How long does it take to train credit analysts to supervise AI agents?

Seven to eight weeks is a sensible plan. Credit analysts need about 18 hours of learning, two weeks of supervised decisions on live credit files and a final exam, plus time to practise checking an agent's financial spreads line by line. That skill separates a reviewer who catches a wrong number from one who approves a confident memo.

What makes credit review slow to learn

A credit analyst reviewing an agent's work is not checking a reply or a refund. They are checking a spread of financial statements, a set of ratios and a drafted credit memo, each built from documents of mixed quality. One misread line, such as a one-off gain treated as recurring income, changes the answer.

Agents are good at assembling these files quickly. The analyst's new job is to test them: which figures came from audited accounts, which from management accounts, and which from a borrower's covering email. That is a provenance question, and it is the core of the training.

For a role that reviews faster, simpler decisions, compare how long it takes to train credit controllers.

How the time divides

What the agent does and what the analyst decides

• Depth 1: the agent reads financial statements, bureau reports and account conduct.
• Depth 2: it drafts spreads, ratio analysis and a credit memo with a recommendation.
• Depth 3, rarely: it may refresh routine annual reviews on low-exposure facilities within strict bounds, where it has earned that.

The analyst keeps the credit decision, the risk grade, any exception to policy, and the covenants. Where the borrower is a natural person, credit scoring and creditworthiness assessment are listed as high-risk under Annex III of the EU AI Act, and in the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used. Analysts must be able to state the reason in their own words, not the agent's.

A worked example

Take a commercial lender's credit team of eight. Foundations in week one, then Operator learning at about three and a half hours a week. In weeks four and five, each analyst re-spreads four files the agent built and compares line by line. Two analysts find the agent took revenue from a borrower's emailed management pack rather than the filed accounts, which were older but signed by an auditor. The agent had flagged the source, but the memo read cleanly enough that neither analyst checked.

That finding goes into the agent's brief, and into calibration on the same files, as described in calibration sessions for credit analysts. Supervised live review then runs for two weeks, and the exam follows in week eight.

Borrowers and relationship managers sometimes lean on analysts for a fast approval, and documents can carry planted instructions. See training credit analysts to spot prompt injection and pressure.

What keeps the analyst's sign-off meaningful

• Re-spread a sample of agent-built files every month, not only during training.
• Keep the agent at depth 2 for any new borrower.
• Treat a memo that relies on declared figures for a decisive ratio as incomplete.
• Know when the agent's certificate lapses; a run that clears no depth drops it to depth 1 the same day, and annual reviews return to the queue.

Fraud analysts work at higher speed with a different kind of judgement; compare the fraud analyst timeline. The wider programme is in training teams for AI agents.

Where the Academy fits

Credit analysts who review agent work day to day take the Operator track of Pryme Intelligence Academy after Foundations. Certified Pryme Operator needs 32 of 40 questions on the written exam and a two-week practicum of real review decisions, signed off by a Pryme assessor. The file-checking practice above is the lender's own addition, and a sensible one.