Key takeaway: When an AI agent completes a purchase, merchants and card issuers now have a new signal to weigh: a probability score, not a verified identity. Fraud and risk teams should treat the score as one input among several. It is not a definitive answer to whether an agent initiated a given transaction.
Mastercard has started testing an Agent Detection Score in the United States. The tool estimates the probability that an AI shopping agent, rather than a human cardholder, initiated a given transaction.
The score sits inside Mastercard’s Agent Pay Trust Framework. That framework has five parts: identity, intent, controls, execution and intelligence. The score draws on three main inputs: Skyfire’s Know Your Agent (KYA) technology for agent identity, web-traffic visibility that Cloudflare provides into AI-agent activity, and Mastercard’s own fraud and behavioral signals. Those signals come from Mastercard’s Decision Intelligence platform, which the company says processes roughly 143 billion transactions a year.
Most current agentic-commerce infrastructure focuses on verifying who an agent is, including Visa TAP, Stripe’s Agentic Commerce Protocol and Google’s AP2. Mastercard’s score instead asks a narrower question first: did an agent initiate the transaction at all?
Cloudflare Chief Strategy Officer Stephanie Cohen said trust will determine how far agentic commerce goes. She added that combining Cloudflare’s visibility into AI-agent web activity with Mastercard’s payment intelligence gives financial institutions and merchants the context to approve legitimate AI-driven transactions with confidence. Skyfire co-founder and CEO Amir Sarhangi said, “Every AI agent that transacts on someone’s behalf should be identifiable, accountable and auditable.”
Mastercard plans to enrich the score over time with additional data, including behavior, merchant risk, transaction patterns, credential risk and consumer propensity. The underlying question of who bears the loss when an agent makes an erroneous purchase remains unresolved.
Why it matters: Merchant hesitation around AI shopping agents has centered on liability. A PYMNTS Intelligence survey commissioned by Visa Acceptance Solutions found that 93% of merchants believe an AI provider should absorb losses from failed agent transactions. The same survey found that only 28% of merchants are currently willing to expose their full inventory to agents. A detection score does not settle who pays when something goes wrong. It does give issuers and merchants a data point for flagging and reviewing agent-initiated transactions. That data point could shape how fraud and dispute teams triage AI-driven purchases as agentic commerce scales.
Sources: Yahoo Finance, Mastercard