Shoppers Let AI Advise, Not Spend

There is a clean split forming in the buyer journey, and most sellers are reading only half of it. On one side, consumers have moved their research into AI faster than almost any behavior change in the history of online retail. On the other, they are slamming the door on letting AI actually spend their money. A report released June 9 puts a hard number on the second half: 24% of consumers say they will never delegate purchases to an AI agent, and 27% say they trust no organization at all to operate one.

That number sits on top of an adoption curve that looks almost vertical. Traffic from AI sources to US retail sites grew 393% year over year in the first quarter of 2026, and in March 2026 that AI-referred traffic converted 42% better than non-AI traffic — a reversal from a year earlier, when it converted 38% worse.

So the same consumer who increasingly starts with ChatGPT or Gemini and lands on your product page primed to buy is the same consumer who will not let that assistant press the button. The discovery layer has flipped. The transaction layer has not. For cross-border sellers, the money question is which side to build for this quarter — and the behavioral research is unusually clear about the answer.

The discovery layer has already flipped

The advisory use of AI is now mainstream, not fringe. An Adobe survey of US consumers found 39% have used AI for online shopping, and 85% of those said it improved the experience. The behavior that's growing isn't autonomous buying — it's assisted deciding: idea generation, review summarizing, price comparison, spec-matching.

And it converts. Adobe's panel data shows AI-referred shoppers spend 48% longer on retail sites and view 13% more pages, with revenue per visit running 37% higher than non-AI traffic. (Adobe sells Analytics and commerce software, so treat the conversion-lift figures as directionally favorable to its narrative — but the underlying traffic data comes from its measurement panel and is consistent with what Chain Store Age, TechCrunch and others independently reported.)

The practical read: the AI is acting as a research analyst that hands a vetted shortlist to a human who still makes the call. That is a very different job to optimize for than "win the robot's purchase."

But the wallet stays shut

Here is where the June data matters. Checkout.com surveyed consumers across six markets and merchants in the UK and US, with fieldwork in early March 2026. By merchants' own count, only 3% of transactions today involve AI agents, even though 89% of merchants are actively preparing for agentic commerce.

The consumer hesitation is specific and instructive. On average, shoppers said they'd let an AI agent spend only £177 per purchase without additional approval — and the conditions they attached read like a control-panel wish list: the top non-negotiables for feeling confident were spending caps (30%), instant revocation of permissions (29%), and easy cancellation (28%). This isn't vague unease. Consumers are telling you the exact levers they need before they'll let go of the wheel.

A skepticism note: Checkout.com sells the payment and fraud infrastructure that agentic commerce will run on, so its forward-looking optimism (a third of consumers expecting 10%+ of purchases to be AI-driven within a year) should be read as an upper bound. But notice that the barrier stats — "never delegate," "trust no one" — cut directly against the company's commercial interest. A vendor with skin in the agentic game has little reason to over-report consumer refusal. That makes those resistance numbers more credible, not less.

Where delegation happens, it isn't your category

When consumers did say they'd hand over spending, the willingness clustered tightly around the boring end of the cart: 41% would delegate groceries and 31% household supplies, while high-consideration buys like financial services sat at just 15%. Delegation is a function of perceived risk and repeatability — people will outsource the restock, not the decision they'd feel embarrassed about getting wrong.

That carries a sharp warning for commodity sellers. 57% of consumers said they would let an AI agent switch brands if it found a better-value option. In categories where the agent does eventually transact, it will behave like a perfectly rational, perfectly disloyal buyer. The protection is differentiation: a product an agent can't cleanly substitute on price.

Why it's happening

The instinct is to call this a technology problem — the agents aren't good enough yet, the standards aren't settled, give it time. That's partly true, but it misreads the deeper mechanism, and the misread leads to the wrong fixes.

The cleanest explanation comes from a decade of behavioral research on what's called algorithm aversion. In a 2015 set of studies, Berkeley Dietvorst, Joseph Simmons and Cade Massey showed that people abandon an algorithm far faster than they abandon a human after seeing each make the same mistake — even when the algorithm is demonstrably more accurate. We forgive human error as bad luck. We treat machine error as proof the machine is broken.

Map that onto shopping. A product you chose yourself that disappoints is a story you can absorb — you took a swing. A product an agent bought for you that disappoints feels categorically worse: you surrendered control and got burned for it. The asymmetry isn't irrational fussiness; it's the predictable shape of how humans assign blame. It's why a consumer happily takes an AI's recommendation but won't let it complete the order. The advisory relationship keeps the human as the decider, and the decider keeps the blame — and the control.

The same researchers found the antidote, which is the part sellers should tattoo on the wall. In a 2018 follow-up, they showed people will use an imperfect algorithm if they're allowed to modify or constrain it — even slightly. A small amount of retained control collapses the aversion. That maps one-to-one onto what the Checkout.com respondents demanded: caps, revocation, cancellation. Consumers aren't waiting for AI to get more accurate. They're waiting to get more control. Accuracy is not the unlock. Agency is.

What it means for sellers

1. Optimize for the advisor moment, not the agent purchase. The volume right now is in AI that researches and hands off. Make your product data machine-readable and unambiguous — structured specs, clear comparisons, reviews that summarize cleanly. Adobe's own finding was that most retail sites still aren't readable by the AI sending them traffic. The retrievable, parseable listing wins the shortlist.

2. Design the page the human lands on after the AI. The last click is still human, and that human arrives pre-researched and high-intent (hence the 42% conversion lift). Don't bury them in re-discovery. Confirm what the AI already told them, remove friction, close fast. You are catching a warm decision, not starting a cold one.

3. Build the control levers in now. Easy cancellation, transparent returns, visible spend and order confirmation, no-drama refunds. The behavioral research says these aren't just service niceties — they are the specific mechanism that converts an AI-curious buyer into an AI-delegating one. Sellers who make "you stay in control" a visible promise will clear the trust bar first.

4. Defend differentiated categories; armor commodity ones. If you sell substitutable goods, the agent era is a price war waiting to happen — 57% will let AI switch brands on value. Invest in genuine product distinctiveness, bundles an agent can't cleanly compare, or loyalty mechanics that survive an algorithm's indifference.

5. Make your authenticity compressible. When an AI summarizes your brand to a shopper, what survives? Trust signals that depend on visual polish or long copy evaporate in summarization. The claim, the proof, the standout review have to compress into a sentence the model will actually repeat.

What to watch next

The single most important early signal is whether that "never delegate" number softens in the next survey waves. Right now it's a hard 24%; if it starts eroding quarter over quarter, the transaction layer is genuinely opening and the build calculus changes. Pair it with the merchant-reported 3% transaction share — when that climbs into double digits, agentic checkout has stopped being a slide in a keynote.

Watch the leading edge specifically in groceries and household restock, where delegation willingness is already highest. That's where consumers will first practice letting go, and the trust habits they form there will bleed into adjacent categories. Sellers in repeatable-consumable niches should treat this as the early-adopter beachhead.

And watch for the break signal: a single high-profile "the AI bought the wrong thing and no one would refund me" story, amplified across social, would hit harder than any accuracy gain could offset — because algorithm aversion means one vivid machine failure undoes a lot of quiet machine success. The infrastructure for trust is being poured right now. Whether it sets depends less on how smart the agents get and more on how much control they're willing to give back.