Voicebots in Customer Service: What Goes Missing After the Call

You call your bank, and a friendly voice introduces itself as an AI, quite correctly, as EU law has required since the summer. Ten seconds later the same voice says, very warmly and after a small pause: “I’m sorry to hear that.” Between those two sentences something gets lost that shows up in no customer service metric, and least of all when those metrics look excellent.

In February 2024 Klarna announced something every head of customer service dreams of. In its first month an AI assistant had handled 2.3 million conversations, two thirds of all service chats, an errand took under two minutes instead of eleven, and customer satisfaction was on par with human agents. A little over a year later Klarna’s CEO Sebastian Siemiatkowski told Bloomberg something quite different, namely that cost had unfortunately been too predominant an evaluation factor and that what you end up with is lower quality. Klarna still relies on AI, but wants customers always to be able to reach a human, and whatever was missing between those two announcements apparently did not appear in any of the numbers.

Voicebots, that is AI systems that talk to customers on the phone, are currently being built toward two goals: they are meant to get faster, and they are meant to sound more human. Both goals are reasonable, and both carry a price that only falls due after the customer hangs up.

The dashboard only sees one end of the line

A voicebot is built for the dashboard, the overview of key figures that shows minutes per call, the share of cases resolved without a human, and the cost per contact. That dashboard is the protected room in which such a system is designed and signed off. The behavior that actually matters, however, happens somewhere else, namely after the call, in the customer’s head.

Ryan Buell and Michael Norton of Harvard Business School studied what happens there in 2011. Their participants searched for flights on two travel sites that returned exactly the same results, one of which answered instantly while the other took 30 or 60 seconds and showed, as it went, which airlines and fares it was searching. Around six in ten chose the slow site, and when the slow site showed nothing but a progress bar for 60 seconds instead of its work, the figure dropped to 23 percent. So it was not the waiting that counted but the visible effort: whoever sees someone working hard on their behalf values the result more, and the two researchers named this effect the labor illusion.

Two kinds of effort

The convenient conclusion at this point would be to simply make the bot slower, and that conclusion would be wrong. In 2010 Matthew Dixon, Karen Freeman and Nicholas Toman analyzed surveys of more than 75,000 customers in the Harvard Business Review. Their finding was clear: what customers mainly want is a quick, easy resolution, and how much effort they themselves had to put in predicts their loyalty better than their satisfaction does.

Both findings fit together once you separate two kinds of effort, namely your own, which the customer wants to get rid of, and the other side’s, which the customer wants to see. A good voicebot takes the customer’s own effort away, which is worth a great deal, except that the effort on the other side disappears from view along with it. The customer hangs up after 40 seconds, the problem is solved, and nobody has audibly gone to any trouble for them.

The company did put in work, it automated processes, connected data and trained a model, only all of that happened months before the call and in a place the customer never gets to see. What stays in the customer’s head is therefore not the work but the impression that nothing much happened for them at all, which is roughly the position of the travel site that shows nothing but a progress bar.

Three ways to keep the effort audible

The bot should keep taking the customer’s own effort away, so you build it in a way that keeps the other side’s effort audible, and I see three ways of doing that. On the first, the bot narrates its work instead of just reporting “done”: “I’m looking at your last three invoices. Here’s the error, in March you were charged twice, I’m refunding that now.” The customer has to do nothing for it, and the call takes barely longer, yet they hear that something is happening for them right now, which of course only works if it is true.

On the second, the customer receives an effort receipt after the call, a receipt not for their money but for the work, and that message states what was checked, what was found and what was changed. The effort stays visible long after the call has ended, and it sits exactly where the customer will find it later, when they ask themselves whether they can trust the outcome.

On the third, the company gives back part of the time it saved, and does so in the right place. At Klarna that was around nine minutes per errand. Part of it could be spent on a human calling briefly after a difficult case to say: “I’ve looked at your case, is everything all right now?” That call becomes more valuable the better the bots get, because a gesture only works if it costs something, as I described in the article on costly signaling, and a person’s time costs something. Human effort does not disappear, then, it becomes scarce.

Sounding more human, and what it costs

The second goal is called more human, and it is worth starting with a figure that is ageing fast. In 2019 Xueming Luo and colleagues showed in the journal Marketing Science, with 6,000 customers of a financial services firm, that the purchase rate fell by 79.7 percent when customers learned before buying that a bot was talking to them. That figure is seven years old, though, and I believe it is ageing quickly, because people get used to things. According to a Visa survey of US consumers, 72 percent have already used an AI assistant, and a bot selling something, or buying something on someone’s behalf, will therefore probably just become normal.

What I find more interesting is something else. Since 2 August 2026, Article 50 of the EU AI Act has required that people be told when they are talking to an AI, unless that is obvious anyway. So the bot says it at the start, and afterwards it often sounds like a person, with small pauses and a warm “I’m sorry to hear that”. The customer’s head then knows it is talking to a machine while their gut hears a human, and this small discrepancy costs twice.

First, the warmth loses its value, because sympathy you know comes from a machine no longer lands as sympathy. Yidan Yin, Nan Jia and Cheryl Wakslak of the University of Southern California showed in 2024 in PNAS that people felt less heard as soon as they believed a response came from an AI, even though the AI responses on their own were actually received better than the human ones. The label was assigned independently of where the response really came from, so the bot’s “I’m sorry to hear that” is devalued before it is even spoken. The researchers suspect, incidentally, that this could change as people grow more familiar with AI, though perhaps it won’t.

Second, the machine loses its own advantage, because there are moments when people are glad nobody is watching, and when they would rather nothing sounded especially warm. Jana Holthöwer and Jenny van Doorn of the University of Groningen studied such a moment in 2022. People turned away at a restaurant because they themselves had forgotten to book were more likely to accept an alternative offer of another restaurant from the same owners when it came from a robot rather than from a human. When the restaurant was simply full, with no mistake of their own involved, that advantage disappeared, and the title of the study sums up the mechanism in four words: Robots do not judge.

There is a related effect with opposing views. In 2025 Louise Lu, Zakary Tormala and Adam Duhachek of Stanford showed, with 2,061 participants, that people listen more readily to a counterargument when it comes from an AI, because they attribute less intent to persuade to an AI.

The Groningen researchers also tested a very human-looking robot, however, and in an embarrassing pharmacy scenario participants felt more judged by it than by a robot that clearly looked like a machine. What they tested was appearance, not voice, but my guess is that a human-sounding voice does the same, even when the label came first. Back in 2000, Clifford Nass and Youngme Moon had already summarized that people respond socially to computers even though they know they are computers.

The label and the voice have to say the same thing

Imagine you have to tell your bank that you forgot your installment and cannot pay it this month either, and ask yourself honestly which voice would make that easier: a warm one that sounds like a nice colleague, or a sober one that is clearly a machine.

My reading of this system is this: faster costs the visible effort, and more human costs the moment when nobody is watching. With an AI label and a warm voice, the customer ends up with neither the sympathy of a human nor the calm of a machine, and that is why the label and the voice have to say the same thing.

When the call is about an overdue installment, a cancellation or a mistake of one’s own, the machine may sound like a machine and say so openly: “I’m an automated system, just tell me what happened.” When someone needs sympathy, the call goes to a human, because only a human can give it credibly, and for routine matters the bot may well sound friendly, because there nobody has anything to hide and nobody needs consoling.

Who is in the room

For heads of customer service this means that a voicebot is never only a question of speed or of the tone of a voice. Whoever builds a voicebot also decides who is in the room when the customer calls: a human who is audibly making an effort, a machine that does not judge, or a machine that sounds like a human and therefore delivers neither.

What remains open is a question every customer service team can only answer for its own calls: on which call would your customer be glad that nobody is watching, and on which would they be glad that someone takes more time?

The idea is also available to watch, in my video on voicebots in customer service (in German) from the format “Wenn keiner hinsieht”.