I designed AI review tools at JPMorgan Chase that help private bankers approve client data they can actually trust.

Context

Private-client onboarding meant a human reading nine document types per client, cross-checking them against each other and against policy. Most of that reading confirmed nothing was wrong. The time went into the reading, not the deciding.

The problem

The team wanted automation but did not trust a model to decide alone, and compliance would not accept a black box. The design problem was not “how do we detect anomalies” but “how does a reviewer stay in charge of a system that reads faster than they do.”

What I designed

An agent that reads the packet first and surfaces only what it cannot reconcile, each flag carrying the two pieces of evidence that disagree. Reviewers confirm or dismiss with one action, and every dismissal teaches the system what a false positive looks like for that client segment.

The interface is built around one question at a time. No dashboard, no scores to interpret. Just: here is what looks off, here is why, what do you want to do.

What changed

Median onboarding time fell 45%. Reviewer confidence, measured by how often they overrode the system, settled at a level compliance was comfortable with. The pattern was picked up by two adjacent teams.