Credit analysts use this log during a two-week practicum in which they review an AI agent's financial spreads, covenant checks and draft credit memos. For each request, the analyst records the figures relied on and where they came from, their recommendation and conditions, a senior analyst's independent call, and any red line. The template and three sample entries follow.
Why is provenance the centre of a credit log?
In credit analysis, the agent's arithmetic is rarely the problem. It spreads accounts and calculates ratios reliably. The problem is what the numbers rest on. Audited accounts filed with a registry, management accounts the borrower prepared, and a projection in a pitch deck can all produce the same ratio, and they deserve very different weight.
So the log asks the analyst to classify each key figure: sealed (from the bank's own ledger, such as account conduct), signed (audited or otherwise attested by a named party), declared (the borrower's own management figures or forecasts), or external (pulled from a web page or an unsolicited email). A recommendation that leans on declared figures is not wrong, but the memo must say so.
How do analysts fill it in?
The agent pre-fills the request, its spread and its draft recommendation from the record at depth 1. The analyst then records:
- which figures they relied on and their provenance;
- any covenant test the agent ran, and whether the analyst agrees with the result;
- their own recommendation, conditions and rationale;
- whether the request sits within delegated authority or must go to committee.
The senior analyst records a recommendation without seeing the analyst's. The match is judged on the recommendation and the key conditions, not on the wording of the memo. How the fortnight is set up is covered in how to run a two-week practicum for credit analysts.
The practicum log template
Header: analyst; senior analyst with a current certificate; portfolio (for example, small business term loans and overdrafts); delegated authority during the practicum; dates; bar of at least 95 correct in every 100 and no missed red line.
Red lines
- Presenting declared or external figures as verified.
- Missing a covenant breach or a near-breach that the facility terms require to be reported.
- Recommending outside credit policy or delegated authority without saying so.
Entry fields
| Field | Entry |
|---|---|
| No. and date | |
| Facility ref and borrower type | |
| Request | New, increase, renewal, covenant waiver |
| Key figures and provenance | e.g. revenue (signed), forecast (declared) |
| Agent's draft recommendation and verdict | |
| Covenant results: agent and analyst | |
| Analyst's recommendation and conditions | |
| Analyst's rationale | |
| Authority | Within delegated limit / committee |
| Senior's recommendation | Recorded first |
| Match | Yes / no / defensible |
| Red line | None / caught / missed |
| Debrief note |
A worked example
Take a sample business bank's credit team. Values are illustrative.
| No. | Request | Key figures | Agent's draft | Analyst | Senior | Match | Red line |
|---|---|---|---|---|---|---|---|
| 22 | Renewal of a 150,000 overdraft for a builders' merchant | Audited accounts (signed); 12 months of account conduct (sealed) | Renew on same terms | Renew; add quarterly management accounts as a condition | Renew, same condition | Yes | None |
| 23 | 400,000 term loan for a restaurant group's new site | Latest audited accounts (signed); new-site forecast (declared) | Approve; debt service cover 1.6 times | Approve with cover on forecast noted as declared; stress case at 1.1 times; committee | Same, plus a personal guarantee | Defensible | Caught |
| 24 | Annual review of a 250,000 loan to a logistics firm | Management accounts (declared) show leverage of 3.4 times | Draft memo says "covenants met" | Approve review as drafted | Flag near-breach: limit is 3.5 times, trend rising; report under facility terms | No | Missed |
Entry 23 shows the skill the practicum is built to test: the agent's ratio was correct, but it was built on a forecast, and the analyst said so in the memo. Entry 24 fails the bar. The draft said "covenants met", which was technically true, and the analyst did not look at the trend.
Credit scoring of natural persons is a high-risk use under Annex III of the EU AI Act. For consumer lending teams, the same log applies with an added column for the reasons the customer would be given.
How does this sit beside fraud and KYC logs?
The three roles often review the same customer at different points. The practicum log for KYC analysts checks who the borrower is; the log for fraud analysts checks whether the activity is genuine; this log checks whether the bank should lend. A lender running all three can schedule them together, using how to run a two-week practicum for KYC analysts alongside the credit version. Teams working towards the Certified Pryme Operator from Pryme Intelligence Academy can use this log as the practicum record; see AI literacy and certification for people who work with agents.
Questions people ask
Should the analyst rewrite the agent's memo?
Only where it is wrong or incomplete. The practicum tests judgement, not drafting style, so log substantive edits and ignore cosmetic ones.
What if the senior and analyst disagree on a committee case?
Record both, mark it defensible if the analyst's reasoning holds, and let the committee decide on the live case as usual.
Does passing change accountability for credit decisions?
No. The credit decision remains with whoever holds the authority. See are credit analysts accountable for an AI agent's mistakes.