Agentic Commerce Trust Gap Pushes Payment Networks to the Center

What happened

AI shopping agents are getting better at finding products, and they are starting to complete purchases. The open problem is whether anyone can trust them. In early August, a Sourcing Journal "Tech Tuesday" column covered two companies approaching that problem from opposite ends. The startup Onton launched Ontology 1, which it calls a trust and authenticity model for agentic commerce. Cloudflare said it was rolling out Cloudflare Wallets and Cloudflare.pay to give AI agents a stable identity and let them buy things within limits set by the people who deploy them.1112

One product looks at what an agent reads. The other looks at who the agent is when it reaches a checkout page. Taken together, they show how the industry now breaks the problem down. In the two months since that column ran, the card networks have spent heavily to own the second part.

Two kinds of trust

Onton is working on the information layer. Its argument is that agents doing product research can pull in synthetic or manipulated data and have no reliable way to separate useful signals from noise. Shoppers then buy on the strength of bad information.12 Alongside the launch, the company released research saying today's leading AI models have trouble filtering out fake content and biased reviews. It also claims its model beat Google Shopping and Amazon on accuracy in head-to-head benchmarks.1112 These are the vendor's own claims and have not been independently checked. Co-founder Alex Gunnarson described the company's focus as working out what agents should trust, as opposed to making agents smarter.11

Cloudflare is working on the transaction layer. The company says the internet was built for people, so a business that sees an agent signing up for a free trial or placing an order has no dependable way to tell a real customer's assistant from someone gaming the system.12 CEO Matthew Prince put it as needing to know who sent an agent when it "shows up at your door," and called identity and payments the infrastructure the agentic web needs.11

Coverage across the sector agrees on that split, but most of the money and press releases have gone to the second problem. A WEF piece argued that agents can already compare and transact faster than a human could supervise them. In its view the missing piece is trust in the form of verified human approval that can be traced cryptographically to the person who gave it.16

The payment networks move in

On September 30, Mastercard expanded Agent Pay with trust and intelligence services. The first is a probability score, now in US testing, that estimates how likely it is that an AI agent started a given transaction. Mastercard says the score is meant to help issuers approve legitimate agent purchases.28 The company tied the release to a white paper describing an Agent Pay Trust Framework built on five pillars: identity, intent, controls, trusted execution and intelligence.27 Mastercard also named Cloudflare as a partner in combining web signals with payment-network data in privacy-preserving environments. It is also working with Skyfire, a "Know Your Agent" provider, to help merchants and banks recognize trusted agents.28 So Cloudflare's August wallet launch is more than a standalone product. Cloudflare is also feeding signals into a card network's risk systems.

Visa has taken a parallel route. In June it announced Agent Score, an Agentic Registry and a Large Transaction Model at its Payments Forum, together with a partnership to bring Visa Intelligent Commerce into OpenAI's products. Under that deal, agent purchases run on tokenized credentials and are limited by spending caps and merchant categories the user sets.29 Earlier, Visa had launched Intelligent Commerce Connect, which lets merchants accept agent-initiated payments under four protocols through one integration: Trusted Agent Protocol, Machine Payments Protocol, Agentic Commerce Protocol and Universal Commerce Protocol.64 American Express went further on liability. Its ACE developer kit comes with what it calls an industry-first Agent Purchase Protection program, which covers cardmembers when a registered agent makes a mistake.29

The networks' basic tool is the scoped credential, a token tied to an agent instead of directly to a card. If an agent goes rogue, its token can be revoked without reissuing the user's card.7 Analysts describe the security question as shifting. Merchants used to ask only whether a payment was fraudulent. Now they also need to know whether an action was authorized, by whom, within what limits, and whether that can be proven afterward.24

Why the networks have the advantage

Consumer surveys explain why Visa and Mastercard are in a strong position. Visa's Trust Index, based on a May survey of 2,065 US adults, found that only 23% trust generative AI to handle payments for them. Sixty-one percent said they would trust Visa to handle agentic transactions.30 An earlier Visa study found Amazon, PayPal and Visa ranked well above generative AI companies as providers people would trust for a shopping agent. One in five respondents said they would not trust any of the brands listed.18

Visa did commission that research, and Visa comes out on top, so it deserves some skepticism. Still, the overall pattern shows up in independent surveys too. A YouGov figure cited by commercetools found that about 65% of Americans trust AI to compare prices, while only 14% trust it to place orders.22 Forrester found that 24% of US online adults trust agents to make routine purchases.23 Worldpay's survey of seven markets found that among consumers open to AI shopping, 80% would give an agent some role in payment, yet 59% still want to approve each purchase.26 Ecommpay's European research found 21.3% would not let an agent spend money at all, and only 6% would allow more than €500.15

The exact numbers differ, but every survey shows the same thing. People are comfortable letting AI do research and much less comfortable letting it pay. The drop-off is steepest at the moment money moves, and a familiar payment brand in the process helps close the gap. The networks are being paid in trust they built up over decades, and they are building products to collect on it.

Where the coverage diverges

The reporting splits on timing. Payment companies frame the story as fast-moving. In December, Visa predicted that millions of consumers would use AI agents to buy things by the 2026 holiday season.4 Mastercard forecasts that AI-driven consumer spending could reach £370 billion a year by 2030.24

Evidence from merchants is far more modest. In Checkout.com's 2026 report, merchants said only 3% of their transactions involve AI agents, even though 89% are preparing for agentic commerce.23 OpenAI's Instant Checkout remains on pause. Anthropic's new commerce blueprints for retailers cover search, comparison and cart-building but stop short of completing purchases, leaving payment to existing checkouts.24 Speaking at Sibos on September 29, Federal Reserve Governor Christopher Waller said market participants broadly agree the field is still in an early phase.23

The more convincing reading is the cautious one. The buildout is real. Stripe made its hosted checkout pages agent-ready for more than 7.8 million businesses, and Meta's Muse agent launched with Link checkout.23 But actual usage is still mostly agent-assisted, with a person approving the purchase, and only rarely fully delegated.

Liability is the real bottleneck

The underlying reason is that nobody has agreed who pays when an agent gets a purchase wrong. A PYMNTS merchant survey found that 93% of merchants think the AI or agent provider should cover the loss, and 80% expect providers to verify an agent's authority.23 Chargeback rules were written on the assumption that a person clicked the buy button. Tokens such as Mastercard's Agentic Tokens and the Shared Payment Tokens used with OpenAI and Stripe's protocol record the agent's identity in the transaction so disputes can be assigned to the right party.21 Juniper Research's April study named trust as the top barrier to deployment, ahead of every technical concern.21 Among financial institutions, 87% of CTOs and heads of payments said the same about agentic payments specifically.22

This is why Amex's purchase protection may matter more than any scoring model. A probability score helps an issuer decide whether to approve a payment. A guarantee helps a consumer decide whether to delegate it in the first place.

What it means for retailers

Onton's point is still relevant to retailers even though the payments story has drawn most of the attention. Visa's research found that 35% of US consumers would be more likely to shop with a retailer that offered an on-site AI shopping agent.18 Worldpay found that some shoppers hesitate because they cannot tell whether an agent's recommendations are sponsored.26 An agent that pays securely but recommends a counterfeit or a product propped up by fake reviews still damages trust.

For merchants, the practical advice in the trade coverage is consistent: make catalogs, inventory, payments and trust signals readable by machines as well as by people, and hold off on autonomous checkout until consumers are ready for it.24 The "Know Your Agent" idea, which applies KYC-style verification to agents, is turning into a requirement on both the merchant and payment-provider side.1528

The most likely outcome is that payment networks become the trust layer for agentic checkout, since consumers already trust them. Information integrity, which Onton is working on, has no comparable incumbent. That makes it the weaker of the two layers, and possibly the more important one, until someone takes responsibility for it.