Visa, Mastercard and Ant International want one way to check whose AI agent is paying, and India is building a register of them

Visa, Mastercard and Ant International said on 9 September 2026 that they will work towards common principles for Know-Your-Agent, the checks a payment network runs to find out which business stands behind an AI agent before it lets the agent pay [1][2]. Each already runs a protocol of its own: Visa’s Trusted Agent Protocol, Mastercard Verifiable Intent and Ant International’s Agentic Mobile Protocol [1][2]. The work will go through BuildFin.ai, an industry platform convened by the Monetary Authority of Singapore, and builds on a framework called Safeguards for Agentic Finance at Runtime [1].

The release sets out three pillars: tracing every agent to a validated operator, certifying agents against shared security and behaviour requirements, and monitoring them continuously on identity and transaction signals [1]. Its verb is a careful one. The three will “explore opportunities to work towards common principles” [1], and TNGlobal notes the participants gave no implementation timetable, pilot volumes or list of the merchants and banks that will test it [3]. The release cites McKinsey research projecting that AI agents will orchestrate US$3 trillion to 5 trillion of global consumer commerce by 2030 [1][2].

Of those pillars, monitoring is the one somebody has to staff. Tracing an agent to a company happens once, at onboarding, and certification is a test the agent passes on a particular day. Watching it afterwards is daily work, and it is the only pillar that catches an agent which was fine when it was certified and starts doing something odd four months later. The release does not say who pays for that.

India wants a register run by the rail

Reuters reported on 10 September, citing three sources involved in the discussions, that the National Payments Corporation of India plans a registry to verify and monitor AI agents that make payments on the Unified Payments Interface [4]. It sits inside NPCI’s planned Unified Agentic Protocol, would vet UPI agents first, and could later cover cards, bill payments and other payment methods [4]. One source said liability for wrong or unauthorised payments will need to be addressed through regulation [4]. ET Small Business, carried by RetailIntel, reports that the first transactions may be lower-value grocery transactions [5]. NPCI did not immediately respond to Reuters [4]. Yesterday’s brief carried the protocol as the Unified Agent Protocol; Reuters calls it the Unified Agentic Protocol.

That makes two models for the same check within a day. The card networks and Ant International are writing shared rules for the agents that touch their own networks, and India’s payments operator is building one register for everyone on its rail. Either way, an agent with no verified operator behind it does not get to pay.

Meta’s agent pays with a card built for one purchase

Meta launched Muse on 8 September, a personal agent that can open a browser, fill out forms and negotiate on the user’s behalf, and which asks before it sends an email or makes a purchase [6]. Checkout runs on Link, Stripe’s consumer network, which Stripe puts at more than 300 million users and more than 1 million businesses [7]. For every purchase the user approves the total in the chat, and Muse never sees the underlying payment details; at a retailer outside Link, Stripe issues Muse a single-use virtual card scoped to the approved purchase [7]. Meta says Muse is the first AI agent covered by Link’s purchase protections [6]. It is rolling out in the United States on iOS, Android and muse.ai [6], and TechCrunch lists a free tier alongside a Power plan at $20 a month and a Maximum plan at $100 [8].

A card that works only for the purchase the person approved is the cleanest answer to the mandate problem I have seen in a consumer product. It turns the dispute from “did I authorise this?” into “is this what I approved?”, and a support desk can settle the second question in a minute, provided somebody kept the approval screen.

Visa’s survey puts trust in AI payments at 23 per cent

Visa’s own research, published on 9 September, found that 72 per cent of US consumers have used an AI assistant [9], and only 23 per cent trust generative AI to handle payment transactions on their behalf [9][10]. Asked about Visa by name, 61 per cent said they would trust it to handle agentic transactions [9]. The Harris Poll ran the survey for Visa from 26 to 28 May 2026 across 2,065 US consumers, and between 1,028 and 1,034 of them answered the payment question for each brand tested [9][10]. It is Visa’s poll, and Visa is one of the brands it asked about.

The fix Visa put its name to that same day, with Mastercard and Ant International, starts a step before trust. Before an agent pays, the network finds out whose it is.