Only downwards, and only when the customer asks. An AI agent can lower a credit limit on request, because that reduces borrowing and can be reversed. Raising a limit extends new credit, so the bank's approved model and policy decide, with referrals to a person. Cutting a limit on the bank's own initiative also stays with a person.
Why is a credit limit increase a new credit decision?
A higher limit lets the customer borrow more. If they cannot afford it, the harm builds over time: a larger balance, higher interest, and possibly arrears. Undoing a limit increase does not undo the debt already drawn against it.
So a limit increase needs the same care as a new application: an affordability check, a look at how the customer has used credit, and a decision under the bank's credit policy. The EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk. GDPR Article 22 and its UK equivalent restrict decisions based solely on automated processing with legal or similarly significant effects; our piece on GDPR Article 22 and AI agents in credit decisions covers what that means in practice. For how scores feed these decisions, see what credit scoring is.
Why does a bank-initiated cut need a person too?
Reducing a limit sounds protective, and when the customer asks, it is. When the bank does it on its own, the customer may suddenly find a payment declined, or their balance over the new limit and charged for it. Cuts can be the right decision, for example when a customer shows signs of financial strain, but they need notice, a clear reason and, often, support. That is a judgement for a person who can weigh the customer's circumstances.
What may the agent do, and what stays with a person?
| Change | The agent | Stays with a person or the approved model |
|---|---|---|
| Lowering the limit on the customer's request | Lowers it at depth 3, if the balance is below the new limit | Requests that would put the balance over the new limit |
| Raising the limit on request | Gathers data and runs the approved assessment | The decision, under policy; referrals to a credit officer |
| Unrequested increase offers | Never initiates | Product governance decides whether offers are made at all |
| Bank-initiated reductions | Identifies accounts that meet policy triggers | The decision, notice and support |
| Annual reviews | Prepares the review pack | Decisions arising from it |
| Explaining a decision | Explains from the recorded reasons | Reconsideration |
Annual reviews of credit facilities are a natural place for the agent's preparation work; see how to prepare annual credit reviews with an AI agent.
A worked example
Take a credit card provider. A customer with a limit of 3,000 asks to lower it to 1,000, saying they want to stop overspending. The balance is 400. The agent confirms the effect, lowers the limit at depth 3 and records the request. It also notes that the customer can ask to raise it again later, but that any rise will need a new assessment.
Another customer asks for an increase from 2,000 to 6,000. The agent runs the approved assessment: a credit search, income verification and a review of the account. The model refers the case, because the customer has paid only the minimum for eight months. The agent holds it for a credit officer, with the account history and the model output. The officer offers 3,000 rather than 6,000, and the agent explains the decision to the customer from the recorded reasons.
Note what the agent did not do: it did not decide that 3,000 was right. It did not suggest a figure. It brought the evidence and carried out the officer's decision.
What controls make it safe?
- Direction-aware bounds. Lowering on request is a separate capability from raising. Raising is never an agent decision.
- Balance checks. A requested reduction below the current balance is held, because it would put the customer over limit.
- Credit file effects. Any change reported to credit reference agencies follows the bank's standard process. The agent does not report changes on its own; see whether an AI agent can update a customer's credit file.
- Reasons recorded at decision time. Each increase decision carries the reasons the model or officer gave, so later explanations are accurate.
- Credit checks under policy. The agent runs the searches that policy requires, and no others; see whether an AI agent can run a credit check.
The wider limits are set out in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent tell a customer how much they might be approved for?
No. An estimate is heard as an offer. The agent can explain how the assessment works and how long it takes.
Can the agent restore a limit that the customer lowered last month?
Restoring a higher limit is still an increase in available credit. It goes through the assessment, though the bank's policy may set a lighter process for recent reductions.
Does lowering a limit affect the customer's credit score?
It can, because scoring often looks at how much of the available credit is used. The agent should mention this when the customer asks to lower a limit, so the customer decides with the facts.