Regulation B requires a lender that denies credit, or offers worse terms, to give the applicant the specific principal reasons for that action. An AI agent's involvement changes nothing about the duty. It changes the method: reasons must come from the factors the decision actually turned on, recorded at the time, never from an explanation the agent writes afterwards.
What does Regulation B require when credit is refused?
ECOA and Regulation B require a creditor to notify an applicant of adverse action and to give a statement of specific reasons, or tell the applicant how to request them. Adverse action is broader than a flat decline. It includes a counteroffer the applicant does not accept, closing an account, an unfavourable change to its terms, and refusing a request to raise a limit.
The reasons must be the ones that actually drove the decision, stated specifically enough that the applicant understands what happened and what might change it. "Did not meet our internal standards" or "insufficient score" are not specific reasons. A checklist reason that is close to the real one but not accurate is not good enough either. The duty holds when a model or algorithm is used, however complex the method. Our explainer on what an adverse action notice is covers the notice itself.
Where a consumer report contributed, separate credit reporting law adds its own disclosures. That sits alongside Regulation B, not in place of it.
Who must give the reasons?
The creditor. That means every bank, credit union, finance company and fintech lender that makes or decides credit, for consumers and businesses. When several parties take part, such as a broker and a lender, the rules let one notice satisfy the duty, but the creditor that made the decision must be able to stand behind the reasons.
Using a third-party agent does not shift the obligation. If a vendor's agent contributed to the decision and you cannot say how, you cannot state accurate reasons. Our guide to AI agents in banking explains why credit decisions stay with a named person.
What makes reasons harder when an agent took part?
With a scorecard, reasons come from the variables that cost the applicant the most points. With an agent, three problems appear.
The agent shapes what the decider sees. It chooses which documents to retrieve, what to highlight in a summary and what to leave out. If its summary flags "irregular deposits" and the underwriter declines on that basis, irregular deposits is a principal reason, whether or not it appears on any scorecard.
The agent can rationalise. Ask a language model to explain a decline after the fact and it will produce fluent, plausible reasons. They may not be the reasons that drove the decision. A notice built that way can be inaccurate while reading perfectly.
The agent's influence is spread across steps. Retrieval, summary, recommendation and a drafted note can each push the outcome a little. We explain why adverse action reasons are harder when an agent helped in more depth.
The answer is to capture decision factors as structured data at the moment they are used, and to generate notices from that record.
How do you give accurate reasons with an agent involved?
Treat the reason record as part of the decision, not as paperwork after it.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Reasons are the ones actually used | The underwriter selects decision factors from a controlled list at the moment of decision; free text cannot replace the selection | Factor record timestamped with the decision |
| Agent influence is captured | Every claim the agent puts in a summary is tied to its source document | Summary with source references, stored with the file |
| No after-the-fact rationalising | The agent drafts the notice only from the factor record, never from its own reconstruction | Notice draft linked to the factor record it used |
| Reasons are specific | Each controlled factor maps to plain, specific applicant wording, reviewed by compliance | Approved wording library with version history |
| A person stands behind the notice | A named person confirms the notice before it is sent | Approver identity and time on each notice |
| Reasons stay accurate over time | Sampled notices checked against files each month | Sample results and corrections |
The agent's work here is real and useful. It can draft the notice, check that each stated reason appears in the factor record, and flag a decline that has no recorded factors. That is depth 2: it drafts and a person sends. Our playbook on how to draft credit decline reasons with an AI agent sets out the steps.
Take one declined file
Take a small personal loan application. The agent retrieves three months of bank statements, notes two returned payments and a falling balance, and recommends decline. The underwriter agrees, selecting "recent returned payments" and "insufficient available income" from the factor list.
The agent drafts a notice from those two factors using approved wording. A compliance check confirms both factors are in the record and both trace to the statements the agent cited. A named person approves and sends. If the applicant later asks why, the file shows the documents, the agent's summary, the factors chosen and who chose them.
Now the failure case. The underwriter declines without selecting factors, and someone later asks the agent to "write the reasons". The agent produces "limited credit history", which sounds right but played no part. That notice is inaccurate, and nothing in the file reveals it.
What will an examiner ask?
- Show us how the reasons on these notices were chosen, and when.
- Where an agent summarised the file, did anything in its summary drive the decision but not appear on the notice?
- Can the agent generate reasons without a recorded factor list?
- How do you map internal factors to the wording applicants see?
- How often do you sample notices against files, and what have you found?
- Who approves each notice?
What does Regulation B not settle?
Regulation B tells you to give specific, accurate reasons. It does not tell you how to validate the agent, which falls under model risk guidance such as SR 11-7. It does not govern fair lending testing of outcomes, which is a separate ECOA duty. It does not cover data protection, and lenders serving people in the EU face separate automated-decision and AI Act duties. Colorado has passed an AI act aimed at high-risk systems that make consequential decisions, which may add requirements.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can an agent send an adverse action notice on its own?
It can prepare and queue one, but a named person should confirm the reasons before it goes. An agent sending notices without review risks inaccurate reasons reaching applicants at scale.
Do we need to mention that an agent was involved?
Regulation B asks for the specific reasons, not a description of the tools. What matters is that the reasons reflect the factors actually used, including any the agent introduced.
What if the agent's summary and the underwriter's factors disagree?
Treat it as a quality finding. Either the underwriter relied on something not recorded or the summary misled, and both need correcting before the notice is sent.