The Den Reviewthe briefing for executives who run AI
Thursday, November 12, 2026

The Denominator

Subscribe
The AI Stack: Cost, Control × RiskAI BudgetsAgentic CommerceCRMOpen-Source ModelsCloud SecurityVertical AI

Agentic Commerce · Agent Liability

Citi Reportedly Wants AI Companies, Not Hotels or Banks, to Pay When Agents Book the Wrong Trip

Visa's study of 22 AI models found 43% of runs failed. About half of those still bought anyway.

· 2 min read

A woman executive in a dark suit stands at the window of a top-floor office at dusk, looking at a tablet beside a walnut desk with a brass lamp. Faint line icons of an airplane, suitcase, bed, payment card, map pin and shield float beside her, joined by thin dotted lines.
Photograph: AI-generated

Abhinav Anand, Citi’s head of value cards, lending and commerce for U.S. consumer cards, said agents might have to accept some liability when they register with Visa or Mastercard, according to Skift.

No rule yet says who eats the loss on a purchase an agent got wrong, and the answer decides where chargeback exposure lands. Citi’s position would move it off merchants and issuers and onto the agent’s maker, which would change what an AI company must accept before its agent can pay.

What does Citi propose?

Anand made the case at a commerce summit in New York, where he said liability should fall “to the agent themselves, the company that is actually coming with the agent.” Hotels, airlines and card issuers would be spared the cost of an agent’s wrong booking.

Disputes are the other half. The industry is working to carry the traveler’s purchase instructions inside the tokenized transaction, Anand said, so issuers and merchants can see the original intent when a charge is challenged. He expects chargeback rates on agent transactions to run higher than average in the short term.

How often do agents get it wrong?

Anand cited a Visa study that he put at more than 75 agents. Visa’s published benchmark put 22 AI models from 12 labs through 15 shopping tasks, such as buying a TV or a laptop. Its top-scoring models passed 82.2% of attempts, and models that passed a scenario often failed when made to repeat it.

Visa’s results put the failure rate at 42.8% of runs, and 54% of failed runs still ended in a purchase, either of a product that missed the customer’s needs or before every requirement was found. By The Denominator’s arithmetic, that is about 23 of every 100 runs. The tests were shopping conversations, and Visa plans to extend them to checkout, payment and order placement.

What is Citi’s own stake?

Anand cited early industry data showing that travelers who arrive through an agent’s search are 40% to 50% more likely to convert, and that they spend more per transaction.

Citi has its own traffic to gain. It was among the first issuers enabled for agentic payments with Mastercard, followed by Visa, and it is building an AI travel concierge inside its own app.

Which agent-made bookings would a travel merchant or card issuer have to refund under current dispute rules?

NVIDIAMicrosoftOpenAIAnthropicGoogleAWSJPMorganCapital OneAT&TCharterComcastT‑MobileUnitedHealthCVS HealthTravelersWalmartSalesforceDatabricksSnowflakeCoreWeaveStripeAccenture