Document classification built on a real data audit

Context

A commercial lender's underwriting team was spending several hours per application on document review — financial statements, contracts, supplier agreements, and regulatory filings. Document inputs were a mix: scanned PDFs of varying quality, native digital, multiple languages, and some legacy formats. Earlier vendor proposals had promised end-to-end automation but skipped over the inconsistency in the source data.

Approach

Before building anything, we ran a six-week readiness assessment. The audit covered document quality, format consistency, classification taxonomy, and how documents flowed through the team. The result was clear: roughly 40% of incoming documents needed pre-processing or normalization before any classifier could perform reliably.

We built the data preparation layer first — an OCR pipeline tuned for the actual document mix, normalization for the long tail of formats, and a small human-in-the-loop step for ambiguous cases. The classification model came after, trained on the cleaned input. Risk and compliance reviewed the audit findings before model work began, which made the project explainable from the start.

Outcome

• 4× throughput on document review per analyst.
• Underwriters now spend their time on judgment cases — covenant interpretation, risk assessment — rather than triage.
• The data foundation set up here became reusable. The same pre-processing layer is now feeding two further AI workflows in adjacent parts of the lending operation.

Principle illustrated

Readiness. Skipping the data audit would have produced a brittle classifier and a project that failed on edge cases. Doing the audit first made the model viable in production and explainable to risk and compliance — both prerequisites for getting AI into a regulated workflow.

Used with permission from the client and shared here for illustrative purposes. Specific commercial details remain confidential and are available on request to qualified counterparties under NDA.

← Back to OrchestrOnLab · More case studies · Discuss an opportunity