How to train the credit team to review AI agent decisions

Credit teams should review the agent's memos, proposed grades and early-warning flags the way they would review a junior analyst's work, only more systematically. Teach how the agent builds a memo, calibrate on real files until grades converge, shadow live reviews with a senior checking each grade, and sign off analysts who can explain every override with evidence.

What will the credit team review?

A credit agent typically spreads financial statements, calculates ratios and covenant headroom, drafts credit memos for new lending and annual reviews, proposes a risk grade, and monitors the book for early warning signs such as overdraft hardcore, missed covenant certificates or adverse bureau changes. Its work sits at depth 1 and 2. Credit decisions stay with people holding delegated authority.

Reviewing that output is different from reviewing operational agents. There is less volume and more judgement. The cost of a weak review shows up months later as a default rather than the same day as a complaint. And the agent's proposed grade can quietly anchor the analyst, who adjusts slightly from it rather than forming an independent view.

The head of credit owns the programme. Credit risk management, or the model risk function where the grading approach is treated as a model, agrees the sign-off standard. Model risk is supervised in most banking markets, so review records matter beyond the credit team.

How does the programme run, step by step?

1. Learning, week 1: analysts complete lessons on how the agent builds a memo: the sources it reads, their provenance labels, how it proposes a grade, and why it flags.
2. Agent, depth 1: assembles a practice set of past memos with the final grade and, where available, how the exposure performed.
3. Calibration, week 2: analysts grade the set alone, without seeing the agent's proposed grade first, then compare. Wide spreads are discussed with the head of credit.
4. Shadow, weeks 3 to 5: analysts review live memos; a senior analyst checks every grade and recommendation before it goes to an authority holder.
5. Agent, depth 2 in live work: drafts memos and proposes grades; flags early warnings.
6. Sign-off: head of credit assesses shadow reviews on grade accuracy and the quality of override reasons.
7. Ongoing: quarterly calibration and annual re-check, plus a refresher when the grading approach changes.

What must analysts learn to check?

Where each number came from. Audited figures, management accounts, the borrower's projections and your own account conduct data carry different weight. The memo labels them. Analysts check the grade does not rest mainly on the weakest.

The grade, independently. Form a view before reading the proposed grade. Then compare. This is the single most useful habit to train, because anchoring is invisible to the person anchored.

Override reasons. Where the analyst departs from the agent's grade, the reason must be specific and evidenced. "Strong management" is not a reason; "management delivered the last three years' budgets within a narrow margin, per audited accounts" is.

Covenant calculations. Check the definitions the agent applied against the facility agreement. Definitions vary, and an EBITDA calculated the standard way may not be the agreement's way.

Early warning flags. Each needs a decision: investigate, watch, or dismiss with a reason. Dismissed flags without reasons are a common audit finding.

What the memo leaves out. Industry conditions, key-person risk, related-party exposure. The agent drafts from what it reads; analysts ask what is missing.

Consider an illustrative file. The agent proposes grade 4 for a manufacturer with a facility of 2 million, citing stable margins from management accounts. The analyst, grading first, reaches grade 5: audited accounts show margins fell in the last full year, and account conduct shows the overdraft at its limit for most of the past six months. The override reason records both sealed and signed sources. That is the habit the programme exists to build.

Which controls leave evidence?

• Independent grade recorded before the proposed grade is shown. Evidence: timestamped analyst grade.
• Overrides reasoned. Evidence: override log with reasons.
• Shadow reviews checked. Evidence: senior analyst's name.
• Sign-off by head of credit. Evidence: assessment record.
• Early warnings dispositioned. Evidence: a decision on every flag.

What should you measure?

• Grade agreement in calibration, over sessions.
• Override rate and direction. Overrides almost always upward is a warning.
• Performance by grade over time, for agent-proposed and analyst-overridden grades.
• Early warning flags undecided past their deadline.

Where does it usually go wrong?

Anchoring. Analysts see the proposed grade first and drift toward it. Hide it until they have graded.

Shadow period too short. Credit volumes are lower than operational ones. Count memos reviewed, not weeks.

Covenant definitions assumed. Train on files where the agreement's definition differs from the standard.

Model risk and credit train separately. Bring model risk into calibration once a quarter so both see the same evidence.

Analysts' individual review is covered in how credit analysts review an AI agent's decisions, and new starters use the reviewer onboarding checklist for the credit team. Credit control reviews a different kind of credit decision; see how to train the credit control team to review AI agent decisions and how credit controllers review an AI agent's decisions. The full programme is in training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's Operator track teaches analysts how a governed agent decides and how to review its drafts, with module tests set in the learner's own industry and role. Banks that build their own credit agents also need people on the Builder track, which ends in a capstone agent certified in Attest to depth 2 or higher, so the people who design the memo agent understand what reviewers need.