Where shoppers stop letting an AI buy: 97 of 146 US online shoppers want to approve the purchase themselves, and not one stops at a risky category or an unknown seller
Offered nine places to stop, 97 of 146 US online shoppers (66%) say their limit is approving the purchase before the assistant completes it, and not one chose a high-stakes category or an unknown seller (n=146, October 2026).
- Who was asked
- US online shoppers, 146 clones, US
- When
- October 2026 (fielded 1 October 2026)
- How
- Predicted responses from Brox clones, each built from a long-form interview with a real person.
- What changed from the obvious question
- The obvious question is whether people would use an AI shopping assistant at all, which returns a number that moves with the news. This asks where willingness stops instead, and offers the consent step alongside eight other places to draw the line, so the answer says which constraint binds rather than how warm people feel.
Section 01: What the panel said
- #
66%
Offered nine places to stop, two thirds choose the consent step.
base 146
- #
51%
A written guarantee against agent error splits them almost exactly in half.
base 146
- #
0
Nobody stopped at a high-stakes purchase or an unknown seller.
base 146 · contrast
- #
30
One in five would give an AI shopping assistant no access to their card at all.
base 146 · contrast
One in five would give an AI shopping assistant no access to their card at all.
02What is the first thing a buyer should take from this?
The constraint is consent, not money and not risk. Offered nine places to stop letting an AI shopping assistant buy on their behalf, 97 of 146 US online shoppers (66%) chose the same one: they want to approve the purchase before the assistant completes it. A further 30 of 146 (21%) would give an assistant no access to their card at all.
The places the industry worries about drew almost nobody. Four of 146 stopped at a spending limit, five at privacy and security, and not one of the 146 stopped at a high-stakes category or at an unknown seller.
03Does a guarantee against the agent's mistakes change that?
For about half. Asked whether a written guarantee from their card issuer covering the agent's mistakes would move their line, 74 of 146 said yes and 72 said no. The split is almost exactly even, so the topline on its own says very little. Cut by where people stop, it says more, and each group below is a count because five of the six are under 30.
- Need final approval: moves 53 of 97
- No card access at all: moves 7 of 30
- Privacy and security: moves 1 of 5
- Low spending limit: moves 4 of 4
- Fit or feel: moves 4 of 4
- Only routine reorders: moves 3 of 3
04What would a product team do with this?
- A reading, not the panel's answer: if two thirds want to approve the purchase, the confirmation step is the product, and how little friction it can carry is the design question. Removing it is not what this audience is asking for.
- A second reading: indemnity against the agent's errors is a comfort lever for one group rather than a general unlock, so pairing it with a visible approval step is likely to reach further than offering it on its own.
- The 30 who would give no card access at all are a different problem from the 97. They are not asking for better terms, and only 7 of them move for a guarantee.
05What would change this answer?
The nine places to stop were this study's list and not the respondents' own words, so a zero means a limit nobody reached first rather than a risk nobody holds. Offer a different list and the distribution moves.
Moving the line is also not the same as removing the need to approve. Of the 26 respondents who want final approval, were coded yes on the guarantee, and whose generated reasoning mentions a guarantee at all, 19 still say they would want to approve each purchase. The even split is better read as a shift in comfort than as permission.
Section 06: What they said, and the interview line it rests on
“My willingness stops right at the point where it would actually place the order without my overseeing. I don't care what kind of guarantee they offer, I would never let an AI make a final purchase on its own.”
Predicted from this line in the interview
“I would not trust it to do anything that's permanent or without my overseeing. It would have to say, this is what I'd like to do. Is it okay? And I can take it or leave it.”
Section 07: How much should you trust this?
What we would defend
- Offered nine places to stop, two thirds choose the consent step.
- A written guarantee against agent error splits them almost exactly in half.
- Nobody stopped at a high-stakes purchase or an unknown seller.
What we would not
- The base is 146 US online shoppers on the Brox E-commerce panel, not the cardholders of any issuer, and every group under 30 is reported as a count.
- The guarantee question was a yes or no, and moving the line is not removing the need to approve: of 26 approval-wanters coded yes whose generated reasoning mentions a guarantee, 19 still say they would want to approve each purchase. Read 51% as a shift in comfort, not an unlock.
The main limitation
The nine places to stop were the study's list and not the respondents' own words, so a zero is a limit nobody reached first rather than a risk nobody holds.
Section 08: How was this run?
Question wording as fielded
- Where does your willingness to let an AI shopping assistant complete the purchase on its own stop?
- Does a written guarantee from your card issuer covering the agent's mistakes move that line?
What was not asked, and why
Nobody was asked what they would do if an assistant bought something they did not want, so this measures where people say they stop rather than how they behave once a mistake has happened. No brand was named, no price was put on any basket, and the guarantee was described in general terms rather than as any issuer's published policy.
What changed from the obvious question
The obvious question is whether people would use an AI shopping assistant at all, which returns a number that moves with the news. This asks where willingness stops instead, and offers the consent step alongside eight other places to draw the line, so the answer says which constraint binds rather than how warm people feel.
Fielded through Ask Brox on 1 October 2026.
More on Online grocery and delivery · Consumer goods & grocery · E-commerce (US) panel
Ask this of your own customers: humans@brox.ai