At 21:40 on a Sunday, a customer rings to change the delivery address on an order. An AI agent answers on the first ring, confirms the order, takes the new address and reads it back. The call lasts ninety seconds. The customer thinks it was great.
On Tuesday, the same customer rings the same company about a damaged item. The same AI agent asks them to describe the problem, then asks again. They say “agent”. It offers to help with their order. They say “agent” twice more, louder. Eleven minutes later they reach a person, who asks for the order number they gave at the start.
Now send that customer a survey asking whether they like AI customer service. Their answer depends entirely on which call they think of first.
Customers don’t hate AI customer service. They hate being stuck. Read past the headlines, and every large survey published this year points the same way, including the ones that seem to contradict each other. People choose the AI when it beats the queue and solves the problem, and they turn on it the moment it stands between them and a person. So the questions that matter aren’t about preference. Does your AI resolve the call? And when it doesn’t, how quickly can the customer get past it? You can answer both from your own calls.
Two questions, four contradictory answers
Ask whether customers want AI, and this year’s research says yes and no. Verint’s State of Customer Experience 2026, based on 5,000 surveys of American consumers, found that 61% prefer speaking to a human agent. Ada and NewtonX surveyed 2,000 consumers who had used AI customer service and found that 59% prefer always-on AI to waiting in a queue for a human.
Ask whether it works, and you get yes and no again. Gladly and Wakefield Research surveyed 1,000 Americans and found that 88% got their issue resolved by AI. Ada’s study found that only 24% of recent AI interactions were fully resolved without help from a person.
Gartner manages both answers in a single release. Of the 3,566 customers it surveyed, half said AI makes their interactions easier, and 87% said companies using it must offer a way to reach a person.
Read only the headlines and customers both love and hate AI, which either solves nearly every problem or fails three times in four. That isn’t noise. The surveys asked different questions, and each answer is correct for the question it was given.
Why the numbers disagree
The alternative changes the answer. “Would you rather speak to a person?” and “Would you rather speak to AI now or a person in twenty minutes?” are different questions. The first asks about taste. The second asks about time, and when time is the question, AI often wins. Verint’s own respondents show it: 69% of them said they would switch from human to automated service if it could fully resolve their issue. Their preference for people is real, but it has a price: the wait.
“Resolved” means whatever the survey says it means. Gladly’s 88% and Ada’s 24% are not measuring the same thing. Ada counts only interactions the AI finished alone, so every call handed to a person counts as a miss. Gladly’s published summary doesn’t define the word. And both numbers are what customers reported, not what happened next. A customer who rang off satisfied and rang back on Thursday about the same problem would count as resolved in both.
Almost everyone asking sells something. Verint, Ada and Gladly all sell customer service software, and so do we. Most of them also surveyed Americans, not Nordic customers. That doesn’t make their numbers wrong. It does mean the findings worth trusting are the ones that hold across them, Gartner’s included.
Where the surveys agree
Set the headlines aside, and four findings hold up.
Customers want the problem solved, by whoever is quickest. In Verint’s data, 78% prioritise the fastest resolution over using their preferred channel. Ada’s respondents ranked accuracy and problem-solving well above empathy. Customers aren’t sentimental about channels. They want to get on with their day.
They don’t dislike AI. They dislike yours. In July, Gartner reported that customers were about three times more likely to use a tool like ChatGPT to solve a service problem than the company’s own chatbot. So the objection isn’t to AI. It’s to an AI that stands between them and a refund.
The way out has to be obvious. Gartner’s 87% is the clearest number of the year. Its analysts went further: service leaders “should not use GenAI as a mandatory first step for every issue.” Gartner also found that customers who are pushed through several failed AI attempts before reaching a person are less likely to use the AI again.
Solving the problem isn’t the same as keeping the customer. Gladly’s 88% came with a second number: only 22% said they preferred the company afterwards. A problem can be fixed and the customer still leave annoyed, and Verint found that 79% would switch to a competitor after a single bad experience.
The cost of getting it wrong
In February, Gartner predicted that half of the companies that cut customer service staff because of AI will hire again for similar work by 2027, often under different job titles. The same research found that only 20% of service leaders had actually reduced staff because of AI, so the wave of layoffs was smaller than the headlines suggested. It’s not hard to guess which companies will be hiring again: the ones that used AI as a wall rather than a door.
None of this is an argument against AI agents. It’s an argument against judging them by the wrong number. An AI agent that keeps 80% of calls away from your team looks like a success. If a third of those customers call back, or leave, it wasn’t one.
Three numbers to measure on your own calls
Surveys tell you how customers feel about AI in general. They can’t tell you how your AI is doing with your customers. For that, measure three things on your own calls, human and AI alike.
1. Resolution, judged from the conversation. Not “the call ended” and not “no transfer”. Did the customer leave with the problem solved, as a careful reviewer would judge it from the transcript? Score it with the same criterion for people and AI agents, so the two numbers can be compared. We’ve written about why the usual AI dashboards miss failed calls and how to hold human and AI agents to one quality bar.
2. Repeat contact. The same customer, with the same reason for calling, within seven days. This is the honest check on any resolution figure, because a customer who has to call back wasn’t helped the first time, whatever the survey said. Look at it by contact reason. AI agents rarely fail evenly. They tend to handle a few reasons very well and fail at one or two.
3. Time to a person, once asked. Start the clock when a customer first asks for a person, and stop it when one is on the line or a callback is agreed. Count how many more times the AI tried to help in between. This is the number behind Gartner’s 87%, and it is where trust is lost. Then listen to a handful of the calls where the customer had to ask more than once. They will show you exactly where your AI is trying to do more than it can.
What to change once you have the numbers
The numbers usually point to the same small set of changes.
- Give the AI the calls it resolves well. Out-of-hours calls, simple changes and status questions are where an answer now beats a person later. That’s the trade behind Ada’s 59%.
- Hand off on doubt, not on demand. Gartner’s advice is to let the AI collect the details and understand the request, and try to solve it only when it’s confident. When it isn’t, it passes the call on, with the details it collected, so the customer never has to repeat the order number.
- Make the first request for a person work. One request should be enough. Every attempt the AI makes after that costs more than it saves.
- Say it’s an AI. Since August, the EU AI Act requires it when a customer might not otherwise know. It’s also the decent thing to do, and it sets expectations before the first question.
So, do customers hate AI customer service?
Ask our customer on Sunday night, and they’ll say it was great. Ask them on Tuesday, and they’ll say they hate it. They’re right both times. The question isn’t whether your customers like AI. It’s which of those two calls your AI is making, and how often. A survey can’t tell you that. Your calls can.
We can show you how your AI calls end. We’ll score a sample of your human and AI calls on the same scorecard, so you can see which contact reasons the AI resolves, which ones come back, and where customers asked for a person and didn’t get one. Read more about Conversational Intelligence.
