An AI Startup Just Moved Into Singapore. Here's What It's Actually Coming For

The market keeps pricing artificial intelligence in financial services as a productivity story. It is actually a labor-supply story, and the supply that is being cut is the one every trading desk, every private equity shop and every hedge fund eventually draws its talent from.

An AI startup built to automate the grunt work of junior investment bankers, the pitch decks, the comparable-company tables, the due-diligence memos, has set up in Singapore. This follows a pattern rather than starting one. Model ML, the company most associated with this push, raised $75 million to build technology aimed squarely at the hours-long tasks that used to define the first two years of a banking career. OpenAI has gone further, hiring more than 100 former investment bankers specifically to train its models on how to build financial models. This is not a lab experiment anymore. It is a hiring decision at the frontier labs themselves.

Singapore is not a random landing spot. The city-state has been running its own version of this experiment inside the banks. Regulators have already sat through demonstrations of agentic models that "could in ten minutes do what used to take a private banker an entire day," as part of a program built to retrain tens of thousands of bank staff rather than simply cut them. HSBC is opening a Global AI Centre in the city in the second half of this year and hiring more than 100 specialists. Singapore-based AI-native firms pulled in $9.3 billion in funding this year alone. The jurisdiction has made itself the place where a bank can run the pilot and the regulator will actually pick up the phone.

The mechanism nobody is modeling correctly

Here is where consensus gets it backwards. The story being told is about cost. Deloitte's own estimate for the largest global investment banks put the productivity gain from generative AI at 27 to 35 percent by this year, worth an additional $3 million to $4 million in revenue per employee. That is the number analysts are plugging into bank margin forecasts. It is the wrong number to watch.

The real transmission mechanism is not cost per employee. It is headcount composition. Junior analyst classes exist for two reasons: they produce the deliverables, and they produce the future dealmakers. Every managing director on every trading floor spent two to three years building models badly before they built them well. That apprenticeship is the actual product being disrupted, not the pitch deck. When the grunt work disappears, the training ground disappears with it, and the pipeline that stocks trading desks, private equity shops and eventually portfolio management seats with people who understand how a balance sheet actually behaves gets thinner at exactly the moment deal complexity is rising.

This is the second-order effect almost nobody is pricing. Strip out the junior tier and you do not just save on comp. You compress the population from which the next generation of risk-takers is drawn. Fewer people learn the mechanics from the inside. Judgment becomes scarcer even as raw analytical output becomes cheaper and more abundant. Scarce judgment gets repriced upward. That repricing shows up years from now, in compensation structures at the senior level and in the premium the market pays for demonstrated deal experience, not in this quarter's cost-income ratio.

The headline says grunt work gets automated. The mechanism that matters is what happens to the people who never get hired to do it.

There is a second layer to this that sits closer to positioning. Jurisdictions that let this automation run without regulatory drag effectively subsidize deal velocity. If model-generated due diligence and comparable-company work clears faster in Singapore than in jurisdictions still arguing about model liability and data residency, deal flow tilts toward wherever the friction is lowest. That is a genuine competitiveness gap between financial centers, and it is not currently reflected in how anyone prices exchange volumes, listing pipelines or regional banking revenue mixes.

The honest counter-case

I would be wrong about the scale of this if history repeats in the way it usually does. Spreadsheets did not shrink banking headcount. Real-time market data terminals did not either. Every prior wave of automation in finance freed up junior time that got reallocated into more deals, not fewer analysts. If deal volume simply expands to absorb the freed capacity, the apprenticeship pipeline does not shrink, it just does different work; less formatting, more judgment, earlier. Compliance and liability concerns inside the banks themselves could also slow real adoption well below what the pilots suggest. A demo that does ten minutes of work a private banker used to spend a day on is not the same as a production system a bank's legal department will let sign off on a live deal memo without a human rebuilding half of it anyway.

I run a market-neutral book at Zentra Asset Management, and this is exactly the kind of structural shift that does not show up cleanly in any single position. There is no clean short here and no obvious long. What there is, is a reason to watch bank cost-income ratios against junior headcount data over the next several reporting cycles, not against revenue per employee. The gap between those two numbers, cost falling faster than headcount, or headcount falling faster than deal complexity would suggest it should, is where this actually gets priced.

What I am watching next is whether the banks that adopt this fastest, the ones running pilots in Singapore right now, start reporting analyst class sizes shrinking faster than deal volume is growing. That divergence, not the funding round, not the office opening, is the number that tells you whether this is genuine structural change or another wave of tools that gets absorbed the way every prior one was.

If the apprenticeship model that built every senior dealmaker on Wall Street is actually being dismantled quietly in a Singapore office right now, ask yourself who is training the person who will be pricing your risk in 2035, and whether anyone currently modeling bank earnings has thought about that at all.