Can an AI agent issue credit notes?

For narrow, rule-based cases, yes; otherwise it prepares and a person approves. A credit note is money given away on paper. An AI agent can issue one at depth 3 only where a written rule fixes the amount, such as a pricing error against the contract. Discretionary credits need an approver independent of the sale.

Why are credit notes a fraud risk?

Credit notes are the quiet way to move value. A salesperson can inflate an invoice to hit a target and credit it back next quarter. A credit controller can pocket a customer's payment and cover the gap with a credit note against the same account. A customer can claim a short delivery that never happened. In each case the ledger stays balanced, which is why credit notes are among the first documents an auditor samples.

So the question is not whether the agent can produce the document. It can. The question is who decides that the customer is owed the credit, and whether that person benefits from the decision.

When can the agent issue one directly?

Only when every one of these holds:

• the reason is on a short list of rule-based cases: a price on the invoice that differs from the signed contract, a duplicate invoice, a quantity billed above what the delivery record shows;
• the evidence is sealed: the contract, the delivery record, the original invoice;
• the amount equals the difference the evidence shows, exactly;
• the amount is within the customer's limit for agent-issued credits;
• the invoice is not already in dispute with a person, and the customer account is not flagged for review.

Everything else, including goodwill gestures, service failures, commercial concessions and returns needing inspection, is drafted at depth 2.

Who prepares, who approves?

Separation of duties sits in the approver column. The person approving a discretionary credit must not have raised the original invoice, must not be paid on that account's revenue, and must not be the person who handles that customer's cash.

An illustrative example

Take an office supplies wholesaler. A customer emails to say they were charged 14.50 per unit for 400 units of printer paper when their contract price is 12.00.

The agent pulls the signed contract, which confirms 12.00, and the invoice, which shows 14.50 because the price list changed and the customer's contract override was missed. The difference is 1,000, inside the agent's limit. It issues a credit note for exactly 1,000, sends it to the billing contact on file, and records the contract clause and invoice as evidence.

The same customer also asks for a further credit "for the hassle". That is goodwill. The agent drafts a small credit with the case history and routes it to the account manager's manager, who approves a lesser amount. The agent then fixes nothing in the price list itself; it reports the missed override so billing can check other customers on the same contract terms.

What controls make it safe?

• A closed list of agent-issued reasons. Anything not on it is drafted.
• Amounts from evidence, not requests. The credit equals the documented difference, not what the customer asked for.
• Independent approvers for discretion. Named per reason and per value band.
• Credit notes linked to invoices. Each one references the invoice it reduces, so unlinked credits stand out.
• Patterns reported. Repeated credits to one account, or many raised by one person, go to the finance lead.
• Case notes on every credit. The reasoning is written down at the time, in the way described in writing case notes after every contact.

How does this differ from a refund?

A credit note reduces a balance owed. A refund sends money back. Where a customer has already paid, a credit note may become a refund, and that step moves money, so it falls under refunds through your payment gateway and its own limits. The rules for the invoices themselves are in whether an AI agent can issue customer invoices, and the wider controls are in AI agents for finance: books you can audit.