Example solution: document AI for faster SME credit decisions

Picture an SME lender losing good applicants to faster competitors. Underwriters spend most of each application reading bank statements and accounts, and decisions take days.

The challenge

Documents arrive in hundreds of layouts: scanned PDFs, phone photos, exports from dozens of banks. Every number used in a credit decision has to be traceable to its source for the regulator, and an extraction error could mean a bad loan.

Three decisions that would make it work

1. Source-linking every field. Each extracted value carries a pointer to its page and bounding box, so underwriters verify in one click and auditors can replay any decision.
2. Cross-document validation. Closing balances must reconcile with opening balances; turnover in accounts must roughly match bank inflows. Mismatches are flagged, not hidden.
3. Confidence-based routing. High-confidence fields flow straight through; low-confidence ones go to a review queue, and corrections feed the evaluation set.

How we'd measure success

• Median time to decision, before and after
• Field-level extraction accuracy on a held-out evaluation set
• Applications handled per underwriter

What we'd do early

We'd involve the compliance team in week one. Their requirements shape the audit trail design, and late input on that is one of the most common causes of rework in regulated projects.