Manual document checks and rule-thin underwriting queues meant 4-day approvals while fintech competitors promised hours. Risk wouldn't accept speed at the cost of control.
A custom origination workflow with document AI (KYC extraction and verification), configurable risk scorecards, straight-through processing for clean files, and human-in-the-loop review lanes for exceptions — every AI decision logged and explainable.
Speed with controls, not instead of them
The brief was explicit: no black boxes. Every extraction, score and routing decision is inspectable, overridable and logged, which is what makes a 100% audit trail a property of the design rather than a reporting exercise run afterwards. AI accelerates the queue; policy still owns the decision, and an underwriter can trace any value back to the document and the page it was read from before accepting it.
Where the time went
Document handling consumed 70% of underwriter time — receiving, reading, keying and cross-checking KYC before any judgment could begin. Automating extraction and verification returned that time to the part of the job that actually needs a person, which is why files per underwriter per day rose 3.2× without the review standard being relaxed to get there.
Straight-through processing is a policy setting
A file that satisfies every rule does not need a human to confirm that it is clean. Which files qualify is configuration that risk owns, not a threshold buried in code — so the automation boundary can be tightened after a bad cohort or widened after a good one without waiting for a release.
Exceptions get lanes, not a queue
The applications that need judgment are not interchangeable: a document mismatch, a thin credit file and a policy exception each need a different reviewer looking at a different thing. Routing them into separate lanes rather than one undifferentiated queue is where much of the 68% turnaround improvement came from.