AI agents can now complete real card purchases through Visa and Mastercard programmes. The harder question is who pays when an authorised agent buys the wrong thing.

For decades, card rules assumed a person pressed the pay button. That assumption is starting to break. AI agents can now search, choose and pay on a customer’s behalf, and every major card network has a programme for it. Agentic payments have moved from demos to live transactions in 2026, but the rules around them have not caught up.

Where agentic payments stand today

The card networks moved first on agentic payments. Visa launched Intelligent Commerce in April 2025, then added its Trusted Agent Protocol with Cloudflare in October that year. Mastercard launched Agent Pay in April 2025. American Express followed with a developer kit in April 2026.

Live use is real but small. In March 2026, Santander and Mastercard completed what they called Europe’s first live end-to-end payment made by an AI agent. Visa says live agent purchases have since run with more than 30 European banks, including Barclays, HSBC UK, Lloyds and NatWest. However, no network has published volumes, approval rates or fraud rates for agentic payments.

The platforms are still finding their model. OpenAI launched Instant Checkout inside ChatGPT with Stripe in September 2025, starting with Etsy. It then retired the in-chat checkout in March 2026, so purchases now finish on the merchant’s own site. Google, Microsoft and Shopify are building their own checkout routes too.

How networks verify an agent

The core design is simple. The agent never sees the card number. Instead, it holds a network token limited to that agent, merchant and amount. The customer approves the token with a passkey or biometric check. Meanwhile, the agent itself is registered with the network, so a merchant can confirm it is genuine.

Identity firms are adding a second layer. Socure launched Agent Intelligence on 27 October 2026, with tools to detect agents, score them and later issue them credentials. It says daily requests from AI agents rose more than 1,700% in a year, Help Net Security reported. Note that only the detection tool is generally available so far, and its accuracy figure comes from lab tests. For more, see our report on Socure’s agent tools.

The liability gap

Fraud is the easier part of agentic payments. When a token is misused, existing card rules largely apply, and Visa says its zero-liability cover extends to agent-initiated purchases. The hard case is agent error. If an authorised agent books the wrong dates or overspends, no network has a dispute rule written for it.

In practice, the merchant usually absorbs the loss through an ordinary chargeback. That matters, because Visa tightened its dispute-monitoring threshold for merchants from 2.2% to 1.5% in April 2026. Agent mistakes count towards it. Unsurprisingly, 93% of merchants told PYMNTS Intelligence that agent providers should bear losses from wrong purchases. So far, American Express is the only network to offer a protection against agent error, and it comes with conditions.

UK and EU rules are still open

Regulators have started asking the right questions. HM Treasury’s payments consultation, which closed on 6 October 2026, asked directly how consent, authentication and liability should work for agentic payments. It also proposed handing strong customer authentication rules to the FCA, so they can be rewritten around outcomes.

Until then, existing rules still apply to agentic payments. In the UK and EU, strong customer authentication covers agent purchases, and how delegated approval fits those rules remains unsettled. In the US, Regulation E and Regulation Z pre-date agents entirely. Therefore, firms running agentic payments today are relying on contracts and network programmes rather than law.

The fraud risks to plan for

New channels attract new fraud. Fraudsters can pose as trusted agents, which is why Visa and Cloudflare now sign agent traffic. Prompt injection can steer an agent towards a fake merchant or a higher price. In addition, if a criminal takes over the account that controls an agent, they inherit its tokens and limits.

Merchants face a balancing act too. Blocking all bot traffic will also block legitimate agents and the customers behind them. As our roundup on AI fraud models found, models catch known patterns well but struggle with new ones. Agentic payments will test that weakness.

What firms should be doing now

For merchants, the first step is to recognise agent traffic rather than block it outright. Next, review terms so they say how agent purchases are treated. Target, for example, updated its terms to treat purchases by authorised agents as made by the customer. Then track agent-initiated disputes separately, so they do not quietly push the business towards network thresholds.

For issuers and payment firms, the priority is clear consent records. Mastercard’s Verifiable Intent links the customer’s identity, instruction and the final purchase in one record. That kind of evidence will decide disputes long before regulators finish their work.

What to watch next

Several signals will shape agentic payments in 2027. HM Treasury will publish its response on consent and liability, and the FCA may then consult on new authentication rules. Card networks may also write the first dispute codes for agent error. Peak season will add pressure, as agent purchases over Black Friday are settled under rules written for humans.

Above all, the technology is ahead of the rulebook. Agentic payments can already verify who sent an agent. The open question is who pays when that agent gets it wrong.

FintechBits covers payments, banking and fintech developments for readers in the US and UK. This article is for information only and is not financial or legal advice. Views expressed are those of the FintechBits editorial team.