This year, I’ve listened back to a lot of AI calls.
Some made the hair stand up on the back of my neck — in a good way. The customer reached the end of the conversation without realizing they had been speaking to a machine.
And there were others I would have preferred to delete immediately.
This post is about the second kind too.
Not because I like sawing off the branch I’m sitting on.
But because, for the past two years, the Hungarian market has been full of promises that “AI solves everything.”
That has one predictable consequence:
A business owner introduces it, gets disappointed, and then doesn’t want to hear about AI again for the next five years.
I would like to avoid that.
So let’s be honest about where an AI phone agent still fails today — and where it easily outperforms a human.
First, the bad news: five situations where AI still fails
1. When you speak at the same time
This is the most common mistake, and the most awkward.
People interrupt each other. That is how we speak; it is completely normal. The machine, however, has to decide in real time whether you have finished the sentence, or just taken a breath before the next clause.
If it decides incorrectly, one of two things can happen. Either it speaks over you, or it waits silently while you are waiting too. The latter is worse: silence on the phone is deadly. After about 0.8 seconds, we start to feel uncomfortable, and above one and a half seconds, our brain automatically says, “the line has been disconnected.”
This area has improved a lot in a year, but if someone says that in their system this is already 100% solved, they are not telling the truth.
2. In a noisy environment
AI is very good on a normal line. In a truck, a machine room, on a busy street, on speakerphone, or amid the noise of children, it is not.
If your target audience is typically on the move — technicians, drivers, construction, agriculture — factor this in. I am not saying it does not work. I am saying there will be a difference between your nice lab test and live calls.
3. Names, addresses, email addresses, order numbers
This is where reality is the harshest.
“Gyöngyössy-Kaszás Bernadett, Nyíregyháza, 14/B Szarvas Street, email: b.gyongyossy@…” — even a human would ask for that again. The machine asks again too, only for it this is not a polite gesture, but a necessity.
The experience: if three pieces of information have to be recorded precisely in a call, that is still fine. If eight, the customer gets tired by the sixth. The solution is not to put a larger model underneath it. The solution is to send the recorded information back after the call by SMS or email for confirmation.
4. When the customer steps outside the script
An AI agent is brilliant within a narrow band. If you step outside that band, the problems begin.
During an outbound qualification call, the customer asks, “and why was there an extra 12 thousand forints on my bill last year?” — the good system says that a colleague will look into it, and records the question. The bad system, however, tries to answer. And if it does not have the data behind it, it makes something up.
This is the most dangerous type of error, because it does not sound like an error. It sounds confident.
5. When there is emotion on the other end of the line
An angry customer. A complaint. Bad news. The subscription of a deceased relative.
AI remains polite even then — but politeness is not enough in these situations. These calls need a human. Not because the technology cannot follow the words, but because at that point the customer does not want information; they want someone to take responsibility.
If someone promises that AI will take this off your shoulders too, ask them whether they have ever listened back to a complaint call.
And now the other side: where AI easily beats humans
If you thought up to this point that this article was the burial of our own product, here comes the twist. Because this same technology does things that a human simply cannot.
Speed. The interested person submits the form, and 20 seconds later their phone rings. Not three hours later, not the next morning. Anyone who has ever worked with lead callbacks or cold calling knows that this is the difference between revenue and a lost lead.
Persistence. After the third unsuccessful attempt, a human crosses the name off the list. AI does not feel uncomfortable even on the fifth call, and it does not have a bad day on Thursday afternoon.
There is another, less spectacular but more important advantage: salespeople instinctively call the leads they like first. And the bottom of the list stays there. AI does not pick and choose — it calls the twentieth name in exactly the same way as the first. Even on Saturday evening, when an e-commerce lead would otherwise go cold by Monday.
And then there is the most boring part, which nevertheless brings in the most money: every call is recorded, summarized, and entered into the CRM. There is no “oh, I forgot to enter it,” no note on the desk. In one of our client campaigns, this meant a 56% appointment-booking rate — not because the machine spoke better than a human, but because it called immediately and called everyone.
And if an advertisement has gone out and 200 leads come in during one day, for AI it is a day just like any other.
The test question that helps you decide whether it is right for you
It is not company size that decides. Nor the industry. This is what decides:
Green light: give it to AI:
- a repetitive, predictable conversation (appointment booking, data collection, callback)
- where speed is worth more than nuance
- where no one is currently calling people back because there is no capacity for it
- where the purpose of the call is to answer a few questions, not to negotiate
Red light: a human is still needed here:
- complaint handling, escalation, compensation
- custom pricing, bargaining, a decision involving multiple participants
- confidential, emotionally charged topics
- complex administration requiring a lot of data
At most Hungarian SMEs, 60–70% of incoming and outgoing calls fall into the first category. This is the part worth automating. The remaining 30–40% would finally receive enough attention — from the same team.
Three rules that eliminate most mistakes
If you take only three things away from this article, take these three:
- One agent, one task. The “AI receptionist that knows everything” is the most common implementation mistake. AI handles one narrowly defined task (e.g. “call the new lead and book an appointment”) excellently. It does not handle ten tasks at once.
- A clear boundary where a human takes over. A good system is not one that tries to solve everything. A good system recognizes in time that “this is no longer my area,” and hands it over — politely, together with a summary of the conversation.
- Listen back to the first 100 calls. This is my most boring advice, and by far the most useful. There is no well-written script that is perfect on the first attempt. The recordings reveal where the conversation gets stuck — and after two weeks of fine-tuning, you get a completely different system.
Original Hungarian article: https://aitrainer.hu/hol-hibazik-az-ai-telefonos-ugynok/