Library · Recipe · Banking

Credit memo first draft from a borrower pack

Two tested prompts, a fictional SME borrower pack with 15 hidden problems, a 25-point answer key, the human checkpoint and the full regulatory fit box.

The story behind this one, and the three problems it fixes: read Srini's write-up.

This is the companion to The RM says approve. The numbers say decline. Read that first. This is the reference you come back to when you run it yourself.

Where each problem is covered

Tuesday's issue works through three problems. Here's where to find the detail for each.

1. The pack

Ferncastle Bank and Brindle & Oak Joinery are fictional. Every name and number is invented. Everything reconciles: the P&L, the balance sheets, the debtor ageing.

Download the pack

2. How we scored it

The answer key has 15 planted problems and 3 traps.

The five that change the decision score 3 points each:

• the existing covenant is breached once you remove an unapproved add-back
• leverage worked on forecast EBITDA instead of actuals
• the revolver excluded from net debt
• pro forma debt service cover of 0.66x
• projections that need second-half EBITDA to almost double

The other ten score 1 point each:

• interest cover
• margin trend
• the margin uplift timed before the machine arrives
• debtor concentration
• debtor days
• the hardcore revolver
• loan term vs lease
• loan-to-cost
• stale KYC
• sanctioning authority

The three traps cost 3 points each:

• flagging a properly discounted screening hit as a live risk
• accepting the add-back
• using the competitor's offer as a reason to approve

That makes 25 the maximum. 17 or more with no traps is a usable first draft.

3. Results

Found by both, not planted by us:

• a £180k balloon. A 12-month capital holiday then £180k a year leaves £180k to repay at the end.
• no funding for the £150k deposit or the VAT
• the covenant breach may already be past its 10-day reporting deadline

Found only by Prompt 2:

• Debt capacity. On last year's actual EBITDA, about £150k of new debt fits a 3.0x limit, and about £420k fits the 3.5x exception ceiling. That turns a decline into a conversation about what the bank could do.
• A flaw in our own policy. Section 3.2 lets a Senior Credit Officer approve leverage up to 3.5x. Section 9 says any exception goes one level up. The model spotted the conflict and asked the policy owner to settle it. We wrote that policy. We missed it.

A few rows from Prompt 2's claims test. It tested 30 claims from the RM (6 contradicted, 11 unsupported) and 7 from the covenant certificate.

4. Prompt 1: the one-line baseline

Run this first if you want to see what the model makes of a case on its own.

I'm a credit analyst. Attached is a borrower pack for [customer], our SME
credit policy and our memo template. Please write the credit memo for this
request using the template.

5. Prompt 2: check, then write

Save it once as your team's memo checker (a Copilot agent, a Claude project, or your prompt library). Don't retype it.

You are a senior SME credit analyst preparing a memo for a Senior Credit
Officer. Attached: the RM's credit application, the customer's financials,
management accounts, projections and aged debtors, the existing facilities
and covenant certificate, the KYC record, our SME credit policy and our memo
template.

New loan pricing: SONIA [x]% + [margin]%. Pro forma on [date] balances.

Treat the RM application and the customer's covenant certificate as claims to
be tested, not facts. Work in two stages.

STAGE 1 — CHECK (show this as tables before the memo)
1. Facts table: pull every number you'll use from the source files, with the
   file it came from.
2. Recompute: net debt / EBITDA, interest cover and debt service cover, actual
   and pro forma, using the policy's definitions and the facility letter's
   definitions word for word. Show the workings. Where you have to assume
   something, say so and use a stated estimate.
3. Claims test: list every factual claim in the RM application and the
   covenant certificate. Mark each Verified / Contradicted / Unsupported, with
   the evidence.
4. Policy test: go through every section of the policy extract and mark Met /
   Not met / Needs info, citing the section.
5. Trend notes: margins, debtor and stock days, RCF utilisation, current-year
   trading vs projections. Explain causes where the files show them; say 'not
   explained in the pack' where they don't.

STAGE 2 — MEMO
Write the memo in the template's structure. The recommendation must follow
from Stage 1, not from the RM's view. Name every policy exception and who can
approve it. If something needs to happen before any approval (e.g. a breach, a
waiver, a KYC refresh), say so in Recommendation and Conditions. Don't raise
issues the files don't support; if a screening hit was properly discounted,
say that and move on. Keep the memo to roughly two to four pages.
Put Stage 1 in an appendix.

What changed from the first version, and how it tested. We added the pricing and pro forma date line, and moved Stage 1 to an appendix. Re-run on 13 Nov with SONIA 4.25% + 3.25% and 30 September balances: 25/25, no traps, 3 min 27 s. Interest cover came out at 2.90x, every pound of it traceable to the stated rate. It also caught one thing earlier runs missed: interest cover on the last twelve months is already 3.82x, below the existing 4.00x covenant.

6. The human checkpoint

The analyst does this before anything reaches the sanctioner. Allow 45 to 75 minutes (our estimate).

1. Tick the facts table against the source files. One wrong number and you can't trust the rest.
2. Re-run leverage, interest cover and debt service cover yourself, on the policy's definitions.
3. Read every Contradicted and Unsupported row. Decide which go back to the RM.
4. Separate facts from inferences. Prompt 1 called the £63k relocation add-back "hard to sustain" because another move looked likely. Prompt 2 said "no relocation is evident". Both are judgements, not facts.
5. Own the recommendation. The model recommends. The analyst and the sanctioner decide.

7. What went wrong

• The one-line prompt launders the RM's story. Unevidenced claims come out as facts. Prompt 2's claims test fixes it.
• Assumptions drift between runs. The interest rate and the pro forma date both varied. Fixed: the prompt now states both.
• It's long. Prompt 2 ran to 5,300 words. Fixed: Stage 1 now goes in an appendix, so the memo reads first.
• Our test was too easy to separate the prompts. Both maxed the key. The next pack will be harder.

8. Time

9. FS fit box

Base regulation

• [ ] Consumer Duty. Not engaged. The borrower is a limited company with £6.2m turnover, outside the FCA's retail customer definition for this lending. Recheck for sole traders or micro-enterprises.
• [x] SM&CR. SMF4 (Chief Risk Officer) owns the credit policy. The sanctioning mandate still decides. The tool sits inside the analyst's existing role.
• [x] PRA SS1/23. It drafts and calculates but doesn't decide. Register it as a tool with a named owner and a human sign-off. If people start relying on its ratios without re-checking, treat it as model-like and bring it into scope.
• [x] UK operational resilience and third-party risk. Your AI provider processes confidential customer data. Contract, retention and exit terms matter. For example, the Claude models released on 9 November need 30-day retention unless zero data retention is authorised (release notes). EU entities: DORA ICT third-party rules.
• [x] EU AI Act. For EU lending, creditworthiness assessment of natural persons is Annex III high-risk. A limited company isn't. Personal guarantor assessments and sole-trader lending would be. The AI literacy duty (Art 4) applies to the analysts using it.
• [x] ICO / UK GDPR. The KYC file holds directors' personal data. You need a documented lawful basis, a DPIA for the use case, and minimisation: redact dates of birth and home addresses before upload. Art 22 isn't engaged, because a human makes the credit decision.

Sector add-ons: none for banking beyond the base set.

Operating controls

• Data allowed: fictional Ferncastle dataset only. Real-world equivalent: confidential customer commercial data, plus personal data (KYC).
• SMF owner: SMF4 Chief Risk Officer.
• Human checkpoint: credit analyst checks the Stage 1 tables against source files. Sanction per mandate.
• Stacks tested: Tested on Claude.

10. Raw outputs

All three memos, unedited, are in the pack download: runs/p1-one-line-prompt.md, runs/p2-analyst-first-prompt.md and runs/p2b-with-pricing-and-appendix.md.

If your credit team wants to run this on its own packs, with the checkpoint built into your process, that's a half-day workshop. See the credit team workshop