Executive Brief

Credit Memo Refinery

One page. Read in three minutes.

The problem

A mid-size commercial bank has 20,000+ credit memos sitting in file shares. Every one of them contains proprietary information — our borrowers, our collateral decisions, our analysts' judgments, our realized losses — and none of it is queryable.

Loan officers lose deals to competitors who price sharper because their comp data is better. Credit analysts re-extract the same fields from the same documents on every annual review. Portfolio managers cannot answer "which of my current borrowers look like the ones that defaulted in 2008?" without a multi-week project. Exam teams rebuild concentration and vintage reports from scratch every quarter.

A fully-loaded credit memo takes 15 to 40 analyst hours to produce. Each annual review re-touches the same data. A regulatory exam cycle is another 40 to 200 hours of manual extraction. Multiply by 20,000 archived memos and the dormant asset is worth eight figures in analyst time alone — plus an unknowable sum in deals lost to faster, better-informed competitors.

Solution pattern

A small team of specialized agents, one orchestrator, one verifier. Not code — a pattern.

Orchestrator reads each page and routes it by content type.
Specialists — form, table, narrative, visual — each handle one kind of content, with multiple methods and fallbacks.
Proofreader runs arithmetic checks against the extracted data: sources equal uses, assets minus liabilities equal net worth, DSCRs reconcile across pages. Nothing ships until the math holds.
Supervisor handles flagged errors — retry with a different method or escalate.

Input: a PDF credit memo. Output: structured JSON with every field linked to a specific source page, plus a verification report. Human checkpoint: one — exception triage when the supervisor cannot reconcile automatically.

Before and after

	Today	With the refinery
Time per memo	15–40 analyst hours	Minutes, unattended
Cost per memo	Hundreds of dollars in analyst time	Cents in compute
Verifiability	"Trust the analyst"	Arithmetic reconciled, page-level provenance
Cross-memo query	File share search, manual linking	Structured query
Exam prep	40–200 hours per cycle	Hours
Error tracking	None — no audit trail	Bounded by an explicit check set

Numbers are representative for a mid-size commercial lender. The shape of the improvement is the point.

Key design decisions

The proofreader extracts nothing. Most teams merge extraction and validation. We separate them, because an extractor cannot catch its own mistakes. A dedicated verification role — running real arithmetic against real expected values — turns raw AI output into data a risk organization will trust.

Content-type routing, not a monolithic pass. A single AI call over the whole document works, but failure modes are hidden inside a confident-looking answer. Routing each page to a specialist by content type exposes failures where they can be caught and recovered. This is what makes the error-recovery story credible.

The contract is arithmetic, not vibes. Every acceptance check in the demo is a specific equation with a specific expected value drawn from the real document. AI output looks trustworthy; arithmetic is the only way to prove it actually is.

Scope — two phases

Phase 1 builds the refinery on one sample credit memo — extraction with arithmetic verification. This proves the hardest and least-proven link: can AI extract data from a complex PDF and can we prove it got the numbers right?

Phase 2 expands to the full loan file. A real credit decision touches many document types — personal financial statements, rent rolls, appraisals, scanned tax returns. Phase 2 asks teams to build a pipeline that classifies documents, routes them to specialized extractors, and verifies data across documents. It also introduces the time dimension: origination documents vs. servicing reviews years later. This is what it takes to scale to 2,000+ loan files.

Both phases deliberately exclude the data warehouse, downstream insight agents, persistent storage, authentication, and compliance workflows. The focus is the extraction pipeline — because that is the credibility anchor for the entire business case.