How can lenders prove staff are trained to oversee AI agents?

Lenders prove it decision by decision. For any credit decision an agent helped prepare, the lender should be able to show the approver, their current certificate checked at source, their right to approve at that depth and amount, and the reasons they confirmed. In credit, rules already expect a competent person behind the outcome, so the evidence will be tested.

Why credit makes the evidence matter

Two strands of rules make lending stricter than most sectors. Under the EU AI Act, creditworthiness assessment and credit scoring of natural persons are listed as high-risk in Annex III. Deployers of high-risk systems must assign human oversight to people with the necessary competence, training and authority. In the United States, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used.

Both create a direct question: could the person who approved this decision explain it, and were they trained to? A lender that cannot answer for a sampled decline has a problem with its supervisor, its complaints process and possibly a court.

What should the evidence set include?

• A credit oversight map. Each agent capability in the credit process, its depth, and the approving role. For example: affordability drafting at depth 2, approved by a credit officer; disbursement held at depth 4 for a named person.
• Authority matrix with certificates. Approval limits by product and amount, each tied to a current certificate.
• Certificate register, checked at source, with expiries.
• Practicum records. Real credit decisions approved under supervision before solo authority.
• Decision records. The agent's draft, the sources it used, the approver's changes, the final reasons and the approver's name.
• Reason quality sampling. Periodic review of decline reasons against the factors that drove the decision.

The AI agent training plan template for lenders lays out the training side of this.

What should credit staff be trained on?

1. Source weight. Income from the lender's own account data is sealed. A payslip is signed by an employer. Income the applicant typed in is declared. Bureau data is from an identifiable party. A decision memo that treats these equally is the commonest flaw an approver must catch.
2. Reasons that match the decision. Decline reasons must reflect the actual drivers, not plausible-sounding text.
3. Fair treatment. Spotting where an agent's proposals disadvantage a group, and escalating.
4. Depth and limits. Knowing why a decision was held, and that a refused action cannot be approved by hand.

An example from a supervisory review

A consumer lender uses an agent to draft affordability assessments and decision memos for personal loans. A supervisor samples 20 declined applications and asks the lender to show, for each, who approved the decline, whether they were trained, and how the reasons were chosen.

The lender produces the filed records. Each shows the agent's memo, the bureau and account data used, the credit officer who approved, their certificate number and status on the day, and the reasons they selected. In two cases the officer had changed the agent's proposed reasons because the main driver was a recent missed payment rather than the agent's stated debt-to-income ratio. Those edits are what show the officers were reviewing, not rubber-stamping.

Which controls keep the evidence sound?

• Approval rights tied to certificates, depth by depth, and removed on lapse.
• Second review of a sample of declines for reason accuracy.
• Briefings logged when credit policy changes, so approvers are trained on the policy in force.
• An append-only record of every verdict, approval and edit.

Market expectations differ; see BaFin expectations in Germany and MAS FEAT principles. For other sectors, see how utilities prove staff are trained and how wealth managers do it. The guide to AI literacy and certification covers the overall model.

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

Where the Academy fits

Pryme Intelligence Academy's Operator track prepares credit officers who approve agent-prepared decisions, with module tests set in lending. The practicum is two weeks of real review decisions with no missed red line, signed off by a Pryme assessor. A lender can adopt the practice of making approval rights depend on a current certificate, depth by depth.