Commercial Bank Cuts Credit-Memo Turnaround 58% with Claude-Powered Document Agents
Agentic document review on Anthropic Claude — every extracted figure cited back to its source page — took middle-market credit memos from eleven days to under five, with the model-risk audit trail designed in.
A commercial bank's middle-market lending group was averaging eleven business days to move a borrower package to credit committee. Analysts read 200- to 400-page packages by hand — audited financials, tax returns, rent rolls, existing loan agreements — spreading figures into templates and hunting covenant language across documents that never used the same wording twice. Two earlier LLM pilots had been stopped by model risk management: neither could show where an extracted number came from, so nothing could be validated under SR 11-7.
Apptad built an agentic document pipeline on Anthropic Claude via Amazon Bedrock, with specialized agents for classification, financial spreading, covenant extraction, and first-draft memo narrative. Scanned and photographed pages pass through Amazon Textract first; every figure the system produces carries a citation to the document, page, and line it came from, and the analyst UI puts the source page beside the extracted value for one-click confirmation. We built the evaluation harness before the assistant — 400 historical memos with analyst-validated ground truth, run as a regression suite on every prompt or model change — alongside PII redaction, no-retention inference, and a full decision log landed in Snowflake. Nothing reaches committee without an analyst's approval: the agents draft, humans sign.
The first line of business went live in 14 weeks. Average turnaround fell from 11 days to 4.6 — 58% faster — with covenant extraction at 94% first-pass accuracy against the analyst-validated set, above the 90% threshold model risk had set as its condition for approval. The artifacts the eval harness produced cleared model risk review with no findings, and the same pattern is now rolling out to annual reviews and KYC refresh, where the document mix is broader but the citation requirement is identical.
The Numbers Behind the Story
Built On
The platforms and partners that powered this engagement. Where Apptad has a dedicated alliance, the chip links to our partnership page.
“Two vendors showed us assistants that sounded right. Apptad showed us one that could prove where every number came from — that's the only version our model risk team was ever going to approve.”
