If you bought a customer service chatbot in the last few years, you may have quietly wondered why usage never really took off. It sits on the site, handles a slice of the easy questions, and the contact volume it was meant to reduce has not moved much.

There is now research that explains it, and the reason is not that customers rejected automated help.

A Gartner survey of 3,566 B2B and B2C customers, fielded in early 2026, found that customers are approximately three times more likely to use third-party generative AI tools than company-provided chatbots when resolving a customer service issue. Use of those tools has nearly doubled in a year. Company-provided chatbot use has not statistically increased since 2022.

Customers went and got automated help somewhere else.

This article covers what those findings do and do not say, why customers behave this way, and what it changes for a service team, which is less about buying different software than it looks.

What the data actually says

Two findings are worth separating, because they get merged constantly and they are not the same claim.

The first is about channel preference. Three times more likely to use third-party GenAI than your own chatbot. Same survey also found 58% of people using those tools have asked one to complete a task on their behalf, rising to 74% among B2B buyers. That last number matters if you sell to businesses: your buyers are not just asking questions, they are delegating.

The second is about where journeys begin. A separate Gartner survey of 5,801 customers, fielded in early 2025, found that 51% of customer service journeys now start on third-party platforms such as Google, YouTube and ChatGPT.

What it does not say

Three corrections, because the sloppy versions of these findings are everywhere.

It is not 51% starting on AI. Gartner's own wording is explicit that search engines are the most popular starting point. A large share of that 51% is people typing into Google, exactly as they have for twenty years. AI is one growing slice of a third-party category, not the category itself.

It is service journeys, not customer journeys. The version that circulates drops the word "service" and expands the claim to cover buying and discovery. That is a different and much bigger assertion, and this survey does not support it.

The two figures are not additive and they are not the same study. One is a channel-preference comparison. The other is about journey origin. Anyone presenting them as a single trend line is guessing.

Neither press release states the geography of the panel, so treat both as global rather than assuming they describe UK behaviour.

Why customers go elsewhere

Nothing about this is mysterious once you look at it from the customer's side.

A third-party assistant has no login. It has no menu of six options that do not include your problem. It does not queue you, it does not ask for your account number before it will speak to you, and it will happily compare four vendors in one answer while your chatbot is still asking whether you would like to check your order status.

It is also often just better. A well-built company chatbot answers questions its designers anticipated. A general assistant will have a reasonable go at almost anything, which for a customer with an unusual question is the difference between an answer and a dead end.

That is the uncomfortable part. The gap is not really a technology gap. It is that customers found something with fewer barriers.

What this changes for service teams

The instinct is to conclude that the chatbot was the wrong purchase. That is mostly the wrong lesson.

The more useful conclusion is that your service content is now an input to systems you do not control. When someone asks an assistant about your returns policy, your delivery times or whether your product does a particular thing, the assistant answers from whatever it can find. If your help content is thin, out of date or contradicts itself across three pages, that is what gets synthesised and presented to a customer as fact, in a conversation you cannot see.

You have always had this exposure through search. The difference is that search showed people your page and let them judge it. An assistant reads your page and speaks on your behalf.

Which makes content quality a service problem rather than a marketing one. We have written separately about why AI answers are only as good as the content behind them, and that argument applies whether the AI in question is yours or somebody else's.

Buyers arrive with the specification already written

We can see something adjacent in our own search data.

Queries reaching our site increasingly read like requirements documents rather than searches. One asked for a WhatsApp CRM with AI automation and an omnichannel team inbox, in 2026. Another asked for tools for identifying customer support quality gaps. A third did not search for a category at all, but described a situation: enquiries arriving on WhatsApp and TikTok, agents working independently, current tools not handling the API properly, and a question about whether a shared inbox would help.

That last one is not a search term. It is somebody explaining their problem to something.

What we cannot tell you is whether they wrote that themselves or an assistant helped them write it. Search Console does not record where a question came from. So we can say buyers are arriving with fully-formed specifications, and we can say the phrasing has changed. We cannot honestly claim to have measured AI-mediated discovery, and neither can anyone else showing you a similar chart.

Questions worth asking, since nobody has clean answers yet

This is a genuinely new problem and the honest position is that the tooling has not caught up. These are the questions we would be asking, rather than a checklist we can pretend is settled practice.

• If a customer asked an assistant about your refund policy today, what would it say? Have you tried?

• Where does your policy content actually live, and is there more than one version of it?

• Who owns updating it when a policy changes, and how long does that take?

• If an assistant described your product incorrectly, how would you find out?

• Does your help content answer questions, or does it restate policy in language nobody searches for?

The first one takes about four minutes and tends to be the most instructive thing a service team does that week.

What not to conclude from this

AI has not replaced search. The same Gartner research that produced the 51% figure says search engines remain the most popular starting point, and there is no evidence in either survey that owned channels have stopped mattering.

Chatbots are not pointless either. A well-scoped one handles high-volume repetitive questions cheaply and predictably, and the distinction between a chatbot, conversational AI and an AI agent is worth understanding before deciding what you actually need. We cover that in chatbot vs conversational AI vs AI agent.

What has changed is narrower and more specific than "everything is different now". A meaningful share of customers now form their view of you through an intermediary, using your content but not your interface. That is worth taking seriously without redesigning your entire service operation around a single survey.

Start with what the assistant would say

The cheapest useful thing you can do this week is ask an assistant about your own refund policy, delivery times or product limits, and read the answer as a customer would.

If it is wrong, the fix is your content rather than your chatbot. If it is right, you have a clearer view of what customers are being told than most of your competitors.