Every company rolling out a customer service chatbot is making the same assumption: that a bot which is faster and cheaper will win customers over, even if it is not perfect. Three recent studies suggest that assumption is riskier than it looks, and the reasons why are more specific than "people don't trust AI."
The first study is the clearest. Researchers set up a simple choice: wait in a line where the outcome is certain, or skip the line with an option that sometimes fails and bumps you back into the line anyway. People should have picked the fast option about half the time if they were only thinking about time saved. Instead they picked it barely a quarter of the time, purely because it had a chance of failing and sending them somewhere else. Label that same fast option a chatbot instead of a person, and the number drops even more. In other words, customers punish uncertainty first, and they punish machines on top of that.
The second study flips a common assumption. When the news is bad, a lowball offer or a rejected claim, customers accept it more easily from an AI than from a person. When the news is good, they prefer to hear it from a person. The reasoning makes sense once you see it: people don't assume a machine is being greedy when it gives them a bad deal, so it stings less. That is a genuinely useful fact for any business that has to deliver refund denials, price changes, or eligibility decisions.
The third study is the one worth pinning to a wall. After combining 163 studies and more than 82,000 people, researchers found that customers only prefer AI when two things are both true: the AI is clearly better at the task than a person, and the task doesn't call for a personal touch. Miss either condition and people default back to wanting a human.
This matches what is already playing out publicly. Klarna cut its customer service staff and leaned hard on AI, then walked part of that back after complaints about quality. Independent industry surveys now show most consumers still prefer a real person for support, and that preference has been climbing rather than fading over the past year. One widely cited report found AI-powered customer service disappoints users at a far higher rate than other uses of AI.
None of this means AI in customer service is a dead end. It means the deployment decision is not "can we automate this." It's whether the task is genuinely something AI does better, and whether customers experience it as needing a personal touch. Bad news delivery, routine status checks, and simple lookups are strong fits. Complaints, disputes, and anything a customer feels emotional about are not, no matter how good the model is.