
knowledge to develop applications that would work in a specific sector. Without sufficient understanding
of industry environment, AI products were disconnected with business scenarios.

THE IMPACT OF COVID-19

The highly competitive and fragmented conversational AI industry had undergone considerable changes
since the COVID-19 pandemic started in 2020. Financing options became much more limited as market
capital dried up. Firms without large benchmark enterprise clients soon ran out of cash and received no
follow-up venture money. Small and medium competitors exited the market due to insufficient funds. At
the same time, major players in the Internet services segment had invested heavily in automatic services,
as market demands increased exponentially for non-human contact technologies. Meituan, China’s largest
food delivery company, used drones to deliver food to designated street kiosks. Customers could use their
smart phones to scan the QR codes on kiosks and pick up their packages.36 Meituan was still testing the
human-robot collaboration model, but it had shown initial success. Another example was Alibaba’s $1.5
billion investment plan to strengthen its AI and Internet-of-things ecosystem around its smart speaker Tmall
Genie. Alibaba launched 100 products in 68 categories, powered by its conversational AI system.37

In 2020, a number of forces significantly affected the supply and demand of the conversational AI industry.
On the supply side, the costs for engineers increased substantially due to a shortage of technical staff from
quarantines. Zhuiyi’s on-premises projects required more on-site engineering hours, which were
significantly impacted by the pandemic. Both project delivery and sales efforts were delayed under the
work-from-home model. All of these factors decreased the supply from Zhuiyi and other competing firms.
At the same time, the pandemic also created new areas of demand. Local governments and other institutions
incorporated conversational AI in response to surging demands of for online public services. Citizens could
check lockdown regulations, ask pandemic-related questions, and use government services. Many
consumer-facing services were shifting online due to a shortage of workers. As life with COVID-19 became
the new normal, an increasing number of Bots and Calls were engaged to answer consumer service inquiries
and actively reach out for sales. Thus, the pandemic expanded both the number of potential use cases and
the demand per customer.

LOOKING AHEAD

Zhuiyi had served more than 300 corporations in China and made its name as preferred vendor in the NLP
market.38 “The market environment changed very fast, and we were not sure what should be our long-term
strategy,” said Wu. “But we hope to make Zhuiyi’s NLP technologies achieve artificial general intelligence
[technological breakthrough in a wider spectrum].” Zhuiyi executives were now asking: What is next?
Should the company invest more to create a wider variety of AI products? What areas and industries should
they explore? Or should they continue to dive into current product bundles and expand into other industries?
How should they maintain growth?
Page 9 W31059

EXHIBIT 1: RANKINGS OF SUPERGLUE COMPETITION

Note: “SuperGLUE Human Baselines” refers to a human level of performance, which serves as a baseline to test whether
algorithms perform faster and better than humans.
Source: “SuperGLUE Benchmark Leaderboard,” SuperGLUE, accessed January 2020, https://super.gluebenchmark.com/leaderboard/.

EXHIBIT 2: FINANCING SCHEDULE

Announcement Transaction Money

Lead Investors
Date Round Raised

China Internet Investment Fund Management

June 8, 2021 Series D Undisclosed
Co Ltd, China Mobile Innovation Industry Fund

April 2, 2019 Series C $41 million China Merchants Capital Investment Co., Ltd.

November 2, 2017 Series B $20.6 million Sinovation Ventures

5Y Capital, Changzhou Gaorong Venture
September 1, 2016 Series A Undisclosed
Capital Co., Ltd.

Source: Zhuiyi Technology, “Deal History for Zhuiyi Technology,” PitchBook, accessed May 20, 2022, https://pitchbook.com/
profiles/company/221951-89.
Page 10 W31059

EXHIBIT 3: DIDI CUSTOMER SERVICE PLATFORM

Source: Created by case authors based on company documents.

Page 11 W31059

EXHIBIT 4: LIST OF TOP 100 BANKS IN CHINA

Ranking Bank Name (Chinese) Bank Name (Translated to English)

1 中国工商银行 Industrial and Commercial Bank of China Limited
2 中国建设银行 China Construction Bank Corporation
3 中国农业银行 Agricultural Bank of China Company Limited
4 中国银行 Bank of China Limited
5 交通银行 Bank of Communications Co., Ltd.
6 招商银行 China Merchants Bank Company Limited
7 中国邮政储蓄银行 Postal Savings Bank of China Co., Ltd.
8 兴业银行 Industrial Bank Co., Ltd.
9 上海浦东发展银行 Shanghai Pudong Development Bank Company Limited
10 中信银行 China Citic Bank Corporation Limited
11 中国民生银行 China Minsheng Banking Corp., Ltd.
12 中国光大银行 China Everbright Bank Company Limited
13 平安银行 Ping An Bank Co., Ltd.
14 华夏银行 Hua Xia Bank Co., Limited
15 北京银行 Bank of Beijing Co., Ltd.
16 广发银行 China Guangfa Bank Co., Ltd.
17 上海银行 Bank of Shanghai Co., Ltd.
18 江苏银行 Bank of Jiangsu Co., Ltd.
19 浙商银行 China Zheshang Bank Co., Ltd.
20 宁波银行 Bank of Ningbo Co., Ltd.
21 南京银行 Bank of Nanjing Co., Ltd.
22 重庆农村商业银行 Chongqing Rural Commercial Bank Co., Ltd.
23 渤海银行 China Bohai Bank Co., Ltd.
24 盛京银行 Shengjing Bank Co., Ltd.
25 上海农商银行 Shanghai Rural Commercial Bank Co., Ltd.
26 徽商银行 Huishang Bank Corporation Limited
27 恒丰银行 Evergrowing Bank Co., Ltd.
28 北京农商银行 Beijing Rural Commercial Bank Co., Ltd.
29 杭州银行 Bank of Hangzhou Co., Ltd.
30 广州农商银行 Guangzhou Rural Commercial Bank Co., Ltd.

Source: “The China Banking Association Released the List of the Top 100 Banks in China in 2021,” Sina Finance, July 14,
2021, https://finance.sina.com.cn/tech/2021-07-14/doc-ikqcfnca6848151.shtml.
Page 12 W31059

EXHIBIT 5: ZHUIYI’S DIGITAL AVATAR USE CASE IN PUBLIC AFFAIRS

Source: Zhuiyi Technology HK, “追一數字人 - Use Case - AI+5G Digital Avatar - Public Affairs,” February 25, 2022, YouTube
video, 2:16, https://youtu.be/t3cQ2mMxJGA.

EXHIBIT 6: PRODUCT MAP

Note: AI = artificial intelligence; Paas = platform as a service.

Source: Company documents.
Page 13 W31059

EXHIBIT 7: COMPARISON OF ON-PREMISES AND PUBLIC CLOUD MODELS

On Premises Model Public Cloud Model

Average lead time Three months Two weeks
Price range 1,000,000 RMB 300,000 RMB
Service frequency Once Subscribed annually
Average percentage of customization 15% 5%

Note: RMB = renminbi.

Source: Company documents.

EXHIBIT 8: MANAGEMENT TEAM BIOS

Name Bio

Master of University of Science and Technology of China. Former head of

Tencent AI Platform Department. Joined Tencent in 2006 and participated in
Joey Yue Wu
and presided over the construction of distributed file systems, big data
Founder & CEO
clusters, large web search engines, and recommendation engines. In 2020,
was selected into the Fortune list of China’s 40 Business Elites Under 40.

PhD, Huazhong University of Science and Technology. Former head of

Glen Yunfeng Liu technology research channel and technical director of Tencent Technology
Co-Founder & CTO Vocational Development Association, focused on research and practice of
search technology and machine learning, Tencent T4 expert engineer.

Master of Xidian University. Former product technical director of Tencent’s

search department, responsible for the R&D and management of systems
Alan Linding Wen such as Qzone photo album storage and Tencent’s unified image storage and
Co-Founder & CPO large-page search engine; presided over and implemented the R&D of
products such as Tencent Cloud Search, Tencent Wenzhi, and Public Opinion
Prediction.

Tsinghua University EMBA. Former Tencent marketing director responsible

Peja Xiao Hu for the marketing of the official website of the Beijing Olympic Games,
Co-Founder & CMO Tencent Group’s mobile Internet products, Sogou Input Method, Sohu Video,
and Sohu News.
PhD, Caltech. Former CEO assistant of Envision Group, senior director of
Jet Jie Cheng Tencent’s Strategic Development Department, senior project manager of
CSO McKinsey. Has rich experience in multi-industry strategic analysis, business
planning, new business development, investment, and financing.

Note: CEO = chief executive officer; Tencent = Tencent Holdings Ltd.; CTO = chief technology officer; PhD = doctor of
philosophy; CPO = chief product officer; R&D = research and development; EMBA = executive master of business
administration; Envision Group = Envision AESC Group Ltd; McKinsey = McKinsey & Company, Inc.
Source: Created by case authors based on company documents.
Page 14 W31059

ENDNOTES
1
Reinforcement learning is a machine learning training method based on rewarding desired behaviours and/or punishing undesired
ones. A reinforcement learning agent can take actions and self-improve the process.
2
Dollar amounts are in US$ unless otherwise specified.
3
IResearch Serial Market Research on China’s AI Industry (Ⅲ), iResearch, 2020, https://report.iresearch.cn/report_pdf.aspx?id=3707.
4
Global Artificial Intelligent Development White Paper, Deloitte, 2020, https://www2.deloitte.com/cn/zh/pages/technology-media-and-
telecommunications/articles/global-ai-development-white-paper.html.
5
Juan Jiao and Yingtao Xia, Three Types of AI Companies Will Compete with Each Other, Essence Securities, 2022,
https://pdf.dfcfw.com/pdf/H3_AP202201161540736143_1.pdf?1642344765000.pdf.
6
China AI Speech Recognition Market Research Report, Frost & Sullivan, 2020, https://pdf.dfcfw.com/pdf/H3_AP202012181441558366_1.pdf.
7
China Chatbot Industry Development Research Report, iResearch, 2021, https://pdf.dfcfw.com/pdf/H3_AP202107041501676711_1.pdf.
8
2020 Tencent Artificial Intelligence White Paper, Tencent Research Institute, 2020, https://ncstatic.clewm.net/rsrc/2020/0710/
14/f361bbdf4dd9b48fb0459c2849093d44.pdf; Chinese Conversational AI Industry Report 2019, LeadLeo, 2019,
http://pdf.dfcfw.com/pdf/H3_AP202010131421070207_1.pdf .
9
Zejing Liu, Conversational AI Track: The Wind Is Coming, Flowers Are Blooming, Huaxi Securities Co., LTD., 2021,
https://dfscdn.dfcfw.com/download/A2_cms_f_20210527113239490929&direct=1&abc2396.pdf.
10
Chinese Conversational 2019, LeadLeo; Michael Parker, “Why Contact Center AI Is a Big Deal for Customer-Focused Brands,”
Cresta, 2021, https://cresta.com/blog/why-contact-center-ai-big-deal/.
11
Chinese Conversational 2019, LeadLeo.
12
Juan Jiao, Develop New AI Infrastructure to Accelerate the Commercialization of Intelligent Driving, Essence Securities, 2021,
https://pdf.dfcfw.com/pdf/H3_AP202109051514350228_1.pdf.
13
Jiao and Xia, Three Types of AI Companies.
14
2020 Tencent, Tencent Research Institute.
15
“Baidu UNIT,” Baidu, accessed July 21, 2023, https://ai.baidu.com/unit/home#/home.
16
“Baidu UNIT,” Baidu.
17
In-Depth Report about Iflytek: Focusing on the Core Track, the AI Voice Leader Enters Strategic Interrogation, Gf Securities Co.,
Ltd., 2021, https://www.vzkoo.com/read/bbf2b6843873ecc3091fd2fe9f0312cd.html.
18
20182023 Report on Chinese Artificial Intelligent Industry Market Competition and Investment Development Prospects, Insights
and Info, 2017, http://wawamiki.com/baogao_/hulianwang/294424294424.html.
19
Yanji Yin and Yingtao Xia, The AI Leader Sets Sail Again, a New Journey for Large-Scale Commercial Use, Huaan Research,
accessed July 21, 2023, https://pdf.dfcfw.com/pdf/H3_AP202008101397540416_1.pdf?1601215122000.pdf.
20
In-Depth Report about Iflytek, Gf Securities Co., Ltd.
21
Deling Wei, “Iflytek: Hard-Core Realization Allows Unprecedented Coupling of AI and Communications,” CCTime, 2021,
http://www.cctime.com/html/2021-12-22/1603352.htm.
22
Wei Lv and Youwen Hu, AI C-Side Breaks the Game: IFLYTEK Apocalypse: Intelligent Bloomming 2021 Series Report No. 3,
Essence Securities, accessed July 21, 2023, https://pdf.dfcfw.com/pdf/H3_AP202103091470134891_1.pdf?1615327429000.pdf.
23
Chao Zhang and Hongtao Wang, Comments on IFLYTEK (002230.SZ) 2019 Annual Report: Artificial Intelligence 2.0 Stage, Starting
a New Journey of AI Large-Scale Landing, Avic Securities, accessed July 21, 2023, https://pdf.dfcfw.com/pdf/H3_AP20200427
1378692506_1.pdf?1587975511000.pdf.
24
“Xfyun,” Xfyun, accesssed July 21, 2023, https://www.xfyun.cn/.
25
China Chatbot Industry Development, iResearch.
26
“Bot Factory21 Conversational AI Platform: Quickly Build Your Own Robot,” Emotibot, accessed July 21, 2023, https://www.emotibot.com/
product/botfactory.html.
27
“Bot Factory21 Conversational AI Platform,” Emotibot.
28
China Chatbot Industry Development, iResearch.
29
China Chatbot Industry Development, iResearch.
30
China AI Speech Recognition, Frost & Sullivan.
31
Future Voice World: China Intelligent Voice Market Analysis, Deloitte, 2021, https://www2.deloitte.com/content/dam/Deloitte/cn/
Documents/technology-media-telecommunications/deloitte-cn-tmt-design-future-of-voice-report-2021dec.pdf.
32
Outbound Telemarketing – Global Market Trajectory & Analytics, Global Industry Analysts, Inc., 2021, https://www.research
andmarkets.com/reports/5030676/outbound-telemarketing-global-market-trajectory.
33
Eileen Yu, “China’s Personal Data Protection Law Kicks in Today,” Zdnet, October 31, 2021, https://www.zdnet.com/article/chinas-
personal-data-protection-law-kicks-in-today/#:~:text=Passed in August%2C the Personal.
34
Steve Lohr, “What Ever Happened to IBM’s Watson?” New York Times, July 16, 2021, https://www.nytimes.com/2021/07/16/
technology/what-happened-ibm-watson.html.
35
Trunk-based development is a version control management practice in which developers merge small, frequent updates to a core
trunk, or main branch. Developers can create short-lived branches with a few small commits compared to other long-lived feature
branching strategies; Kev Zettler, “Trunk-Based Development,” Atlassian, accessed July 24, 2023, https://www.atlassian.com/
continuous-delivery/continuous-integration/trunk-based-development.
36
Rita Liao, “How Meituan Is Redefining Food Delivery in China with Drones,” TechCrunch, December 29, 2021,
https://techcrunch.com/2021/12/29/meituan-food-drone-delivery-china/?guccounter=1#:~:text=The drones deliver to designated,up
its aerial delivery ambitions.
37
Jingli Song, “Alibaba Opens Its Conversational AI Tech, Expanding Its Use beyond Smart Speakers,” KrAsia, October 23, 2020,
https://kr-asia.com/alibaba-opens-its-conversational-ai-tech-expanding-its-use-beyond-smart-speakers.
38
“Emerging AI Enterprises: 6 Chinese Companies Including Zhuiyi Technology Were Selected as the Top 100 AI Startups,”
Chinanews, 2022, https://www.chinanews.com.cn/business/2020/03-05/9114394.shtml.