AI has won the shopping half of commerce and lost, for now, the paying half. Visa's Trust Index for agentic commerce, released on 9 September 2026, found that 72% of US consumers have used an AI assistant but only 23% trust generative AI to handle payment transactions on their behalf. The day before, Visa CEO Ryan McInerney named the barrier in one word: trust. Agentic checkout, where an AI agent completes the purchase rather than handing the shopper back to the merchant, is the step consumers are least willing to delegate, and the industry's product decisions this year show the platforms have already accepted that.
What Did Visa's Trust Index Find?
The Visa Trust Index is a Harris Poll survey of 2,065 US adults, fielded from 26 to 28 May 2026 and weighted to the US Census adult population. Its headline findings are measured at the person level, and they describe attitudes, not behaviour.
| Finding | Figure |
|---|---|
| US consumers who have used an AI assistant | 72% |
| US consumers who trust generative AI to handle payments on their behalf | 23% |
| Respondents who would trust Visa to handle agentic transactions | 61% |
| Visa trust among consumers aged 18 to 34 | 68% |
| Visa trust among frequent AI users | 71% |
The gap between 72% and 23% is the whole story. Nearly three in four Americans are already comfortable asking an AI for help, and fewer than one in four will let it spend. Oliver Jenkyn, Group President at Visa, framed the finding in the release: "Trust will be foundational to driving agentic commerce adoption."
Why Do Consumers Trust AI to Shop but Not to Pay?
Independent research points the same way. PYMNTS Intelligence, in its report on consumers moving from assistive to agentic AI, found that 56% of consumers would allow an AI agent to search and compare products, but only 37% would allow one to authorise payments and 35% would give one access to saved payment methods.
The drop sits exactly where the consequence changes. A bad recommendation costs a second search. A bad payment costs money, a dispute, and a chargeback process nobody has yet explained for a purchase a machine made. McInerney described the pattern at the Goldman Sachs Communacopia + Technology Conference on 8 September: "We are seeing adoption for shopping, but not yet for autonomous payments." Consumers use large language models to compare products, then move to the seller's website to complete the transaction.
That behaviour already shows up in traffic data. Adobe Digital Insights reported that AI-referred visits to US retail sites grew 393% year on year in Q1 2026 and, by March 2026, converted 42% better than other traffic at the session level. Those are shoppers who used AI to decide and then bought on the merchant's own site. The conversion is happening; it is happening after the agent hands over.
How Should the 61% Figure Be Read?
The survey's most quoted number deserves the most caution. Visa commissioned a survey that found consumers trust Visa more than any other brand to handle agentic payments. That does not make the figure wrong, but it does make it self-interested, and the release does not publish the scores of the other brands tested, so the size of Visa's lead cannot be checked.
There is also a question-wording effect worth naming. "Do you trust generative AI to handle payments" and "would you trust Visa to handle agentic transactions" are different questions. The first asks about an unfamiliar actor; the second attaches a familiar, regulated payment brand to the same act. The jump from 23% to 61% is best read as evidence that a known intermediary with fraud protection changes consumer comfort, not as a measure of Visa's specific advantage. The finding that holds up regardless of sponsor is the gap itself: people will delegate the search and hesitate at the payment.
Why the Leading AI Assistants Have Stepped Back From Checkout
The trust data explains a pattern in product decisions that otherwise looks like retreat.
| Company | Decision | When |
|---|---|---|
| OpenAI | Retired Instant Checkout in ChatGPT, about six months after launch, returning the purchase step to merchants' own sites | March 2026 |
| Anthropic | Released open-source commerce agents that build the cart but never place an order or charge a card | September 2026 |
| Visa, Mastercard, Ant International | Announced a shared Know Your Agent framework to verify agents before they pay | September 2026 |
Each decision answers the same consumer signal. OpenAI kept discovery and gave up the transaction. Anthropic built the agent up to the edge of payment and stopped. The networks are working on the precondition for agent payment, which is proving who the agent is and who stands behind it. None of the three is betting that consumers will hand over the payment step soon without that proof in place.
The pattern applies most clearly to standalone AI assistants. Assistants built into a phone or a platform that already holds a trusted, biometrically secured wallet start from a different position. The Visa finding that trust jumps when a familiar payment brand is attached suggests their gap is smaller, although no survey cited here measures it directly.
What Would Close the Trust Gap?
The survey measures an attitude, but the attitude breaks down into three questions a shopper needs answered before handing an agent the payment step. Each maps to infrastructure at a different stage of readiness.
| Shopper question | Infrastructure answer | Status, September 2026 |
|---|---|---|
| Is this agent who it says it is, and who is behind it? | Agent identity: Visa Trusted Agent Protocol, Ant's Agentic Mobile Protocol, and the new Know Your Agent framework linking them | Protocols live; cross-network framework announced without a specification |
| Did I actually authorise this purchase? | Verifiable authorisation: Mastercard's Verifiable Intent, co-developed with Google, which records the user's instruction and the agent's action in one tamper-resistant log | Open-sourced March 2026 |
| What happens if the agent gets it wrong? | Dispute rules for purchases where the disagreement is about what the agent was told to do | Not addressed in any of the announcements reviewed for this article |
The first two questions now have technical answers, even if they are not yet joined up. The third is only partly covered. Card dispute rules exist, but they were written for purchases a person made, and none of the announcements reviewed here say how a dispute is settled when the argument is about what the shopper instructed the agent to buy. That question has no public answer yet, and it is the one a consumer is most likely to ask. The 61% of respondents who said they would trust Visa with agentic transactions are, in effect, pointing at the answer they already know: a payment brand with established fraud protection and chargeback rights. Until agent purchases carry equally clear recourse, the gap between shopping and paying is unlikely to close on technology alone.
Where the Data Is Thin
Three limits on the evidence should shape how far any merchant leans on it.
- Attitudes, not behaviour. Both the Visa and PYMNTS figures record what people say they would allow. No independent source cited here measures how many purchases agents actually complete today, only how much traffic AI sends to merchant sites.
- United States only, fielded in May. The Visa survey covers US adults and was fielded four months before release. Attitudes in markets with high mobile-wallet use, where Ant International operates, may differ.
- Different grains. The Visa and PYMNTS figures are measured per person. Adobe's conversion figures are measured per session. They support the same direction, but they cannot be combined into a single rate.
What This Means for Merchants
The contrarian conclusion is that agentic checkout is not the near-term priority for most merchants, and the investment should go one step earlier.
There is a merchant-side reason to be unhurried as well. A purchase completed inside an assistant skips the merchant's own pages, and with them the cross-sell, the bundle and the account sign-up that happen there. Consumer hesitation and merchant incentive currently point the same way.
If consumers use agents to decide and then buy on the merchant's site, the commercial contest this holiday season is about being the product the agent recommends and then converting the shopper who arrives already decided. Two practical consequences follow.
The first is that product data an agent can read is worth more than a checkout an agent can complete. Accurate stock, prices that match across feed and page, and returns and delivery terms held as structured fields decide whether an agent recommends a product at all. None of that depends on the consumer trusting an agent with a card.
The second is that the handoff page matters more than it used to. A shopper arriving from an AI assistant has already compared options. The page they land on should confirm what the agent told them, product, price and availability, and get out of the way. Any mismatch between what the agent said and what the page shows is the fastest way to lose a pre-decided buyer.
The third is measurement. AI-referred shoppers behave differently from search or social visitors, so they should be reported as their own segment. Most analytics platforms record the referring domain of assistants such as ChatGPT, Perplexity and Gemini, which is enough to build a separate channel grouping and compare conversion, basket size and returns at the session level. Without that split, the effect this article describes is averaged into general organic traffic and becomes invisible to the people deciding where to invest.
Agent-completed checkout will come, and the identity work now under way between the networks is the path to it. The data says to prepare the catalogue and the landing experience first, and to treat agentic checkout as the step that follows trust rather than the one that creates it.
