Shoppers just handed AI a $250 budget. Can it read your store?

The study, titled "Who's Buying? Consumer Trust in the Age of Agentic AI," surveyed 1,840 shoppers across the US, UK, France and Japan in June and July 2026, fielded with research partner Cint. Retail Dive's coverage pulled out the headline: 68% of consumers used at least one AI platform for shopping in the past three months.

The delegation numbers are where it gets interesting for anyone who sells things. Asked whether they would give an AI agent a budget of up to $250 to make a purchase on their behalf, with a seven-day return window guaranteed, 42% of US millennials said yes. Gen Z came in at 35%, Gen X at 29%, and Baby Boomers at 22%. Strip out the return guarantee and US millennial willingness falls from 42% to 34%. Outside the US the appetite is thinner across the board: 27% of non-US millennials would take the same deal.

Read those conditions again, because they are doing all the work. The cap is $250. The return window is a precondition, not a nice-to-have. This is not a customer who has handed their wallet to a robot. This is a customer running a controlled experiment with a refund policy as the safety net.

Every write-up of this study led with AI beating social media. Fair enough: in the US, AI tools now out-rank influencers by 20 points, newspapers by 16, and TikTok by 8 as trusted shopping advisors, per PPC Land's breakdown of the data. Google AI Overviews and ChatGPT each sit at 43% trust among US shoppers. Claude sits at 23% and Grok at 21%.

But look at the top of the ladder. Friends and family still hold 59%. And when shoppers were asked what safeguard they want before an agent buys anything, approval before checkout was the single most requested control globally. Roughly a third of respondents want a human review before any AI agent completes a transaction.

So the trust is real and the trust is bounded. Shoppers will let an agent research, compare, and shortlist all day. The purchase itself still runs through a human thumb on a confirm button, for everyone except the most comfortable cohort, and even they capped the experiment at $250 with an exit hatch.

One more number worth sitting with: 42% of US respondents say AI tools have made their purchase decisions take longer, not shorter, because the tools surface more brands and alternatives to consider. Among Gen Z it is 48%. AI is not compressing the path to purchase. It is widening the consideration set, which is precisely how 59% of US shoppers say AI has introduced them to brands they did not know, and how 46% of US millennials bought from one of those newly discovered brands this year. If your store has decent products and no household name, that is the best distribution news you have heard in a decade.

The survey measures intent. The transaction data measures behavior, and behavior is further along than the survey suggests.

Salesforce's 2025 holiday recap, built on activity from 1.5 billion shoppers across 89 countries, put global online holiday sales at $1.29 trillion, and attributed 20% of it to AI and agents: $262 billion in influenced revenue. Shoppers arriving from AI-powered search converted nine times more often than shoppers arriving from social referrals. And retailers that had deployed their own branded shopper agents grew 59% faster over the season than retailers that had not, 6.2% year over year against 3.9%.

Adobe's traffic data tells the same story from the referral side: AI-referred traffic to US retail sites grew 693% year over year through the 2025 holidays, those visitors converted 31% more often than other traffic, stayed 45% longer, and revenue per visit from AI referrals ran 254% ahead of the season's non-AI baseline.

None of that requires a single fully autonomous purchase. Almost all of it is the discovery half of agentic shopping: a person asks an assistant, the assistant reads the open web, and the person clicks through and buys the normal way. Which is exactly what the RTB House trust ceiling predicts. The buying is still human. The finding is already machine.

Three stages, and a store can fall out at each one.

Discovery. The agent reads what is machine-legible: product schema, feeds, sitemaps, and the plain text of your product and policy pages. Your hero banner means nothing to it. A store whose prices, variants and availability live inside JavaScript that never renders into anything parseable simply never gets seen, and nothing tells you it happened.

Evaluation. The agent compares. Price, stock, shipping cost, delivery time, and, per the RTB House conditions, the returns policy. That seven-day return window that moves millennial delegation from 34% to 42% makes your returns policy ranking criteria, read by software, at the exact moment a purchase gets decided. A generous policy buried in a PDF is a policy an agent cannot weigh in your favor.

Checkout. The stage everyone obsesses over and the one that matters least this quarter. Stripe's agentic commerce guide describes where the plumbing is heading: agents querying structured product data over Model Context Protocol instead of scraping, and paying with single-use virtual cards so the agent never holds real card numbers. It is real infrastructure and it is arriving. But the survey just told you the customer wants an approval step before checkout anyway. An agent that shortlists three options and hands a human the confirm button satisfies the 35% who demand review and loses almost nobody else.

The two-sided version, because this cuts both ways.

If it goes well for you: discovery moves into AI surfaces, your products are legible there, and you inherit traffic that Adobe says converts 31% better, from shoppers 59% of whom are actively discovering brands they did not know. For a mid-size store with good products and no ad budget to outbid incumbents, agents are the first channel in years where data quality beats domain authority.

If it goes badly: your store is invisible to agents and the revenue moves to competitors without a single alarm going off in your analytics. Or agents can see you fine, and what they see is a stale price, a wrong stock count, and a returns policy they cannot parse, so they shortlist you out. And for commodity SKUs, price-comparing agents will compress margins with a patience no human bargain hunter ever had. That last one has no fix in this post. The first two do.

The work that pays this quarter is unglamorous: complete product schema on every product page, brand, GTIN, price, availability and ratings included. A feed that agrees with your site. Stock and price data that is actually true, because an agent burned by a phantom listing does not come back to argue. Returns and shipping policies in plain page text. And a check that your robots rules are not blocking the AI referrers you want, because plenty of stores blanket-blocked every bot in 2024 and forgot.

This is a bounded, checkable project, not a platform migration. It is the kind of thing we build as a fixed-price starter, and most of it is verifiable in an afternoon: fetch your own product page the way an agent does, as text and structured data with no JavaScript, and see what survives. If what survives is a nav bar and a cookie notice, that is your answer.

What can wait: agent-native checkout. The protocols are moving fast and they are worth watching, but the customer just told you, in four countries at once, that they want the approval step. Until the delegation numbers clear their $250 cap, the store that wins is the one agents can read and shortlist, not the one that rebuilt its cart for buyers who have not arrived yet.

One survey, 1,840 people, four countries, run by an adtech firm with a product to sell against this exact narrative. That is worth saying out loud. The reason we treat the finding as load-bearing anyway is that three independent measurement systems agree with it: Salesforce's transaction data, Adobe's referral data, and the survey all describe the same shape, heavy AI involvement upstream of the purchase, human hands on the purchase itself. When intent data and behavior data from different vendors converge, the direction is usually right even when any single number is soft.

What we would not do is extrapolate the curve. Most writing about AI agents for retail treats the delegation number as destiny, and nobody knows whether the $250 cap becomes $2,500 next year or stalls for five. How many companies have actually deployed AI agents is a case study in how far survey enthusiasm can run ahead of production reality, and consumer delegation could stall the same way enterprise deployment did. Prepare for the discovery shift that is measurably happening, not the autonomy shift that is still a survey answer.

If you want the same discipline applied to your own operation, a 10x audit starts with which of your workflows the data actually supports automating, and says so when the answer is none yet.