How to train the lending team to review AI agent decisions

Loan officers should leave training able to defend any decision to the applicant, an auditor or a supervisor. Teach how the agent assesses applications and why it refers, calibrate on real referrals and declines, shadow live decisions with every one checked, and sign off officers whose decline reasons are specific, accurate and fair. Exceptions to policy stay with named people.

What does the lending team review?

A lending agent typically gathers applications and documents, verifies income and identity, checks affordability against policy, runs the scorecard or decision rules, drafts a decision with reasons, and refers anything outside its bounds: borderline affordability, thin credit files, self-employed income, policy exceptions, and signs of vulnerability. In many lenders it may approve small, clean applications within tight bounds at depth 3, and a person makes every decline and every exception.

Lending carries particular weight in regulation. In the EU, the AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk. Under GDPR, decisions based solely on automated processing with legal or similarly significant effects are restricted. In the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used. Whatever the market, the loan officer's review is what makes a referral or a decline a considered human decision, and the reasons they record are what the applicant sees.

The head of lending owns the programme. Credit risk sets the policy the agent applies. Compliance reviews the decline standard.

What does the programme look like, step by step?

1. Learning, week 1: officers complete lessons on how the agent decides: the policy and scorecard as bounds, the sources it uses and their provenance, and why it refers.
2. Agent, depth 1: builds a practice set of past referrals and declines with outcomes, anonymised.
3. Calibration, week 2: officers decide alone, writing the reasons they would give the applicant. Then compare decisions and the reasons.
4. Shadow, weeks 3 and 4: officers decide live referrals; a senior underwriter checks every decision and every reason before it is issued.
5. Agent, depth 2 and 3 in live work: drafts decisions; approves small clean applications within bounds; refers the rest.
6. Sign-off: head of lending and compliance assess shadow decisions on accuracy, reasons and fairness.
7. Ongoing: weekly sampling of agent approvals and officer declines; monthly calibration.

What must officers learn to check?

Income evidence. Payslips and bank statements uploaded by the applicant are external until verified; data from a verified feed is stronger. Officers check what the affordability calculation rested on.

Affordability judgement. The agent applies the policy's numbers. Officers consider what the numbers miss: irregular income, a known upcoming change, commitments the applicant mentioned that do not appear in the data.

Decline reasons. Specific, accurate and understandable. "Does not meet our criteria" is not a reason. The real factors, in plain words, are.

Exceptions. Requests to lend outside policy go to named people with authority. Officers prepare them; they do not grant them unless their delegation allows.

Vulnerability. Signs that the applicant may be in difficulty, may not understand the product, or may be under pressure from someone else. The agent flags what it sees; officers act on it before deciding.

Consistency. Similar applicants should get similar outcomes. Calibration exposes officers who are stricter or looser than their peers.

For an illustrative calibration case, take a self-employed applicant asking for 8,000 over four years. The agent refers it because the most recent year's income came from the applicant's own figures rather than filed accounts. Some officers decline; others approve on the strength of two years of verified bank credits that match the declared figure. The agreed standard names the evidence that makes declared self-employed income acceptable, and the agent's brief is changed to ask for it up front.

Which controls leave evidence?

• Every decline decided by a named officer. Evidence: officer's name and reasons.
• Lessons before shadow. Evidence: completion dates.
• Shadow decisions checked. Evidence: senior underwriter's name.
• Exceptions approved within delegation. Evidence: approver and authority level.
• Agent approvals sampled. Evidence: weekly sample log.

What should you measure?

• Shadow accuracy and reason quality per officer.
• Declines overturned on complaint or appeal.
• Arrears on agent-approved and officer-approved loans.
• Variance between officers on calibration cases.
• Time to decision on referrals.

Where does it usually go wrong?

Reasons are copied from the draft. Officers accept the agent's reasons without checking they reflect what decided the case.

Referrals pile up. Officers clear the queue by agreeing with the draft. Watch approval times; seconds per referral means no review.

Exceptions creep. Officers start granting small exceptions informally. Sample for it.

Vulnerability training is separate. Teach it on the same cases.

How support team leads and paralegals review shows the same habit in different settings; see how support team leads review an AI agent's decisions and how paralegals review an AI agent's decisions. New officers use the reviewer onboarding checklist for the lending team, and IT staff supporting the platform follow how to train the IT team to review AI agent decisions. The full programme is in training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's Operator track covers how a governed agent decides and human review, with module tests set in the learner's own industry and role, so officers work lending cases. The written exam has 40 questions with a pass mark of 32, and the practicum needs two weeks of real decisions with no missed red line. A lender can make decline authority depend on a current certificate.