Agentic Commerce: Fed's Waller Flags Trust as Top Payments Hurdle

A central banker draws the map

Federal Reserve Governor Christopher J. Waller has given one of the most detailed official accounts yet of AI shopping agents. His message was mixed. Agentic commerce could reshape how money moves, but only if the industry first solves a set of trust problems that are still open. He spoke on September 29 at Sibos 2026 in Miami, in an address titled "Payments in the Age of AI Agents." The first half covered AI in cross-border payments. The second turned to what he called a potentially transformative topic for the industry: AI agents buying things online for people and businesses.11

Waller's main point came from a Fed roundtable on agentic commerce that he hosted on May 26. Participants there broadly agreed that the market is still early but could significantly reshape commerce and payments if adoption scales.11 He called trust between buyers and sellers the biggest barrier to that growth, especially when software acts on its own.1113

Waller announced nothing. He set no rule, launched no program and endorsed no product.17 Still, a sitting Fed governor has now publicly laid out the problems that the card networks, AI labs and retailers are racing to solve. That gives the debate a policy reference point it lacked before.

Two models, very different risks

Waller split agentic commerce into two categories. In the agent-assisted model, the AI mostly handles search and discovery, while the person makes the decisions and pays. In the agent-delegated model, the person gives the agent authority to shop and pay, sometimes within set limits, and the agent then acts on its own.11 His example was a household that gives an agent a pre-funded virtual card and tells it to restock groceries based on past orders.11

He expects consumer-to-business purchases to come first, with assisted shopping arriving before full delegation. The reason is that delegation carries a higher risk of unintended purchases and needs more trust infrastructure.1114 This matches what retailers are seeing. As of October 2026, one analysis of the merchant landscape found that ChatGPT shopping mostly works as a discovery channel. Shoppers there are usually sent to the retailer's own checkout instead of paying inside the assistant.2

This is the most useful context for reading the speech. A year ago, Visa predicted that millions of consumers would complete purchases through AI agents by the 2026 holiday season, and it presented 2025 as the last year people would shop and check out alone.3 Payment executives told CNBC that commercial agent transactions could start as early as the first quarter of 2026. Visa's Asia-Pacific products head also admitted it was less clear exactly when the model would scale.5 Waller's description of the market as early-stage is a more cautious timeline from a regulator. It points to adoption that is real but still at an early-adopter stage.

The three walls: authentication, liability, fraud

Waller named three specific obstacles, and every outlet that covered the speech treats them as its core.121520

Authentication. In card payments, the usual question is whether the buyer is authorized to pay. With agents, the question becomes whether the agent has authority to pay for the buyer, and Waller said that will require new authentication methods.1614

Liability. He put it simply: who is responsible when an agent buys the wrong thing? He suggested that current e-commerce frameworks, such as network rules and consumer protection standards, might be adapted. He also left room for new approaches, including technical standards that record what the buyer intended and how the agent carried it out.1120 Earlier coverage made the same point from the industry side. A traditional card dispute involves four parties, and an AI agent adds a fifth.7

Fraud. Fraud models are tuned to human behavior. Waller argued they may not work for agents, which run around the clock at machine speed, and will need to be recalibrated around agent payment patterns.1114

The industry is already building in these areas, which makes Waller's list look less like an agenda and more like a confirmation of work in progress. Mastercard's Verifiable Intent, launched in March, links a shopper's identity, instructions and transaction outcome into one tamper-resistant record for use in disputes. That is close to the kind of intent-capturing standard Waller described.2 American Express said its agentic developer kit comes with a promise to cover errors made by AI agents. It can make that promise because it runs a closed loop, acting as card issuer, network and merchant processor at once.2 Visa's Trusted Agent Protocol is meant to give agents cryptographic identities so merchants can tell legitimate agents from malicious bots.31

Where the coverage diverges

Wire reports and specialist outlets agree on the substance of the speech but frame it differently. Several syndicated versions of the ANI wire story say Waller spoke about "superintelligence agents."1318 His published text refers throughout to AI agents and LLM-based agents.11 The difference matters. Waller described an incremental, infrastructure-heavy change, not a jump in machine capability.

Crypto- and fintech-focused coverage read the speech as handing the job to the private sector. Forkast described it as a market-structure problem rather than a Fed mandate, and wrote that the industry, not the central bank, is expected to write the authentication and liability rules.19 That outlet also said Waller named Visa Intelligent Commerce and Mastercard Agent Pay.19 The Fed's published text refers only to card networks publishing technical specifications, including ways to register agents and log cardholder approvals.11 The general point holds either way. But readers should not take the speech as backing any particular network's product, and one recap says plainly that it did not.17

The timing also shaped how the speech was received. Mastercard expanded Agent Pay the next day, adding probability scoring for agent-initiated transactions along with trust and intelligence signals.15 Just before the speech, six banks, including Bank of America, Capital One, ING and NatWest, published voluntary principles for agentic commerce. Those principles are non-binding and have no timetable.19 Read together, these moves show the industry setting its own rules before regulators step in. Waller's speech neither endorsed that approach nor discouraged it.

Open or closed: the market-structure fork

The most consequential part of the speech for retail technology may be its discussion of standards. Waller asked whether agentic commerce will settle on platform-specific or interoperable standards. He noted that interoperable standards, which work across e-commerce systems, agent interfaces and payment methods, could help level the field for smaller merchants and payment providers.1116 He also raised a related question about open versus closed systems. A large retailer could require shoppers to use its own agent and block outside ones. That would give it more control but less buyer choice.11

He said it is unclear which model will win, while noting that whichever one gains traction could significantly influence how the market develops.15 The current landscape shows why the question matters. Google's Universal Commerce Protocol launched in January with retailers including Etsy, Shopify, Target, Wayfair and Walmart.1 Stripe and OpenAI co-authored the Agentic Commerce Protocol, and Stripe separately released the Machine Payments Protocol with Tempo.2 Visa's Intelligent Commerce Connect accepts agent payments across several of these protocols and supports non-Visa cards. Visa presents it as a vendor-neutral on-ramp.8

The incumbents are hedging, and that tells its own story. Card networks are keeping their identity and tokenization layers as their own while staying compatible with multiple protocols. The networks would rather be the trust layer underneath every agent than bet on a single agent ecosystem.

B2B and micropayments: the quieter frontier

Waller suggested that business purchasing may suit agents better than consumer shopping. Corporate buying is often recurring and follows rules such as approved-supplier lists and budget caps, which act as natural guardrails. Agents might also negotiate terms and manage working capital.1116 He warned that higher transaction values raise the cost of mistakes. He also noted that B2B agents would need to work across ACH, wires, instant payments and cards, not just the card rails that dominate consumer e-commerce.16 Some of this is already being tested. Ramp has used Visa Intelligent Commerce for B2B bill pay in closed beta.3

He also raised machine-to-machine micropayments, the small fees an agent might pay for LLM queries, price feeds and API calls before making a purchase. At machine speed these could add up to large volumes, which would favor payment rails with low flat fees.11 He added that agentic transactions could materially change how often and when payments happen.18 That is a central banker's way of signaling interest in settlement infrastructure and not only in consumer protection.

The takeaway

Waller did not dampen the enthusiasm around AI shopping agents, and he did not endorse it either. He described agentic commerce as a real development that is still early. In his account, scale depends on unglamorous work on authentication, liability and fraud models, and on whether the market settles on open or walled-off standards.11 He closed by calling for a proactive approach that balances innovation with the safety and stability behind trust in payments.14

The practical upshot for retailers and payment firms is that, for now, the Fed is watching and setting the questions, not writing the answers. The companies that can show buyers and sellers who authorized an agent, what it was allowed to do and who pays when it gets something wrong are likely to set the terms of agentic commerce.