Supervisors have adopted AI faster than almost anyone expected. Central Banking's 2026 supervision benchmarks found AI in use at three quarters of supervisors, up from a quarter a year earlier. The BIS reports that more than two thirds of central bank AI projects are already live.
Look closer, though, and most of that AI is doing one job: summarising and drafting. Useful, but it sits beside supervision rather than inside it.
Adoption has outrun governance
The BIS survey shows governance frameworks expanding, yet only 4% of central banks rate their ability to monitor compliance with them as high. Nearly 80% name IT costs as a risk. Authorities have the tools; fewer have the controls and the evidence to show how those tools behave.
Summaries are the easy win, not the destination
Summarisation and document preparation are the most common supervisory uses of AI. They save time and carry little risk. But the work that defines supervision is checking: is this return complete, does it reconcile, is it plausible against history, and what should happen next.
Singapore shows where the bar is heading
On 7 October, MAS issued guidelines on AI risk management for the financial sector, taking effect from October 2027. Firms must keep an inventory of their AI, assess how material each use is, keep human oversight and remain accountable for AI supplied by third parties. A further consultation on agentic AI follows in 2027. The direction is clear: AI that can be inventoried, evidenced and overseen.
Where the next gains are
The next step is putting AI inside the work: reading every return as it lands, with deterministic rules making the calls that must be repeatable, and an evidence trail behind every finding. That is how supervisors get the speed without losing the predictability their decisions depend on.
- 1Inventory every place AI is used today, including ad hoc chat tools.
- 2Separate reading tasks, where AI helps, from decision tasks, where rules decide.
- 3Pick one high-volume return and move validation to the point of submission.
- 4Require every AI-assisted finding to be traceable to its data and rule.
- 5Measure analyst hours on that workflow before and after.
Sources: Central Banking, Supervision Benchmarks 2026; BIS Irving Fisher Committee, AI readiness in central banks, via Poland Insight; MAS, media release, 7 October 2026. Try our readiness check to see where your team stands.