Summary
- Production observations published on 11 November show AI agents already performing customer account operations and transactions.
- An October UK survey found 84% of respondents would give AI agents moderate, low, or no control over everyday activities.
- Consent, financial liability, and human oversight create separate challenges from authenticating an agent's access to an account.
AI agents are already completing customer transactions, bringing practical urgency to questions about what people have authorised automated systems to do and who is responsible when those systems act beyond their intended role.
Research published by Cequence Security on 11 November identifies agents signing into accounts, making purchases, and managing customer services. Those production observations give fresh operational context to an earlier British survey examining how much authority consumers are prepared to delegate to AI.
The survey, commissioned by Eskenzi PR and conducted by Censuswide among 2,000 UK adults in September, was first publicised in October 2026. It found that 84% would permit moderate, low, or no control over their daily activities, while 34% would grant no control whatsoever.
The November observations do not measure the same population or establish that public attitudes have changed. They demonstrate that some of the account-access and transaction scenarios considered in the earlier survey are already occurring in production applications.
Delegating assistance without surrendering control
Consumers distinguished between asking an AI agent to prepare work and allowing the same system to complete actions independently. Twenty-seven per cent of survey respondents would permit an agent to prepare tasks but wanted to approve every action before completion. Another 23% would allow routine tasks to proceed without repeated approval while reserving consequential decisions for themselves.
Only 11% were comfortable granting AI agents high or very high control. Age was associated with different preferences: 47% of respondents aged over 55 would give no control whatsoever, compared with 21% of those aged 18 to 24.
These are stated preferences, rather than measured use of deployed AI agents. The survey was conducted from 16 to 18 September, and its findings circulated publicly before Cequence released its account of observed automated transactions.
That chronology matters. The new development is the operational evidence that agent-led account actions are happening, not a new UK poll or a change in the recorded percentages.
Data access and financial decisions raise distinct concerns
When asked about the risks of agents using their accounts, 36% of the UK respondents identified personal information being hacked or stolen. Thirty-one per cent feared mistakes, while 30% worried about an agent having excessive access to their information.
Unauthorised spending or purchases concerned 28%, and 22% feared that an agent might communicate with another person without their knowledge. Fourteen per cent worried that an agent could implicate them in illegal activity, rising to 21% among respondents aged 18 to 24.
The figures describe concerns rather than confirmed incidents. They nonetheless identify different points at which an agent’s technical access may exceed the authority a user intended to grant. A consumer can permit an agent to sign in to a service without necessarily authorising a purchase, an account change, or the disclosure of confidential material.
Cequence’s production analysis examined more than 250,000 agent visits across five industries during one week in October, finding that about one in seven involved an account action or transaction. The sample reflects traffic visible to that provider; it is not a measurement of all online services.
Business trust depends on decision boundaries
The earlier UK survey also asked how people would react if a bank, insurer, or utility used an AI agent to advise its board on important decisions. Thirty-seven per cent wanted reassurance that a human would retain the final say, while 31% wanted to know which decisions involved AI. Twenty-four per cent would investigate how personal data was protected.
Twenty per cent would consider switching providers, and 18% said they would be less likely to purchase additional products or services. Those are hypothetical reactions, not observed customer departures.
Yvonne Eskenzi, co-founder of Eskenzi PR, said: “What really stands out is that people aren’t rejecting AI. They’re setting boundaries.”
The boundaries are not confined to identifying an automated client. An agent may pass ordinary authentication checks by using a customer’s credentials, but the receiving application still needs a basis for determining whether each requested operation falls within delegated authority.
Responsibility does not automatically follow public opinion
The survey found that 40% believed the AI developer should be primarily responsible if an agent was hacked, deceived, or acted improperly. A further 25% selected the company whose service the agent accessed, and 18% selected the individual using the agent.
These responses describe public expectations, not the legal allocation of liability for any particular transaction. Contractual terms, consumer protections, authorisation records, and the circumstances of the action may all be relevant.
Jonathan Armstrong, Partner at Punter Southall, warned that “In most cases you’ll be responsible for the AI you use and the consequences – and sometimes even if it took you by surprise.” His comments are general legal commentary rather than a determination of liability under a specific law or contract.
Permission models that distinguish preparing, approving, and executing actions give service providers a way to make delegated authority more explicit. Accurate activity records can also help establish which decisions were requested, which were permitted, and which required human approval.
Eskenzi PR has published its survey findings and methodology, while Cequence’s 11 November research provides the newer account of transaction activity. The emerging challenge is to make those boundaries enforceable when autonomous software acts through ordinary customer accounts.
