The problem
The client is a credit union with around 120,000 members. New members apply online and upload identity documents and proof of address. Every application was reviewed manually by a small operations team: checking that documents were legible and matched the application, running identity verification and sanctions screening, and requesting missing items by email. Median time from application to open account was five business days, and a significant share of applicants abandoned the process before completing it.
What we actually did
Mapping the process before automating it
We sat with the operations team, mapped every step and decision in the review, and classified each one: mechanical checks (is the document legible? does the name match?), checks performed by existing third-party services (identity verification, sanctions and watchlist screening), and judgment calls that require a trained person under the credit union's compliance policy. Only the first two categories were candidates for automation.
Document intake and extraction
Uploaded documents are classified by type and checked for image quality, and their fields are extracted with a document understanding model. Extracted fields are compared against the application. When something is missing or unreadable, the applicant gets an immediate, specific request ("the photo of your license is blurry; please retake it in good light") instead of an email days later.
Orchestration and routing
A workflow engine runs each application through extraction, the existing identity verification and screening services, and a set of rules the compliance team wrote and owns. Clean applications move to account opening. Anything with a potential screening match, a mismatch, or a low-confidence extraction goes to a reviewer with the specific issue highlighted and all evidence attached.
Challenges and tradeoffs
- Document variety. Members submit licenses and IDs from many states and countries. Unsupported document types route straight to a reviewer rather than being forced through extraction.
- Audit requirements. Every step, extracted value, rule outcome, and reviewer decision is logged with timestamps and the rule version in force, so any account opening can be reconstructed for an examination.
- Change management. The operations team's work shifted from reviewing everything to handling exceptions. We redesigned their queue around exceptions with the most context up front, and they tuned the rules during the pilot.
Results
Median time from application to open account fell from five business days to one. 78% of applications now complete with no manual step, and application abandonment dropped 40% compared with the prior six months, which the credit union's growth team values more than the labor savings. Every decision has a complete audit trail, and the operations team now spends its time on the cases that actually need judgment.
What we'd do differently
We would add the instant "please retake this photo" feedback in the first phase instead of the second. It had the largest single effect on abandonment and depended on very little of the rest of the system.