The objection turns up early in almost every rollout, and it usually comes from someone who knows the customers well. The wording barely varies: our customers aren't like everyone else's, they want a person on the other end, and they will hang up on a machine.
The claim is worth taking apart. Not because it is nonsense, but because it is imprecise in a way that leads to bad decisions.
Why the claim feels obviously right
It has three good reasons behind it.
The first is twenty years of phone menus. When someone says “automation on the phone”, that is the picture that appears: the menu with nine options, none of which is your problem, where zero is the only route to a person. That is not a prejudice. It is a description of what most people have actually met.
The second is that the people saying it often have numbers to point to. Complaints about automated phone systems exist, they are logged, and they are real. But they are logged against the system that was there — not against the category it belongs to.
The third is new, and it sharpens the objection. Since 2 August, the EU AI Act's requirement that a voice agent disclose that it is AI has applied. You can no longer try it quietly and see how it goes. The customer is told in the opening line, which turns the objection into a prediction tested in the open from the first call.
That is a real change. But it is about disclosure, not about whether the customer gets helped.
What gets said, and what set it off
Complaints about a phone call are rarely precise. They are quick summaries of an experience, and the summary points at the most visible thing — not necessarily at what went wrong.
| What the customer says | What actually happened |
|---|---|
| “I want to speak to a person” | The customer had already explained the problem once |
| “It's so impersonal” | Whoever answered could do nothing about the case |
| “I hate those bots” | The way forward ran through a menu instead of a question |
| “It didn't work” | A callback was promised and never came |
None of the lines on the right is about who was speaking. They are about whether the problem was solved, and how many attempts it took.
What is actually true in the claim
Two things, and neither should be explained away.
Some calls should have a person regardless of how good the technology gets. Bad news, complaints, anything requiring judgement or the authority to depart from a rule. That is not a transitional phase that passes once the models improve. It is a property of the cases themselves.
And: a customer who is told she is speaking to an AI and then gets nowhere has been given two reasons to be annoyed instead of one. The disclosure requirement raises the price of a bad exit. That is why the route out to a person is the part of the setup that is most expensive to build halfway. An agent that says what it is and can hand over predictably makes the disclosure harmless. Without that route out, the disclosure is just a warning.
The test that settles it
The objection cannot be settled in a meeting room. Both sides argue from their own anecdote, and both remember best whatever confirms it.
It can be settled on one number: what share of customers call back about the same thing within twenty-four hours. That number does not care who answered, and it is hard to dress up. A customer who does not call back got what she came for. One who calls back did not — and then it matters little whether the call was pleasant along the way. It is the same reason the share of calls answered says so little: it counts what was picked up, not what was finished.
Soon the caller may not be a person either
The premise is also shifting under the whole discussion.
Google's own documentation for businesses describes the search engine calling a business on a customer's behalf to confirm price, availability or an open appointment, and sending the answer to the customer by email or text. The feature is US-only for now. But it moves the question: a growing share of inbound volume will have no human preference to accommodate at all.
Rewrite the objection, don't dismiss it
“Our customers want to talk to a person” is a good objection, badly phrased. Rewritten, it becomes two answerable questions: which case types require a person, and what does the customer need to leave with so that she does not call again tomorrow?
Those two questions have concrete answers. The original one has only opinions.
Threll.ai builds voice agents in Norwegian, Swedish and Danish. Almost nobody hangs up thinking it went well because there was a person on the other end. They think it went well because they did not have to call again.
