How to handle chargeback disputes with an AI agent

Most of a chargeback is evidence and deadlines, which suits an agent. It takes the cardholder's dispute, sorts it by reason, checks the scheme's time limits, and gathers the evidence. It can raise standard chargebacks within bounds once certified. Large provisional credits, merchant challenges, arbitration and debits back to the cardholder all stay with a named disputes analyst.

What is the process, and who owns it?

A chargeback is the card issuer's route to recover a disputed payment from the merchant's bank, under the card scheme's rules. For an issuer, the dispute process runs from the cardholder's complaint to a final outcome: refund to the cardholder, or the charge stands. On the way there are strict deadlines, defined reason codes, evidence requirements and, if the merchant pushes back, further rounds that carry fees.

The head of disputes owns the process. Disputes analysts own individual cases. Fraud owns any dispute where the cardholder says they did not make the transaction; those go to a separate route. For the wider frame, see our guide to AI agents in banking.

What happens at each step, and at what depth?

1. Agent, depth 1: takes the dispute and classifies it. Goods not received, not as described, cancelled but still charged, duplicate, wrong amount. "I didn't make this" goes to the unauthorised transaction route. Evidence: reason category and the cardholder's words.
2. Agent, depth 1: checks eligibility. Time since the transaction against the scheme limit for that reason; whether the cardholder contacted the merchant first, where the reason requires it; transaction type. Evidence: eligibility result with dates.
3. Agent, depth 2: requests evidence from the cardholder. Order confirmations, cancellation emails, delivery tracking, correspondence with the merchant, using the chargeback dispute evidence checklist for that reason. Evidence: request sent and items received.
4. Agent, depth 4 within limits: applies a provisional credit where your policy gives one. Above the limit, held. Evidence: verdict and credit reference.
5. Agent, depth 3: raises the chargeback for standard reason codes, below the value limit, where evidence meets the checklist. Otherwise drafted for an analyst. Evidence: submission and reason code.
6. Agent, depth 2: reviews the merchant's response. If the merchant challenges, the agent compares their evidence against the cardholder's and drafts a recommendation. Evidence: comparison and draft.
7. Person, disputes analyst: decides the next round. Accept the merchant's case, or go to pre-arbitration or arbitration, which carry fees and are hard to reverse. Evidence: decision and reasons.
8. Person, disputes analyst: approves any debit back to the cardholder if the dispute is lost. Evidence: held verdict, approval, and the message sent.

Whether an agent should raise chargebacks at all is answered in whether an AI agent can raise a chargeback for a cardholder.

Which controls matter, and what evidence do they leave?

Deadlines as bounds. Each reason code's time limit is configuration, and the platform refuses a submission after it. Evidence: dates on every verdict.

Fraud separated at intake. Unauthorised transactions follow different rules and often different liability. Evidence: routing decision recorded.

Evidence from the cardholder is declared. It supports the claim but has not been verified. Evidence: provenance class per item.

Provisional credit limits. Evidence: running totals per cardholder and per day.

Debits back are always held. Taking money from a customer who believed the matter settled needs a person and a clear explanation. Evidence: held verdict and message.

One claim per transaction. A cardholder who disputes in the app and again by phone must not produce two chargebacks or two provisional credits. Evidence: duplicate claims linked to one case.

Merchant contact is not the agent's job. The agent does not contact the merchant directly outside the scheme process. Evidence: no outbound merchant contact in the log.

What should you measure?

• Submissions after the deadline. Target zero.
• Disputes won and lost, as counts by reason code and by whether the agent or an analyst raised them.
• Time from intake to submission.
• Provisional credits later debited back.
• Cardholder complaints about the dispute process.

Where does it usually go wrong?

Wrong reason code. The dispute is filed under the easiest code rather than the right one and fails on technical grounds.

Merchant contact skipped. For reasons that require the cardholder to have tried the merchant first, the chargeback is rejected.

Billing disputes treated as chargebacks. Many disputes, especially about the bank's own charges, are not card scheme matters at all; see how to handle customer billing disputes with an AI agent.

Arbitration by default. The agent escalates every merchant challenge, and fees pile up on cases the issuer cannot win.

Silent debits back. A lost dispute is debited from the cardholder without explanation, producing a complaint worth more than the transaction.

Where Pryme fits

In Pryme, intake, eligibility checks and evidence gathering run at read and draft; raising standard chargebacks and applying provisional credits are separate capabilities, each certified on your own past disputes and bounded by your limits. Scheme deadlines are enforced at the moment of action, and debits back are held for a named analyst.