AI in Public Services: Assist, Don't Decide

Founders pitching AI into a public service usually start with the wrong question — 'where can the model decide?' In govtech the right question is 'where can the model assist while a human or a clear rule still owns the decision?' The reason is trust. A citizen service has to survive public scrutiny, and an opaque AI making consequential calls about benefits or eligibility is exactly the kind of thing that ends up in a headline. So my AI strategy here is deliberately unglamorous: use it where it lowers effort without raising stakes. Plain-language explanations of dense policy. Helping someone find the right form. Pre-filling and validation that the user can always see and correct. Triage that routes to a person faster. What I keep AI away from is anything the user can't understand or appeal. Every AI-touched step needs a visible, human-legible path: what it did, why, and how to override it. That fits the layered-complexity principle perfectly — AI smooths the simple core path, while the deterministic, auditable logic stays underneath for anyone who needs depth or contests an outcome. The mature move isn't the flashiest AI. It's the AI a regulator, a journalist, and a stressed citizen can all live with.