How to handle fee refund requests on bank accounts with an AI agent

Split every request into two questions. Was the fee charged correctly under the tariff? If not, the agent refunds it within limits and records the cause. If it was, the request is for goodwill, and the agent applies your goodwill policy exactly. Repeated fees that suggest money trouble, and errors that look systemic, go to a person.

What does the process involve, and who owns it?

Customers ask for fees back for many reasons: an unarranged overdraft fee after a salary arrived late, a returned payment charge, a foreign transaction fee they did not expect, a monthly fee on an account they thought was free. Some of these fees were wrong. Most were right, and the customer is asking for discretion.

Those are different decisions. An error refund corrects the bank's mistake and should be automatic. A goodwill refund is a concession that should be consistent: two customers in the same position should get the same answer. A third group sits underneath both: customers who are being charged repeatedly because they are struggling, for whom a refund is not the main need.

The head of customer operations owns the process and the goodwill policy. Finance owns refund limits. The financial support team owns cases showing difficulty. Whether an agent should refund charges at all is answered in whether an AI agent can refund bank charges; the wider limits are in our guide to AI agents in banking.

What does the agent prepare, and what does a person approve?

1. Agent, depth 1: identifies the fee. Type, date, amount, the event that triggered it. Evidence: the fee record.
2. Agent, depth 1: checks it against the tariff in force on that date. Was the fee due under the account terms? Was it the right amount? Was any grace period or cap respected? Evidence: tariff version and the check.
3. Agent, depth 4 within limits: refunds errors. If the fee was wrong, it refunds it, tells the customer plainly, and records the cause. Evidence: refund verdict and cause code.
4. Agent, depth 1: checks for a pattern. If the same error appears on many accounts, it raises an incident for a person rather than refunding one customer at a time. Evidence: incident reference.
5. Agent, depth 1: screens for financial difficulty. Several fees in a short period, a balance repeatedly below zero, or the customer saying they are struggling. These cases go to the support team. Evidence: screening result.
6. Agent, depth 4 within policy: applies goodwill. Only within the written policy: which fees, how often per customer, up to what amount. Evidence: policy version and the rule applied.
7. Person, team leader: decides goodwill outside policy. Evidence: held verdict and decision.
8. Agent, depth 2: drafts the reply for refunds outside its bounds and for refusals. Evidence: draft and sender.

Which controls matter, and what evidence do they leave?

Tariff by date. Fees are checked against the terms in force when they were charged, not today's. Evidence: tariff version on each check.

Error and goodwill kept apart. They are recorded with different codes, so error refunds can be traced to causes and goodwill can be checked for consistency. Evidence: refund type on every refund.

Goodwill as rules, not discretion. The policy is configuration the platform checks, not guidance the agent interprets. Evidence: rule cited on each goodwill verdict.

Running totals. Per customer over the policy period, and per day across the agent. Evidence: totals on each verdict.

Difficulty routed. Refunding fees for a struggling customer without addressing why they keep arising helps once. Evidence: referral to support.

Special accounts excluded. Accounts of deceased customers, accounts under power of attorney and accounts in a restructuring arrangement follow their own process. A fee query on a deceased customer's account, for example, belongs to the bereavement team; see whether an AI agent can handle a deceased customer's accounts. Evidence: refused verdicts on flagged accounts.

What should you measure?

• Error refunds by cause. The point is to fix causes, not keep refunding.
• Goodwill refunds per customer group, to check consistency.
• Requests refused and later overturned by a person or on complaint.
• Customers referred to financial support from fee requests.
• Incidents raised for systemic errors, and the number of customers each covered.

Where does it usually go wrong?

Goodwill becomes the default. The agent is generous because customers are pleased, and the policy is quietly exceeded. Enforce it as a bound.

The tariff check is skipped. Every request is treated as goodwill, so real errors are never traced to a cause.

Persistent fees mask difficulty. A customer asking for the third overdraft fee refund in a few months needs a conversation about their finances, possibly about overdraft options or, for loans, restructuring.

Systemic errors are refunded one by one. Only the customers who ask get their money back.

Refusals are blunt. "This fee was correctly applied" without explaining why leaves the customer with nothing. Explain the trigger and how to avoid it.