Who's Liable When Your Agent Overpays?

Your AI agent books a flight for $1,200. The same flight was available for $340 on another airline. The agent picked the expensive one because its context window didn't include the cheaper option, or because the merchant's structured data surfaced it first, or because the agent optimized for departure time instead of price.

Who pays the $860 difference?

I think the answer should be the entity that deployed the agent. Not the consumer, the merchant, or the bank. The developer or platform that put the agent into production with insufficient constraints is the party that created the risk, and they should carry the cost when the agent acts within its technical permissions but outside the consumer's reasonable expectations.

That position puts me at odds with how the industry is currently handling it, which is to say, not handling it at all.

The gap in every framework

The chargeback problem is already real. Visa and Mastercard have no reason code for an AI mistake. An agent that buys the wrong hotel or overpays for a flight gets disputed under the same codes designed for damaged packages and undelivered goods. The dispute system assumes a human made the purchase, and the evidence framework assumes a human can describe what went wrong.

The liability question sits one layer above the chargeback question. Chargebacks determine who refunds a specific transaction. Liability determines who absorbs the systemic cost when a class of transactions fails.

The Consumer Bankers Association published a white paper in January 2026 after convening a symposium on agentic payments. The paper, co-authored with Davis Wright Tremaine, warns that consumers may bear costs for errors their agents make. It also tells banks to expect customers to seek help when agent transactions go wrong.

That's a forecast of the mess, not a framework.

Reg E wasn't built for this

Regulation E covers unauthorized electronic fund transfers. If someone steals your debit card and buys a TV, Reg E limits your liability to $50 if you report it within two business days.

But an agent that overpays for your flight isn't an unauthorized transfer. You authorized the agent to buy a flight, and it did, just the wrong one, or the expensive one, or the one that doesn't match what you would have chosen.

The Electronic Fund Transfer Act's protections assume a binary: either you authorized the transaction or you didn't. An agent operating under delegated authority falls into a gap between those two states. You authorized the agent to act, but you didn't authorize this specific outcome.

It's like giving your house keys to a friend and asking them to let in the plumber. The friend lets in the plumber, a locksmith, and a painter. You authorized access, but not that much access. Current law has no clean way to say "the delegation was valid but the scope was exceeded" for an electronic transfer. Reg E has no vocabulary for that distinction.

Credit cards offer a better path, at least mechanically. A cardholder can dispute under "goods not as described" if the agent bought something that doesn't match what was requested. But the evidence burden is awkward. The consumer has to explain what they asked the agent to do, what the agent did instead, and why the difference constitutes a dispute. That's harder to prove than "I didn't receive my package."

Stablecoin and crypto payments have no chargeback rights at all. A stablecoin transaction is final. If an agent overpays on a stablecoin rail, the consumer's recourse is whatever the platform's terms of service provide, which in most cases is nothing.

Where the money actually goes

When a chargeback lands on an agent-initiated purchase today, the merchant absorbs it. Existing card network rules don't distinguish between a human purchase and an agent purchase. The merchant pays the chargeback fee, loses the revenue, and eats the operational cost.

This is backwards. The merchant filled the order. If a customer's agent selected the wrong product, the merchant didn't fail to deliver. They delivered exactly what was ordered.

But the card networks haven't updated their liability rules for agent commerce. Visa's Trusted Agent Protocol and Mastercard's agent payment framework are emerging programs, mentioned in industry coverage but not yet reflected in updated dispute rulebooks. Amex announced a commitment to Agent Purchase Protection in April 2026, designed to cover charges from registered agent errors, though no launch date has been confirmed. Chargebacks911 launched an AI agent purchase indemnification product in October 2026.

These are vendor-specific patches that don't answer the systemic question.

Who should pay

I think liability should follow the failure, not the transaction.

The agent exceeded its delegation. The consumer authorized a flight under $500. The agent booked one for $1,200. The agent platform or developer bears the loss, because the agent violated the constraints the consumer set.

The delegation was poorly defined. The consumer said "book me a flight to Denver" without specifying a budget. The agent picked the most convenient option, which happened to be expensive. The entity that designed the delegation interface, the platform that let a consumer deploy an agent without requiring budget constraints, bears the cost.

The bank controlled the delegation and failed to enforce it. If a bank issues a card linked to an agent with spending limits and the agent exceeds them, the bank's authorization system failed. The bank absorbs the loss.

The consumer explicitly accepted the risk. If the consumer set no constraints and acknowledged that the agent has full purchasing discretion, the consumer owns the outcome. But this should require an explicit opt-in, not a buried terms-of-service clause.

Whoever controlled the constraint that failed is liable for the failure. The consumer is only liable when they explicitly chose to have no constraints.

Why this matters now

Visa and Mastercard are building different authorization models for agent payments. AP2's cryptographic mandates let consumers pre-authorize specific constraints. If the agent exceeds the mandate, the mandate itself becomes evidence in the dispute.

That's the right direction. A signed mandate that says "flights under $500, economy class, departing after 6 PM" creates an auditable record of what the consumer authorized. When the agent books a $1,200 business class flight departing at noon, the mandate proves the agent exceeded its authority. The liability question has a clear answer.

Without mandates or equivalent delegation records, every dispute becomes a credibility contest between the consumer and the agent platform, adjudicated by a card network using reason codes designed for a world where the buyer is human.

The card networks, the CFPB, and the banks are all watching each other, waiting for someone else to define the framework. Meanwhile, merchants are absorbing chargebacks from agent errors, consumers have no clear recourse for overpayments, and the developers deploying these agents have no financial skin in the game.

That last part is what needs to change first. When the entity deploying the agent bears the cost of the agent's mistakes, the incentive to build better constraints follows immediately.