← All rollout articles

Bank: trade finance workflow on documents that refuse to die

Problem

Trade finance operated through email chains, PDFs, and spreadsheets with heroic analysts. Digitization attempts failed when OCR hallucinated fields on low-quality scans and nobody trusted the machine output.

Constraints

Legal hold requirements, sanctions screening integration, and strict segregation between teams handling different client tiers. False negatives were unacceptable; false positives were expensive but tolerable if routed well.

Approach

We treated extraction as probabilistic: every field carried confidence and lineage. Low-confidence paths always landed in human queues with structured decisions, not freeform email replies. Workflow became the system of record for “what we believed when.”

Rollout

Corridor-by-corridor rollout by document type (invoices vs bills of lading vs packing lists). Analysts co-designed exception taxonomy so metrics meant something operationally. Shadow mode compared machine suggestions to human final decisions for months.

Risks mitigated

• Silent automation errors: dual control on releases that changed extraction models
• Staff deskilling: rotated analysts through model evaluation to keep judgment sharp
• Audit panic: immutable event log with export formats legal already accepted

Outcomes (illustrative)

Straight-through processing rose for high-quality scans while exception queues remained bounded because capacity planning matched arrival rates. Mean cycle time for standard instruments improved without increasing compliance incidents in monitoring samples.

Lessons

Document workflows are queueing theory with feelings. The breakthrough was honest uncertainty in the UI—not higher OCR accuracy alone.