Imagine your AI agent had gone shopping for you this morning. A new camera, 400 euros, chosen by itself, paid by itself. What's your first impulse?
Mine: I want to look at it first. I use AI all the time to research and compare. I stop at paying. That's not a technology problem, it's a feeling. And apparently I'm not the only one.
Search yes, buy no
In early 2026, Riskified surveyed around 2,000 people in the US and the UK. About 61 percent use AI for product search and recommendations. And 55 percent don't want an agent to shop for them.
A survey by ACI Worldwide and YouGov from June 2026 shows the same pattern: only 19 percent of respondents in the UK trust an AI assistant to follow the rules on everyday purchases. 55 percent trust a human expert. Half trust AI to find the best price, but only 17 percent to protect their payment details. And six in ten would stop using a shopping agent after a single mistake.
People delegate the search, not the decision.
The agents have long been able to compare ten toasters. What's missing is a reason to believe that someone will put things right when something goes wrong.
Who's liable?
Riskified also asked who's responsible when an AI purchase goes wrong. Around half see the AI platform as responsible, a good fifth the retailer, less than a fifth themselves.
There's already a ruling pointing in the same direction. In 2024, a tribunal in British Columbia ruled against Air Canada. The airline's chatbot had wrongly told a customer after a bereavement that he could apply for the special fare retroactively. Air Canada essentially argued that the chatbot was responsible for its own actions. The tribunal saw it differently: the bot is part of the website, and the company is liable for what its website says.
The amount was small. The message wasn't: Whoever lets an agent speak for them answers for it.
Where the human decides
I work with AI agents every day and notice that I behave exactly the same way. I let them draft, research, sort and suggest. They're not allowed to send, publish or delete without me. Emails go out as drafts. Anything public needs my yes.
At the same conference, one speaker said a sentence that stuck with me: the question isn't how autonomous an agent should be, but where it has to involve you. That's exactly what customers are asking of shopping agents right now. They draw their own decision boundaries.
What I take from it
I don't think this gap will last forever. The surveys come from payment and fraud providers with their own interests, and they cover the US and the UK. I haven't found comparable numbers for Austria or Germany. But for now, if you sell something, it means:
- Make responsibility visible. Who's behind it? What happens in case of a mistake? How do I reach a human?
- Confirmation moments are trust, not an obstacle. The step where someone decides is part of the product.
- Don't rebuild everything for agents on a hunch. Clarify the basics: one sentence for your offer, a reachable person, a traceable trail.
Building has become cheap. Being someone people trust hasn't.