Most voice AI projects don't fail because the technology isn't good enough. They fail earlier - often in the first serious management meeting, when someone asks a question that sounds perfectly rational. In Central and Eastern Europe, where many companies still run support on layered systems, local exceptions and informal workarounds, that question is especially dangerous. A CFO put it to me a few weeks ago, in a meeting about introducing a voice agent into customer support. Not in our company - in the client's. The head of support was at the table, along with two people from IT.
For half an hour, the discussion had been exactly where it needed to be. What should the voice agent handle? Which systems should it access? Where can a call be resolved automatically? When should a human step in? Where are the risks? Where is the business case? Then the CFO leaned in, and his tone changed. "So tell me straight, Igor. How many people in the call centre can we cut with this?"
It was a fair question. An expected one, too. Any serious leadership team has to care about cost, capacity and productivity. But as the starting point for a voice AI project, it is dangerous. Not because saving money is wrong, but because this question almost always pushes the project towards the wrong design.
Start by asking how many people you can remove, and you will probably build a cheaper service. It may even look excellent in a spreadsheet. But customers will sense very quickly whether the voice agent was built to help them - or to keep them away from someone who can. And in customer support, an obstacle is a mortal sin.
Let's say what often stays unsaid in steering committee meetings: most AI projects in customer support begin as cost-saving projects. Not service redesign projects. Not customer experience projects. Not "let's make support better" projects. Cost-saving projects. That is why the CFO's question did not surprise me. It reflects the pressure many support leaders are under. How many calls can we deflect? How much work can we shift away from agents? How many more interactions can we handle without hiring? How much demand can we absorb without increasing headcount?
The problem is that customers can feel this. They do not see your board deck. They do not know who sponsored the initiative internally. They have no idea whether the project came from finance, operations, IT or customer experience. But they know when a bot is standing between them and a solution. That is where many voice AI projects lose trust before the technology even gets a fair chance. The speech recognition may be good. The integrations may be stable. The response time may be impressive. Technically, everything may work. But if the customer feels the bot was built to shield your budget from their problem, the experience is already broken.
Klarna is a useful cautionary tale. In February 2024, the company announced that its AI assistant was doing the work of 700 agents. A year later, reports emerged that Klarna was bringing human agents back into customer support because of service quality concerns. The company publicly emphasised making sure customers could always reach a person when they needed one. The bot may have worked technically, but the more complex cases exposed the weakness of generic answers. The technology was not the whole story. The starting question was.
I used to think people resisted bots because they resisted technology. I no longer believe that. The more uncomfortable truth, especially for those of us who sell and implement this technology, is that people do not reject AI because it is AI. They reject the feeling that nobody is listening. A few weeks ago, I called a delivery company. The bot told me three times in a row to "please state your issue in one sentence". On the fourth attempt, I shouted "agent" and waited nine minutes. That was not a technology problem in the narrow sense. The system could hear me. It just could not deal with me.
Nobody complains about automation when it quietly solves the problem. Booking a table, confirming an appointment, checking order status, changing a delivery time — these are exactly the kinds of interactions where automation can feel faster, cleaner and less painful than waiting for a human agent. In those moments, customers are not writing essays about the loss of human touch. They just want the job done. The trouble starts when the issue is urgent, emotional, ambiguous or unusual. You explain your situation. The system asks you to repeat it. You choose from a menu that does not include your case. You say "operator" three times. Eventually you reach a human being, who asks you to start from the beginning. At that point, the customer is no longer judging your bot. They are judging your company.
SurveyMonkey’s US research reports that 79% of Americans strongly prefer dealing with a human rather than an AI agent in customer support. YouGov found that in the UK, only 1% of people chose a chatbot as their favourite support channel, even though 18% use one. That gap between "I use it" and "I want it" tells us a lot: we keep pushing people into channels they would rarely choose for themselves. But here is the twist. EY UK research found that 74% of people in the UK have used AI in the last six months. So this is not simply fear of machines. The same person who happily uses AI at home can still spend the morning swearing at their bank's bot. This is not an AI adoption problem. It is a bad experience problem.
One of the fastest ways to destroy trust in customer-facing AI is fake empathy. A customer says, "I urgently need to change my flight. My mother is in hospital." A weak bot recognises the words "change my flight", serves up the standard fee policy and adds, "I'm sorry to hear about your situation." That is not empathy. That is automated politeness with nothing behind it.
Customers do not need a machine to pretend it has feelings. They need it to understand the task well enough to move them towards a solution. And when it cannot do that, they need it to hand them over quickly, cleanly and with the full context intact. This distinction changes the whole design philosophy. A company asking "how do we save money?" builds a bot that minimises the cost of the interaction. A company asking "how do we improve the experience?" builds a bot that minimises friction. On a project plan, those two initiatives may look similar. In real life, they are completely different machines.
At FrodX, we work with voice agents every day. In April, our Kinetara service passed 10,000 production calls for the first time, across twelve clients in six industries. These were not lab demos or friendly internal tests. They were real calls from real people: bookings, appointments, changes, questions, complaints, confusion and frustration.
People do not call customer support because they have time to spare. They call because something needs to be solved. That makes production data far more useful than any demo environment, because it quickly shows where voice AI helps, where it fails and where companies are fooling themselves. Three findings stood out.
First, trust grows through repeated good experiences. In November 2025, 27% of callers hung up within the first fifteen seconds, as soon as they realised they were speaking to a machine. By April 2026, that number had fallen to 8.5%. That is not a cosmetic improvement. It is a shift in behaviour. It means people learn from experience. If the voice agent wastes their time once, they punish it. If it solves their problem once, twice, three times, resistance starts to fall. Not because people suddenly fall in love with bots. They do not. But they learn that a well-designed voice agent is not necessarily another dead end. Trust in AI is not built by a launch announcement. It is built one resolved call at a time.
Second, the end of the call matters more than the channel. When the voice agent achieved the goal of the call, 76% of conversations were positive. When it failed to do so, only 50% were positive. That may sound obvious, but many companies still behave as if the channel itself is the problem. It is not. Customers are not angry because a machine answered. They are angry when the machine stalls them, confuses them or sends them in circles. For bookings and appointments, we saw success rates between 78% and 90%. For simple transactions, in some cases the rate reached 100%. That is exactly where voice AI works well: structured, predictable, repetitive interactions with clear rules and clean data.
Third, handover is not the problem. Bad handover is. Some managers treat a handover to a human agent as a failure. I think that is the wrong way to look at it. A handover is not a failure when it happens at the right time, for the right reason and with the right context. The failure is making the customer explain everything twice. That happened to me with an insurance company. First, I explained everything to the automated system. Then I finally reached a person. Then I had to start from scratch. At that point, I no longer cared whether they had a sophisticated system. I only cared why I had to repeat myself.
When a customer gives all the relevant information to the voice agent, and the human agent then asks the same questions again, you have not created a hybrid service. You have created two poor experiences back to back. A good handover should feel like continuity. The human agent should already know why the customer is calling, what has been said, what the system tried to do and where the issue got stuck. Without that, voice AI is not support. It is another obstacle.
This is where many leadership teams still get it wrong. A voice agent is not a worse, cheaper version of a human being that you tuck into the process to save on salary. It is a different capability, with different strengths and different limits. It is very good at clearly defined, repetitive, rule-based work: confirming appointments, checking availability, updating basic information, answering standard questions, handling status requests, processing simple changes and capturing structured information outside business hours. Most companies have more of these calls than leadership usually thinks. But voice AI does not work everywhere. It does not work when most calls are unique and complex. It does not work when company policy requires a human conversation. It does not work in emergencies where even a short delay is unacceptable. And it definitely does not work when the company has not sorted out its own data, processes and decision rules.
That last point matters especially in Central and Eastern Europe. Many companies in our region are not short on ambition. They are short on operational clarity. Customer data sits in several systems. Exceptions live in people's heads. Processes have been patched over years. Local habits matter. Workarounds become policy without anyone formally admitting it. Voice AI will not magically fix that. It will expose it. A voice agent can only be as good as the process behind it. If your data is messy, your rules unclear and your escalation logic political, the bot will not hide the problem. It will make the problem audible.
So I did not answer the CFO's original question directly. I did not tell him how many people he could let go. I answered the question he should have asked: which calls can we resolve better, faster and more consistently - and which ones must keep a human being human?
That is the difference between a weak voice AI project and a strong one. A weak project starts with cost. A strong one starts with experience. The irony is that the second approach often produces a better financial result as well. When customers get answers faster, when repetitive calls are handled consistently, when agents spend less time on routine work and more time on complex cases, efficiency follows. But it follows as a consequence, not as the design principle. That is why with Kinetara we do not sell contained calls and we do not sell cut jobs. We sell the outcome: how many inbound calls are actually resolved, at what level of satisfaction, and how cleanly unresolved cases move to a human agent.
Let the voice agent take the dull, repetitive and predictable work, including the calls that come outside business hours and would otherwise go unanswered. Give humans back the conversations where it genuinely pays to be human: the angry customer, the exception, the tangled case, the situation where rules are not enough and judgement is needed. That is where the real value is. Not in replacing people, but in using them where people still make the biggest difference. Start a voice AI project by asking, "How many people can we cut?", and you will probably build a worse service. Start by asking, "Which customer experience can we improve?", and you have a much better chance of building something customers will accept, use and trust. The savings come later. But only after you stop treating the customer as the cost.
Read more about how voice AI can improve customer support.