A customer who has a bad experience with a company's chatbot rarely gives it another go. In a survey published on 2 September 2026, Gartner found that only 27% of customers would try a chatbot again after a negative experience, which means that for most people the first failed conversation is also the last one. That finding is about chat, not voice. It still describes the risk that every automated customer contact programme carries, and it is worth reading carefully before anyone draws a larger conclusion from it.

What Gartner actually found#

The source is a Gartner press release titled "Only 27% of customers would try a chatbot again after a negative experience", dated 2 September 2026. The survey behind it covered 3,566 B2B and B2C customers and was fielded in February and March 2026. Two figures carry the story. The first is the 27%: the share of respondents who said they would try a chatbot again after a bad experience with one. The second is that only 7% of respondents had used a company chatbot in their most recent customer service interaction.

Read together, the two figures say something more pointed than either one alone. Bots are still a small part of how customers actually get service, and the people who have tried one and been let down mostly do not come back. The channel is young and already losing customers it never properly won.

The survey at a glance
ItemWhat the release states
PublisherGartner, press release
Date2 September 2026
Sample3,566 B2B and B2C customers
FieldworkFebruary–March 2026
Headline finding27% would try a chatbot again after a negative experience
Context finding7% used a company chatbot in their last service interaction
TypeCustomer survey, not a forecast

Source: Gartner press release, 2 September 2026. Figures as stated by the publisher; we have not seen the underlying dataset.

What the finding does not say#

It is a survey of stated intentions, not observed behaviour. "Would try again" is what people say they will do; what they actually do when the only open channel is a bot may differ. It is also a study of chatbots in general. It does not separate text from voice, simple FAQ bots from transactional ones, or consumer brands from regulated industries. Nothing in it tells us how a well-designed voice process in a bank or an insurer is received.

The same Gartner survey population of 3,566 customers also produced two other releases this year: one on customers expecting access to a human when a company uses generative AI for service, and one on customers turning to third-party generative AI tools rather than company chatbots. They are the same data seen from different angles, not three independent confirmations. We read them as one survey and count it once.

Why a phone conversation has even less room for error#

If anything, the second-chance problem is sharper on the phone. A chat user who hits a wall can scroll, rephrase or open the website. A person on a call has no screen and no patience reserve. In outbound contact the situation is harder still: the customer did not ask for the conversation at all. Every second of an unwanted call has to earn the next one.

That is why the second rung of the KPI ladder we use is continuation: the inverse of hang-ups after the company name is said. A customer who has had a poor automated experience with your brand, in any channel, arrives at the next call already sceptical. Gartner's 27% gives a rough sense of how few customers extend the benefit of the doubt. In regulated contact the consequence is concrete. A debtor in early delinquency who hangs up is not reached, and a debtor who is not reached drifts into the next bucket.

Clients often assume outbound is simpler than inbound. Someone who calls you will cope even with an IVR. Someone you call may not even want to pick up, let alone keep talking. – Dawid Wójcicki, Founder & CEO of Vee

What we have learned about first conversations#

We have run automated telephone contact since 2012, and in production at scale since 2017. To 31 October 2026 our AI has handled 48,626,959 cases and 38,325,799 connected conversations (cumulative, Vee operating data; a case is a record assigned for handling, not a person or a call). Over that volume the lesson that matches Gartner's finding most closely is that customers judge the conversation, not the technology. A person will not forgive a weak conversation just because it is automated.

As early as March 2020, we said publicly that systems built on a narrow set of commands were not ready for real conversations and might discourage customers. That was a view from listening to calls, not a survey. Gartner's 2026 data is consistent with it: a bad automated experience costs more than the single contact, because it closes the door on the next one.

Four design choices that protect the second chance#

  • Say who is calling. Vee always introduces itself as AI, in line with Article 50 of the AI Act. A customer who discovers the truth halfway through is a customer you lose for good.
  • Design the route to a person before launch. Asking for a human, failing identity verification, withdrawing consent to recording, financial hardship or a legal dispute each has an agreed path: transfer, stop or referral to the client's own staff.
  • Keep regulated replies in a closed, versioned library. The conversation can understand free speech, but what is said back comes from approved content, so a customer never hears an improvised answer about their debt or policy.
  • Listen every day. In every project our people listen to selected calls, especially unusual ones, to hear what the dashboards cannot show.

The Gdańsk Lech Wałęsa Airport assistant we launched in 2017 is an early example of the second point. In 2018, roughly 70% of calls ended with the automated assistant and around 30% were transferred to a consultant (historical, one inbound deployment). The transfer was part of the design, not a failure of it. Gartner's companion finding, that customers want access to a human when a company uses generative AI for service, points the same way.

What it means for regulated enterprise contact#

Banks, insurers, healthcare providers, telecoms and utilities cannot treat a customer as a test subject. The 27% figure turns a technical question into a governance one: who decides when an automated process is good enough to meet real customers, and on what evidence?

Our answer is to start on a sample, never on the whole database. A controlled pilot exposes a limited part of the base, lets the team listen, compares results with a comparator on the same buckets and portfolios, and fixes what does not work before scale. Anything on our side that needs correcting is corrected straight away, and in regulated set-ups every content change is versioned and approved. That discipline matters more when, as the survey suggests, most customers who are let down do not come back.

There is also a measurement point. Containment, meaning the share of conversations that never reach a person, looks good on a dashboard and can hide exactly the failures Gartner describes. We measure reach, continuation, obtaining the key information and the business effect, in that order. A conversation that ends without a transfer but also without a result is not a success, and a customer who will not pick up next time is a cost that appears weeks later.

Where the finding might mislead#

Two over-readings are tempting and both are wrong. The first is "customers hate bots, so avoid automation." The survey does not say that. It says that a negative experience is rarely forgiven, which is an argument for quality and for a designed exit, not for keeping every call with a human agent. The second is "voice is different, so this does not apply." The channel differs, but the mechanism of lost trust is the same, and on the phone it can show up as unanswered calls rather than complaints.

Nor should anyone treat 27% as a benchmark for return rates in their own process. It is a stated intention across a mixed population of customers. Your own data on repeat contact, hang-ups after the greeting and transfers requested in the first seconds will tell you far more about your customers than a global average.

Questions to ask a supplier#

  1. How does the conversation disclose that it is AI, and at what point?
  2. Which situations trigger a transfer, a stop or a referral, and who agreed them on our side?
  3. Will the launch start on a sample, with a comparator on the same buckets and portfolios?
  4. Which metric shows a customer who did not hang up but also got nothing done?
  5. Who listens to calls after launch, how often, and how quickly is a confirmed problem fixed in every live conversation?

Gartner's survey is a reminder that automated contact is judged once. Our view, after more than a decade on the phone, is that the first conversation should be designed as if there will not be a second. That is how we run the service we provide.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS