Onboard a new credit analyst by having them check the agent's financial spreading line by line before they comment on its credit view. Week one: how the agent spreads statements, tests covenants and drafts memos, and where its sources come from. Week two: real files, with a senior analyst reviewing every recommendation. The credit judgement is theirs from day one.
What is the agent doing for the credit team?
In a commercial lending or credit risk team that runs agents, the credit agent typically extracts figures from borrowers' financial statements and management accounts, spreads them into the bank's template, calculates ratios, checks covenant compliance, gathers bureau and registry data, and drafts sections of the credit memo. It works at depth 1 and 2. The risk grade, the recommendation and any approval stay with people.
A new analyst therefore starts with most of the mechanical work done. Their contribution is checking that the spread is right, understanding the business behind the numbers, and forming a view. The shift is described in what changes in a credit analyst's day when AI agents arrive. The credit team head owns onboarding; a certified senior analyst coaches.
What does the two-week plan contain?
- Day 1. Person (credit team head): explains the agent's bounds: what it extracts, what it drafts, what it may never decide.
- Day 1 to 2. Person (new analyst): learns the provenance of credit data. Audited accounts are signed by the auditor. Management accounts are declared by the borrower. Bureau data is signed by the bureau. Account conduct from the bank's own core system is sealed. Ratios built on declared figures carry less weight.
- Day 2 to 3. Agent, depth 2: spreads statements and drafts memo sections on live files, reviewed by the coach while the new analyst follows each adjustment.
- Day 4 to 5. Person (new analyst): re-spreads two past files in a sandbox workspace from the source documents, then compares with the agent's spread to find where it went wrong.
- Week 2. Person (new analyst): prepares live files, checking and adjusting the agent's spread and writing the credit view, with the coach reviewing every recommendation.
- Week 2. Person (new analyst): records reasons for each spreading correction.
- End of week 2. Person (credit team head): signs off the analyst to prepare files for approval without line-by-line review, with sampling for the first quarter.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Extract | Pulls figures from statements | 1 | Checking against source |
| Spread | Maps to template, calculates ratios | 2 | Adjustments and normalisation |
| Covenants | Tests against facility terms | 2 | Breach assessment |
| Memo | Drafts factual sections | 2 | The credit view |
| Grade and recommend | Nothing | none | Analyst and approver |
What must a new analyst check every time?
Classification. Agents commonly misclassify one-off items: a gain on asset sale counted as operating income, a lease reclassified under new accounting rules, related-party balances buried in other debtors. The analyst checks these first.
Period alignment. Management accounts covering eleven months set beside audited accounts covering twelve will distort every ratio.
Covenant definitions. The facility agreement defines terms like EBITDA in its own way. The agent should read the actual definition; the analyst confirms it did.
What the numbers do not say. Customer concentration, a key person, a lost contract. The agent gathers facts; the analyst asks the borrower and the relationship manager.
Group structures. Where a borrower sits in a group, the agent may spread the borrowing entity alone. The analyst checks guarantees, intercompany balances and whether the consolidated accounts tell a different story.
How should corrections be recorded?
Each spreading correction gets a reason: misclassified item, period mismatch, covenant definition, source error, or analyst adjustment. The last code marks a legitimate judgement, such as normalising an exceptional cost, so it is not taken for an agent error. Clear reasons let the agent owner change the agent's brief or sources where the error is real, rather than every analyst fixing the same thing.
Which controls and evidence should onboarding leave?
- Re-spread exercise. Evidence: the two practice files and the differences found.
- Reviewed recommendations in week two. Evidence: coach comments on each.
- Staged sign-off. Evidence: dated authority and the sampling plan.
- Reasons on corrections. Evidence: reason codes by analyst.
What should the team head measure?
Spreading errors the new analyst caught in week two. Recommendations changed by the coach and why. Time per file, expected to fall. After sign-off: a sample of files each month, and any later credit issues on files the analyst prepared.
Where does it usually go wrong?
Trusting the spread. It looks finished and the ratios calculate. A new analyst moves straight to the narrative, and a misclassified gain flatters the borrower's earnings.
Skipping the re-spread. Doing it once by hand shows the analyst what the agent does and where it struggles.
Memo by agent. The factual sections are drafted well, so the analyst's view shrinks to a sentence. The memo should show the analyst's thinking.
Analysts joining neighbouring teams follow related plans; see those for a new fraud analyst and a new KYC analyst. The wider approach is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy's Operator track teaches how a governed agent decides, human review and teaching the agent, with tests set in the learner's own industry and role, so a credit analyst practises on lending cases. Labs run in a free Build workspace, which suits the re-spread exercise above.
Questions people ask
Should a new analyst learn to spread manually first?
Yes, briefly. Re-spreading a couple of files teaches what the agent does and where it errs, which makes checking faster.
Can the agent propose a risk grade?
Some banks let it draft one with reasoning. Assigning the grade stays with the analyst and approver, and many lenders keep the agent out of grading entirely.
What if the borrower's figures change after the memo?
Re-run the spread from the new figures, and note the change in the memo. The record shows both versions.