AI agents in banking: five cases where they held up in production

An agent that works in a demo and an agent that survives production are two different things. The difference is not the model but everything around it: which tools it can reach, a spending limit, a record of every call, and a point at which a person says yes or no.

Across the five cases observed — AML screening, credit pre-assessment, document handling, answering branch queries and payment reconciliation — the agents held up wherever the process was already described precisely and there was an objective test of success. Where the rule was implicit, the result stayed ungovernable.

The pattern that worked was always the same: short, checkable tasks; explicit tools instead of free text; the case memory kept separate from the model memory; and a log that lets you retrace how any decision was reached.

One thing is routinely underestimated: variable cost. Without a spending cap per case and an estimate of how many steps it takes, the cost of the process can swing widely. Measure it before the pilot, not after.