Can lenders use AI agents before staff are trained?

Yes, for the paperwork. A lender can start an agent collecting documents, checking them against each other and drafting affordability calculations while credit staff train. Credit decisions, declines and disbursements must wait until the people reviewing them are trained, because lending to individuals is heavily regulated and an untrained approver cannot explain a decision the law says must be explained.

Why is lending a strict case?

Lending decisions about individuals attract specific duties in most markets. The EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk. In the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is involved. Conduct rules almost everywhere expect lenders to lend responsibly and treat borrowers in difficulty fairly.

These duties land on the people who approve. If a reviewer cannot tell why an agent recommended a decline, they cannot give the borrower an accurate reason.

What can agents do first?

1. Document collection. Chasing missing payslips or statements with standard messages a person sends (depth 2).
2. Verification. Matching declared income to bank statements and flagging mismatches (depth 1 and 2).
3. Affordability drafts. Calculating income, commitments and headroom, showing the source of each figure (depth 2).
4. Collections preparation. Summarising an account in arrears before a call (depth 1).
5. Policy questions. Answering staff queries from the current credit policy with citations (depth 1).

What must wait?

• Approving or declining an application without a person.
• Writing the final decline reasons.
• Disbursing funds (depth 4), which stays held for a named approver even after training.
• Forbearance and payment arrangements for borrowers in difficulty.
• Any action on a customer flagged as vulnerable.

What does a sensible start look like?

Take a consumer lender offering personal loans online, with a credit team of fifteen. It starts an agent drafting affordability assessments. Underwriters review each one.

In the first month, the credit manager samples approved drafts and finds that the agent treated a regular transfer from a partner's account as the applicant's income. Two underwriters accepted it. The figure was declared by the applicant and not supported by a payslip, and the agent's draft showed that provenance, but nobody looked. The manager runs a short session on reading provenance and updates the brief so that income from a third party is always flagged. The whole team completes foundational training; the senior underwriters who approve the most complex cases work towards operator certification. A full plan is in how lenders should train staff before deploying AI agents, and a ready structure is the AI agent training plan template for lenders.

Which controls fit a lender?

• The lender's own limits keep agents at depth 2 for all credit work until trained reviewers and a certified agent are in place.
• Declines held for a named underwriter at every depth.
• Provenance shown for every figure; declared income never treated as verified.
• Sampling of approvals as well as declines, since a generous agent is as dangerous as a harsh one.
• Every verdict filed, so a decision can be explained to a borrower, a complaints body or a supervisor.

Lenders make some familiar errors, such as training only on the application journey and not on collections, covered in what lenders get wrong about AI agent training. Utilities face a similar challenge with payment plans; see whether utilities can use agents before staff are trained. Lenders that want an external standard for reviewers can use a certification such as Pryme Intelligence Academy's Operator track. The topic guide on training teams for AI agents covers the order of training.