Cutting loan onboarding from 6 days to under 8 hours

The situation#

Applicants submitted identity documents, income proof and bank statements through a web portal. Three analysts then extracted the data by hand into the loan origination system, cross-checked it against registry sources, and flagged discrepancies for an underwriter.

The cycle took six days at median and stretched past ten when volume spiked. Drop-off during that window was the single largest source of lost qualified applications.

What we did#

Week 1 — mapped the real process. The documented process and the actual process differed in an important way: analysts had developed informal rules for handling statement formats from a dozen banks, none of which were written down. Those rules were the actual specification.

Weeks 2–3 — built extraction with validation first. We assembled a labelled set of 600 historic applications, including the awkward ones — poor scans, unusual statement layouts, mismatched names across documents. Accuracy was measured against that set from the first commit.

Week 4 — shipped behind a confidence threshold. Applications where every field cleared validation and confidence passed threshold flowed straight through. Anything below went to an analyst with the source page and the flagged field side by side.

Weeks 5–6 — governance package. Model risk documentation, a DPIA under the DPDP Act, decision-level audit logging and the oversight design, delivered alongside the build rather than after it.

The boundary we kept#

Credit decisions stayed with underwriters. The automation assembles and validates the evidence; it does not decide whether to lend. That boundary made the internal risk approval straightforward and it is the right design regardless.

Where it went next#

The client moved to a managed retainer covering the automation and the surrounding infrastructure. Statement-format coverage has since been extended twice, both times by their own team using the evaluation suite we handed over.