Smart AI, Dumb Jobs: Making Mission Possible

Please click the links below to view the full papers:

NUMA Monthly (Ver: 2026-11-03)

"Hong Kong FinTech Week has come and gone, and one subject ran through all of it: AI. Every panel found its way back there, and I'm glad it did. Here are a few thoughts I took away." 

Charles Li

3 November 2026

Two currents ran through almost every conversation. The first was genuine amazement. AI has already changed how research is written, risk modelled and trades executed, and nobody believed the change was anywhere near finished.

The second was more sober, and voiced everywhere: compliance, regulation, and how far AI can go before it meets the wall of accountability. Finance rests on a simple principle. When something goes wrong, a person has to answer for it. Not the system, not the vendor, not the model, but a named individual with authority and responsibility. Regulatory trust is built on that principle, and so is client confidence. AI can deliver remarkable breakthroughs, but it cannot itself be held to account.

Wonderful progress at one end of the table, serious concern at the other, and everyone looking for a way to thread the needle: how to put AI to full use while still delivering accountability. AI can advise and execute within clearly defined authority; key decisions and ultimate responsibility must remain with people. Whether AI is advising on a landmark deal or trading at machine speed, the aim is not to avoid every mistake, which no system can. It is a chain of accountability, so that when something goes wrong, someone answers for it, has the authority to halt the system and can put things right. Nobody disputes that machines can fail; there are well-documented cases of automated systems damaging markets because no person was accountable in the moment, or able to intervene in time.

How to harness AI's genius while keeping that chain intact is a great challenge for big finance, and one it must answer on its own terms. We wish it every success, because every market depends on it.

Built from scratch, built for the workhorse

I walked away impressed, and more determined. Our challenge is a slightly different one. We need accountability just as much, but we do not have to fit AI into institutions and rulebooks designed for people. We start from a blank sheet of paper. We get to build small finance from scratch, deciding from the first line how humans and machines divide the work and how accountability runs through it.

That design begins with a simple observation: AI has a dual personality. One side is the genius, capable of smart wonders. The other is the workhorse, capable of getting vast numbers of unglamorous things done, at scale, without rest. FinTech Week talked mostly about the first. Our model is built on the second.

Financing small contracts, not companies, and why now

Wall Street finances companies; we apply the same model to contracts. A data centre that equips an AI start-up and earns by the hour of computing. A landlord who funds a restaurant's fit-out in return for a share of its revenue. Each contract carries real, observable cash flow, and there are hundreds of millions of them. That is small finance: not less important finance, but finer.

Nobody ever disliked small. It was simply unaffordable. Examining a small contract costs nearly as much as examining a large one, so inclusive finance stayed trapped in an impossible triangle of cost, risk and scale: its own mission impossible. Digitalisation and AI change that arithmetic. And because each unit of work is small, AI can be deployed in a way that meets our regulatory responsibility rather than straining against it.

Smart AI, dumb jobs: HAT Over MASS

Small finance does not need a genius. It needs an enormous number of simple tasks done right: reading clause 14 of every contract, reconciling a restaurant's daily takings against its bank receipts, noticing that roadworks are about to close the street outside.

That is what I call big AI. Big in number: a month ago I wrote about 500 agents. Today, more than 7,000 are at work in our simulations and internal testing, more than fourteen times as many. The marginal cost of adding one more agent is close to zero, though verification and human oversight still have costs. Big in frequency: the same checks, on every contract, every day. And big in endurance: around the clock, seven days a week, without tiring.

How many agents we can use depends on how small we can make the work. More AI cannot divide a major financial decision into independent pieces; it can only help people work through it. Small finance already consists of individual contracts that AI can work on separately.

We still use the most capable models available. But intelligence is not why we deploy them; scale is. However brilliant the model, we give it small jobs. At the panel I said our agents do dumb jobs, and it drew a laugh. The point is serious. A dumb job is narrow, short and verifiable, which makes AI's occasional unreliability manageable. Each task is small, so an error is contained. The tasks are independent, so one fault does not cascade into the next. For verification, each task goes to two or three agents on different models, whose blind spots differ. One may catch what another misses, but all may miss the same thing. Checks must be specified and evidence retained. Disagreements require further checks; even when models agree, key items need human review and other tasks need spot checks. A missed item should trigger a review of similar tasks.

Those extra checks also have a cost. We need to test whether fewer errors and less rework justify it.

We call this architecture HAT Over MASS: Human-commanded, AI-operated Trust over Micro-Agents with Short and Simple tasks. The MASS is the workforce, each agent confined to a single task. The HAT is human governance: people set the rules, draw the boundaries, sign the decisions that matter, and answer for the outcome in their own names.

How can a small team be accountable for millions of tasks? Precisely because every task beneath them is narrow, verifiable and bounded. AI does the work; people bear the ultimate responsibility. That is our chain of accountability. It is also how mission impossible could become mission possible, with a chance to improve cost, risk and scale significantly at the same time: AI takes on repetitive work to lower processing costs per contract; ongoing checks and human judgement at key points help control risk; and many agents working in parallel can serve many contracts. Those gains still need to be tested.

The future of Hong Kong's market, big and small

The moderator asked what Hong Kong must do to lead AI-native finance by 2030. My answer came down to two words: big and small.

Today Hong Kong plays one set of big and small: BIG MONEY, SMALL NUMBERS. Big in capital deployed, small in the number of transactions; big in global impact, small in the number of players. Hong Kong plays this game extremely well, and should keep competing with New York and London for every inch of it.

It can also open a second front, with a second set: BIG NUMBERS, SMALL MONEY. Big in the number of transactions, small in the size of each; big in impact on local economies, small in the downside of AI, because every task the machines perform is narrow, verifiable and answered for by a person.

With AI, we now have an opportunity to divide the revenue rights under business contracts into units that can be assessed and financed individually. Each may be small, but together they could form a substantial market.

For China, Southeast Asia and the other large economies of the Global South, this second game has only just begun. Hong Kong has a real chance to lead it, helping more small businesses access capital markets and making finance serve local economies better.

We set out the practical arrangements in the first issue of NUMA Monthly, launched this month. On the 3rd of each month, we will report where we have reached, what problems we have found and what we have changed.

Smart AI, dumb jobs. That is how small builds big.

The descriptions of design outcomes, testing and future arrangements in this article reflect plans at the time of publication. They do not constitute commitments regarding launch dates, service availability or investment results. Actual business activities and any related changes must comply with applicable laws, regulatory requirements and the applicable rules of the relevant licensed trading venues, with approvals, filings or other necessary procedures completed as required.

Public access to this article does not constitute the provision of investment services to readers or confirmation of their eligibility to invest. Investment participation is restricted to eligible professional investors outside mainland China who meet the requirements of the applicable jurisdiction, and must take place through appropriately licensed intermediaries. It is not open to retail investors. NUMA does not conduct fundraising activities in mainland China or accept funds from mainland China for investment.