Commercial lending intake, four days to one

Every commercial loan application arrived as a stack of PDFs — financials, tax returns, rent rolls, entity documents — emailed to a shared inbox and rekeyed into the origination system by a six-person intake team. Four days of median intake latency sat in front of every credit decision.

Compliance made the problem harder to fix than to staff. Any automation had to satisfy the bank's model-risk group under SR 11-7: documented, monitored, validated, and explainable to an examiner before it touched a single live file.

We spent the scoping phase inside the intake queue, classifying a month of real submissions by hand alongside the team. The finding: 80% of documents fell into eleven types with stable structures, and most of the four days was routing and rekeying, not judgment.

Acceptance criteria were signed before the build, alongside a model-risk documentation package drafted to the bank's SR 11-7 template: classification accuracy thresholds by document type, mandatory human review below a confidence line, and a complete action log for every file.

The system reads each incoming package, classifies the documents, extracts the fields the origination system needs, and routes exceptions — a missing schedule, an ambiguous entity name — to the intake team with the reason stated.

Every automated action writes to an audit log as it happens: what was read, what was extracted, what confidence it carried, who reviewed it. The audit file examiners used to reconstruct after the fact now assembles itself during the work.

Median intake time dropped 75% — four days to one — measured against the baseline from scoping. The intake team now handles exceptions and quality review, and application volume has grown into the freed capacity without new hires.

The model-risk group approved the validation package without a second review cycle, and the system reports its own accuracy monthly under managed operations.