A credit memo is mostly assembly. Someone pulls the borrower’s financials, reads the latest filings, checks internal exposure and covenant history, then spends hours arranging all of it into a document that an approver can read in twenty minutes. The thinking happens at the end. Most of the clock is spent before it, copying numbers between systems and reconciling figures that three sources report three different ways.
That assembly is where AI earns its place, and it is also where it gets dangerous if you are careless. The risk is not that the model writes a bad sentence. The risk is that it writes a confident sentence with a wrong number in it, and the wrong number reads exactly like a right one.
What the draft can do, and what it cannot
Credit memo automation, done properly, produces a first draft and nothing more. The system reads the audited statements, the most recent filings, internal limit and utilisation data, and prior memos on the same name. It extracts the spreads, computes the ratios the desk actually uses, drafts the business description and the risk narrative, and lays it out in the bank’s own template. What lands on the officer’s desk is a starting point that already has the mechanical work done.
It does not make the credit decision. The grade, the structure, the conditions, the call on whether a weak quarter is a blip or a trend, all of that stays with the officer. An underwriting copilot is a useful framing here: it sits next to the person doing the work and speeds up the parts that are clerical, without touching the part that requires judgement and carries accountability. The model proposes; the credit officer disposes, and signs.
That line matters for more than comfort. Under most supervisory regimes the human sign-off is what makes the decision attributable. An adverse action notice has to rest on reason codes a person stands behind. Automating the draft does not move that obligation, and it should not pretend to.
Every figure cited back to its source
The single feature that makes this safe is mundane: citation. Every number in the draft carries a link back to where it came from, the exact filing, page, line item or internal record, with the date it was pulled. The officer does not have to trust the model. They click the EBITDA figure and see the income statement it was lifted from. They check the leverage ratio and see both inputs and the period each belongs to.
This is the same lineage discipline that makes any finance AI system auditable, applied at the document level:
- Filings extraction that records the source document, page and field for every extracted value, not just the value
- Reconciliation when sources disagree, with the discrepancy surfaced in the draft rather than silently resolved
- Point-in-time correctness, so a memo dated at quarter-end uses the figures that were actually filed by then, with no later restatement leaking in
- Entity resolution, so the subsidiary in the filing and the obligor in the internal system are confirmed to be the same legal entity
Without citation, document automation just relocates the manual work: instead of assembling the memo, the officer re-verifies every figure from scratch, which is slower than writing it themselves. With citation, verification becomes a scan, and the audit trail is a by-product of how the draft was built rather than something reconstructed afterwards.
Why this cuts time without cutting accountability
The time saving comes from the boring part. Pulling, formatting and cross-checking figures is the bulk of the hours, and it is exactly the part a machine does well when its outputs are traceable. The judgement part was never the bottleneck, so leaving it untouched costs nothing in speed.
Accountability survives because nothing about the decision changed. The same person reviews, the same person signs, and now they spend their attention on the credit rather than on data entry. When a reviewer, a model validation team or a regulator later asks why the memo said what it said, the answer is in the document: each claim points to its source, the human sign-off is recorded, and the reasoning is the officer’s own.
We would treat anything that blurs that line as a defect, not a feature. A drafting system that cannot show its sources, or that nudges toward a grade, has stopped being a copilot. The goal is narrow and worth stating plainly: take the assembly off the officer’s plate, and leave the credit entirely on it.