53 percent of US businesses would let AI agents negotiate on their behalf. The payments stack is not ready for that.

Visa dropped its B2AI survey last week, and one number landed harder than anything else in the report. Over half of US businesses say they would let AI agents negotiate prices and terms with other AI agents. Not in some future pilot. Now.

That is not a concept paper. It is not a keynote slide. It is procurement appetite, measured across thousands of decision-makers, and it should change how we think about the next 18 months in payments infrastructure.

The problem is simple. The appetite is here. The plumbing is not.

The Visa/Morning Consult survey covered more than 15,000 consumers and business leaders across 10 markets. The headline figure, 53 percent of US businesses willing to let AI agents negotiate on their behalf, is striking. But the supporting data is what makes it real.

88 percent of businesses said they would share pricing and inventory data with AI systems to enable faster transactions. 77 percent are already using or piloting AI in some form. These are not hypothetical "would you consider" responses. A lot of them are already building.

The consumer side tells a different story. Only 36 percent of consumers trust bank-backed AI agents to act on their behalf. Independent agents? 28 percent. The generational split is stark. Gen Z trust sits at 48 percent. Boomers at 20 percent.

Here is the number that caught our attention most. 40 percent of Americans say they have already bought something they would not have considered without an AI tool recommending it. That is not trust in agents making decisions. That is agents already shaping decisions, whether consumers frame it that way or not.

Survey data is useful. Live transactions are better. And we are now seeing both.

Visa and Ramp announced a multi-year issuing agreement to automate corporate bill pay using agentic AI. Ramp has over 50,000 customers. This is not a sandbox experiment. It is AI agents handling real invoices, matching them to purchase orders, and initiating payments without a human clicking "approve" on each one. We covered the details in our earlier piece on Visa and Ramp's agentic bill pay partnership.

Mastercard went further. In Hong Kong, it completed its first live agentic transaction: an AI agent booking a ride through hoppa, paying with tokenized credentials, Payment Passkeys, and Mastercard's new Agentic Tokens. A real customer. A real ride. A real payment. No human touched the checkout.

Santander and Visa have agentic payments live in production. eBay is hiring an Agentic Search Product Designer in Toronto, a role that did not exist 12 months ago.

These are not demos. That is a point worth repeating because we spent most of 2025 watching network after network announce "agentic commerce initiatives" that turned out to be slide decks with timelines. The shift from demo to live happened faster than most people expected. For a broader look at how we got here, see our explainer on what agentic payments actually are.

Here is the thing. Every live deployment we just described runs on bespoke rails. There is no shared standard for how an AI agent authenticates, pays, or resolves disputes. Each implementation is a custom build.

Start with authentication. When a human pays online, Strong Customer Authentication (SCA) asks them to confirm with a fingerprint, a PIN, or a one-time code. That entire framework assumes a person is on the other end. An AI agent does not have a fingerprint. It does not receive SMS codes. So who proves the agent is authorised to spend?

Visa's answer is the Trusted Agent Protocol. Mastercard's answer is Agentic Tokens plus Payment Passkeys. Both are real. Both work. They are also completely incompatible with each other, and neither has been adopted as an industry standard.

Liability is the next problem. When a human clicks "confirm" and the purchase goes wrong, the dispute path is clear. Chargeback rights, consumer protection law, card scheme rules. When an AI agent negotiates a price, accepts terms, and initiates payment without human confirmation, who is liable? The consumer who delegated authority? The platform that built the agent? The merchant who accepted the transaction?

Nobody has answered that question yet. Not the card networks. Not the regulators. Not the courts.

Dispute resolution is especially messy. Current chargeback and dispute mechanisms require a cardholder to assert they did not authorise a transaction or did not receive what was promised. If an agent acted within its delegated authority but made a poor decision, that is not fraud. It is not an unauthorised transaction. It is a bad call made by software the consumer chose to trust. Existing dispute categories do not cover that scenario cleanly.

The Visa survey exposes a split that could define how agentic commerce develops over the next two years. Businesses are ready. Consumers are cautious. That mismatch matters.

On the business side, the enthusiasm is broad. 53 percent willing to let agents negotiate. 88 percent willing to share data. 77 percent already piloting. These numbers suggest B2B agentic commerce will move fast, because both sides of the transaction, buyer and supplier, are willing participants.

Consumer commerce is different. Only 36 percent trust a bank-backed AI agent. 28 percent trust an independent one. The trust gap in agentic commerce is real and it is not closing quickly.

The generational divide makes this more complicated, not less. Gen Z at 48 percent trust looks promising until you consider that Gen Z also has the least purchasing power and the smallest average transaction size. Boomers at 20 percent trust control a disproportionate share of consumer spending. The cohort most comfortable with AI agents is the cohort least able to drive volume.

Here is the part that should worry payments executives most. If B2B agentic commerce scales while consumer agentic commerce stalls, we get a two-speed market. Business payments infrastructure evolves rapidly. Consumer payments infrastructure stays largely the same. That is not necessarily bad. But it means the billions being invested in consumer-facing agentic checkout experiences might not pay off for years.

The 40 percent figure on AI-influenced purchases suggests something subtler. Consumers may not trust agents to buy for them, but they already trust agents to tell them what to buy. That is a different kind of delegation. Less visible. Possibly more powerful.

Three things to watch.

Protocol convergence. Visa and Mastercard are building competing authentication and authorisation frameworks for agentic transactions. The payments industry has been here before. Competing standards eventually converge, but the process is slow and expensive. The question is whether a third party, a W3C working group, a banking consortium, or a regulator, forces convergence before the market fragments further.

Regulatory response. No major regulator has issued guidance on agent-initiated payments. Not the FCA. Not the CFPB. Not the European Banking Authority. That silence will not last. The moment a high-profile agent-initiated transaction goes wrong, probably a consumer dispute involving a significant sum, regulators will move. The industry would be better served getting ahead of that.

The consumer trust trajectory. The 36 percent trust figure for bank-backed agents could go either way. A few successful, visible deployments (a major bank launching agent-managed bill pay that actually saves people money) could shift sentiment quickly. A single publicised failure could set it back years. The asymmetry between building trust and destroying it has never been more relevant.

The Visa survey is not a prediction. It is a snapshot of an industry in mid-leap. Businesses have decided they want this. Consumers have not. The infrastructure is being built in parallel by competing networks with incompatible approaches. And not one regulator has weighed in.

That is four facts. Together, they describe a market that is going to move fast and break things, whether we are ready or not.

The question is not whether AI agents will participate in commerce. They already do. The question is whether the rules will be written before or after the first big dispute.