Visa is preparing European payment infrastructure for a new kind of customer: software that can search, decide and purchase on a person’s behalf.
Its Agentic Ready programme, involving banks including Commerzbank and DZ Bank, tests how existing systems handle AI-initiated transactions. The work is not primarily a consumer assistant. It is an effort to define authentication, consent, approval and disputes when the party starting a payment does not hold a physical card.
The change could be comparable to the arrival of online payments. The underlying rails remain, but assumptions about where intent originates must be redesigned.
Identity and intent must travel with the agent
Today’s payment flow begins with a person authorising a purchase. Banks verify identity and assess whether the transaction matches expected behaviour. An agent complicates both questions.
The system needs to prove which user the agent represents, what permission was granted and whether the purchase falls within that permission. A broad instruction to “keep supplies stocked” cannot become unlimited authority to spend.
Visa imagines agents handling repeat or routine purchases under user-defined rules. Software might monitor inventory, compare prices and complete a transaction when quantity, supplier and budget conditions are met. Human input becomes a policy set in advance rather than a confirmation at every checkout.
That makes the audit trail essential. A bank should be able to reconstruct the goal, options considered, rule applied and identity behind the agent. Without that evidence, fraud review and customer disputes become difficult to resolve.
Banks are testing control before convenience
Commerzbank and DZ Bank are examining how agents fit existing fraud, compliance and consent requirements. These controls cannot disappear simply because software creates the instruction.
Agent authentication must distinguish approved software from an attacker pretending to be an agent. Spending limits and merchant restrictions can reduce exposure. Exceptions should return to a person when the request differs from the established pattern or contains insufficient information.
Dispute handling also needs a clear allocation of responsibility. If an agent chooses the wrong product, follows an outdated rule or misunderstands a price condition, the network must determine whether the consumer, agent provider, merchant or bank bears the loss.
The source cites evidence that banks already face costly AI-related incidents. That strengthens the case for testing infrastructure before agentic payments become a mass-market feature.
Enterprise procurement may arrive first
Corporate purchasing contains repeated orders, defined suppliers and formal approval thresholds. Those qualities make it a natural early environment for payment agents.
An agent could assemble a routine order and complete it within a department’s approved limit. More expensive or unusual purchases would still move to a human approver. The result could compress administrative steps without removing the organisation’s procurement policy.
Clear boundaries are the difference between useful delegation and uncontrolled automation. Companies need named owners, logs, monitoring and the ability to revoke an agent immediately. Permissions should be narrow enough that a compromised system cannot move freely across budgets or accounts.
Visa’s programme signals that payment networks are treating agents as a distinct participant rather than disguising them as human cardholders. That recognition allows controls to be designed around machine behaviour.
The everyday act of paying may eventually require less direct input, but responsibility will not disappear. Banks still need to monitor, networks still need to authenticate and consumers still need meaningful control.
Agentic commerce will scale only if convenience is matched by an answer to a simple question: when software presses the invisible purchase button, can every party prove who allowed it, what was allowed and how to recover when the action was wrong?
That proof, rather than autonomy alone, will make a machine-initiated purchase feel like a legitimate payment.
