69% Would Let AI Buy for Them. 39% Check It Against Four Sources.

The agency Croud released a consumer study on August 11 with a number built for a headline: 69% of Americans are open to letting AI purchase on their behalf without approval. It surveyed more than 2,000 US consumers, nationally representative, under the title “The Expansion of the Validation Economy.”

Four paragraphs down the same release sits the number that contradicts the mood of the first one. Among people already using AI to shop, 39% check an AI recommendation against four or more sources before buying.

Both are real. They’re answers to different questions.

One asks what someone would be willing to do. The other asks what they did last time. Surveys are good at the first and poor at the second, and the gap between them is where a lot of AI commerce planning is currently being done.

The tell is in what the study is named after

Croud’s report is called the Validation Economy, and its own executives lead with the checking rather than the buying. Val Davis, Croud’s US CEO, put it this way:

“Ironically, AI is making the consumer journey more human, not less. As consumers increasingly rely on AI to discover more options, they’re placing significant value on human validation before making a purchase.”

That’s a more useful sentence than the 69%. It says AI widened the top of the funnel and then handed the shopper a verification job they didn’t have before, because a recommendation from a system with no accountability has to be checked against something.

Rajiv Ragu, VP of Growth at the supplement brand Thorne, was quoted in the same release and landed on the same split:

“What stood out to us wasn’t just that consumers are using AI to discover products, but that trust still has to be earned beyond the AI recommendation.”

We published the opposite headline four days earlier

On August 10 we covered RTB House’s finding that AI stretches out purchase decisions for 42% of US shoppers, and noted that the company measuring the longer consideration window sells advertising into exactly that window.

So within one week: a survey saying most people would delegate the purchase, and a survey saying AI makes people take longer. Neither company is doing anything improper. Croud advises brands on AI-mediated journeys and its data says those journeys are getting more complex. RTB House sells ads that run during consideration and its data says consideration is getting longer. Croud’s own numbers show AI-assisted fashion shoppers spending 56% more than non-users, which is a fine reason for a brand to hire an agency that understands AI-assisted shoppers.

Vendor research is aimed. Nobody had to falsify anything to get there. The measurement gets designed around a question the sponsor already has a product for, and the number that travels is the one that makes the strongest case.

The mechanism that makes both true

Stated willingness and observed behavior come apart most sharply where risk is uneven, and this study shows exactly where. The categories where people accepted AI-automated checkout were low-risk and routine: groceries, household essentials. Half of consumers said they were comfortable with AI purchasing across three or more categories, which means half weren’t.

Reading “69% would let AI buy for them” as “69% of your buyers are about to stop reading your site” gets the mechanism backwards. A shopper handing over a repeat grocery order and a shopper choosing a $200 supplement stack are not the same shopper, and the second one is the one running four sources against the AI’s answer.

That checking behavior is the actionable half. If 73% of AI users research through a large language model before settling on a brand, and a large share then verify what it told them, the brand’s job splits in two. You have to be present in the answer, and you have to survive the check that follows it. We’ve written about the confidence gap between AI visibility and actual sales, and this is the missing middle of it: the verification step nobody is measuring.

The verification surfaces are ordinary. Reviews, Reddit threads, a comparison someone wrote, the brand’s own site read with suspicion. Which is to say the work that earns the check is the work marketers were already supposed to be doing, now with a shorter window and a more skeptical reader.

One thing to do with any AI-shopper statistic that crosses your desk this planning season: find the question that produced it. “Would you be open to” and “how many sources did you check” produce different worlds, and the release will usually print both if you read past the headline. The 69% sells a report. The 39% tells you what to build.