A chargeback is the user (via their card issuer) reversing a charge. Even when the agent's transaction was legitimate, the burden of proof is on the merchant. Agent-initiated payments add a new dimension — proving the agent acted within authorized scope — and require new evidence categories.
The chargeback flow
User disputes a charge with their card issuer. Issuer pulls funds from merchant immediately. Merchant has ~30 days to provide evidence. Issuer reviews; usually decides based on documentation quality. Lost dispute: merchant loses the money + dispute fee ($15-25). Won: funds returned.
Standard evidence packages
Proof of delivery (tracking). Proof of service (login records). Customer communication (emails, support tickets). Receipt of consent (terms acceptance). Lots of templates exist; populate from your actual data.
Agent-specific evidence
Signed agent consent record showing user authorized agent X to spend up to Y for Z. Agent's transaction trace including LLM reasoning. User's confirmation message if any. This documentation is structurally new for agent payments; build the capture from day 1.
Friendly fraud vs real fraud
Friendly fraud: user genuinely bought it, then disputes ('I don't recognize this charge'). Real fraud: stolen card. Friendly fraud is winnable with good evidence. Real fraud isn't worth fighting; just refund and improve detection.
Dispute rates as a signal
Above 1% dispute rate, payment processors start charging fees and threatening termination. Above 2%, you're in the merchant blacklist tier. Track per-merchant, per-agent, per-product. Spikes signal product or fraud issues.