.png)
The Network Will Now Score the Agent: Why Mastercard's Probability Model Changes Authorisation Economics
Meta title: Mastercard’s Agent Score: The New Economics of Authorisation Meta description: Mastercard is testing a score to identify AI initiated payments. Discover what it changes for authorisation, fraud, payment success rates and merchants. Slug: mastercard-agent-score-authorisation-economics
For decades, card authorisation has largely been built around a human shaped assumption: the person making a purchase behaves in ways that fraud models can learn. An AI agent challenges that pattern. It can book a flight, hotel and taxi across three merchants in seconds, following a single consumer instruction. To a model trained on human behaviour, that burst of activity can look like account takeover.
Mastercard’s new probability score makes a machine’s role an explicit input to the authorisation conversation. The shift is not just that a network can identify agentic payments: the network is positioning itself to interpret that signal where it already has influence, at the authorisation decision.
For merchants, that creates opportunity and exposure. A better signal could help legitimate agent purchases through. But it does not prove a customer consented, replace the issuer’s decision or decide who is responsible when an agent buys the wrong thing. Merchants still own the operational outcomes: approval rates, mandates, evidence and disputes.
What the score actually is and is not
Announced on 30 September 2026, the first intelligence service in Mastercard’s Agent Pay Trust Framework is a probability score indicating how likely it is that an AI agent initiated a transaction. Testing is US only. Mastercard has not announced a timeline for other markets.
The score is one part of a framework covering identity, intent, controls, execution and intelligence. Planned signals include agent behaviour, merchant risk, transaction patterns, credential risk and consumer propensity. Work with Cloudflare aims to combine web traffic and payment network signals in a privacy preserving environment. Skyfire contributes its know your agent, or KYA, approach to agent identity.
The critical distinction is that agent origin is not the same as authority. A score that suggests a machine initiated a payment does not certify the user’s consent or establish that the agent stayed within its mandate. Nor does it approve or decline the payment on an issuer’s behalf. As ID Tech Wire explains, transaction identification and permission to spend are separate parts of the trust framework.
There are still unknowns. Mastercard has not published score thresholds, accuracy metrics, latency or pricing, and there is no public evidence yet of performance at production volume. The score is a new signal to evaluate, not proof that agentic payments are now safe.
The two failure modes it is designed to break
The first failure mode is declining legitimate activity. Fraud systems trained on human patterns may treat rapid, cross merchant purchases as suspicious. An agent’s itinerary can look nothing like a consumer’s ordinary shopping session. Mastercard’s added context could help issuers distinguish an authorised agent journey from an account takeover attempt.
The second is loosening controls too far. If legitimate agent transactions keep getting declined, teams may relax rules, which can leave genuine fraud hiding behind a bot label.
This matters to payment success rate as well as fraud prevention. A PYMNTS study reported that 56 per cent of consumers had experienced a false decline in the previous 90 days. The global e commerce false decline average is around 1.51 per cent of sales, and false declines may cost ecommerce merchants more than US$231 billion in 2026, according to Datos Insights data summarised by Signifyd. Agent led journeys add a measurement problem: agents can abandon a decline or silently switch merchants rather than retry, so standard reporting never shows the sale that disappeared.
Demand remains early. PYMNTS found 56 per cent of consumers would let agents compare products, but only 35 per cent would let them use saved payment methods. Mastercard projects that one in ten consumers will routinely use agents by 2030, a forecast rather than observed behaviour.
Why the scoring layer is the strategic prize
Mastercard launched Agent Pay in April 2025 with agentic tokens, tokenised card credentials tied to a particular agent so an assistant can check out without handling the card number. Later work, including Verifiable Intent with Google, addressed what credential is used and whether the user instructed the agent to act.
The probability score asks a different question: was a machine on the other end of this transaction, and should that change the risk assessment? That places Mastercard at the point where its network is strongest, without requiring it to own the agent’s identity system. As Mastercard president and CTO Ed McLaughlin told Fortune, the ecosystem needs to know which agent is buying, who permitted it and what it may buy. He also stressed the importance of keeping people able to intervene: “Part of it is bringing humans back into the loop. So even if the agent goes rogue and tries to do something, you can say no.”
The wider stack is contested. Skyfire and Baselayer are working on KYA, while Visa’s Intelligent Commerce and Stripe’s agentic tooling develop their own approaches to trusted execution. Visa and Mastercard have also joined Stripe’s Tempo settlement network as launch partners. If network signals are incompatible, merchants may inherit the integration work across providers, checkout paths and payment methods. Forkast describes consumer protection as the least mature part of that stack.
For merchants, a network score is a signal they may not control. The commercial outcome remains theirs to manage.
What merchants and issuers should do now
Separate agent detection from permission checks. Record what the customer authorised, including spend limits, merchant categories, time windows and cancellation rights.
Map where agent transactions enter your payment flows. Include direct checkout, saved credentials, pay by link, subscriptions and merchant specific agents.
Ask providers what the score means operationally. Clarify whether it is an input, recommendation or decision, and how it combines with existing risk rules.
Request performance evidence before changing policy. Seek accuracy, false positive and false negative measures, latency and availability when they become available.
Measure agent related approvals separately. Track approval rate, fraud, disputes, conversion and abandonment by signal, issuer, market and channel.
Preserve evidence of intent and execution. Keep the instruction, the agent’s actions, any changes to the basket and the final payment record.
Set escalation paths for high impact purchases. Decide when an agent should pause for customer confirmation, particularly for unusual, irreversible or high value orders.
Review dispute and liability processes. Identify who investigates when a purchase followed a valid instruction but produced an unwanted outcome.
Plan for multiple rails and standards. In Australia, consider how agent origin and behavioural signals could work across cards and account to account payments, including PayTo and the NPP.
The Australian questions extend beyond cards. The RBA’s payments system review drew 75 submissions, with stakeholders raising uncertainty about chargeback liability when agents act beyond their authority. PayTo and the NPP have no card chargeback equivalent, so machine readable origin and behavioural signals could matter there too while redress stays unsettled. The RBA’s review summary and the BIS analysis of payments in the age of AI agents underline that authority, fraud and liability are connected problems.
The bottom line
Mastercard’s score moves agentic commerce from an identity question towards transaction level judgement. That could help issuers distinguish unusual but legitimate machine behaviour from fraud. It does not certify consent or resolve liability, and its real value will depend on evidence of accuracy, reliability and interoperability.
For merchants, the priority is to connect agent identity and intent with authorisation outcomes, operational evidence and dispute handling. Quantum Payments’ unified commerce and orchestration capabilities are designed to help businesses manage payment flows across channels, while agentic payment capabilities support more automated checkout and payment workflows. The strategic requirement is not to trust every agent. It is to make every decision traceable, measurable and manageable across the payment journey.
For more on the operational changes ahead, read our analysis of why fraud models built for humans struggle with AI agents, the agent liability gap and why payment stacks need to prepare for agents. You can also explore the emerging agent exception queue, the implications for payment routing and what early x402 activity means for merchants.
Frequently asked questions
What is Mastercard’s Agent Pay probability score?
It is a score indicating how likely it is that an AI agent initiated a transaction. Mastercard is testing it in the United States as part of the intelligence layer of its Agent Pay Trust Framework.
Does the score confirm that a customer authorised the purchase?
No. It indicates the likely origin of a transaction, not whether the customer consented or the agent stayed within its authority.
Will the score approve or decline a payment?
The score does not replace the issuer’s authorisation decision. It is an extra signal to help payment participants assess agent activity and decide whether to approve, decline or step up checks.
Why could agent detection improve payment success rates?
Agent behaviour may look unusual to fraud systems trained on human shopping patterns. Better context could help issuers avoid declining legitimate purchases while retaining controls against suspicious activity. Its real world effect on approval rates has not yet been published.