Credit analysts remain the deciding control because an AI agent can assemble an affordability case but cannot answer for it. The agent spreads the figures, checks policy and drafts the memo; the analyst decides whether the numbers are true, whether the risk is acceptable and whether the reasons will stand up when a customer, auditor or supervisor asks.
What does an agent actually take off a credit analyst's desk?
A lot of the assembly. At depth 1 the agent reads the application, the bureau file and the account history. At depth 2 it drafts: it spreads income and expenditure, calculates the affordability ratios the lender uses, flags policy breaches and writes a first credit memo with a recommendation. In some lenders a narrow band of small, clean applications may later be certified for depth 3, where the agent executes a routine decision within tight bounds.
What it does not take is the judgement. A recommendation is a draft. Approving or declining credit for a person changes their finances and, in many markets, their legal position. That decision sits with someone who can be named in the record.
Why is a complete-looking memo the main risk?
Agent-drafted memos look finished. The ratios are calculated, the headings are filled, the recommendation is confident. A tired analyst at the end of a long queue sees a tidy page and signs. That is the failure mode, and it is a training problem, not a technology problem.
The analyst's real job on a drafted memo is to ask where each number came from. Under the Pryme Context Protocol every fact reaching the agent carries its provenance. A salary figure taken from the lender's own ledger of incoming payments is sealed. An income statement from the applicant is declared. A payslip sent as an attachment arrived from outside and is external. The agent's ratio is only as good as its weakest input, and only a trained analyst will check that before agreeing with it.
Which judgements must stay with the credit analyst?
The list below is where analysts earn their place once agents prepare the file:
- Deciding which income is real. Bonus, overtime and self-employed income need judgement about stability that a policy table cannot fully capture.
- Reading the story behind the bureau file. A missed payment during an illness and a pattern of rising arrears look similar in a score and very different to an experienced eye.
- Granting or refusing exceptions. Policy exceptions are the analyst's call, with a reason recorded.
- Writing reasons that stand up. Where credit is declined, the reasons need to be specific and true. In the United States, ECOA and Regulation B require specific reasons for adverse action even when a model is used.
- Spotting what the agent was never told. Two applications from the same address within a week, a business account behaving like a personal one: context the agent's sources do not hold.
How does the regulatory picture raise the stakes?
Creditworthiness assessment and credit scoring of natural persons is listed as high-risk in Annex III of the EU AI Act, which brings human oversight under Article 14 and, for deployers, AI literacy under Article 4. GDPR Article 22 constrains decisions based solely on automated processing that have legal or similarly significant effects. Regulators in most banking markets also supervise model risk.
None of that is a reason to keep agents out of credit. It is a reason to make sure the person in the loop is trained to oversee, not just to click. An analyst who cannot explain why the agent recommended a decline is not providing oversight in any meaningful sense. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What does a credit analyst need to learn?
Four things, mostly. First, depth: what the agent may do on each capability and why a draft memo is depth 2 however good it looks. Second, provenance: how to read the source stamp on every figure and when to send a declared or external fact back for verification. Third, verdicts: what it means when the agent holds, refuses or has an action withheld, and which bound decided. Fourth, the record: every verdict and every human decision is filed in an append-only record, so the analyst's approval and their stated reason will be read later, possibly by someone hostile.
The detailed review method is in how credit analysts review an AI agent's decisions. For team-level planning, the competency matrix for a credit control team shows how skills can be graded.
How does this differ from fraud and KYC review?
Neighbouring roles review agent work under different pressures. A fraud analyst decides in minutes whether to release money; a KYC analyst decides whether to let a customer in at all. Credit sits between: slower than fraud, more numerical than KYC, and with a decision that the customer will often challenge. See how fraud analysts and KYC analysts carry the same responsibility, and the overview in AI agent skills by role.
Where the Academy fits
Pryme Intelligence Academy's Operator track covers how a governed agent decides, human review and keeping knowledge current, with module tests set in the learner's own industry and role, so a credit analyst is tested on credit cases. Certified Pryme Operator adds a two-week practicum of real review decisions, signed off by a Pryme assessor. A lender can make approval rights depend on a current certificate.
Questions people ask
Can an agent decline a credit application on its own?
Only if that capability has been certified to execute and the lender's own limits allow it, and many lenders choose never to allow it. Declines are often held for a person because the reasons will be challenged and must be specific.
Do credit analysts need to understand how the model works?
Not its internals. They need to understand what the agent was given, how far it may act and what the record shows. That is enough to oversee it; model design belongs to builders.
What happens if an analyst disagrees with the agent's recommendation?
They decide differently and record why. The disagreement is filed with the verdict, and patterns of disagreement are useful evidence for improving the agent's brief or sources.