TLDR;
Mastercard has started testing a score that estimates whether a card transaction was initiated by an AI agent. The US test is meant to give banks and merchants more context before they approve or challenge an automated purchase. It is useful plumbing for agentic commerce, but it is still a probability signal, not proof that the customer authorised the agent or that the transaction is safe.
What happened
Mastercard announced the new Agent Pay trust and intelligence services on 30 September. The first service is a probability score that indicates how likely a transaction was initiated by an AI agent. It is rolling out for testing in the US, so this is not a global production standard and Mastercard has not published a performance benchmark for the score.
The company says later versions will add signals covering behaviour, merchant risk, transaction patterns, credential risk and consumer propensity. Those inputs are intended to help an issuer tell the difference between a legitimate delegated purchase and activity that only looks unusual because an agent is moving faster or buying across several merchants.
Mastercard places the score inside an Agent Pay Trust Framework built around identity, intent, controls, trusted execution and intelligence. That distinction matters. Identifying that software probably initiated a payment answers who or what acted. It does not, by itself, show what the customer approved, whether the agent stayed within a spending limit or whether the product and delivery terms remained unchanged.
Independent coverage from PYMNTS confirms the score is the first of the new services and is being tested rather than generally released. Mastercard is also working on shared signals with partners and on ways for merchants and financial institutions to recognise trusted agents. The direction is clear, but the operating model is still being assembled.
Why it matters
Today, an automated purchase can resemble account takeover, card testing or an impossible sequence of transactions. A cautious issuer may decline it, while a permissive system may approve an agent that has exceeded its mandate. Agentic commerce needs a middle layer that carries enough context to make a better decision without exposing raw customer data to every participant.
That creates a commercial dependency outside the AI assistant itself. A shopping agent can find the right product and construct a valid cart, yet conversion still fails if the payment chain cannot recognise delegated intent. Mastercard’s earlier Agent Connect announcement focused on connecting merchants and agents. The new score addresses what happens when that activity reaches risk and authorisation systems.
Enterprise brands should also resist treating a network score as a complete control. A high probability of agent initiation says nothing about whether the product description was accurate, the shopper accepted a substitution or the agent used a valid delivery address. Payment trust, customer consent and order integrity remain separate checks.
How your brand can benefit / be affected
Map the agentic purchase journey from discovery to refund. Record the agent or platform, the user approval event, cart version, price, shipping promise, payment credential, final order and any later change. Keep stable identifiers across those steps so customer service and fraud teams can reconstruct what happened without relying on a conversational transcript alone.
Ask payment providers and acquirers which agent-origin signals they receive, how they expose them in authorisation and reporting, and what happens when the signal is missing or uncertain. Test step-up checks for high-value, regulated or unusual purchases. A blanket challenge on every agent order will destroy conversion, while blanket approval turns delegation into a new fraud path.
Finally, separate customer permission from transaction probability in your controls and dashboards. Use the network signal as one input alongside mandate evidence, product rules, velocity and existing fraud models. Brands that can explain why an automated order was accepted, challenged or rejected will be better placed to scale agentic checkout without creating avoidable disputes.
News date: 30 September 2026. Editorial review: 2 October 2026.