Riverbend Bank

Client / Project

Riverbend is a regional commercial bank with about 1,500 employees. Their compliance department reviews every outbound customer communication that mentions specific regulated topics — investment advice, lending decisions, fee structures — under standing regulator expectations.

Problem

The review process was entirely manual. Four reviewers worked through 1,400 communications a week, with a backlog that ran to 3-5 business days. The bank's seniormost reviewer — the one who knew the edge cases and informally trained the team — had filed retirement paperwork. The compliance officer was looking at a knowledge loss event with no realistic hiring plan.

Approach

A 2-engineer AI Discovery Sprint that converted into a 6-week build engagement. The system has two stages: a classifier that scores each communication on 7 regulatory risk dimensions, and a second-stage agent that for any flagged communication produces a written rationale explaining the specific clauses of concern and the relevant policy reference.

Critically, we treated the senior reviewer as the primary subject-matter expert during build, not just at the end. Over six weeks she labeled roughly 1,200 historical communications with us, walked through her edge-case heuristics, and reviewed every prompt change. Her tacit knowledge ended up encoded in the eval set and the prompt — durable in a way it never was when it only lived in her head.

Outcome

The system handles 89% of weekly communications without human review and flags the remaining 11% with a written rationale that the human reviewers either confirm or override. Backlog dropped from 3-5 days to under 4 hours. The senior reviewer's retirement is no longer a continuity crisis; her judgment is encoded in artifacts the compliance team owns and can update. The bank's regulator was briefed and accepted the architecture as part of an updated supervisory submission.