Mastercard Now Scores Whether an AI Agent Paid: What the Agent Trust Layer Means for Providers
TL;DR: Mastercard’s machine-probability score estimates whether an AI agent, rather than a person, initiated a payment. It is not a new checkout product: it is an input to the authorisation and underwriting decision. As agent payment rails divide between open, high-velocity x402 payments and enterprise session-based flows on the Machine Payments Protocol (MPP), payment providers should prepare to identify agent traffic, measure its distinct risks and define recourse before volume scales.
What did Mastercard actually build?
Mastercard announced its new machine-probability score on 30 September 2026, with the news widely syndicated on 9 October. The score tells an issuer or merchant how likely it is that an AI agent, rather than a person, initiated a transaction. Testing is US-only, with no timeline announced for other markets.
The score is the first piece of the intelligence layer in Mastercard’s Agent Pay Trust Framework. The framework has five parts: identity, intent, controls, execution and intelligence. Mastercard is working with Cloudflare to combine web-traffic and payment-network signals in a privacy-preserving environment, and with Skyfire on know-your-agent identity and verification.
Planned extensions cover agent behaviour, merchant risk, transaction patterns, credential risk and consumer propensity. None is live, and Mastercard has published no thresholds, accuracy, latency, API schemas or pricing. That makes the score a risk input, not proof that a transaction is legitimate.
Agent Pay has developed in stages since April 2025, around agentic tokens, then merchant-side acceptance tooling and, with Google, Verifiable Intent: an open-source way to link a user’s instruction to what an agent actually did.
The distinctions matter. Tokens answer “what credential is being used”. Verifiable Intent answers “did the user actually ask for this”. The new score asks “was this a machine, and should that change the risk decision”.
Why does a machine probability score matter more than a checkout integration?
Because the payment rail can carry the transaction, but the authorisation decision determines whether it proceeds, and on what risk terms.
Card-fraud models learn human patterns: typing speed, location, time of day, basket mix. An agent that books a flight, a hotel and a cab in seconds off a single consumer approval looks, to those models, like account takeover.
Two failure modes follow. Issuers decline legitimate agent purchases as fraud, or they loosen controls and let genuine fraud hide behind a bot label. A machine-origin signal helps only when it is combined with context: the agent’s identity, the user’s instruction, spending permissions and the transaction pattern. On its own, it cannot establish that the agent had authority, stayed within budget or bought the right thing.
The broader shift is from winning the agent checkout to scoring it. The checkout is only one interface; the scarce capability is a trusted decision inside authorisation. As explored in Know Your Agent Is the New Know Your Customer, agent identity and permission are becoming infrastructure that networks can buy or partner for, rather than necessarily own outright.
What do the numbers say about the two agent payment rails?
The rails are developing quickly, but early transaction counts do not yet prove meaningful commerce.
The landscape is splitting into two patterns: open, high-velocity micropayments on x402, and more controlled, session-based enterprise payments on the Machine Payments Protocol.
As at 11 October 2026, agenteconomy.to reported 191,751,598 cumulative x402 transactions across 12 chains and US$41,917,495 settled, with Base the largest chain at 85,055,098 transactions. Analysts estimate 25–30% of that activity may be test or leaderboard noise. x402 V2 added wallet-based identity, service discovery and reusable sessions, and the x402 Foundation is now hosted by the Linux Foundation with more than 50 members. AWS Bedrock AgentCore Payments supports x402, and Bedrock Bazaar lists more than 10,000 endpoints.
MPP, launched by Stripe and Tempo in March 2026, reports 46,198 cumulative events and counts Anthropic, DoorDash, Mastercard, OpenAI, Ramp, Shopify, Standard Chartered and Visa among its launch partners. Visa’s early data, reported via Artemis, showed roughly 115,000 transactions and about US$25,000 settled.
Our calculation: x402’s reported settlement works out at roughly 22 US cents per transaction, and Visa’s early MPP figures land in the same place. The datasets are not directly comparable, but the direction is the same.
The strategic point is that transaction count can be noise. The authorisation decision is the business. A rail becomes cheap and widely available; a reliable way to assess machine identity, permission and risk does not. This is the argument behind Visa Just Put Payments Inside OpenAI: infrastructure matters, but trust and control shape who can use it.
Who pays when the agent gets it wrong?
A probability score may inform approval; it does not resolve liability for an agent’s mistake.
A September 2026 PYMNTS study found 93% of surveyed merchants believe the AI or agent provider should bear losses when an agent makes an incorrect purchase. An April 2026 Product.ai report found 42% of consumers would not trust AI to make purchases over $25.
That exposes the gap between a valid payment and a valid outcome. An agent may have been authorised to act and still misread the request. A risk score can help an issuer approve. It cannot determine who pays when the instruction was valid but the agent got it wrong.
Consumers delegate discovery more readily than payment authority: PYMNTS found 56% would let agents compare products, but only 35% would let them use saved payment methods, and Mastercard projects one in ten consumers will routinely buy through agents by 2030. That is why agentic commerce is constrained by liability and recourse more than by payment technology, as discussed in Agentic Commerce Is Stuck Below 1% of US E-Commerce.
What does this mean for Australia?
Australia’s key question is whether an agent preserves merchant choice over the payment route.
In its June 2026 issues paper on payments system regulation, the RBA asked whether AI agents that default to a particular network at checkout could limit merchant choice and quietly undermine least-cost routing. Submissions closed 7 August 2026, and the RBA has said it will publish its regulatory priorities by the end of 2026.
The issue is practical: who chooses the rail, the merchant, the consumer, the agent platform or the payment provider? Visa’s Agentic Ready programme includes Australian financial institutions. With cost of acceptance now visible inside the merchant’s price, the routing question in an agent-led purchase is a margin question, not a technology one.
What should payment providers do now?
Build operational readiness around agent traffic before assuming that a network score will solve the problem.
Treat agent traffic as its own segment. Give machine-origin transactions their own risk thresholds instead of reusing human-fraud rules.
Instrument the declines. Record machine-origin volume, decline reasons and whether a retry or tighter control would have saved a legitimate sale.
Split the two failure modes. Track false declines of good agent activity and fraud hiding behind a bot label separately.
Take a position on delegated authority. Decide whether you support budget grants and revocable permission tokens, and what evidence you require.
Build the recourse path now. Agree how a customer disputes an agent error, who reverses the payment and who holds the loss.
Write down your routing position. If an agent or platform picks the network, say so, and say what the merchant still controls.
For operators assessing how trust, payment rails and merchant control fit together, The First Integrated Agent-Commerce Stack Is Live and Ant International Just Priced Agent Error offer related perspectives.
Frequently asked questions
What does Mastercard’s machine-probability score do?
It estimates how likely a transaction is to have been initiated by an AI agent rather than a person. It is intended to inform issuer and merchant risk decisions, not independently verify permission or guarantee a safe transaction.
Is the score available globally?
No. Testing is US-only, and Mastercard has not announced a timeline for other markets.
What is the difference between x402 and MPP?
x402 is an open rail for high-velocity machine payments, often using stablecoins. MPP is built for session-based enterprise flows. Neither, by itself, resolves who has authority to transact.
Does an agent score settle who pays for an incorrect purchase?
No. A score can support an authorisation decision. Liability and recourse depend on who authorised the agent, what it was permitted to do and the rules governing the transaction.
Why does least-cost routing matter to Australian AI agent payments?
If an agent or platform selects the network by default, a merchant may lose a lower-cost eligible route. The RBA has raised that question in its payments-system regulation review.
The strategic takeaway
The contest is shifting from checkout integration to risk decisioning. The rails are multiplying and their unit economics already point to low-value automated transactions. What stays scarce is the ability to tell legitimate machine activity from suspicious activity, and to connect that judgement to user intent, permissions and liability.
The providers best positioned for agentic commerce will not process the most transactions. They will make machine activity legible to authorisation systems, preserve merchant choice and give every participant a clear route when an agent gets it wrong.
Sources