When I moved to Singapore in 2022, many of the enterprise conversations I was having about AI centred on a fundamental question: can we trust the model?
Eighteen months later, generative AI had changed the conversation. The question was increasingly: how do we deploy these models while keeping them reliable, governable and useful to the business?
During those years, I worked on machine-learning and AI deployments with banks, insurers and public-sector organisations across Asia-Pacific. At the same time, I led the Data, AI & Analytics community at La French Tech Singapore and co-founded the Singapore Machine Learning meetup.
Those two experiences complemented each other. One gave me a view from inside enterprise AI projects; the other created a space to compare experiences with practitioners, technology companies, institutions and researchers.
Looking back, three lessons stand out.
I joined DataRobot in Hong Kong in 2018 as its first customer-facing data scientist there. At the time, machine learning was already becoming significantly easier to build and deploy.
Yet technically successful projects could still stall.
The reasons were often surprisingly mundane: a model-risk team had not seen the necessary documentation; a business owner did not fully understand what would change once the model was deployed; or nobody had been given clear responsibility for taking a successful pilot into a real business process.
Over time, this changed the way I thought about AI adoption.
A good model is necessary. But it is not enough.
One approach that worked particularly well was pairing the customer's own data scientist—someone who understood the data, the organisation and its constraints—with an experienced field engineer who knew how to take models into production.
The technology and the organisation had to move together.
That experience left me with three questions I would now ask before launching an enterprise AI pilot:
If those questions do not have clear answers, improving model performance alone may not solve the adoption problem.
My second lesson came partly from the Singapore ecosystem itself.
In early 2023, I joined La French Tech Singapore and took on its Data, AI & Analytics community. At the time, practitioners were often hearing two very different AI conversations: what technology providers said was becoming possible, and what their own governance, risk and compliance teams said was acceptable.
There were relatively few places where those perspectives could meet.
We tried to create some.
An “AI in Banking” event with Google in March 2023 put banking practitioners into the discussion. Later that year, we organised a generative AI experts panel. In November 2024, IMDA hosted “Generative AI, Year 2: Look Back and Forward”, where I moderated a discussion with Huawei, Microsoft and K&L Gates about what had survived the first year of enterprise experimentation.
In April 2025, we co-hosted “Scaling with Intelligence” with AI Singapore at NUS, looking at another practical constraint: how organisations could build the teams needed to scale AI.
What I found valuable about Singapore was not simply the number of AI events. It was the willingness of different parts of the ecosystem to participate in the same conversation.
Public agencies and professional bodies were often directly involved. I spoke about IMDA's Project Moonshot, its open-source toolkit for evaluating large language models, at the Singapore FinTech Festival in 2024. I also participated in the Singapore Actuarial Society's Data Analytics Committee and later joined the programme board of A*STAR's Advanced Remanufacturing and Technology Centre.
Across these different experiences, a pattern emerged.
Among the teams I worked with that successfully moved generative AI toward production, evaluation and governance were increasingly treated as part of the product—not as compliance exercises added once the technology was ready.
That distinction matters.
If governance only enters after the pilot, it can become a gate. If questions about evaluation, risk, explainability and accountability are considered from the beginning, they can become design parameters.
For organisations experimenting with AI, the practical lesson is simple:
Do not wait until the end of the pilot to ask whether the organisation can deploy what you are building.
The third lesson took me longer to recognise.
The extraordinary progress of large language models naturally focused attention on text: assistants, copilots, search, summarisation and increasingly autonomous agents.
But ask a bank, insurer or industrial company about some of its highest-value predictions and many of the answers are not paragraphs. They are rows and columns.
These problems depend heavily on structured, or tabular, data: the transactions, customer records, claims, prices, sensor readings and operational data stored in enterprise systems.
I had spent years working on these kinds of machine-learning problems before following much of the industry toward language models. Conversations with enterprise customers eventually brought me back to a simple question: could some of the advances behind foundation models also change how we work with structured data?
That question is what eventually took me from Singapore to San Francisco.
In January 2026, I joined Fundamental to lead Applied AI. The company is developing foundation-model approaches for structured data, and my role is focused on translating those capabilities into enterprise applications.
For me, however, the broader lesson is more important than any particular technology.
Generative AI is not synonymous with enterprise AI.
Organisations should start with the decision or business problem they are trying to improve and then determine which technology is appropriate—not start with the technology and search for somewhere to deploy it.
Moving from Singapore to San Francisco has given me a new perspective on what I learned during those years.
One characteristic I value more now is Singapore's ability to bring different parts of the ecosystem into the same conversation: enterprises, technology companies, researchers, practitioners, public agencies and professional bodies.
Communities such as La French Tech Singapore can play a useful role precisely because they sit between these worlds. Their value is not simply networking. At their best, they create a neutral space where practitioners can compare what works, what does not, and what they are learning.
For anyone working on enterprise AI today, I would take three principles from my Singapore experience:
Design for adoption, not just technical performance.
Identify the business owner and the path to production before the pilot begins.
Build governance into the product.
Define evaluation, risk and accountability requirements early rather than treating them as final-stage approvals.
Start with the business decision, not the AI trend.
The right solution may involve a language model, traditional machine learning, a model designed for structured data—or a combination of approaches.
AI technology will continue to change quickly. The harder question for enterprises remains remarkably consistent: how do we turn technical capability into something people can trust, deploy and use to make better decisions?
That is the question Singapore taught me to focus on.
Alexandre Gerbeaux is Head of Applied AI at Fundamental in San Francisco. He led the Data, AI & Analytics community of La French Tech Singapore from 2023 to 2025 and co-founded the Singapore Machine Learning meetup.