Case study · Fintech
AI-assisted document processing for lending
An LLM-powered pipeline that reads, classifies and validates loan documents, with human review only where it matters.
Results
- Average review time, down from hours
- 8 min
- Field extraction accuracy on eval set
- 96%
- Customer data used for model training
- 0
The challenge
Loan officers spent most of their day reading bank statements, payslips and IDs, copying figures into a core banking system. Volumes were growing faster than the team, and errors were expensive.
Our approach
We treated AI as one component in a well-engineered system, not as magic.
- Evaluation first. We assembled a labelled set of real (anonymised) documents and agreed accuracy thresholds before writing pipeline code.
- Structured extraction. An LLM extracts fields into a strict schema; deterministic rules validate totals, dates and cross-document consistency.
- Human in the loop. Low-confidence fields are routed to reviewers in a purpose-built interface, and corrections feed back into the eval set.
- Compliance by design. Data stays in-region, providers are configured to exclude data from training, and every decision is logged.
The outcome
Officers now review exceptions instead of re-typing documents. The evaluation suite runs in CI, so model or prompt changes cannot ship if quality drops.