Fifty-eight percent of consumers in a recent U.S. survey said they were open to ordering through an AI assistant. Only 6% had actually done it. That gap is the AI-agent market in one line. Interest is real. Trust has not caught up.
The obvious response is to make agents look more capable: longer task lists, more integrations, more autonomy. The research points somewhere else. People already understand that an AI can search, compare, click, and buy. What they want to know is whether it will stop at the right moment, show its work, and leave the final decision where it belongs.
Interest and hesitation can both be true
The 58% figure comes from two Radial-commissioned surveys of 1,000 U.S. consumers each, conducted by Dynata in late 2025 and early 2026. The same research found that 53% wanted approval before any purchase went through.
A separate Commerce and PayPal study surveyed 3,000 online shoppers across the United States, United Kingdom, and Australia. Roughly two-thirds were interested in trying agentic shopping tools. Their preferred role for AI was research and comparison, not quietly taking over checkout.
There is no contradiction. People want the time back. They do not want to discover later that the wrong item was ordered, the wrong appointment was booked, or the wrong account was changed.
The trust cliff appears at the final decision
Gartner found the same boundary from another angle. In a January 2026 survey of 322 U.S. consumers, willingness to let AI make a purchase decision topped out at 11% across the categories studied. Openness was higher when the job was to narrow choices: 31% for household supplies and 28% for personal electronics.
That difference matters. Searching a catalogue, checking availability, comparing prices, and filtering bad options are useful but reversible. Pressing buy is consequential. A product that treats those steps as one undifferentiated block of “autonomy” misses how people actually think about risk.
The better interaction is simple:
- Do the reversible work. Search, compare, fill the boring fields, and prepare the action.
- Show the receipt. Make the item, price, time, recipient, and account visible.
- Ask once. Stop before the action spends money, sends something, books something, or changes an account.
This is not the same as asking permission for every click. That creates approval fatigue and turns the person back into the operator. The gate belongs at the consequence, not at every step.
One mistake can erase the benefit
Trust is not only hard to earn. It is easy to lose. A June 2026 YouGov survey commissioned by ACI Worldwide asked 2,080 UK adults about AI shopping agents. Sixty percent said they would stop using one after a single mistake.
That is a brutal quality bar, but it makes sense. A chatbot mistake gives you a bad paragraph. An agent mistake can create a charge, a commitment, a message, or a cancellation. The cost is not measured only in money. It includes the time required to notice the error, understand what happened, and undo it.
This is why an honest failure state matters. If a speech service is unavailable, the interface should say the service is unavailable, not blame the microphone. If a booking cannot be confirmed, the agent should show where it stopped, not smooth the uncertainty into a confident “done.” Reliability includes telling the truth when the task did not finish.
The first winning tasks will be inspectable
Radial's research found the strongest openness in post-purchase work such as delivery monitoring and problem solving. Those tasks are a useful starting point because the intended outcome is clear, progress can be checked, and most steps are reversible.
The same logic applies beyond shopping:
- find appointment options, then ask before booking;
- prepare a subscription cancellation, then ask before confirming;
- compare reorders, then ask before payment;
- draft a message, then ask before sending;
- track a delayed delivery and prepare the support request.
These are not the flashiest agent demos. They are the jobs where trust can accumulate. The person can see the result, understand the boundary, and decide whether to hand over more next time.
Autonomy should be earned one boundary at a time
Vuto is being built around that rule. You should be able to say the outcome once and let the assistant handle the technical middle. When the task reaches money, messages, bookings, personal information, or account changes, Vuto should show the exact action and wait for one unmistakable approval.
The product is still pre-launch. A narrow voice prototype is live, and the current build is deliberately smaller than the full promise. That is part of the point. Trust will not come from claiming that an agent can do everything. It will come from showing what it did, where it stopped, and why the next tap belongs to you.
For more on the mechanism, read the approval-before-action rule and how Vuto is designed to handle real tasks.
If you want to try Vuto as the product moves beyond the prototype, join the waitlist at vuto.ai.