KYC Accelerated
with Document Analysis

Situation

A financial services provider processed client onboarding largely by hand. Analysts read through extracts, financial statements, identity documents and UBO declarations to complete each file. Turnaround times kept growing, and quality varied by analyst and by week.

Challenge

AI support was the obvious answer, but KYC is a regulated process. Every automated step had to be explainable, reproducible and traceable to a human decision. A model that is right 95% of the time is not automatically usable in this domain: everything depends on what happens in the remaining 5%.

Approach

We started with the workflow, not the model. Which steps in handling a file are extraction (pulling facts out of a document) and which are judgment (weighing whether that fact represents a risk)? Only the extraction steps were in scope. For every extraction we set a confidence threshold: above it, the system proposes a value; below it, it goes to the analyst unfiltered.

Decision

The architecture was deliberately conservative: the system proposes, the analyst confirms. Every confirmation and every correction is logged, giving both an audit trail and a measurable quality loop. There is no step where a file is closed out without a human check.

Result

Turnaround time per file dropped 60%, mainly because analysts no longer had to search but only had to check. The quality score in sample audits rose 35%, because the system consistently surfaced every field instead of only the ones an analyst happened to think of. The internal audit raised no findings.