Specialty lender: clearing the exception backlog
A mid-size specialty lender’s risk analysts spent most of their week re-checking policy hits that were obvious clears. We built a Python review service on GCP, grounded in their written policies — with a human still signing off on anything ambiguous.
Every flag looked the same in the queue
Volume had grown; the exception queue had not. Analysts clicked through hundreds of near-identical policy hits each week. Monthly compliance packs were late because someone was still clearing last month’s leftovers.
Score and draft — don’t auto-post money
We indexed their policy PDFs in a retrieval store, then built a Python service that scores each exception, drafts a short rationale, and either proposes a clear (with audit trail) or routes to a person. Nothing posts without their existing core-system controls. Ambiguous cases still land with a human.
Main pieces of the work
Policy corpus
Versioned procedures and rule docs used as the only retrieval source.
Scoring service
Python workers classify exception type and confidence against that corpus.
Analyst queue
Low-confidence items with the draft rationale attached.
Audit export
Monthly pack of decisions for compliance review.
- Python
- Google Cloud / Vertex AI
- Document retrieval (RAG)
- Core banking APIs
- About 70% of routine exceptions clear without analyst time
- Multi-week backlog gone within the first month of production use
- Compliance still gets a full decision log each month
- No change to who can post or reverse in the core system
Buried in exception review?
Bring a sample of your queue and policies — we will tell you quickly if automation is worth the build on your cloud.