Governance intelligence for the era of public AI models

We put high-intent retail questions about Malaysian banks to a set of public AI systems and checked each answer against the bank's own published service charter. The pattern was degradation, not invention: a 3-working-day card commitment told as 2–3 weeks; a 5-day mortgage as ~30 days; a 14-calendar-day complaint decision as 20 working days. A preliminary, single-month reading of a risk one step outside most governance frameworks.

Public AI assistants now answer customers' questions about your institution — and often get the product, price, complaint timeline, or fraud channel wrong. You can't see or edit those answers, yet the consequences land at your door. This explainer makes the case that their accuracy is a governance surface you already own, and lays out the loop for governing it.

We put high-intent retail questions about Singapore banks to a set of public AI systems and checked each answer against the bank's own published facts. Many aligned with the published record; a recurring few didn't — and they missed by a lot, where a customer acts: a 28.5% cash-advance rate quoted below it, a 27.9% card rate quoted at 15–18%, a two-business-day complaint window quoted as five. A preliminary, single-month reading of a risk that lives one step outside most governance frameworks.

We put high-intent questions about Singapore's insurance schemes to a set of public AI systems and checked each answer against the official record. Other tested systems were often correct; on three rules one system each missed where a consumer acts — CareShield Life's 3-of-6-ADL trigger said as 1; FIDReC's S$150,000 limit said as S$500,000; FIDReC-NIMA's non-injury, sub-S$3,000 scope described as broadly available. A preliminary, single-month reading of a risk one step outside most governance frameworks.