How to prepare credit memos for underwriters with an AI agent

The agent drafts the memo; the underwriter owns it. At depth 1 the agent reads the file, spreads the financials, summarises the score and tests each policy rule. At depth 2 it drafts every section in your template with each figure cited. The underwriter rewrites what they disagree with and makes the decision; the agent never recommends approval.

What is a credit memo, and who owns it?

A credit memo is the document an underwriter or credit committee relies on to decide a lending request. It sets out who the borrower is, what they want and why, how they will repay, what security supports the loan, what could go wrong and which policy rules the request meets or breaks. A good memo lets someone who has never met the borrower make a sound decision.

The underwriter owns the memo once it reaches them. In many banks the relationship manager or a credit analyst prepares it first. Either way, the decision belongs to whoever holds the lending authority. The agent replaces the hours of gathering and formatting, not the judgement.

What are the steps?

1. Agent, depth 1: reads the file. Application, financial statements, bank statements, valuations, the bureau report, prior memos for existing customers.
2. Agent, depth 1: spreads the numbers. Income statement and balance sheet into the bank's template, with ratios. Each figure points to its page in the source.
3. Agent, depth 1: reports the score. It states the score your credit scoring model produced and the main factors the model returns. It does not re-score or explain the score in its own terms.
4. Agent, depth 1: tests policy. Each lending policy rule that applies is listed as met, not met or exception required, with the evidence.
5. Agent, depth 2: drafts the memo. Borrower, purpose, repayment source, security, risks, policy position. The risks section is mandatory and comes before any mitigants.
6. Person, underwriter: edits and decides. Approve, decline, approve with conditions, or refer.
7. Agent, depth 2: drafts a referral note when the request sits outside the underwriter's mandate, using the bank's credit referral note template.
8. Agent, depth 1: records the edits. What the underwriter changed, section by section.

Which controls apply, and what evidence do they leave?

Citation for every figure. No number appears in the memo without a source reference. The evidence is the memo itself: a reviewer can click through from any figure to the page it came from.

Provenance labels. Your own account data is sealed. Accounts and payslips supplied by the borrower are declared. Bureau data is external. The memo shows the class next to key figures, so an underwriter knows which numbers rest on the applicant's word.

No recommendation field. The template the agent fills has no "recommend approve" slot. The agent describes; the underwriter concludes. Any draft text that reads as a recommendation is caught at certification under the scope gate.

Edit log. Every change the underwriter makes is filed. This is how you learn where the agent is weak, and it is evidence that a person engaged with the memo rather than signing it unread.

Bureau access as a separate grant. Pulling the bureau report is its own capability, with recorded consent. Whether an agent can run a credit check is answered capability by capability, not by the memo grant.

What should you measure?

• Sections rewritten by underwriters. Track by section. Rewrites of the risks section matter most.
• Policy breaches the underwriter found that the agent missed. Each one is a certification case for the next run.
• Time from complete file to memo ready.
• Memos returned for missing evidence.
• Decisions reversed on later review where the memo was agent-drafted, compared with those that were not.

Where does it usually go wrong?

Advocacy creep. Memos drift into sales documents. "Strong management team with a proven track record" is a claim, not a fact. The agent should write only what the evidence shows, and leave character judgements to the RM and underwriter.

Invented mitigants. Asked to list mitigants, a model will find some. "Personal guarantee provides comfort" is only true if the guarantor has means. Each mitigant must cite evidence, or it is left out.

Spread errors that propagate. One misread figure changes every ratio. The agent should reconcile the spread back to the totals in the statements and flag any line it could not read with confidence.

Prior memo bias. For an existing customer, the agent reads the last memo and repeats its view. Mark what is carried forward, as you would in annual credit reviews.

Memo before file. A memo drafted from an incomplete file looks finished and hides the gap. The file should pass completeness checks first, which is the job of preparing applications in the origination system.

Where Pryme fits

In Pryme, memo drafting is certified through Attest on your own past files, scoring every section against your underwriters' final versions. The grounded gate checks that each claim traces to a source; the scope gate checks that the agent stops short of a recommendation. The broader limits are in AI agents in banking.