Fintech
Growth-stage fintech underwriting team
~30% less drafting time
on the sections covered by the assistant, measured during a limited pilot
0 uncited figures reached a draft
across the pilot period, enforced structurally rather than by review alone
Still in pilot
expanding coverage gradually as the evaluation set grows
The Challenge
Underwriters spent hours per file assembling risk memos from financial statements, credit reports, and prior correspondence.
An earlier internal experiment with an ungrounded LLM draft produced a memo with a fabricated figure that nearly went to a credit committee, so trust in AI-assisted drafting was low going in.
My Approach
Constrained the assistant to only cite figures that trace back to a specific line in a supplied source document: no number appears in a draft without a citation.
Built an evaluation set of past memos to check the drafts against known-correct figures before any underwriter saw the tool.
Positioned the output explicitly as a first draft: every memo requires underwriter sign-off before it moves forward, with no exceptions.