In brief
- 01Agentic payments are payments an AI agent starts on a customer’s behalf, within limits the customer has set. The hard problems are proving the customer authorised it, limiting what the agent can spend, and settling disputes.
- 02Three mechanisms are emerging: payment provider tokens (such as Stripe’s Shared Payment Token), card network agent credentials (Visa Intelligent Commerce, Mastercard Agent Pay) and signed mandates (Google’s AP2).
- 03This is live in Europe. In 2026 both networks reported live agent-made purchases with European banks, including UK issuers such as Barclays, HSBC UK, Lloyds, NatWest and Revolut on Visa’s programme.
- 04Merchants mostly configure rather than build. The work is with your payment provider, your fraud rules, the evidence you keep and how you handle disputes.
This is a deep dive from my agentic commerce architecture blueprint. For which checkout protocol carries the payment, see ACP vs UCP.
Why agent payments are different
Card payments rest on an assumption: the cardholder is present, or has clearly set up a recurring payment with this merchant. An agent breaks both. It might buy from a merchant the customer has never visited, at a time the customer is asleep, for a total the customer never saw.
That raises four questions that existing payment flows do not answer well:
- 01Authority: did the customer really authorise this agent to make this purchase?
- 02Limits: was the purchase within what they allowed: amount, merchant, category, timing?
- 03Credential safety: how does the agent pay without holding the card number?
- 04Accountability: when something goes wrong, who carries the loss, and what evidence settles it?
UK law adds a fifth. Strong customer authentication (two of: something the customer knows, has or is) applies to most online card payments and assumes the customer is there to authenticate. Every serious agent payment design is, in part, an answer to how SCA is satisfied when the customer is not.
Exhibit 1
Three ways an agent pays, and what each one gives the merchant
1. Payment provider tokens
The simplest model for a merchant is a token issued by the agent’s payment provider. With the Agentic Commerce Protocol, Stripe’s Shared Payment Token lets an agent platform pass the merchant a credential that is limited to that merchant, an amount and a time window. The merchant charges it much like any other payment and never sees the card.
Stripe has since widened the model. At its Sessions conference in April 2026 it announced an agent wallet in Link, where customers grant agents permission to pay with spending approvals; single-use virtual cards that businesses can issue to their own agents; and the Machine Payments Protocol, co-authored with Tempo, for agent micro-transactions and recurring payments, accepted through the same token. Other payment providers are building equivalents, and Visa’s Intelligent Commerce Connect is designed to accept payments started through several of these routes with one integration.
What it means for merchants: the lowest-effort path, provided your payment provider supports it. The trade-off is that the evidence of what the customer authorised sits largely with the platform, not with you.
2. Visa Intelligent Commerce
Visa announced Intelligent Commerce on 30 April 2025 with AI platform partners including Anthropic, Microsoft, OpenAI, Samsung and Stripe. It packages five capabilities as APIs:
- 01Agent-ready credentials: the customer’s Visa card is tokenised and provisioned to a specific agent, so the card number never leaves the network.
- 02Authentication: Visa Payment Passkeys confirm the cardholder and that the agent is acting for them. Visa says the passkeys support issuers’ compliance with European strong customer authentication rules.
- 03Payment instructions: the customer sets limits, merchant categories and when the agent must ask before buying.
- 04Signals: transaction data flows back to Visa in real time for risk control and disputes.
- 05Personalisation, with consent: the agent can use spending patterns to improve recommendations.
Two further pieces matter to merchants. The Trusted Agent Protocol (October 2025), with an agent directory, lets a merchant verify that a request comes from a recognised agent; it is covered in detail in Web Bot Auth and the Trusted Agent Protocol. Intelligent Commerce Connect (8 April 2026) is a single integration through Visa’s acceptance platform that supports the Trusted Agent Protocol, the Machine Payments Protocol, ACP and UCP, and handles non-Visa cards where available.
In July 2026 Visa reported live agent-executed purchases across Europe under its Agentic Ready programme, with more than 30 issuers involved. Those named include Barclays, HSBC UK, Lloyds Banking Group, NatWest and Revolut, alongside merchants including Frasers and lastminute.com.
3. Mastercard Agent Pay
Mastercard announced Agent Pay in April 2025. It rests on the same building blocks: agentic tokens tied to a specific agent, passkey authentication, and consent that the customer gives in advance. Mastercard describes the goal as making sure any action an agent takes “can be traced back to explicit consumer authorisation”, an idea it calls verifiable intent.
In 2026 Mastercard said all its issuers in Europe were enabled for Agent Pay at network level, and that 36 European banks had completed controlled live agentic transactions using passkeys. With Worldline and ING it completed a live transaction in which every part of the flow ran in Europe. Mastercard also worked with Cloudflare on the signed-agent standard behind agent recognition.
What the two networks mean for merchants: agentic card payments arrive as tokenised card transactions, flagged as agent-initiated, with controls applied by the network and issuer. Your payment provider has to support them; your fraud rules and dispute processes have to recognise them.
4. AP2: mandates as evidence
Google announced the Agent Payments Protocol on 16 September 2025 with more than 60 organisations, including Mastercard, American Express, PayPal, Adyen, Worldpay, Coinbase and Revolut. It is payment-method agnostic: it works with cards, real-time bank transfers and, through the x402 extension built with Coinbase and others, stablecoins.
AP2 does not move money. It produces signed, tamper-evident records called mandates, built on verifiable credentials, that prove who authorised what.
Exhibit 2
The AP2 evidence chain and who signs each mandate
What a UK merchant needs to change
Most merchants will not integrate with Visa, Mastercard or AP2 directly. They will receive agent payments through their payment service provider and through whichever checkout protocol the agent uses. The changes are real but mostly operational.
| Area | What to do | Why it matters |
|---|---|---|
| Payment provider | Confirm support for agent tokens (PSP and network), agent flags in authorisation data and AP2 mandate pass-through; get the roadmap in writing | Your provider is where agent payments arrive; gaps here block everything else |
| Firm quote | Make the quote include every mandatory charge and sign it if your protocol supports it | The cart mandate and the DMCC Act drip-pricing rules both depend on a complete total |
| Fraud rules | Score by agent and by mandate, not only by card and device; tune velocity limits for agents | Device fingerprints and behavioural signals built for humans misfire on agents |
| Authentication | Support step-up (a secure link to the customer) when the issuer or your risk engine asks for it | Some purchases will still need the customer to authenticate |
| Evidence | Store quote, mandate references, agent identity and a correlated log with each order | Liability for agent mistakes is not yet settled in UK guidance; evidence is your protection |
| Disputes | Add agent-specific dispute reasons and a playbook for “my agent bought the wrong thing” | These disputes will grow faster than your current categories expect |
| Customer communication | Send order confirmation and cancellation rights to the customer directly, not only to the agent | Consumer Contracts Regulations obligations still apply |
Strong customer authentication without the customer present
UK lawyers point to three workable patterns under current rules: authenticate the customer once when they provision a card to the agent, so the resulting token carries that authentication; rely on exemptions such as trusted beneficiaries, where the customer has whitelisted a merchant with their bank; and step up to the customer for higher-risk purchases. Passkeys make the first and third far less disruptive than one-time passcodes. What is not yet clear is how liability is shared when an agent misreads an instruction, which is why the evidence in the table above is not optional.
Agent payments do not remove the need for trust. They move it from the checkout page into tokens, mandates and logs.
Abubakr Akhtar
Which mechanism should you prioritise?
- 01If you are on a major payment provider: start with what it supports today, usually PSP tokens and network agent tokens. Ask specifically about Visa Intelligent Commerce, Mastercard Agent Pay and AP2.
- 02If you sell high-value or regulated goods: prioritise mandates and evidence. The ability to prove what was authorised is worth more to you than the lowest-friction checkout.
- 03If you sell digital goods or subscriptions: watch the Machine Payments Protocol and x402, which are designed for small, frequent, machine-initiated payments.
Payments are only one layer. Getting recommended comes first; that is covered in generative engine optimisation for ecommerce. The full architecture is in the main agentic commerce guide.
About the author
Abubakr Akhtar
Enterprise business development & partnerships
Abubakr Akhtar is a technology expert who has worked with global companies on martech and solutions.
LinkedInFrequently asked questions
Agentic payments are payments an AI agent initiates on a customer’s behalf, within limits the customer sets. They use delegated credentials such as payment provider tokens or card network agent tokens, sometimes with signed mandates that record what was authorised.
They solve the same problem with similar building blocks: tokens bound to an agent, passkey authentication and customer-set controls. Each works on its own network. For merchants, both arrive as tokenised card payments through the payment provider.
No. AP2 produces signed mandates that prove what a customer and merchant agreed. The payment itself can be made by card, bank transfer or, with the x402 extension, stablecoin.
Yes, unless an exemption applies. Common approaches are authenticating the customer when they link a card to the agent, using exemptions such as trusted beneficiaries, and stepping up to the customer for higher-risk purchases.
Ready to put AI to work in your business? Let's talk.
Book a free 30-minute introduction call with our team. No obligation.
- PYMNTS: Visa launches Intelligent Commerce (April 2025)
- Visa: Intelligent Commerce Connect (8 April 2026)
- The Paypers: Visa expands agentic commerce access for merchants across Europe (July 2026)
- Mastercard: Mastercard unveils Agent Pay (April 2025)
- Mastercard: Europe is building the foundations for trusted agentic commerce (2026)
- Payment Expert: Google’s AP2 to standardise AI payments (September 2025)
- Vellum: Google’s AP2, a new protocol for AI agent payments
- Stripe: Everything we announced at Sessions 2026
- Pinsent Masons: Agentic AI challenges online payments (2026)




