Document intelligence for a commercial bank's credit operations

A commercial bank's credit operations team was spending the bulk of its cycle time on manual document review. Every corporate loan file arrived as a bundle of scanned financial statements, board resolutions, collateral documents, and KYC records — often hundreds of pages — which analysts read line by line before a facility could move to committee.

Three constraints shaped the brief. First, customer financial data could not leave Nigeria, ruling out naive use of offshore model APIs. Second, the output had to be auditable: a credit decision influenced by an extraction error is a regulatory and commercial risk, so every extracted figure needed a traceable source page. Third, the system had to degrade gracefully — scanned documents in Nigerian banking are frequently low quality, and a pipeline that silently mis-reads a smudged balance sheet is worse than no pipeline at all.

The engagement followed our standard sequence: a two-week discovery to map the document taxonomy and failure modes, a scoped pilot on a single document class, then staged expansion.

Key design decisions included an in-country inference architecture with redaction before any external call; a retrieval-augmented extraction pattern in which every extracted field carries a citation to the source page and bounding region; confidence scoring with a human-review queue — fields below threshold route to an analyst rather than passing silently; and an evaluation harness built from a hand-labelled golden set before any model was selected, so vendor and model choices were made on measured accuracy rather than demonstrations.

Model and vendor selection was conducted neutrally against the evaluation harness. The bank's own team was trained to own the golden set and re-run evaluations as models evolve — capability transfer, not dependency.

The pilot demonstrated that first-pass extraction could remove the majority of manual reading time on the pilot document class, with every figure traceable to its source page. The human-review queue caught low-confidence extractions rather than letting them pass silently — the property the credit risk function cared most about.

The bank's internal team left the engagement owning the evaluation harness, the golden dataset, and the review workflow, positioning them to extend the pipeline to further document classes without external dependency.