Three accounts published in the past week put a number on how far US shoppers trust AI shopping agents, and the numbers look like they contradict each other. One says 3%. Another says more than half. The surveys behind them agree more than the headlines do, because each one asked about a different action.
What the three accounts said
AdExchanger’s Monday roundup frames retailers as divided on agentic shopping and consumers as not divided at all. It says some large retailers have partnerships with Meta’s Muse agent, which includes an autonomous shopping feature, while others, including Tapestry, let agents browse their sites but not complete transactions. It then cites a survey by the payments firm NMI to say that only 3% of US adults would trust an agent to buy something for them.
CX Dive reported on a different survey, from Synchrony and Oxford Economics, of 2,000 US consumers. It found that 82% say data security matters most in an AI-powered shopping experience and 77% want clarity on how their data is used. It also found that more than half of consumers would trust an AI shopping agent provided by a brand they already deal with, and that three-quarters would shop less or stop shopping with a brand after an AI recommends a poor product.
Search Engine Journal came at the same market from the merchant side. Columnist Slobodan Manic described a Shopify setting, “Allow Shopify to manage for me”, that auto-enrolls eligible stores in each new agentic channel. By his account that is how eligible stores’ products reached Meta on September 8, the day Muse launched. He also wrote that direct checkout is switched on by default on Meta, Google and Microsoft Copilot for stores that qualify, which means a customer can pay inside the chat app without opening the store.
Read together, the first two accounts disagree sharply. A reader who sees “3%” in one place and “more than half” in another would conclude that one survey is wrong. The third account explains why that matters: the supply side is already wiring agents into checkout, so the trust question has a deadline.
Where the accounts disagree, and why
The NMI release is the primary for the 3% figure, and it words the number differently. NMI’s “Embedded Payments: Agentic Commerce Reality Check” surveyed 1,000 US adults. It says “only 3% currently describe AI as a personal shopper that manages every stage of the shopping journey.” That is a question about how people describe AI today. It does not ask whether they would trust an agent to buy.
The NMI release has other figures that sit closer to a trust-to-buy question. Ten percent would give AI full control over a purchase. Forty-five percent would not trust AI to make purchases on their behalf at all, and 76% are uncomfortable with AI completing purchases for them. Seventy percent say they must be able to review or override an AI decision before a purchase is made. So NMI’s data does show low appetite for autonomous buying. The 3% figure measures something narrower than the summary suggests.
The Synchrony report, the primary behind the CX Dive story, asks about a ladder of tasks instead of a single yes or no. Among respondents who had not yet used AI for a given task, 78% were willing to let AI apply discounts automatically and 75% were willing to use it to compare prices. For making a purchase with their payment information, only 40% said they were comfortable. The Synchrony results as CX Dive reports them add a condition, a brand the consumer already deals with, and CX Dive built its headline on that condition.
NMI’s usage numbers fit the same picture. In the past month, 53% of respondents used AI for research or information searches and 33% used it for shopping or product discovery. Only 11% had used it to complete a transaction.
Laid next to each other, the surveys do not conflict. Comfort is high for tasks that save money or time and low for tasks that move money. A survey that asks about discounts will find a majority. A survey that asks about handing over the purchase will find a minority. The condition of a brand the shopper already deals with in the Synchrony results is a third variable, and the NMI release does not test it.
Our read: three variables decide the number
This section is our analysis, and none of the three outlets makes it. The survey results sort along three variables, and any trust figure you read should be checked against all three.
The first is the action. Applying a coupon, comparing prices and recommending a product sit on the low-risk rung. Paying sits on the high-risk rung. The two primaries agree on that ordering even though their headline numbers differ by an order of magnitude.
The second is who provides the agent. The Synchrony result is about agents from brands consumers already use. The NMI figure is about AI in general. A retailer launching its own assistant is in the first situation. A retailer whose catalog appears inside a third-party agent through a default platform setting is in the second, where the shopper may have no existing relationship with the agent.
The third is whether there is an approval step. NMI found 70% want to review or override a decision before purchase, and Synchrony’s CX Dive coverage reports that three-quarters would stop or pause using AI if it bought the wrong items without first checking. Both point to the same design feature, and neither survey says how many people would accept an agent that asks before it pays.
This also explains the retailer split that AdExchanger describes. A retailer that lets agents browse but not check out is holding the action variable at the low-risk rung. A retailer that partners with an agent that shops autonomously is betting that its own brand covers the second variable. Neither choice is wrong on the evidence available. They are bets on different rungs of the same ladder.
What it means for the marketing leader
Marketing and growth teams now sit between two clocks. Platforms are turning on agent channels and direct checkout by default, and consumers have not moved past the assist rung. NMI’s Chief Product Officer, Tiffany Johnson, put the mismatch this way: “The industry is designing for autonomy, while consumers are still asking for assistance.”
Start with what your own team is reading. When a deck cites an agent trust number, ask which action the question covered, whose agent it was, and whether a review step was part of the scenario. A 3% and a 53% can both be correct, and they point to different launch plans.
Next, launch where the surveys say comfort already exists. Discount application, price comparison and product recommendation are the tasks a majority of Synchrony’s respondents were willing to hand over. Put a review step in front of any payment, since that is the feature both surveys keep returning to. We covered a related point in Own the Approval Step Before Agents Own Your Stack, and the consumer data here supports it from the shopper’s side.
Then check the supply side. If you sell through Shopify, the setting Manic describes decides which agent channels carry your catalog, and a store that has not reviewed it may already be listed in a new one. Earlier coverage of Meta’s Muse agent connecting to Klaviyo and Shopify and of ad buying moving to software gives more background on both sides of that connection.
Finally, plan for measurement that matches the ladder. Count assisted actions and completed purchases separately. NMI’s 11% completion figure and its 33% discovery figure are different measures, and a dashboard that merges them will overstate agent-driven sales or understate agent-driven influence.
Johnson also said payment infrastructure will need to carry more context that connects identity, authentication, permissions and credentials. Marketing teams that plan agent-led offers will need to ask their payments counterparts what that context looks like before launch.
Source: NMI
