Your customers and your chatbots: it's complicated

Observations

In 2025, the Observatoire des Services Clients run by Ipsos bva asked 5,000 European adults a simple question: would you rather wait for a human agent, or speak to a virtual one right away? In France, 90% chose to wait. And 64% turned down even a human agent assisted by AI, despite being promised faster handling in exchange.

You could stop there and conclude that people do not want artificial intelligence anywhere near customer service. The same survey shows that the chatbot has become the second most used channel, just behind the phone. It also comes last on satisfaction: 50% of users satisfied, against 72% for the phone and 68% for email.

Both findings hold at once because they measure different things. One records a preference, the other records behaviour that is often constrained. A channel can be heavily used because it is the only one open at 10pm, or because the phone number has been buried three levels deep in the help centre.

Usage tells you nothing about buy-in

This is the first misreading to avoid when you run an automated service. A chatbot handling 60% of inbound contacts might signal good design, or it might signal a funnel with no exit. Both situations produce the same number on the dashboard.

The Pega/YouGov survey run in late 2025 among 4,748 adults in the US and the UK gives a sense of the gap. 46% say they rarely or never get a satisfactory outcome when they go through AI. 66% prefer human-led support. Only 2% want an interaction handled entirely by generative AI. The study was commissioned by a software vendor, which calls for some caution, though the order of magnitude matches the independent sources.

What actually drives acceptance

The most complete meta-analysis available on the subject pools 561 correlations from 56 studies, covering 25,511 respondents. It produces a ranking of what genuinely pushes someone to use a chatbot: trust and expected performance come out clearly on top, ahead of enjoyment, social influence, and finally ease of use.

That last point deserves a pause, because it cuts against how design effort tends to get allocated. Teams invest heavily in interface fluidity, response times, the styling of the chat bubble. Those things matter. The question the customer asks first is still whether this thing will fix their problem. A flawless interface that ends in "I didn't understand your request" loses on the only criterion that counts.

The stakes of the task change everything

The Qualtrics XM Institute surveyed 28,400 consumers across 26 countries. 73% are comfortable using a chatbot to check the status of an order. 81% want a human as soon as medical advice is involved.

The dividing line does not sit exactly between simple and complex tasks in the technical sense. It sits at the cost of an error for the customer. Tracking a parcel costs little if the bot gets it wrong, since you can simply try again. Disputing a payment, filing an insurance claim or asking for more time to pay puts the customer in a vulnerable position, and uncertainty becomes much harder to tolerate.

That suggests a different way of mapping customer intents. Ranking requests by volume pushes you mechanically towards automating the most frequent ones. Ranking them by consequence of failure gives a better shortlist.

The gatekeeper reflex

A recent set of experiments by Kagan, Hathaway and Dada documents something most service leaders will recognise immediately: users avoid the chatbot even when it should objectively cut their waiting time, as soon as they suspect it is only a first step before a human.

The reasoning is sound. If I expect the bot to ask me six questions before handing me to an agent I will have to brief from scratch, then waiting twelve minutes on the phone becomes the better option. The effect grows as the stakes of the request go up.

Gartner's surveys tell the same story from another angle. In 2024, among 5,728 customers, 64% would have preferred companies not to use AI in customer service at all, and 53% said they would consider switching provider over it. The main fear they cited was no longer being able to reach a person. In 2026, on a fresh sample of 3,566 customers, half now find interactions easier when the company uses generative AI. Yet 87% still consider access to a human essential.

Opinion on technical capability is improving. The demand for a way out has not moved.

Making the bot more human is not a universal fix

Anthropomorphism has become a design reflex: a first name, an avatar, a warm register, empathetic phrasing. Some studies do show that these cues increase perceived social presence and user cooperation.

Crolic and colleagues found the opposite effect under specific conditions. Their work combines controlled experiments with an analysis of 461,689 real sessions at a telecoms operator. When the customer arrives angry, humanising the bot degrades satisfaction rather than improving it. The measured effect on satisfaction at high anger levels is negative and statistically significant.

Their explanation fits the wider literature: human cues raise expectations. A system that presents itself as an attentive interlocutor then gets judged against that promise. When it cannot deliver, the disappointment runs deeper than if the exchange had stayed openly functional.

In practice, this argues for a persona that shifts with context. A delivery update can carry some lightness. A complaint following three failed attempts calls for restraint, precision, and as few stock phrases as possible.

The transparency paradox

72% of French respondents in the Ipsos survey said they would feel misled, or would find the practice unacceptable, if a company failed to tell them upfront that they were talking to an AI. The normative expectation is overwhelming, and European regulation points the same way.

A field experiment by Luo and colleagues, run with more than 6,200 consumers, complicates the picture. In structured outbound sales calls, a chatbot that was not identified as such performed on par with experienced human salespeople. As soon as its artificial identity was disclosed at the start of the call, purchase rates dropped by nearly 80%, mainly because customers immediately credited it with less competence and less empathy.

The lesson is obviously not to hide what the system is. The useful observation is that the "AI" label on its own triggers an unfavourable heuristic, so it needs something alongside it to push back. That means separating several forms of transparency that tend to get treated as one:

• identity: you are talking to an automated assistant;
• capability: here is what I can handle, here is what I cannot;
• process: here is what happens next, and how long it will take;
• data: here is what I can see and what I record.

Disclosing identity without disclosing capability leaves the customer holding the stereotype and nothing to correct it with. The same research on chatbot avoidance shows that communicating wait times, limits and the route to a human improves adoption of the channel.

Three modes instead of one dial

What the evidence supports is dropping the question "how much of our contact volume can we automate?" in favour of a narrower one: for which intents, under which conditions, and above which confidence threshold does automation actually give the customer something?

That produces three operating modes.

Full autonomy when the intent is clearly identified, the risk is low and model confidence is high. Order tracking, opening hours, rescheduling an appointment, reissuing an invoice.

Assisted when AI drafts, summarises or proposes, and an agent validates. This is often the most profitable mode and the least visible one, because it works on handling time rather than contact volume.

Human first as soon as there is financial or medical exposure, ambiguity, vulnerability, anger, or simply a second failure on the same request.

In all three, escalation to a human works better as a feature of the system than as its failure state. It should be visible before the customer gets frustrated, it should carry over the context already given, and the bot itself should be able to trigger it when it detects that it is out of scope. A chatbot that can say "I'm not going to solve this, let me get someone" protects the relationship far better than one that rephrases the same question a fourth time.

Measure something other than containment

Containment remains the most closely watched metric in most organisations. On its own it pushes in the wrong direction, since it rises whether the bot resolves the issue or the customer gives up.In short

To sum up

Customers are not asking for a human at every step. They are asking for a reliable outcome, minimal effort, and the certainty that they can take back control when they need to. The three travel together: once the first is dependable, the other two stop being sore points.

The companies getting good results from AI in customer service are the ones that placed it where it is genuinely competent, and kept the route to a human fast and legible. The ones using it to filter contact end up with flattering automation rates, falling satisfaction scores, and a slow erosion of trust in the brand.

Key sources: Observatoire des Services Clients 2025 (Ipsos bva / ESDCA, 5,000 respondents across France, the UK, Germany, Italy and Spain); Gartner customer surveys, 2024 and 2026; Qualtrics XM Institute, 2023; Pega/YouGov, 2026; Rohit and Vishwakarma, meta-analytic structural model of service chatbot adoption; Luo, Tong, Fang and Qu (2019); Crolic, Thomaz, Hadi and Stephen (2022); Kagan, Hathaway and Dada on gatekeeper aversion.