Turning compliance from a cost centre into a data product.
A regional bank was reviewing 40,000 documents a month by hand across three product lines. We built a document intelligence layer that reads, classifies, and escalates — cutting review load by 38% without adding headcount.
The situation on arrival.
Compliance headcount had grown 4× in three years while transaction volume grew 11×. Every new product launch required a linear increase in reviewers, making the unit economics of growth progressively worse.
Regulators had also flagged inconsistency between reviewers as an emerging concern — different analysts were reaching different conclusions on structurally similar files.
How we structured the work.
We began with a two-week readiness assessment covering data quality, document taxonomy, and reviewer decision patterns. Rather than automating the review outright, we designed a triage architecture: the model handles classification and routing, humans retain all adjudication authority.
This kept the bank comfortably inside its regulatory perimeter while removing the highest-volume, lowest-judgement work from senior desks.
The goal was never to remove the human from compliance. It was to stop spending senior judgement on document sorting.
Systems now in production.
- A RAG-based document classification pipeline over the bank's existing document store.
- A confidence-scored routing engine that escalates ambiguous cases to senior reviewers.
- An audit trail capturing every model decision and human override, plus a governance dashboard for the compliance committee.
What changed for the client.
Manual review load fell 38% within two quarters. Reviewer-to-reviewer consistency improved measurably.
The bank launched its next product line without adding compliance headcount for the first time in its history.
Considering a similar programme?
We take on a limited number of engagements each quarter. If the shape of this work is familiar, we should talk.
Start a conversation →