Who Pays When AI Agents Make Costly Mistakes?

Payment companies, card networks and technology firms are creating rules and tools that let AI assistants make purchases while limiting what the software can buy and who is responsible when a transaction causes problems.

The work includes protocols, token systems and signed records that link a user’s instruction to a specific purchase. Google’s Agent Payments Protocol uses digitally signed ‘mandates’ that can record a maximum total price and the exact product approved. A signed mandate provides a documented record connecting an instruction to a payment if an assistant later books something different.

Card networks are developing agent-focused programs that combine authentication and authorization with controls on what an assistant may do. Card tokens replace raw card details with identifiers used to process payments. Some proposals use dollar stablecoins to move value between participants.

The payment method affects both control and redress. Crypto wallets can enforce strict, programmatic spending rules. Basic token transfers and some crypto payments lack card-style dispute mechanisms, so refunds depend on seller policy, escrow arrangements or contractual terms. Regulators including the Consumer Financial Protection Bureau outline circumstances in which a credit-card issuer may reverse a charge, but those processes do not guarantee refunds for every disputed automated purchase.

AI assistants also make many small, programmatic purchases while completing larger tasks. The x402 web payment standard lets software pay for specific data or inventory lookups with exact web requests: a service replies that payment is required, the assistant submits payment, and the seller returns information. That model avoids accounts with many tiny suppliers but creates a risk of repeated microcharges if a request is retried many times.

Standards and implementations include per-payment ceilings, total budgets for a task, limits on repeated attempts and receipt formats that link payments back to the task. Under x402, executed transfers are irreversible unless the seller returns funds; batch settlement and escrow models provide structured refund options.

A completed payment does not prove the purchase met the user’s expectations. Resolving a disputed booking usually requires evidence of the advertised product, cancellation terms and the instruction that authorized the spend. Signed mandates, receipts that identify the merchant and include fees and taxes, and permissions that expire at task completion contribute to that record.

Commercial arrangements can affect outcomes. An assistant’s ranking of options may reflect user preferences, payments from sellers or a mix of factors. Disclosure of whether seller payments influence recommendations and the costs of intermediate services such as availability checks can show how an assistant consumes a user’s budget.

Products now in development include signed mandates to link instructions and purchases, tokenization to protect card data, spending ceilings and task budgets, attempt limits, task-linked receipts and escrow or batch-settlement options for refunds. Final liability and remedies will depend on the payment rails, seller policies and contractual arrangements used in each transaction.