Only 20% of UK Adults Want Chatbots - Citizens Advice Calls for a ‘Right to Talk to a Human’

Citizens Advice has called on UK essential service providers — energy, banking, phone and internet firms — to guarantee customers a “right to talk to a human”. The charity, which gave one–to–one help to more than 2.7 million people in England and Wales last year, found that only 20% of UK adults polled in July 2026 considered chatbots their preferred way to communicate with those providers. Phone calls, email and instant messaging were all preferred instead. Almost half of the people who had to use a chatbot found it unhelpful, and some concluded the system seemed rigged to “make me give up”.

That last phrase is the one worth sitting with, because it describes a design outcome rather than a technical fault. Tom MacInnes, the charity’s policy director, put it directly: “Going round in circles with a chatbot isn’t just annoying, it can be a blockade to essential services.” The charity says overreliance on chatbots and online forms has left people unable to fix billing problems, sort out debt, or access homelessness prevention services. Most UK SMEs are not energy suppliers or banks, and will read this as a story about somebody else. It is not. The mechanism Citizens Advice has identified — a deflection system that looks successful on a dashboard while quietly costing you the customers it deflects — is available to any business that deploys AI support without designing the exit route first. This piece looks at what the research found, why the measurement creates the harm, and how to deploy AI in customer service without building a blockade.

What the research actually found

The headline finding is a preference gap. Only 20% of UK adults polled on behalf of Citizens Advice in July 2026 named chatbots as their preferred channel for dealing with essential service providers. That is not a rejection of digital — email and instant messaging both ranked higher, and both are digital. It is a rejection of a specific interaction pattern: the scripted, unescalatable, intent–matching bot that stands between a customer and a resolution.

The harm figures are more pointed, and they apply specifically to people who found it difficult or impossible to reach a human to resolve an issue. Among that group, 49% felt stress or frustration, 27% felt powerless, 44% experienced delays to resolution, and 14% gave up trying to resolve the issue altogether. Read those in order and they describe an escalating sequence rather than four separate complaints: irritation, then loss of agency, then material delay, then abandonment. The final figure is the one with a commercial reading, and we will return to it.

The context that makes this a public policy question rather than a customer satisfaction question is exclusion. Citizens Advice puts the number of UK citizens affected by digital exclusion at at least 14 million, and around 1.7 million UK households have no laptop, tablet or desktop computer. For that population the chatbot is not a slower route to resolution; it is frequently no route at all. The charity’s energy billing example makes the disparity concrete: submitting a meter reading to correct a wrong bill can take under an hour online, and as long as six weeks through offline channels when digital access is poor. Same task, same company, same outcome required — a difference of roughly three orders of magnitude in elapsed time, determined entirely by whether the customer has a working device and the confidence to use it.

The government’s response acknowledged an existing duty rather than proposing a new one. A spokesperson noted that businesses have “a legal obligation to provide contact information in a manner which is clear and easily accessible”, and pointed to an £11.9m digital inclusion innovation fund. That obligation is worth reading carefully by anyone who has recently redesigned a contact page, because the phrase is about accessibility of the contact route itself — not about whether a support channel exists somewhere behind three clicks and a chatbot that will not hand over.

How the chatbot became the front door

The through line is that nobody set out to build a blockade. Each individual decision — add a bot to handle simple queries, route the rest, measure how many contacts avoid the queue, improve that number quarter on quarter — is defensible on its own. The blockade is emergent. It is what you get when a deflection metric is optimised for several cycles by people who never see the customers it deflects, and it is the reason this story is worth the attention of businesses far smaller than the ones Citizens Advice is addressing.

Where AI support genuinely works, and where it does not

The useful question is not whether to use AI in customer service but which contacts to point it at. The chart below is an indicative planning model — not survey data — ranking how well automated chat typically serves different kinds of enquiry.

The shape of that distribution explains the research findings almost entirely. Automated chat performs well where the customer’s need is a lookup — a fact exists, the system holds it, the job is to retrieve and present it. It performs poorly where the need is a judgement: something has gone wrong, the situation does not match a defined case, and resolving it requires someone with the authority to make an exception. Citizens Advice’s examples cluster precisely at the bottom of the chart. Billing problems, debt, homelessness prevention — these are not lookups. They are the cases where the standard path has already failed, which is definitionally the category a system built on standard paths handles worst.

The commercial trap follows directly. The contacts at the bottom of that chart are also, generally, the highest–value and highest–risk ones: a billing dispute is a customer considering leaving, a complaint is a regulatory event in waiting, a hardship case is a person in real difficulty. A deflection strategy applied uniformly across all contact types will therefore deflect hardest exactly where deflection costs most, because those conversations are the longest and most expensive to handle and thus the most attractive to automate. The measurement points the system at the worst possible target.

The number that should settle the channel debate

Most internal arguments about customer service channels are conducted on assumptions about what customers want. On this particular question there is now a figure, and it is not ambiguous.

The temptation is to read that as evidence that customers are behind the technology and will catch up. The composition of the remaining 80% argues against it. If the preference had gone to the telephone alone, the “they will adapt” reading would be reasonable — a generational attachment to a familiar channel, eroding over time. But email and instant messaging also outranked chatbots, and both are asynchronous digital text channels. Customers are not expressing a preference for analogue. They are expressing a preference for text conversations in which there is a person at the other end, which is a judgement about the interaction rather than about the medium.

That distinction is the most useful thing a business can take from this research, because it points at a solution that is neither expensive nor regressive. The problem is not that customers were asked to type. It is that they were asked to type at something that could not help them and would not pass them on. An AI assistant that answers confidently where it can, and hands over cleanly and immediately where it cannot, sits in the preferred category rather than the rejected one — and it does so without abandoning automation or adding headcount for the contacts it genuinely resolves.

It is worth adding the honest caveat. This was a poll about stated preferences for dealing with essential services — energy, banking, phone and internet — where contacts are disproportionately about problems rather than purchases. A poll about pre–sales enquiries on a retail website would very likely produce different numbers, and a business should not assume the 20% transfers directly to its own context. What transfers is the mechanism: satisfaction tracks whether the customer’s actual problem was solvable through the channel they were given.

Where SME AI support deployments go wrong

The failures Citizens Advice describes occur at large providers, but the design errors behind them are not a function of size. The rows below reflect where they typically appear in a smaller organisation, with badges indicating how much attention each usually needs.

The first two rows are the whole story in miniature, and they interact. An escalation route that depends on the assistant recognising its own limits fails precisely in the cases where the assistant has misunderstood — which is to say, in every case where escalation was most needed. Meanwhile a deflection metric records that failure as a success, because the contact did not reach an agent. A business can therefore run this configuration for a year, watch the numbers improve every quarter, and be steadily accumulating the 14% who gave up. Nothing in the reporting line will show it.

The seventh row belongs to a different discipline but deserves a place here, because it is the one that turns a customer service decision into a data protection one. If your AI support tool is a third–party service, customer messages — which in a support context routinely contain names, addresses, account numbers, and sometimes health or financial circumstances — are being processed by that provider. That engages the ordinary UK GDPR questions about lawful basis, processor terms, retention, and whether the content is used for model training. These are answerable questions with standard answers, but they need asking before the tool is live rather than after a subject access request arrives.

What getting this right costs

The bands below are indicative planning figures for UK businesses deploying or repairing an AI customer service capability — not quotes. The variable that moves them most is the complexity of the systems the assistant needs to read from, not the number of customers.

The second row is deliberately cheap and it is where most businesses reading this actually sit. The failure Citizens Advice describes is not usually a failure of model quality requiring significant investment to fix. It is a missing button, a removed phone number, and a metric pointing the wrong way. Those three things can generally be put right in under a week, and doing so converts the assistant from something customers suspect of trapping them into something they simply use or bypass as they see fit.

Two ways to deploy an AI assistant

What separates those columns is almost entirely design intent rather than technology — the same underlying assistant can sit in either. That is worth emphasising because the instinctive response to this research is to conclude that the AI was not good enough and that a better model would have prevented it. A better model would have resolved a few more contacts at the top of the suitability chart and changed nothing at the bottom, because the bottom of that chart is not a comprehension problem. It is a category of situation that requires someone empowered to depart from the standard case, and no assistant has that authority by design.

The obligations behind the headline

There is no new regulation here, and businesses should not respond as though there were. What exists is a restatement of something already true: the government’s comment that businesses have “a legal obligation to provide contact information in a manner which is clear and easily accessible”. That phrasing is worth parsing, because it addresses the accessibility of the contact route itself rather than the existence of a support function somewhere. A contact page that offers only a chat widget, with no address or telephone number reachable without engaging it, is in tension with the spirit of that obligation even where it satisfies the letter.

Accessibility is the sharper issue, and it applies to businesses of every size. A chat–only support route can present a genuine barrier for customers with certain disabilities — cognitive, visual, motor or otherwise — and UK equality law expects service providers to make reasonable adjustments so that disabled customers are not placed at a substantial disadvantage. We would not attempt to state where any particular arrangement falls, and a business with a single automated contact route should take proper advice rather than rely on a news article. But the practical point requires no legal opinion: if your only support channel excludes some of your customers, you have a problem that is simultaneously commercial, ethical and potentially legal, and maintaining an alternative route resolves all three at once.

For firms in regulated sectors the bar is higher again. Financial services firms operate under obligations to deliver good outcomes for retail customers and to give particular care to those in vulnerable circumstances; an automated front door that makes it hard to raise a complaint or disclose hardship sits badly against that. Energy and telecoms providers carry their own consumer protection requirements. Most readers of this article are not in those sectors — but many supply businesses that are, and a support model that would embarrass a regulated client is increasingly the sort of thing that surfaces in a procurement questionnaire.

Finally, the data protection dimension deserves a mention that it rarely gets in coverage of this topic. Support conversations are unusually rich in personal data — identifiers, account details, and frequently information about someone’s financial or health circumstances offered voluntarily in the course of explaining a problem. Where an AI assistant is provided by a third party, that content is being processed by that party, and the standard UK GDPR questions apply: what is the lawful basis, what do the processor terms say, how long is it retained, and is it used to train models. These have ordinary answers. They just need to be established before the tool is live.

The story at a glance

This story connects to several we have covered recently, and the common thread is what happens when an organisation stops being reachable. The BT email password reset flood is the closest parallel: customers with a real, urgent problem, front–line advice to ignore it, and no effective escalation route into a very large supplier. Cambium Networks entering administration makes the same point from the supplier side — support capacity halved at the moment customers most need it. The reporting on UK police data held on Microsoft Azure is about a different kind of unexamined dependency, accepted once and never revisited. The Gemini autonomous hacking research and the NCSC’s adversary simulation work sits on the other side of the same AI adoption question: capability advancing faster than the governance around it. And the PSTN switch–off is a reminder that the phone line many customers still expect to use is itself being rebuilt underneath them.

Frequently asked questions

CloudSwitched

London-based managed IT services provider offering support, cloud solutions and cybersecurity for SMEs.