hoomaraThe temptation
Credit is where the business case for AI looks most obvious: every decision follows documents, ratios and precedents, and every day of delay costs conversion. So the temptation is to automate the decision itself — score in, verdict out.
In regulated lending that is the wrong boundary. Not because models are weak, but because a credit decision is an accountable act: someone must be able to explain, later and under scrutiny, why this loan was made. Accountability does not delegate well.
What the agent should own
Everything before the call. Document intake and extraction, so the file is complete before a human opens it. The first draft of the credit memo, with figures, covenants and precedents assembled and sourced. Risk surfacing, so the exceptions and anomalies are on top of the file, not buried on page forty. Monitoring after the decision — covenants, payment behaviour, exceptions — so review happens when something changes, not when the calendar says so.
Done well, the agent removes the hours that were never judgment in the first place: searching, retyping, cross-checking, formatting.
What must stay human
The decision, its rationale, and the appetite behind it. A named person approves, adjusts or declines — and the system records what they saw when they did. The reasoning that goes into the file is theirs, informed by the draft, not generated by it.
This is precisely what makes the setup auditable: the trail shows what the agent prepared, what the human decided, and why. Speed comes from preparation; defensibility comes from the boundary.
Where we start with lending clients
Credit memo drafting, document intake and extraction, and covenant and exception monitoring — the three places where preparation time dominates and the judgment boundary is easiest to hold. From there, the same pattern extends across the credit lifecycle.