Buyers are starting to let AI agents shop with them.
Not all the way. Not everywhere. And not with an unlimited mandate.
They are asking agents to research products, compare options, find better prices, narrow a shortlist and sometimes recommend what to buy. But the final step — spending money without human approval — is still different. It feels different because it is different.
There is a particular discomfort in the idea of software buying things for you. Not researching. Not comparing. Buying. Spending your money, on your behalf, while you are not watching.
It feels like handing your wallet to a stranger and walking away.
That discomfort is not irrational. It is exactly the trust curve every serious form of delegation has to climb.
Consumers are interested, but conditional
The data looks contradictory at first.
Accenture’s 2026 Consumer Pulse Research, based on 25,590 consumers across 16 countries, found that 74% would trust a personal AI agent more than their best friend to make a purchase on their behalf. The same research says 74% would let an agent handle routine tasks, 32% would let an agent decide what to buy within boundaries, and 9% are open to fully autonomous purchases.
That sounds like a dramatic shift.
Then Gartner’s May 2026 survey pulls the enthusiasm back to earth: consumers are more willing to let AI help with discovery, research and comparison than to let it make the purchase decision. In that survey, willingness to let AI make purchase decisions reached only 11% even in lower-stakes categories such as household supplies and personal care.
Checkout.com’s 2026 agentic-commerce research shows the same tension from another angle. A third of consumers expect at least 10% of purchases to be AI-driven within a year, but 24% say they would never delegate purchases to AI and 27% trust no organization to operate an AI shopping agent.
These numbers are not really in conflict. They describe a boundary.
Consumers are not saying “yes” or “no” to agentic commerce. They are saying: yes to help, yes to convenience, yes to better information, maybe to recommendations, and only later — under strict conditions — to payment.
Delegation starts where mistakes are cheap
People delegate the parts of shopping that feel like work before they delegate the parts that feel like responsibility.
That is why low-stakes, repetitive and reversible tasks come first:
- finding better deals;
- comparing prices;
- narrowing options;
- restocking household items;
- tracking orders;
- handling returns;
- surfacing alternatives when the usual brand is poor value.
Checkout.com found higher willingness to delegate repeatable categories such as groceries and household supplies than high-consideration categories. Accenture found that many loyal consumers would even let an agent switch brands if it found a better fit.
This is not indecision. It is a rational architecture of trust.
People delegate what is easy to verify and easy to undo. They keep control of what is expensive, personal, risky or hard to reverse.
The wallet goes last because the wallet is where delegation becomes authority.
We have seen this pattern before
This is not the first time people have been asked to trust a machine with something that used to require a human.
There was a time when putting money into a wall and receiving cash from a machine felt reckless. The ATM had to earn trust.
There was a time when typing a card number into a website felt unreasonable. E-commerce had to earn trust.
There was a time when getting into a stranger’s car because an app told you to felt like something parents warned you not to do. Ride-hailing had to earn trust.
There was a time when tapping a card without a signature or PIN felt like fraud made easy. Contactless payments had to earn trust.
None of these became normal because people were persuaded by a slogan. They became normal through repeated small successes, clear limits, visible receipts, easy reversal and enough value to justify the loss of direct control.
The experience did the work.
Agentic commerce will follow the same pattern.
Not because agents are magic, and not because consumers suddenly become reckless, but because successful delegation compounds. If the agent saves time, avoids a bad choice, finds the right product and keeps the buyer in control, the buyer is more likely to delegate a little more next time.
The honest part: it is still early
It is tempting to say the revolution has arrived. That is too easy.
Most people who use AI while shopping are not yet letting an agent buy for them. They are asking questions, comparing products, checking reviews, looking for alternatives and reducing the effort required to choose.
Gartner’s point is important here: passive exposure to AI is not the same thing as active adoption of AI-led purchasing. Seeing AI features inside search engines, apps and marketplaces does not mean consumers are ready to hand over decision-making authority.
So the accurate description is this:
Demand is forming. The habit is forming. The infrastructure is forming. But trust is still catching up.
That gap matters for merchants, platforms and anyone building agentic commerce tools, because trust will not be built at the payment screen alone.
It starts much earlier.
Before the wallet, there is the catalog
Before an AI agent can safely buy a product, it has to answer simpler questions correctly:
- What is this product?
- Which variant is actually available?
- What is the current price?
- Is it in stock?
- What are the delivery constraints?
- What are the return rules?
- Is there a compatible alternative?
- Where should the buyer be sent to complete the purchase?
If those answers are stale, incomplete or ambiguous, the agent has to guess.
And guessing is poison for trust.
An agent that recommends the wrong variant, quotes a price that has changed, sends a buyer to an out-of-stock product or misses a crucial condition does not merely fail one transaction. It confirms the buyer’s fear that delegation creates more work, more risk and less control.
One bad delegated purchase can undo many good assisted searches.
This is why the first practical layer of agentic commerce is not the wallet. It is the computable catalog.
Not a beautiful product page designed for human eyes. Not a generic feed built for advertising. Not an internal API that exposes data but leaves meaning to be inferred.
A computable catalog is product information structured so an agent can discover it, filter it, compare it and act on it without scraping pages or inventing missing context.
For a merchant, this is less glamorous than autonomous checkout. It is also more immediate.
Because long before a buyer lets an agent pay, the buyer may let an agent choose.
And before an agent can choose well, it needs reliable product truth.
“Last” is not “never”
The wallet goes last, and it should.
Spending caps, revocable permissions, confirmation steps, clear receipts and easy cancellation are not obstacles to agentic commerce. They are the conditions under which it becomes normal.
But last does not mean never.
People once said they would never put a card number into a website. They said they would never trust a machine to dispense cash, never get into a stranger’s car, never pay by tapping plastic on a terminal.
Then the systems became useful, bounded and boring enough to trust.
AI shopping will take the same road if it earns the right to do so.
Not through hype. Not through “agents will buy everything” predictions. Through one correct, reversible, transparent step at a time.
The first merchants to benefit will not necessarily be the ones who automate payment first.
They will be the ones whose products agents can understand first.