What AI Users Fear Most: Lessons for Ecommerce

In March 2026, Anthropic published what it calls the largest and most multilingual qualitative study ever conducted: interviews with 80,508 Claude users across 159 countries and 70 languages, run over a single week the previous December. The aim was to understand not just what people use AI for, but what they want from it, and what worries them.

The headline is more optimistic than the discourse around AI would suggest. Globally, 67% of interviewees expressed net positive sentiment. But the more useful finding for anyone deploying AI in a store is buried in the concerns, and it is unambiguous.

Getting Things Wrong Is the Problem People Actually Have

Unreliability topped all thirteen concern categories, ahead of job displacement at 22.3% and loss of autonomy at 21.9%. More striking than the ranking is what sits underneath it: of everyone who raised reliability as a concern, 79% were describing something that had happened to them personally. This is not an abstract fear absorbed from headlines. People have watched an AI state something false with complete confidence.

Anthropic singles it out as the one tension in the whole study where negative experiences outweighed positive ones. Everywhere else, the good outweighed the bad.

One participant described having to photograph something to convince the model it was wrong. Anyone who has argued with a confident chatbot will recognise the feeling, and it is worth sitting with what that means for a shop.

What This Means for the AI on Your Store

Four of the study's findings translate almost directly into decisions about the agent on your storefront.

1. An invented answer is a return waiting to happen

If a customer asks whether a jacket is waterproof and your agent invents an answer, you have not created a support ticket. You have created a return, a refund, and a shopper who now distrusts everything else on the page. Returns are expensive enough already: the National Retail Federation puts the online return rate at 19.3%, and an agent that invents product details is a machine for manufacturing more of them. A quarter of people are already primed to expect exactly that failure, and four in five of them have seen it happen.

This is the argument for an agent grounded in your actual catalogue rather than one improvising from the open internet, and it is the difference between the tools that work and the ones that quietly fail their stores. An agent that says "I do not have that detail for this product" is doing better commercially than one that guesses, because the guess is the thing customers have learned to fear.

2. Privacy peaks exactly where your European customers are

Surveillance and data exploitation ranked as a concern for 13.1% of people globally, but Western Europe was the standout region at 17%, the highest anywhere. If you sell into the EU, that is your market telling you data handling is a feature rather than paperwork, which is worth reading alongside the actual GDPR obligations.

3. People dislike being flattered

Sycophancy, meaning AI that is too agreeable and reinforces whatever the user already thinks, was raised by 10.8%. In a shop this is the agent that tells every customer every product is perfect for them. Shoppers can smell it, and an agent willing to say "this one is probably not right for you" buys more trust than one that never disagrees.

4. People want outcomes, not features

Asked what they hope AI makes possible, participants did not describe capabilities. They described lives. Professional excellence led at 18.8%, then personal transformation at 13.7%, life management at 13.5% and time freedom at 11.1%. Roughly a third of all responses came down to making room for life; another quarter to doing better work rather than escaping it.

The pattern holds for shopping. Nobody wants to converse with your store. They want the right thing, at the right price, without spending their evening on it. An AI sales agent earns its place by shortening the distance to that outcome, not by being impressive in a demo.

The Takeaway

The bar is not "can your AI answer questions". Most can. The bar is whether it knows the difference between an answer it has and an answer it is constructing, and whether it will admit which is which.

That is an unglamorous standard. It will not demo well. But it is the one thing the largest study of AI users ever run says people are most worried about, and the stores that clear it will be trusted in a category where trust is currently the scarce resource.