OCBC asked 33,000 staff to build a one-person company with AI agents

Most bank AI training starts with a licence and a lunchtime webinar. People learn where the button is. Very few of them learn what it feels like to hand a whole piece of work to an agent and live with the result.

OCBC tried the second thing. The bank built a hackathon around the idea of the one-person company, the founder who runs a business with AI agents doing the work a team used to do. Employees could enter alone or in teams of up to three, and the stated aim was blunt: help people get over the fear of using agents at all. The finals ran on 22 October at the Singapore Institute of Technology, and the competition was open to the bank's 33,000 employees.

What came back

Thousands of submissions arrived within weeks. Close to three quarters came from people outside technology functions, which is the number I would pin to the wall if I ran learning at a bank. The usual pattern is that engineers play and everyone else watches. Here the operations, sales and risk people did most of the building.

The screening is the part I keep coming back to. OCBC used its own internally built agents to sort the first pile and surface the promising ideas. Its AI Champions then assessed more than 400 prototypes and cut them to 10 finalists. So the bank used agents to judge work made with agents, and kept people on the final calls. That is a sensible split, and a slightly unsettling one.

The judging panel was not a token one. Group CEO Tan Teck Long sat on it, alongside Group Chief Risk Officer Carina Lee, Sunny Quek (Head of Global Consumer Financial Services), Praveen Raina (Head of Group Operations and Technology) and Tan Chor Sen, CEO of OCBC Malaysia, with Associate Professor Indriyati Atmosukarto from SIT. Prototypes were scored on innovation, business impact and fit with the bank's strategic priorities. Four winners were named: the top three teams and the best individual entry.

Why it matters

Tan's line was about culture rather than tools. "Organisations must build a strong test-and-learn culture that empowers employees to experiment boldly, push boundaries, learn quickly and fail fast." Plenty of chief executives say that. Fewer of them spend an afternoon judging prototypes built by people who have never written code.

The hackathon also does not end with a trophy. SIT will pick promising prototypes for further work at the SIT x NVIDIA AI Technology Centre, where researchers, faculty and students will refine them and test how they might apply to the bank's operations. Professor Susanna Leong, Deputy President (Academic) and Provost at SIT, said the aim is to move promising ideas closer to implementation. That second stage is where most corporate hackathons quietly die, so it is worth watching whether any of the 10 finalists reach a branch or a back office.

It sits inside OCBC's Next Frontier strategy, which puts AI, digital and data at the centre of the bank's plans. Singapore's banks have been retraining staff at scale for most of this year. This is a different instrument. It asks people to build something, not to finish a module.

What this means for L&D

A hackathon is an honest test of whether training landed. If three quarters of entries come from outside technology, the floor has moved. If they do not, the course did not do its job. The design choice worth copying is the fear framing: a safe, time-boxed place to hand real work to an agent before anyone has to do it on a live customer file. An AI design sprint is a smaller version of the same idea, run on one problem a team already owns.

The harder question is what happens to the 390 prototypes that did not make the final. Somebody built each of them. If nobody reads them again, the bank has taught people that experiments go into a drawer.

Sources: Human Resources Online, 30 October 2026.