The AI Landscape in

Southeast Asia
August 19, 2024
Overview
At Kaya Founders, we believe that AI will be a fundamental technology shift that will shape business
and society for the coming century. Although we are still in the relative early innings of this shift, we
believe it is important to contribute our point of view on the topic for our investors, our companies,
and our partners to inform their own opinions. This paper aims to speak to a broad audience and
assumes no previous expertise on the topic of artificial intelligence.

For those looking for an introduction to the topic, sections 1 and 2 provide a brief grounding on
definitions and a framework on how to think about the components that make up AI as a technology.
Thank you to the Amazon Web Services team for their contribution to this section. For those already
familiar with the topic, sections 3 and 4 provide specific analysis on the Southeast Asian region and
what we’re seeing based on interviews with companies in the Philippines and our database of over
500 startups working on AI across the Philippines, Singapore, Vietnam, Malaysia, and Thailand.

Lastly, as testament to the power of the technology shift, AI tools were used across the research, note
taking and synthesis, analysis, writing, editing and graphic construction in this paper allowing us to
cut the time required for most tasks by at least 50%.

If you want to continue the conversation, you can reach out to us at [email protected] or
kayafounders.com.

Page 2
Section 1: Defining AI and GenAI
Intelligence can be defined as the ability to learn and perform suitable techniques to solve problems
and achieve goals, appropriate to the context in an uncertain, ever-varying world. A fully
pre-programmed factory robot can solve specifically identified problems and achieve scoped goals
but would not be considered to be intelligent due to its inability to problem solve and understand
context (even identifying an incorrect part on an assembly line is a task beyond the intelligence of
most factory robots).

Artificial Intelligence is then often used to refer to technologies that mimic or exhibit human-like
intelligence to learn, reason, perceive, and act autonomously. The term was first coined by Stanford
Professor John McCarthy in 1955, who already at the time had proposed and was conducting
research into a prominent use case of AI of 2024: self-driving cars.

So while AI itself has already passed its quinquagenary, in just the past 7 years there have been
large new breakthroughs in the field of Generative AI (“GenAI”) using GPT (Generative Pre-trained
Transformer) models. Previous AI research focused on performing specific tasks or making
predictions based on existing data and predefined rules or algorithms (often referred to as predictive
AI). Generative AI in contrast focuses on creating entirely new data or content that resembles
human-generated content across mediums such as text, images, videos, audio, etc. The capability of
GPTs to use natural language to successfully respond to a variety of prompts in different disciplines
led Bill Gates, upon trying OpenAI’s ChatGPT (one of market-leading GPTs currently on the market),

Page 3
to comment “I knew I had just seen the most important advance in technology since the graphical
user interface”. In 2024, GenAI models already can outperform humans on various simple tasks and
are approaching the baseline on others.

AI Index Technical Performance Benchmarks vs Human Performance

If GenAI is the next revolution in humankind’s development, the world will soon look very different
with categories being disrupted or forged and winners and losers created in each. We are optimistic
and excited about the potential the technology has for Southeast Asia and the Philippines. This
document aims to communicate how we currently see the world of GenAI in 2024 in the region and
what opportunities we are excited by.

Page 4
Section 2: The AI Value Chain (“The
Stack”) in 2024
The AI value chain is quickly evolving with a mix of commercial, academic, and government players
participating in almost every part of the stack. We prefer to simplify the complex value chain to a set
of four components that need to exist for AI to be used. These components are:

1. Hardware foundation: the physical silicon that is being used to run the various software
layers and the energy required to power the hardware. Compute provided directly by chips
such as NVIDIA GPUs or cloud services such as AmazonSagemaker are common 2024
examples.
2. Software foundation (algorithms): the specific model architecture (the most common in 2024
being “transformer” models) and the instance of that architecture that is being used.
Facebook’s Llama model or ChatGPT-4o (specifically the model) are common 2024 examples
of instances of transformer models.
3. Data: the information being used to train the model and how it is collected, processed,
stored, and integrated. Scale AI and Pinecone are examples of large but recent businesses
who started explicitly to power AI applications serving the processing/labeling and storage
subsegments of the data stack, respectively.
4. Applications: the end-user facing software or hardware being built on top of the rest of the
stack in order to solve the specific end use case. The Humane AI pin and the ChatGPT 4o
mobile app are examples of popular consumer applications in 2024; however, B2B focused
applications for businesses and developers are also present in the market.

Large industry players, usually with existing technology solutions, also have begun to provide
products aimed at making application development as simple as possible by providing integrated
solutions that provide computing power via hardware and infrastructure, access to foundational
models, as well as application development solutions. Amazon Web Services (AWS), for example,
offers Amazon Bedrock, which allows companies to build generative AI solutions with existing
foundation models and LLMs; AWS Trainium and AWS Inferentia – purpose-built AI chips to train
and deploy foundation models on the AWS Cloud; and applications like Amazon Q – a generative
artificial intelligence (AI)-powered assistant for accelerating software development and leveraging
companies’ internal data.

Page 5
To date, most of the investment has been in building the “infrastructure” parts of the stack that will
support the industry: chips that will provide computational power to train and run models, data
storage and management that will capture data for models to use, and the foundational models
themselves which will be leveraged for specific applications. With the recent emergence of early
options for this infrastructure (e.g. NVIDIA for chips, Scale.ai for data management, and ChatGPT
for model), there has been investment in more obvious use cases such as customer support,
agent-based chats, and health care (mainly in diagnostics). However, investment in other use cases
has still been limited.

Page 6
Hardware foundation
Although consumer attention has primarily been directed toward free-to-try applications such as
ChatGPT (a text-based general assistant) and Midjourney (an image generator based on text input),
the largest amount of investment since the start of the publication of Google’s transformer paper in
2017—which many regard as the beginning of the GenAI industry—has been into the hardware that
powers the training and inference of models. Similar to the gold rush in San Francisco in the late
19th century, when entrepreneurs such as Levi Strauss made their riches selling the picks, shovels,
and jeans to the various outfits, the likes of NVIDIA and GROK—which manufacture hardware adept
in parallel processing, a key requirement for efficient model training—have been some of the early
winners in the recent AI boom. While NVIDIA’s GPUs were initially designed for use cases such as
video games, which happened to have overlapping requirements with AI training (namely the ability
to do heavy amounts of matrix multiplication), GROK is an example of an investment into AI-specific
hardware optimized specifically for training and inference of AI models. In the short term,
performance of hardware, regardless of use case optimization, has room to grow as companies are
able to continually add more computing power into the same-sized chips. However, industry experts
expect that based on the historical pace of innovation, we will begin approaching the physical limit
of how many transistors (what determines compute power) can fit into a GPU, after which continued
hardware performance will be harder to achieve.

Page 7
Companies face a strategic decision between buying their own chips or utilizing cloud services for
their computational needs. Owning hardware offers full control and potentially lower long-term costs,
but demands significant upfront investment and ongoing maintenance. Conversely, cloud services
like Amazon Web Services and Microsoft Azure provide flexibility, access to technology upgrades,
and lower initial costs (a single new NVIDIA GPU costs around $40,000 currently and multiple units
are required even for small AI startups), but they can become expensive over time and introduce
dependency on third-party providers.

In 2024, increasing regulations and security concerns are prompting more businesses and
governments overall to consider local hardware solutions. In China, for example, Huawei has
developed its own AI chips, such as the Ascend series, aimed at reducing dependence on foreign
suppliers. Alibaba's DAMO Academy is focused on AI research and has introduced the Hanguang
800, a highly efficient AI inference chip. Baidu has also launched its Kunlun AI chips, designed for
various AI applications including autonomous driving and cloud services. Shifts to onshore or
on-premises demand greater investment in physical infrastructure and skilled personnel to manage
these systems.

Lastly, though not an AI specific technology, a critical consideration is the substantial energy burden
associated with powering the sophisticated hardware required for generative models. Advanced
chips consume significant amounts of electricity, contributing to higher operational costs and
environmental impact. By 2036, the amount of power consumed by the AI industry is expected to
double, and constitute around 6% of electricity used in the United States. To lower costs and limit
environmental impact, data center operators are experimenting with more efficient cooling and heat

Page 8
reuse technologies with private companies like Aligned and Stack launching high-density,
water-cooled racks as an example.

Algorithms
The software foundation of the GenAI industry revolves around advanced algorithms that power AI
models. At the forefront in 2024 are transformer models, renowned for their efficiency in handling
large-scale data and complex tasks. Transformer architectures, introduced in 2017, revolutionized
natural language processing (NLP) and other AI fields due to their ability to handle sequential data
more effectively than previous models. Prior to transformer models, training and inference times
were magnitudes of order longer which, in addition to impracticality, meant much higher cost.
Transformers also outperform when it comes to capturing context and relationships in data, leading
to superior performance in tasks such as language translation, text generation, and image
processing.

Prominent instances of transformer models in 2024 include Facebook's LLaMA 3 (Large Language
Model Meta AI) and OpenAI's ChatGPT-4o. Both illustrate key trends occurring in the algorithm
space today. ChatGPT is a closed source model that cannot be cloned and built upon by outside
developers. In contrast, LLaMA’s source code is open-source meaning it is available for free for
anyone to download and build on top of. On average across benchmarks, ChatGPT is the best
model available on the market but comes at the cost of dollars for usage (the latest ChatGPT models
are behind a paywall as of the writing of this paper) and opacity of the underlying technology.
LLaMA 3, on the other hand, is slightly behind ChatGPT 4o in performance but the open source
nature of LLaMA 3 makes it easier for other companies to deploy AI of their own (benefiting
Facebook as less companies need to rely on purchasing from their competitors) and also makes it
easier for Facebook to improve the model using feedback from the open source community.

Page 9
The latest iteration of both models also demonstrate the ability of the models to use not just text as
input but images and audio as well (known as “multimodal”), unlocking use cases such as real time
translation (which relies on information such as pauses or inflection that are harder to pick up from
just text) or live math tutoring using a whiteboard for diagrams.

Lastly, the fact that both the closed source and open source leaders are from US institutions reflects
the current dominance of the country in terms of model development. Countries such as China,
whose political parties have a strategy of controlling information access within their borders, have
reacted by limiting the consumer and business use of foreign AI models (particularly closed source
ones such as ChatGPT) and by funding the development of models both by private and state entities.

Page 10
Data
The data portion of the AI stack encompasses several critical stages, including raw data collection,
data processing, storage, and access. While fossil fuels were the driver of many of the products and
service innovations that defined the 20th century, data has been called the oil of the 21st century
due to its value in driving future technological innovation. Assuming similar quality of data and
power of hardware, the more data you have, the better your model can perform (the tradeoff
between these factors has been a topic of debate among researchers).

For the first iterations of GenAI models that led to the models we see today, the internet provided a
free and easily accessible data set for companies to train on (the data a model is trained on is often
referred to as a “corpus”). Increasingly, there is a growing concern from researchers that most of the
quality, publicly available information has already been used up. The implication is that the huge
increases in performance from models that we’ve seen, which have been in huge part to being able
to train on more and more data, will be constrained by the lack of new data to train on. This has led
to four emerging trends:

1) Licensing of private data: Model developers signing (typically nonexclusive) licensing deals
to access data typically behind paywalls or not indexed online. OpenAI’s licensing deals to
use the content of the likes of Vox media, Atlantic media, and Reddit are the early examples
of what will become mainstream.

Page 11
 2) Improving the quality of available data: Good model performance relies on good quality
data. Less data of a higher quality typically outperforms a large quantity of incorrect or
difficult to interpret data. Various companies offer services to transform low quality data into
high quality data using humans to review and edit datasets. New entrants such as Scale AI
that moved early on the space have been accompanied by traditional outsourcing firms like
Telus which quickly pivoted their service lines towards AI clients.
3) Improving capture of proprietary data: Companies turning attention and resources to
capturing data from within their existing business by digitizing manual processes, increasing
logging of online interactions, and investing in data infrastructure to power both of the
above. For example, Tesla leverages data from the cameras and sensors in its vehicles to
enhance its autonomous driving algorithms. This vast and proprietary dataset allows Tesla to
continuously improve its models, outpacing competitors who lack access to such
comprehensive data sources.
4) Synthetic generation of data: Using computers (and increasingly dedicated AI models) to
create simulations to produce data rather than collecting data from the real world.

As data collection expands, so does the risk of exposing sensitive information both through security
breaches to access stores of data but also from manipulating models to reveal sensitive data
themselves (a nefarious use of prompt engineering). The growing concern over data privacy has led
to increased investment in secure data storage solutions and privacy-preserving technologies as well
as the creation of internal and external teams with the mandate of setting and testing privacy policies
for models.

Because models make statistical predictions using their underlying training data, biases in the
composition of the underlying data can result in biases in the results of models. Given that much of
the training data being used is documented, historical data from the open internet, various
publications have cited biases in models indexing towards caucasian faces for facial detection or
gender associations based on historical circumstances (e.g. associating women to be more likely a
nurse or teacher in outputs involving generating a profession). Companies have created specific
mandates to provide a conscious attempt to include more diverse sets of data or to develop
techniques to monitor or adjust for bias.

Page 12
Applications
Applications are solutions to use cases and generally can be organized into pillars of automation,
augmentation, and autonomy. Which solution is appropriate is determined based on requirements of
speed, reliability, cost, complexity and explainability. We illustrate the difference in application types
below using business process outsourcing (BPO) as an example domain for use cases. So far,
contact center automation has been one of the most common early applications for industry players
looking to create value with GenAI.

An organization looking to implement a solution like the ones above a couple years ago would have
had to build capabilities in-house, but as companies like Stripe did with payments, an increasing
number of B2B companies which offer technology that reduces barriers to entry by providing full
stack (or specific parts of the stack) solutions (including low and no code solutions) to build
applications are an available option.

With the rapid pace of deployment of applications, organizing them into three subcategories is
helpful in order to understand the method through which the application attempts to solve the use
case.

Page 13
Subcategory 1: Automation
Automation involves taking a well-defined process with known inputs and outputs, and using AI to
perform it faster, more reliably, and/or at a lower cost. There is typically a low amount of
decision-making required for the success of the process, making it low complexity. And because the
inputs and outputs are known and easy to audit, explainability of how the AI system works usually is
not a high priority. Core implementation issues are usually to do with ensuring the technology fits
seamlessly into existing workflows. Integrations with existing tools or software as well as clear handoff
points for when the software takes over and when it creates the desired output, are crucial.

In the BPO field, use of natural language processing to turn paper records into digital records, and
automatically sending these records to the relevant location based on the record type, has been a
use case in place prior to GenAI. The capability of GenAI to better understand written context has
led to better performing models that can successfully translate a larger amount of inputs even if they
are not in the same format or can write certain outputs in natural language, enabling automation
further down the process chain.

Subcategory 2: Augmentation
In contrast to automation that focuses on solving low complexity problems, augmentation focuses on
assisting humans with tasks that are high complexity. Models in these domains need to have a higher
amount of explainability in order for human reviewers to make a judgment on whether to use their
output, but can be less quick given the problems being solved often already have long solving times.
Because results are reviewed by humans who use judgment on whether to apply the output or
deviate from it, models can be less reliable than those for automated tasks where the outputs of
models are relied upon to be correct.

“Copilot” applications designed for specific types of knowledge workers have been the most
common examples of augmentation thus far. For a BPO agent, a copilot that is listening in on voice
calls and is able to bring up customer information, relevant policies, recommended actions just by
listening to the voice interaction between the customer and agent is a possible application that can
already be deployed today.

Subcategory 3: Autonomy
While automation focuses on low complexity problems with a known output, and augmentation
focuses on complementing human ability to solve high complexity problems where the output is not
necessarily known, autonomy focuses on solving problems end-to-end, where the objectives and

Page 14
constraints (“rules of the game”) are known but the outcome is not necessarily known. Often referred
to as “agents”, these models often use a conjunction of different models to produce different outputs
combined with a “reward model” which tells it how to use the outputs in order to maximize its
“score” within the rules of the game. These applications most resemble interacting with humans,
often creating outputs that are multiple steps removed from the initial input, and thus have a high
burden of explainability for humans to be able to trust and verify the resulting output. Cost and
speed matter less than the ability for the model to create high quality answers where the “thought
process” behind them can be audited.

Chat agents for BPOs are the most common use case today. These agents are able to automatically
reply with seemingly “intelligent” and bespoke answers regardless of inputand improve upon
previous chat agents which could only handle inquiries about specific topics and give pre-written
responses.

Section 3: Zooming in on SoutheEast

Asia
Both at the top down government level as well as bottom up at the individual corporation level,
players in Southeast Asia generally have understood the importance of prioritizing GenAI research
and adoption.

On the public side, countries such as Indonesia and Singapore have labeled GenAI as a national
priority and created government programs that involve setting national strategies and regulatory
frameworks as well as direct investments to encourage the development of the technology within
their borders. The Association of Southeast Asian Nations (ASEAN) has put forward a set of
guidelines for governments and businesses to follow as they develop and use AI systems in the
region. This non-binding approach aims to provide oversight of AI technologies while ensuring they
remain accessible and beneficial to the region's growing digital economy.

On the private side, most company CEOs are moving quickly to at a minimum explore how to
incorporate GenAI into their organizations. In a recent PWC CEO survey, 2 in 3 APAC CEOs
reported they had already adopted GenAI across their companies in the last 12 months. In the
majority of cases, the technology was not expected to immediately replace people, with only 26% of

Page 15
CEOs predicting a reduction in headcount due to GenAI. The topic of headcount reduction has
been a focus in countries like the Philippines given the size of the BPO industry (making up around
8% of Philippine GDP) with the conclusion so far being that countervailing forces of increasing
offshoring from foreign firms focusing on profit margins are masking the effect of reduced need for
human labor due to AI. Real-world proofpoints such as Klarna which laid off staff and froze hiring
directly attributing part of the motivation to the utilization of AI suggest that there is a structural shift
happening which may outlive the current wave of increased offshore activity.

Impeding GenAI implementation in the region (with the exception of Singapore) are challenges
around data infrastructure and human capital, which are two of the main factors identified by
research firm Oxford Insights as barriers to implementation in East Asia specifically. This fact has
been recognized by many global players in the space. Major providers of cloud data infrastructure
have been expanding into the region due to lagging penetration of modern data infrastructure. More
recently in 2024, Microsoft announced >$4B in commitments across Malaysia, Thailand, and
Indonesia to support digital transformation, AI skilling, and developer communities with the goal of
equipping 2.5 million people to participate in AI. In the development of this paper, multiple AI
professionals interviewed echoed the same concern that companies were putting the “cart ahead of
the horse” and trying to develop GenAI applications without having solved the data part of the stack
first.

Page 16
Regulatory Landscape
Singapore: Top down, national strategy. Creating governance frameworks. Has budgeted and
funded direct development.
Singapore has launched 25 initiatives, including the National AI Strategy, to foster a trustworthy
environment for AI development. This approach extends to the proposal of a Model AI Governance
Framework for Generative AI in January 2024, addressing the rising use of this technology globally.
Singapore's Infocomm Media Development Authority (IMDA) has also consulted on a draft Model AI
Governance Framework for Generative AI, outlining areas where future policy interventions may take
place and options for such intervention. Above frameworks, the government also announced it had
budgeted spend of US$500M over the next 5 years for AI.

Indonesia: Top down, national strategy and heavily encouraging foreign investment. Focused on
applications in healthcare, education, bureaucratic reform, food security, and smart mobility.
Indonesia has taken a proactive stance towards AI by publishing a national strategy for artificial
intelligence. The strategy focuses on four key missions: Unleashing the Power of Collaboration for a
Transformed Nation, Empowering Exceptional Minds with Character and Competence, Building an
Ecosystem for Data and Infrastructure, and Upholding Ethics and Values for a Harmonious Digital
Future. This comprehensive approach is expected to drive innovation and growth across various
sectors, including health, government reform, education and research, food security, and smart
mobility. The government has sought to partner with external parties to provide funding for AI
initiatives in the country with private players like Microsoft and NVIDIA both planning to invest
heavily in software and hardware in the country and the technology ministers approaching the
Chinese government to open doors for collaboration on the topic.

Page 17
Malaysia: Announced various initiatives and has private partnerships focused on upskilling.
Malaysia created a digital ministry in 2024, recognizing the need for a dedicated brand of
government to deal with AI in the country. One of its first initiatives is the “AI For The People ''
initiative, developed with Intel, which offers free online AI courses to upskill the workforce, aiming to
future-proof the economy. The strategic partnership between YTL and Nvidia is establishing a
US$4.3B AI cloud and supercomputer infrastructure to support local startups and digital
transformation. Additionally, under the “Bersama Malaysia” initiative, Microsoft is committed to
upskilling 1 million Malaysians by the end of 2023. These efforts, along with developing
regulatory frameworks and public-private collaborations, aim to leverage AI for enhancing
productivity and economic growth.

Vietnam: Top down, national strategy. Announced various initiatives and has private partnerships
focused on upskilling.
The Vietnamese government is implementing a comprehensive strategy to advance AI development
and integration, aiming to position Vietnam as a leading AI player globally by 2030. The Ministry of
Science and Technology (MoST) is promoting a national AI strategy focused on research,
development, and application, targeting a top-four ranking in ASEAN and top 50 worldwide. This
involves creating a robust legal framework, improving data management, and fostering international
collaboration. The National Data Strategy, approved in February 2024, aims to enhance connectivity
among data centers and support high-performance computing. Key initiatives include a collaboration
with Google, launched in July 2024, providing 40,000 scholarships and training courses like
"Google AI Essentials" to develop AI skills and support startups. To drive AI innovation, the
government has established AI Centers of Excellence and offers tax incentives for AI adoption. Legal
frameworks for AI applications, including virtual assistants, are being developed to ensure
responsible AI use. Plans include creating 10 AI brands, developing three national centers for big
data, and establishing 50 open datasets. Vietnam is also engaging globally, co-sponsoring AI
resolutions and learning from developed countries.

Philippines: Early strategy and initiatives focussed on upskilling. Limited funded programs on
technology development thus far.
The Philippines was quick to announce plans in reaction to the recent AI boom. In 2021, the
Department of Trade and Industry (DTI) announced a national AI strategy that aims to accelerate AI
adoption, boost industrial development, and generate high-quality entrepreneurship and job

Page 18
opportunities. Key components include the establishment of the National Center for AI Research
(NCAIR) and promoting the country as a hub for data analytics and AI services. At the writing of this
paper, most initiatives have yet to be funded. This strategy was updated in July 2024 branded as the
National AI Strategy Roadmap 2.0 and aims to transform the Philippines into a premier destination
for AI-driven innovation. It emphasizes strategic imperatives such as improving data access,
nurturing future AI talent, and fostering an ethical AI ecosystem. Multiple upskilling initiatives and
frameworks have been announced, some with private organizations like Microsoft, that focus on
developing IT and AI skills. The Philippines Skills Framework from the Analytics and AI Association
of the Philippines serves as one example of such a framework, developed in collaboration with the
Department of Education.

Southeast Asia: Current Company Focus Areas

Commercial activity for GenAI in Southeast Asia so far has existed across large enterprises,
mid-stage start ups incorporating AI into their product offerings, and new early stage start-up
creation. As a venture fund, we wanted to focus on empirical insights specifically within the startup
sector. In order to do this, we created a database of private startups that had recent news or filings
related to their developing AI products from the past 3 years. Our analysis found that the vast
majority of companies (87%) were developing applications, specifically in the B2B space. 70% of
these companies were headquartered in Singapore, with Indonesia being the next largest country of
residence at 15%.

The skew away from pure play AI infrastructure companies developing chips or foundational models
has been due to the high price to develop such technologies. Taking algorithms as an example, the
development of Google’s latest version of Gemini (their flagship foundational model) was estimated
to have cost US$100M+ due the increasing amount of processing power and data required to train
improved foundational models. Given the large upfront cost and high amount of risk due to rapid
evolutions in technology, private players are unlikely to try and compete to create a foundational
model without the sponsorship of the state. Instead we find that most companies are focusing on
applications, where initial products are cheaper and faster to develop and rely on “as a service”
providers to allow them to leverage the work being done at the cutting edge of the infrastructure
stack.

Page 19
Across the companies we spoke to deploying GenAI in production to real customers, the majority of
efforts fell into one of eight buckets:

Applications
1. Data processing automation: using GenAI to extract both structured and unstructured
information (usually in the form of paper or digital forms/text) and convert it into a
standardized set of information needed as an output from the process. Fr8Labs for example
is using AI to automate the process of creating booking based on complex booking requests
or creating invoice records in a database from images of invoices using different formats to
reduce the amount of time spent by freight forwarders on repetitive tasks.
2. Sales and marketing augmentation and autonomous agents: automating the asynchronous
portions of the sales and marketing process as well as augmenting sales reps on high stakes
touchpoints. Gathering and scoring leads, triaging inbound leads and generating automatic
responses, surfacing relevant information about leads during sales calls, marketing content
and copy creation were common use cases. Sprout Solutions has created an AI agent that
autonomously qualifies and nurtures leads through email and messaging then once leads
book demos, hands those high quality leads over to human salespeople.
3. Customer service augmentation and autonomous agents: helping primarily voice agents be
more effective on calls or create fully automated chat agents to respond to customers without
human intervention. Augmenting reps with data from a customer database during calls and

Page 20
 autonomous chatbots were the common use cases. Trusting Social has created an
autonomous support agent focused on specific industries (such as banks) that are able to
answer customer questions at all hours of the day and create tickets/escalations for human
review when required.
4. Knowledge worker augmentation via insight generation: Internal tools trained on relevant
internal and external knowledge bases to make finding and analyzing information easier.
Thinking Machines has developed a service line specifically to help enterprises fine-tune and
deploy AI assistant agents to augment frontliners and knowledge workers.

Data
5. Data labeling service providers, both from traditional BPO providers and upstarts.
Mettamatch focuses on training communities of Women in semi-urban and rural parts of the
Philippines on tasks such as computer labeling image datasets for self-driving car companies
and sourcing projects from companies that need those forms of datasets labeled.
6. Government sponsorship for creation of “local” data sets. The Philippine Department of
Science and Technology – Advanced Science Technology Institute (DOST-ASTI) has
launched the iTANONG project which focuses on creating a dataset that can be easily used
as a corpus by AI model developers to improve the ability of models to respond to queries in
Filipino and “Taglish” which blends words and phrases in both English and Filipino.

Algorithms
7. Fine-tuned models on top of open source algorithms. Companies who have proven product
market fit are beginning to migrate from foundational models to their own implementations
of open-sourced models due to cost concerns at scale, the lack of a need for a larger model
vs a smaller more specifically trained model, and data privacy concerns related to sharing
private data with 3rd parties.
8. Government sponsorship for country specific models optimized for local use cases.
SEA-LION (Southeast Asian Languages in One Network) is a Singapore government-led
initiative by the AI Singapore organization trained in the region's languages and cultural
norms. Trained on data in 11 Southeast Asian languages including Vietnamese, Thai and
Bahasa Indonesia, the open-sourced model is a cheaper and more efficient option for the
region's businesses, governments and academia.

Hardware was not a production use case for any of the companies we surveyed in SEA.

Page 21
Section 4: So What? Our Perspective
First, for almost all companies, there is already an opportunity to incorporate GenAI to drive real
business value. For medium- sized and larger companies, GenAI can improve the ability to operate
at scale. The easiest starting place to incorporate GenAI is replacing repeated tasks and helping to
generate insights from internal information. Companies may have tools such as ERPs or analytics
software already implemented for these use cases, but there are opportunities to both supplement
these tools with GenAI to further save employees on time or even replace these tools with a lower
cost solution. For smaller startups who don’t have expensive processes yet that they need to operate
at scale, GenAI can add the most value by helping to give founders leverage in areas of work they
usually would spend cycles hiring for or freeing up time from bespoke tasks that are one-off and not
worth hiring for. For founders spending a lot of their time generating marketing or sales leads,
multiple free/low cost solutions exist for generating content quickly to quickly experiment and iterate
on content and tools that plug into email clients + LinkedIn to automate the up-funnel portion of the
sales funnel.

Page 22
Second, for established companies implementing GenAI in SEA, the easiest way to deploy
applications quickly and validate/prove the AI use case will be first leveraging external service
providers to provide almost all parts of the stack. Use cases that don’t involve high-risk data such as
querying public data sources (e.g. aggregating the latest price information from competitor
websites) or providing responses to customers based on your public information (e.g. replacing
extensive FAQs with a chatbot that is trained on FAQ data) are good initial applications if internal
data security and privacy policies still need to be evaluated. As the business gains familiarity with
and gives sponsorship to integrating GenAI, you can evaluate whether moving to more advanced
architectures or in-house parts of the stack make sense from a cost, efficiency and security
perspective. Keep in mind however that the industry is still seeing a rapid pace of innovation which
has led to “free” increases in performance for customers who have remained on third-party parts of
the stack. Another local nuance we observed was driven by the fact that data infrastructure in
companies implementing GenAI tended to be less advanced than those outside SEA, leading to
local applications that relied on “humans as the API” (human-based input or response as part of the
process) rather than regular software APIs. As a result, chat software integrations that allow
employees to interface with the AI layer in the same way they would interface with a coworker or
customer, are a unique differentiator for SEA specific applications.

Page 23
Third, for new startups in the region, based on our survey of startups across SEA, a lot of the
attention has been building B2B software to solve problems either across the stack for specific
industry verticals or to build for specific horizontal functions that exist across many industries. Early
GenAI applications within businesses have tackled areas such as customer service, sales,
accounting, legal, and HR, where documentation and rules about what should happen in different
scenarios is more likely to have preexisted.GenAI applications for job functions such as field
operations and sales, where the end user may not have a laptop and high speed internet, we believe
are viable businesses that are still rare in the global market. Making use of smaller, cheaper
specialized models designed to run “on the edge” (on the device using the model rather than on a
remote server) would be uniquely valuable for the SEA context. There is also a large amount of
whitespace to be captured within the B2C segment, with less companies focused on building for
consumers directly. The ones that have been built for consumers are focusing on making it easier for
them to generate content such as videos or blogs.

Illustrative SEA companies per category

Page 24
Lastly, in terms of investment opportunities, we believe there is currently an opportunity within the
data component of the stack and the opportunity to build low cost applications to provide large scale
access to services that traditionally have relied on either rare or expensive human capital.

In the data stack, companies that can partner (or already have partnered) with large incumbents with
pools of historical data or can integrate (or already have integrated) across fragmented players in
large industries can create valuable data sets that can be used for local applications. This looks like
1) helping companies capture their proprietary information (by leveraging IoT, improving data
logging online, or introducing processes to ensure data capture by humans, 2) converting stores of
information whether in physical, digital image, or messy database form into secure and performant
databases, and 3) setting up analytics and measurement within key parts of the business to illustrate
opportunity and upselling GenAI applications that solve those opportunities.

The ability for GenAI models to replicate human ability to synthesize information, communicate, and
understand questions at a low cost remove the bottleneck of needing high cost human capital such
as teachers or doctors, unlocking the potential of products for verticals that have been traditionally
difficult to monetize at scale such as education or primary and preventative care. As social media
companies from the past decade can attest to, digital penetration and engagement in the region is
extremely high by global benchmarks, although this is primarily delivered by low cost mobile
devices consuming low amounts of bandwidth at limited speeds. Companies that can develop low
cost, high efficiency models and user interfaces will be able to serve millions of consumers with

Page 25
high levels of demand but are traditionally ignored by local and international companies due to
unworkable unit economics.

Because of the ease and low cost to deploy an AI application, our point of view is that early stage
investing will take the shape of a larger number of smaller sized bets, primarily on applications,
accompanied with a higher expectation of development velocity given the rapid pace of the
underlying technology and the increases in productivity that AI unlocks generally for teams. On the
other side of this coin, our data shows that almost half of startups in the AI space in SEA have been
able to bootstrap their product. Startups that raise funding will likely need smaller checks to cover
compute costs of initial pilots and maybe spend on traditional sales/marketing in order to validate
their idea in the market. Making sure that ample funds for follow-up funding are lined up from the
initial check will be important. As startups move from pilot phases to scaling and taking a pilot
product to market, companies will increase spend on their technology stack in-house as well as
hiring human capital, both of which will likely be pushed back relative to the usual trajectory for
SaaS startups.

Page 26