Insights

Agentic commerce and banks: what it is, the issuer and acquirer roles, and the questions still open

Techtone Group · Oct 19, 2026 · Last reviewed Oct 20, 2026

Agentic commerce means AI agents that search, decide and pay on behalf of a person or a company. Live agent payments have been completed in Europe since March 2026, but the rules that decide who carries the risk were written for a human at the checkout. This article sets out the basics, the two bank roles and the questions issuers and acquirers will have to settle, as an opening for discussion rather than a set of answers.

What agentic commerce is

An AI agent is software that acts on an instruction such as "reorder printer toner when stock runs low". In agentic commerce the agent searches, compares, decides and pays within limits the user has set. The user may approve each purchase or delegate a standing mandate. For companies it starts to look like procurement run by software.

The industry moved from announcements to live payments in under 18 months:

• Mastercard Agent Pay, announced on Apr 29, 2025, uses Agentic Tokens and registered, verified agents. Santander and Mastercard reported Europe's first live end-to-end agent payment, a controlled pilot, on Mar 2, 2026.
• Visa Intelligent Commerce, announced on Apr 30, 2025, added the Trusted Agent Protocol on Oct 14, 2025, so merchants can recognize approved agents by cryptographic signatures. After an issuer readiness programme launched in Europe on Mar 17, 2026, Visa reported live agent transactions with 31 European issuers on Jul 2, 2026.
• Agent Payments Protocol (AP2), launched by Google in September 2025, records what the user asked for as signed mandates, and was contributed to the FIDO Alliance on Apr 28, 2026.
• Agentic Commerce Protocol, released by OpenAI and Stripe on Sep 29, 2025, passes the agent a payment token limited to one merchant and one basket total.

How an agent payment differs from card-not-present

Today's online card payment is card-not-present, but the person is present: they choose the goods, enter or confirm the card and pass strong customer authentication when asked. Fraud models, 3-D Secure and dispute rules assume a human made the final choice at the moment of payment.

With an agent, three things change. The decision is separated from the payment: the user authorizes in advance and the agent executes later. The party at the checkout is software, which merchant systems block as a bot. The evidence of what the user wanted sits in the agent platform, not in the payment message. Agent tokens, agent identification and signed mandates each cover part of the problem, not all of it.

The issuing side: the bank whose customer delegates

The issuer holds the customer relationship and, under PSD2, the refund obligation when a payment is unauthorized. Its questions fall in five areas:

• Consent and mandates. How the customer grants, limits and withdraws an agent's authority, and where that record is kept.
• Spending limits. Per payment, merchant and category, and who enforces them: the bank, the token or the agent.
• Authentication. How strong customer authentication works when the customer authenticates once, for a mandate, and the agent pays many times.
• Tokenized credentials for agents. A credential per agent, with its own controls and life cycle.
• Liability, disputes and fraud. Telling "my agent did this and I did not want it" apart from fraud, and both apart from a change of mind.

The acquiring side: the merchant's bank

The acquirer carries the merchant's risk and sits between merchant and issuer in a chargeback. Its questions differ:

• Recognizing agent traffic. Merchants built defenses against automated traffic and now need to let good agents through.
• Telling good agents from bots. Agent identification helps only if merchants, gateways and acquirers can read and trust the signal.
• Chargebacks and liability shift. Whether an authenticated agent payment shifts fraud liability to the issuer in the way an authenticated payment by a person does today.
• Merchant onboarding and risk. Which merchants accept agent payments, under what monitoring, and how agent platforms are assessed.

The open questions

None of these questions has a settled answer. PSD2 has no separate regime for agent payments, and its idea of consent was written for a human payer. PSD3 and the Payment Services Regulation, politically agreed on Nov 27, 2025, are expected in the Official Journal around the turn of the year, but were drafted before agent payments went live.

Two questions run across both sides. First, who is the customer of the agent platform, and does the platform itself provide a regulated payment service? Second, refunds: PSD2 lets a payer reclaim an authorized payment initiated through a payee when no exact amount was set and the amount exceeds what the payer could reasonably expect. Whether that logic covers an agent that spends within its mandate but against the user's wishes is open.

An invitation to discuss

Agent payments are live in pilots, and European banks are testing them now. The answers on consent, liability and economics will likely come from the schemes, regulators and the banks that move first. We would welcome a conversation with issuers, acquirers and payment teams working through these questions, to compare notes.

This article is general information, not legal advice. Programmes and protocols are described from their public announcements, and their rules and specifications are changing quickly.

Related service: Strategy and digitalization