Can an AI agent refund bank charges?

Yes, inside a written refund policy. An AI agent can refund a charge the bank applied in error, and give goodwill refunds within limits the policy sets per customer and per period, at depth 4 within those limits. Refusing a refund, anything above the limits, and customers whose charges point to financial difficulty go to a named person.

What decides whether a refund is the agent's to give?

Bank charges fall into three groups, and each is decided differently.

Charges applied in error. The tariff says the fee should not apply, or should be smaller, or a grace period was ignored. This is a factual check against the terms in force on the date. If the charge was wrong, the customer is owed the money, and the agent should refund it without making them argue.

Goodwill refunds. The charge was correct, but the bank's policy allows refunds in set circumstances, such as a first unarranged overdraft fee in a year. The agent can apply the policy exactly as written.

Everything else. Repeated requests, large sums, a pattern of charges or a complaint. These need judgement about the customer's situation and the bank's wider duties.

Why does consistency matter more than generosity?

Goodwill refunds handled by people often follow the loudest voice. Customers who complain firmly get money back; customers who do not ask, or who give up, keep paying. That is unfair, and under rules such as the FCA's Consumer Duty in the UK, the outcomes customers get are what supervisors look at.

An agent applying a written policy treats every customer who meets the conditions the same way. That only helps if the policy itself is fair. If the bank's rule allows one goodwill refund a year, the agent gives one, to anyone eligible, whether they ask politely, angrily or not at all, where the bank chooses to refund proactively.

What may the agent do, and what stays with a person?

The detailed process is in handling fee refund requests on bank accounts with an AI agent, and the questions customers ask before they request a refund are covered in answering questions about fees and charges with an AI agent.

An example: three overdraft fees in a week

Take a customer who messages the bank about three unarranged overdraft fees in one week. They ask for all three back.

The agent checks each fee against the tariff. The first was correct. The second was charged on a day when a salary credit had arrived before the cut-off time, so the account was not overdrawn at the end of the day: an error. The agent refunds it and records the cause. The third was correct.

The policy allows one goodwill refund of this fee type per year, and the customer has had none. The agent refunds the first fee as goodwill. That leaves the third. Rather than drafting a refusal, the agent notes the pattern: three overdrafts in a week, a balance below zero most days this month. It sends the case to the support team, flagged as possible financial difficulty. A person calls the customer, refunds the third fee under the difficulty policy, and talks through an arranged overdraft and a budgeting plan.

What controls make it safe?

• The tariff in force on the date. Fee checks use the version of the terms that applied, held in the policy library.
• Limits per customer and per period. Goodwill refunds are capped in the customer's own limits, which the agent cannot exceed; see how to set refund limits for an AI agent.
• Refusals held. Telling a customer no is drafted by the agent and sent by a person.
• Difficulty screening before any refusal. Patterns of charges route to support.
• Error patterns become incidents. The same error across accounts is fixed at source, not refunded one complaint at a time.

The same principles apply to other refunds; see whether an AI agent can approve a refund. Where fee refunds sit among other capabilities is shown in the bank operations depth matrix, and the general limits are in AI agents in banking: what they may and may not do.