Field notes: what I look for in AI-fintech infrastructure

A running list from the diligence chair. Not a rubric — just the questions I keep coming back to.

1. Permissions before features. Before I ask what the agent does, I ask what it’s allowed to do — and who decided. If that model is an afterthought, everything downstream is fragile.
2. Failure is a first-class citizen. What happens when the model is wrong? The best teams have a crisp answer and have already designed the blast radius down to something survivable.
3. The human-in-the-loop is intentional, not decorative. Trust rises sharply when the review step is designed, not bolted on — and it craters when it’s an afterthought. Where the human sits is a product decision.
4. Boring where it counts. Audit logs, idempotency, reconciliation. Unsexy, and exactly where consumer-finance companies live or die.

None of this shows up in the demo. All of it shows up in the diligence.