A lender that denies credit must still give the applicant specific reasons, whether an underwriter, a scoring model or an AI agent shaped the decision. Agents make this harder: the real reasons spread across many steps, and the agent's own account may sound right without being true. The answer is design: know which facts decided, and record them.
What do the rules require?
In the United States, the Equal Credit Opportunity Act and Regulation B require a creditor that takes adverse action, such as denying an application, to give the applicant a statement of specific reasons. The requirement does not fall away because a model or complex system was used; the reasons must still be specific and must reflect the factors that actually drove the decision. The CFPB supervises and enforces these rules for many lenders. Our explainer on what an adverse action notice is covers the basics, and our page on ECOA and Regulation B reasons when an AI agent is involved goes into the requirement.
Other jurisdictions reach a similar place by different routes. In the EU, creditworthiness assessment and credit scoring of individuals are listed as high-risk under the EU AI Act, which brings duties on transparency and human oversight, and GDPR Article 22 gives people rights where a decision with significant effects is based solely on automated processing. The common thread is that a person refused credit is entitled to understand why.
Why does an agent make the reasons harder to give?
A traditional scoring model produces a score and, usually, a ranked list of the factors that lowered it. Those become the reasons. The chain is short and stable.
An agent's involvement can be longer and less tidy. It might read bank statements, extract income and outgoings, flag irregular deposits, compare the application against policy, summarise the file and recommend a decision. A person or a model then decides. By the time the decision is made, the facts that drove it have passed through several steps, and any of them could have shaped the outcome.
Three specific problems follow.
Diffuse causes. The real reason for a denial might be a combination of the agent's reading of income and the policy rule it applied. If nothing records which facts the decision turned on, the reasons have to be reconstructed after the fact.
Plausible but untrue narratives. Ask an agent why an application was declined and it will produce a fluent explanation. That explanation is generated text, not a trace of what actually drove the decision. It may name factors that sound right but were not decisive, or omit the one that was.
Generic reasons. Under pressure, organisations fall back on broad categories like "insufficient income" or "credit history". The rules expect reasons specific enough to tell the applicant what actually counted.
What should the agent be allowed to do in a credit decision?
Its role should be designed so the reasons stay traceable.
| Step | Agent | Person or decision model |
|---|---|---|
| Gather documents and data | Reads and extracts, depth 1 | Checks extraction on a sample |
| Assess against policy | Drafts an assessment with the facts it relied on, depth 2 | Reviews the assessment |
| Decide approve or decline | Does not decide | Decides, using defined criteria |
| Identify principal reasons | Drafts reasons from the recorded deciding factors, depth 2 | Confirms the reasons match what decided |
| Send the notice | Drafts the notice | A person sends, or a certified process within strict bounds |
The key line is between assembling facts and deciding. If the agent recommends and a person or a defined decision model decides, the reasons can be tied to the criteria that were applied. If the agent decides on its own, the reasons depend on reconstructing its reasoning, which is the weakest position to defend. Our piece on whether adverse action notices apply when an AI agent declines credit explains why they do.
How do you keep the reasons accurate?
Record the deciding factors at the time of the decision, not afterwards. Every fact the agent extracted should carry its source. Every policy rule applied should be named. The decision record should show which rules and facts led to the decline, in an order that reflects their weight. The reasons on the notice should be drawn from that record, not written fresh.
Then test it. Take a sample of declined applications and check that the stated reasons match the recorded deciding factors, and that changing those factors would have changed the outcome. Where they do not match, either the record or the notice is wrong. This is also good preparation for explaining decisions under other regimes; our guide to EU AI Act explainability and giving reasons for an agent's decision covers the European angle.
What will an examiner or auditor look for?
Whether reasons are specific, whether they are accurate, and whether the lender can show how they were derived. Expect requests to trace a sample of declined applications from the data the agent extracted, through the policy applied, to the decision and the notice. Expect questions about who decided, what the agent contributed, and how its contribution was checked. A lender that can produce that chain from its records is in a far stronger position than one that asks the agent to explain itself afterwards.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What is the trade-off?
Designing for traceable reasons limits what the agent may do. It may not be allowed to decide, even where it could, and its assessments must be structured and sourced rather than free-form. That reduces some of the efficiency lenders hoped for.
It also avoids a failure that is hard to repair: notices whose reasons cannot be tied to what actually happened. Before an agent touches a credit file, decide where its role ends, make sure the deciding factors are recorded at the moment of decision, and draw every adverse action reason from that record. The wider mapping of rules to controls is in our guide to AI agent regulation, mapped to controls.
Where Pryme fits
In Pryme, every fact reaching an agent is stamped with its provenance, and every verdict is filed with the sources relied on and the bound that decided. A credit-related capability can be held at draft so a named person decides, and the record shows which facts and rules led to each outcome.
Questions people ask
Can the agent's own explanation be used as the adverse action reason?
Not on its own. An agent's explanation is generated text and may not reflect what actually drove the decision. Reasons should come from the recorded deciding factors, with any agent-drafted wording checked against them.
Does a human reviewer remove the need for specific reasons?
No. Whoever decides, the applicant is still owed specific reasons. A human decision-maker using an agent's assessment must make sure the stated reasons reflect the factors that actually led to the denial.
Should the notice say that an AI agent was involved?
The core duty is to give specific, accurate reasons, and naming the tool does not replace that. Some rules separately require people to be told when they are interacting with an AI system. Whether and how to mention an agent in a notice depends on the jurisdiction, so confirm it with counsel.