Srijith Ravikumar is a Principal Engineer at Amazon building AI-powered recommendation systems at scale. Published researcher at AAAI.getty​Your customers are about to stop shopping. Their agents will do it for them.The shift from AI that recommends to AI that buys is already here. In January 2026, Google introduced the Universal Commerce Protocol at the NRF retail conference, an openly licensed standard built with Shopify, Etsy, Wayfair, Target and Walmart. OpenAI and Stripe launched a competing open standard, the Agentic Commerce Protocol, that Microsoft's Copilot Checkout now runs on. Morgan Stanley projects that agentic shoppers could account for 10% to 20% of U.S. e-commerce spending by 2030, as much as $385 billion.But there's a line your customers won't cross yet, and most roadmaps are ignoring it. In an April 2026 survey of U.S., U.K. and Canadian online adults, Forrester found about three-quarters were uncomfortable letting an AI agent complete a purchase and pay on their behalf, even with spending limits set in advance. They'll let an agent find and compare products. They stop at checkout. Their worries: loss of control, errors and liability, distrust of the agent's judgment and data security.That hesitation is rational, because the real question underneath agentic commerce is not whether the AI is good enough. It's the question of who's responsible when it's wrong?The Trust Gap Is Really A Liability Gap​Consider the failure modes a deployed shopping agent has. It buys the wrong variant, or at the wrong moment. It gets manipulated into checking out somewhere it shouldn't by a malicious product listing, an attack known as prompt injection. Or, it gets impersonated by a bot posing as legitimate. The card networks are not treating this as hypothetical. Visa now runs an "Agentic Directory" of agents and merchants it has verified, and its threat researchers measured a 25% rise in malicious bot-initiated transactions over six months, 40% in the U.S. The IMF devoted a 2026 note to the subject, organizing it around three layers: intent, authorization and settlement.Two announcements this spring started to answer the question. In April 2026, American Express published what it called an industry first. It will back Card Member purchases made by AI agents, on three conditions: the agent has to be registered on its network, the Card Member has to be authenticated, and the agent has to transmit authenticated purchase intent with the transaction. A month earlier, the U.K.'s Competition and Markets Authority told businesses that the same consumer law applies whether a customer deals with a human or an AI agent, and that the business remains responsible even when a third party built the agent.Read together, a pattern is forming, though not a settled rule. Liability tends to follow authorization. Where the agent is registered, the customer is authenticated and the intent is recorded, responsibility can be assigned. Where it isn't, the loss lands on whoever cannot produce the evidence, and payments specialists expect that to be the merchant until the rules catch up.The Infrastructure Is Not Ready Yet​Now notice the gap, the part the press releases don't advertise. Of the three conditions the American Express protection depends on, two shipped in April. Agent registration, the condition that decides which agents qualify at all, is published as a specification and still listed as being under development, with no announced date. The FIDO Alliance took in Google's and Mastercard's contributions and routed them to its payments working group, which has not published a standard yet. The rule is arriving ahead of the plumbing that makes it checkable. In that gap, the burden of proof falls on whoever kept the better records, and that is you.The rest of the industry is converging on the same shape. Mastercard's "Agent Pay" binds a payment credential to a specific agent, merchant scope and consent policy. Its "verifiable intent" standard, built with Google, produces tamper-resistant proof that the consumer actually asked for what the agent did. Experian's "Agent Trust" ties an agent's action back to a verified human. None of it is about making the model smarter. It is about identity, authorization and proof.Build The Trust Layer Before You Need It​For more than a decade, I've built search, personalization and generative AI systems for consumer platforms at scale. From that work, I've learned the lesson is that the trust layer is an engineering problem, not a legal afterthought. The teams that win will treat four things as first-class infrastructure:1. Agent IdentityEvery agent acting for a customer needs a verifiable identity, the way every API caller does. An anonymous agent is an unauthorized agent.2. Server-Side LimitsSpending caps, category rules and approval thresholds must be enforced on your servers, not requested politely of the agent. A rule that lives only in the prompt isn't a rule.3. Authorization LoggingFor every agent action, record the request as the customer expressed it, the scope they granted, the agent's identity and the time. That record is what lets you demonstrate later who authorized what. It's cheap now and impossible to reconstruct after the dispute, and it's what both the American Express protection and the CMA guidance turn on.4. Standards Over BespokeYou don't have to invent agent trust from scratch, but you do have to tell the difference between what has shipped and what has been announced. Visa's Trusted Agent Protocol is published with a public specification and reference code, so a merchant can verify today whether an agent on its site is authorized rather than a scraper or a card tester. The FIDO work is a working group, not yet a standard. Build against what you can integrate now, and leave seams for the rest.The Bottom Line​Almost everyone is investing to make their AI shopping experience smarter. Almost no one is investing to make it accountable. Customers will let an agent pay only when they trust who is on the hook if it's wrong. Build the proof layer and agents will be allowed to buy from you. Skip it and you'll have the best demo and no transactions.When an AI agent buys the wrong thing, someone pays. Build the system that can show, with evidence, that it isn't you.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?