Are credit analysts accountable for an AI agent’s mistakes?

Yes, for the analysis they sign. When an AI agent spreads the financials, calculates the ratios and drafts the credit memo, the analyst who puts their name to the recommendation owns it. They do not own a spreading error they could not have seen, or a credit decision taken by a committee above them. The record makes each of those boundaries visible.

Where does the analyst sit in the decision chain?

Credit work usually has three layers. The agent prepares at depth 1 and 2: it reads financial statements, extracts figures, runs ratios and covenant tests, and drafts the memo. The analyst reviews, corrects and recommends. A credit officer or committee decides. The agent should never approve credit.

Lending to individuals is also sensitive in law. The EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk. In the United States, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used. An analyst who signs a memo is signing the reasoning that may later need to be explained.

What may the agent do, and what stays with the analyst?

• The agent may: extract figures from statements, normalise them, calculate leverage, cover and liquidity ratios, compare against policy thresholds, draft the memo and list the risks it finds.
• The agent holds: anything where sources conflict, where figures come only from management accounts, or where a policy threshold is close.
• The analyst keeps: judging the quality of the figures, adjusting for one-off items, weighing management and market risk, and the recommendation itself.

What does the record show?

Provenance is the heart of credit accountability. Audited accounts filed by the borrower's auditor are signed: a named party answers for them. Management accounts sent by the borrower are declared. A bank statement pulled from the lender's own ledger is sealed. The record shows which figure in the memo came from which source.

It also shows what the analyst changed. If the agent built a cover ratio from declared management accounts, marked it as such, and the analyst signed without adjusting or querying it, the record shows that. If the agent took a figure from the wrong period and presented it as audited, the record shows that instead, and accountability moves towards the agent's design.

An example: the covenant that looked comfortable

Take a mid-sized business lender. A borrower requests a facility increase of 500,000. The agent spreads the last audited accounts and the latest management accounts, and calculates interest cover at a comfortable level, mainly from the management figures. It marks those figures as declared.

The analyst, under deadline, signs the memo without asking for the debtor ageing behind the management accounts. Six months later the borrower breaches. The record shows the declared label, the cover calculation and the analyst's sign-off. The analyst owns the recommendation. The committee owns the decision it made on that memo.

What controls keep this fair?

1. Label provenance in the memo itself, not just in the background.
2. Hold memos that rest mainly on declared figures for a senior reviewer.
3. Train analysts on the lender's own files first, through a two-week practicum for credit analysts, logged in a practicum log template for credit analysts.
4. Keep up with the agent. Agent certificates lapse after 90 days and spreading rules change; see how to prepare credit analysts for 90-day recertification.

The question plays out differently for people who chase overdue accounts; compare credit controllers and AI agent mistakes and fraud analysts and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.