Six in ten UK adults would stop using an AI shopping agent after a single mistake, according to a YouGov survey commissioned by payments software company ACI Worldwide and published on June 29, 2026. The poll of 2,080 adults found that 19% trust AI assistants to make everyday purchasing decisions within rules they set themselves, against 55% who would trust a human expert or adviser with the same job.

The numbers arrived at an awkward moment for the companies building the plumbing. Over the preceding nine months, Google, OpenAI, Visa, Mastercard and a string of retailers had moved AI agents from recommending products toward paying for them. ACI's survey asked the people who would be doing the buying, and the answer was cool. Half the respondents trusted AI to find the best price. Only 15% trusted it to sort out a problem when something went wrong.

ELI5

A YouGov poll of 2,080 UK adults for payments software company ACI Worldwide found that most people do not trust AI agents to buy things for them, and 60% would stop using one after a single mistake. That matters to the retailers, banks and AI companies building tools that take AI from suggesting products to paying for them. People were happy to let AI hunt for low prices, but they want to approve each purchase, cap what the agent can spend and know who refunds them when something goes wrong.

How the survey was run

According to ACI, all figures come from YouGov Plc. The total sample was 2,080 UK adults aged 18 and over, with fieldwork carried out between June 19 and June 22, 2026. The survey was conducted online, and the results were weighted to be representative of the UK adult population. The body of the announcement rounds the sample to "more than 2,000 UK adults."

ACI defines the product in question narrowly. AI shopping agents, in the company's words, are "tools that can search, compare and, with a consumer's permission, complete purchases on their behalf." The announcement places the research against a specific industry shift: merchants, payment providers and technology platforms moving these tools "beyond product discovery and closer to checkout and payment."

Several things are missing. ACI did not publish the questionnaire, the full list of answer options for each question, a margin of error, or any breakdown by age, region, income or prior use of AI tools. The sentence describing the headline trust question also contains a typographical error - it says 19% of consumers "trust AI assistants to follow rules to set up make every day purchasing decisions" - so the exact wording put to respondents cannot be established from the announcement. Some percentages, where they can be added up, leave a residual that the announcement does not explain; those gaps are noted below as PPC Land's own arithmetic.

The commissioning company also has a stake in the subject. ACI Worldwide, listed on Nasdaq as ACIW, describes itself as a provider of software that powers "intelligent payments orchestration in real time so banks, billers and merchants can drive growth." Banks came out of the survey better than any other type of organisation. Retailers did not.

Help with the search, not the decision

The central finding is a sharp divide between AI as an assistant and AI as a decision-maker. Respondents were most comfortable with tasks that resemble a search engine's job. According to ACI, 50% trust AI to find the best price available, and 43% trust it to follow spending limits.

Confidence fell away once money, data and judgment were involved. Only 18% trust AI to act in their best financial interest. Just 17% trust it to keep personal and payment data secure, and 15% trust it to handle problems when something goes wrong. By PPC Land's calculation, that is a gap of 35 percentage points between the most trusted task - price comparison - and the least trusted one, which is fixing what went wrong.

ACI's own reading of the data was blunt. "Consumers are comfortable with AI assisting decisions, but deeply uncomfortable with AI making them," the announcement states.

The broader trust question produced a similar picture. Seven in ten respondents (69%) said they do not trust AI even when it follows rules they set. The 19% figure and the 69% figure appear to come from the same question; if they do, 12% of respondents gave neither answer, by PPC Land's arithmetic. The announcement does not say whether those respondents chose a neutral option or answered "don't know."

And then there is the headline number. Six in ten (60%) said they would stop using an AI agent after just one mistake. The announcement does not define what counted as a mistake - a wrong size, a duplicate order, an overspend - or whether the question specified one.

"The findings clearly show that consumers are open to AI helping them shop smarter, but only if they remain firmly in control of both the decision-making and their money," said Adriana Iordan, head of merchant and payments intelligence at ACI Worldwide. "They're telling us very clearly that they won't hand control of their finances to an autonomous agent without safeguards. This isn't a capability gap; it's a trust and confidence gap. If the industry wants adoption, it must prioritise control over capability: explicit approvals, hard spending limits, protected payment details and clear accountability when things go wrong."

Savings do not buy trust

The obvious commercial pitch for a shopping agent is money: an agent that compares every retailer should, in theory, land a lower price. ACI's data suggest that argument has limits. According to the survey, 44% of respondents would not trust an AI shopping agent regardless of how much it saved them. One in four said it would need to save them more than 15% before they would trust it.

The tension is worth spelling out. Price search was the task respondents trusted AI with most. Yet savings - the output of that search - were not enough to persuade almost half of them to let an agent complete the purchase. ACI attributes the resistance to "perceived risk and loss of control" rather than value. The announcement does not state whether the 15% threshold question was put to all respondents or only to those who did not rule out an agent entirely.

Autonomy is where resistance peaks

Three scenarios tested how far respondents would let an agent go. Seven in ten (70%) said purchases made without asking would affect their willingness to use an AI shopping agent. Linking the agent to a bank account would affect the willingness of 61%. Tracking everything they browse online would affect 54%.

Those three concerns map closely onto design choices the payments industry was making in the first half of 2026. On April 28, 2026, the FIDO Alliance formed an Agentic Authentication Technical Working Group and a Payments Technical Working Group, with Google donating its Agent Payments Protocol and Mastercard contributing Verifiable Intent. Version 0.2 of the Agent Payments Protocol, released the same day, added support for "Human Not Present" payments - purchases an agent completes without the user present at the moment of payment. That is close to the scenario 70% of ACI's respondents flagged as a deterrent.

Other projects pulled the opposite way, toward proof that a human approved the purchase. Mastercard and Google introduced Verifiable Intent on March 5, 2026, a specification built on FIDO Alliance, EMVCo, IETF and W3C standards and meant to verify what a user intended an agent to buy. In a pilot announced on April 30, 2026, Mastercard and Rabobank completed the first AI agent-initiated payment in the Netherlands, in which the purchase was confirmed with Mastercard Payment Passkeys and, according to Mastercard, the agent never received the card number. Visa's approach, set out when it backed Google's Universal Commerce Protocol in January 2026, relies on three layers: tokenization, bank-verified provisioning and biometric checks at the moment of purchase.

The bank account figure deserves a separate note. ACI's respondents were asked about linking an agent to a bank account, not to a payment card or a tokenised credential. The announcement does not show whether the 61% concern would change if the agent never saw account details at all - which is precisely what the card networks' token-based designs claim to offer.

Who pays when the agent gets it wrong

On accountability, respondents were unambiguous. According to ACI, 54% said the technology or AI company that built the agent should be accountable for refunds. Just 9% would hold the retailer responsible, and only 3% would blame banks or card issuers.

That runs against how refunds work in conventional online retail, where returns are processed by the retailer that took the order and disputed card payments go through the card issuer. The two parties that run the existing refund machinery received the fewest votes. ACI's announcement describes the result as "a challenge many emerging commerce models have yet to address."

Those three answers add up to 66%. The remaining 34% of respondents chose answers the announcement does not list, by PPC Land's arithmetic.

Regulators had begun mapping the same question before the survey went into the field. In Singapore, the Infocomm Media Development Authority published a 36-page discussion paper in May 2026 on legal responsibility for AI agents, naming seven categories of actor - from model developers and tooling providers to deployers, end users and third parties - without recommending who should carry liability. In the UK, the Digital Regulation Cooperation Forum, which brings together the Competition and Markets Authority, the Financial Conduct Authority, the Information Commissioner's Office and Ofcom, published a foresight paper on agentic AI on March 31, 2026. It set out five levels of agent autonomy and placed most deployments at Levels 2 and 3, with Level 5 described as "still largely theoretical."

No organisation inherits trust

Perhaps the most uncomfortable figure for the industry concerned who should run AI-powered shopping and payments at all. According to ACI, 59% of respondents would not trust any organisation to manage it. Banks and other regulated financial payments providers were chosen by 20%. Technology companies and retailers were each chosen by 4%.

Those four answers account for 87% of respondents, leaving 13% whose choices the announcement does not describe, by PPC Land's arithmetic.

Set the two accountability questions side by side and a contradiction appears. A majority (54%) want the AI company that built the agent to pay refunds when it errs. Only 4% would trust a technology company to manage the service in the first place. Respondents, in other words, assign liability to the group they trust least.

The survey's least-trusted categories are also the ones that had shipped the most visible products. OpenAI announced Instant Checkout in ChatGPT on September 29, 2025, built with Stripe on the Agentic Commerce Protocol and initially limited to single-item purchases from US Etsy sellers. Google launched the Universal Commerce Protocol on January 11, 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than 20 companies including Adyen, American Express, Mastercard, Stripe, Visa and Zalando. Mastercard set out Agent Pay at the National Retail Federation conference the same day, requiring agents to register on its network using tokens.

UK respondents were answering largely in the abstract

Most of the agent checkout products in market at the time of fieldwork were not available in the UK. When Google previewed Universal Cart at I/O on May 19, 2026, it said the product would roll out in the US across Search and the Gemini app "this summer," and that checkout through the Universal Commerce Protocol would extend to Canada and Australia in the coming months, then to the UK. OpenAI's Instant Checkout launched as a US-only feature.

That context matters for interpreting the data. ACI's survey measured attitudes, not behaviour, among a population with limited direct experience of agents that pay. The 60% figure describes what people expect they would do after a mistake, not what users of a live service actually did.

Where behaviour has been measured, the signal pointed in a similar direction. Walmart executive vice president Daniel Danker told WIRED in March 2026 that in-chat checkout through ChatGPT converted three times lower than click-out links, even though Walmart had made about 200,000 products available inside the chatbot. Shoppers, it seems, were willing to let the AI find the product but preferred to complete the purchase on the retailer's own site. ACI's 50% trust in price search, set against 15% trust in problem handling, describes the same split from the other end.

Earlier UK surveys pointed in different directions

ACI's numbers sit uneasily next to some earlier UK research. In November 2025, PSE Consulting released a survey of 2,000 UK online shoppers, fielded by OnePoll between November 13 and 17, 2025, which found that 85% of those planning AI-assisted holiday shopping would trust the tool to place orders and pay. The difference in base explains much of the gap: PSE's 85% applied only to people who already intended to shop with AI, while ACI's figures cover a weighted sample of all UK adults. Even so, PSE's respondents named concerns that echo ACI's - data handling (49%), fraud (46%) and receiving the wrong item (41%).

A preference for people over machines is not new either. A survey of 2,000 UK adults commissioned by 8x8 and fielded in February 2025 found that 83% preferred speaking to human agents for customer support, while 4% preferred virtual agents or chatbots. ACI's 55% versus 19% split between human advisers and AI assistants follows the same pattern in a different setting.

Analysts had also questioned the core assumption. In October 2025, shortly after ChatGPT's checkout launch, analyst Andrew Lipsman assigned a 50% probability to shoppers trusting agents with their financial information and 25% to shoppers accepting products without inspecting them. ACI's finding that 17% trust AI to keep payment data secure comes in well below even that sceptical estimate - though the two figures measure different things.

Why the marketing community is watching

For advertisers and retailers, agent checkout is not an abstract payments question. It determines where a purchase is completed, which data the merchant receives and who controls the customer relationship afterwards. If AI agents become the interface where shoppers choose and pay, retail media, search advertising and affiliate models all have to adjust to a buyer that is software acting on a person's behalf.

ACI's data suggest the adoption curve may be steep in places and flat in others. Discovery - comparing prices, surfacing options - appears to be territory consumers are prepared to cede. Checkout, and especially unsupervised checkout, is not. A product that loses 60% of its users after a single error has very little margin for the kinds of mistakes that automated systems make at scale.

Fraud adds a further complication. HUMAN Security's 2026 benchmark report, released on April 9, 2026, found that agentic AI traffic grew 7,851% year over year from a very low 2024 base, with checkout pages accounting for 2.31% of agentic activity and only half a percentage point separating benign from malicious automation in its data. Identity schemes such as Experian's Agent Trust framework, announced on April 30, 2026 with Visa, Cloudflare and Skyfire, aim to bind each agent to a verified human. Whether such machinery reassures the 17% who trust AI with their payment data, or the 59% who trust no one to run the service, is a question ACI's survey raises but cannot answer.

ACI's prescription - explicit approvals, hard spending limits, protected payment details and clear accountability - reads as a checklist of features the industry was already building in June 2026. The survey's contribution is to show how far consumer sentiment lagged behind that infrastructure. It also shows who consumers think should pay when the system fails: not the retailer, not the bank, but the company that built the agent.

Timeline

Summary

Who: ACI Worldwide, a Nasdaq-listed payments software company, commissioned YouGov to survey 2,080 UK adults. Adriana Iordan, ACI's head of merchant and payments intelligence, commented on the findings. The results bear on technology companies, retailers, banks and card networks building AI shopping agents, and on the advertisers and merchants whose sales would pass through them.

What: The survey found that 60% of UK adults would stop using an AI shopping agent after one mistake, 19% trust AI assistants to make everyday purchasing decisions within rules they set (against 55% for a human adviser), 44% would not trust an agent regardless of savings, 54% say the AI company should be accountable for refunds, and 59% would not trust any organisation to manage AI-powered shopping and payments.

When: Fieldwork ran online from June 19 to June 22, 2026. ACI Worldwide published the results on June 29, 2026.

Where: The survey covered the UK adult population, a market where most agent checkout products available at the time, including OpenAI's Instant Checkout and Google's Universal Cart, had launched or were planned for the US first.

Why: Payment networks, AI companies and retailers had spent the preceding nine months building systems that let AI agents complete purchases. The survey indicates that UK consumer trust lagged well behind that infrastructure, with respondents accepting AI for price comparison but resisting autonomous purchases, bank account access and data tracking, and placing responsibility for errors on the companies that build the agents.