Banks across Asia were eager to use generative AI, but no supervisory rules existed for it yet. Hong Kong's central bank set up a controlled space where banks could test real AI tools under direct supervision, with the compute power provided by a local innovation hub.

The dilemma

By 2024, banks in Hong Kong and across Asia were under pressure to start using generative AI. They wanted to test it for customer service, for spotting fraud, and for routine back-office work. The problem was that no one knew yet what the rules should look like. Generative AI behaves differently from earlier types of AI. It can produce confident answers that are completely wrong, it can be hard to explain after the fact, and the technology itself is changing every few months. The Hong Kong Monetary Authority had two choices. It could write rules now, before anyone really understood how banks were using the technology in practice, and risk locking in the wrong rules. Or it could let banks experiment first, and write the rules later based on what actually happened.

The regulatory gap

What was needed was a space where banks could test real generative AI tools on real banking data, under direct supervision, so the regulator could learn alongside them. Most banks did not have ready access to the kind of computing power needed to run generative AI experiments safely. So the space also had to come with the technical resources to make experimentation possible, not just permission to try.

The approach

In August 2024, the Hong Kong Monetary Authority launched its Generative AI Sandbox. It worked with Cyberport, Hong Kong's main innovation hub, which provided the computing infrastructure for the participating banks. Licensed banks could test AI tools for risk management, fraud detection, and customer service inside the sandbox, with technical support from Cyberport and direct feedback from the Authority. In 2025, a second round of the sandbox added a new earlier stage, called the Collaboratory, where banks could work through what they wanted to test before formally entering the sandbox.

Results achieved

The sandbox produced supervisory insights that the Hong Kong Monetary Authority has said will shape its future guidance for banks using generative AI. It also created a shared language between regulators and banks on practical questions such as how to test whether an AI model is performing as expected, how to handle the risk that an AI tool gives false but confident answers (sometimes called hallucination), and when a human should review the AI's output before action is taken.

The Collaboratory addition in the second round reflected a clear lesson from the first. Many banks did not yet know how to describe what they wanted to test. They needed help framing the problem before they were ready to build a pilot. Adding an earlier stage gave them that space without forcing the regulator to design the experiment for them.

Two limits are worth noting. The sandbox is explicitly experimental. Participation does not amount to regulatory approval of the AI tools being tested. And the formal supervisory guidance that the sandbox is meant to inform has not yet been published.

The transferable lesson

The most useful design choice in this sandbox is the one other regulators are most likely to miss. Adding an earlier stage for problem definition turned out to matter as much as the sandbox itself. The first cohort showed that banks could not always describe clearly what they wanted to test, and that this was a problem the regulator could help solve without having to write the experiment for them. Other regulators setting up generative AI sandboxes should expect this challenge from the start, rather than waiting for participants to ask for help.