Thursday afternoon, half past four. Headphones on, a spreadsheet in front of me, and one job for the next two hours: listen back to 100 calls made by our AI agent.
Not to pat ourselves on the back. I wanted to know what really happens when a machine calls a person. Where does it work? Where does it get stuck? And what does the person who answers think of it?
These are the five things I learned.
1. Timing matters more than the voice
Everyone asks about the voice: does it sound natural enough, will people notice? But after listening to the calls, success did not come down to the voice. It came down to when the phone rang.
Dr James Oldroyd of MIT analysed data from more than 15,000 leads. He found that calling a lead within 5 minutes makes you 100 times more likely to reach them and 21 times more likely to qualify them than waiting half an hour. Harvard Business Review researchers later tested the same point with one web enquiry sent to each of 2,241 US companies. The result was bleak: only 37% responded within an hour, while almost a quarter never responded. The average response time was 42 hours. Not minutes. Hours.
Why does this matter so much? Because a prospect rarely contacts only your company. They fill out forms with two or three competitors as well, then usually start talking to whoever responds first. If you call back hours later, you are no longer competing to make the first impression. You are competing for what is left.
You can hear it in the recordings. Call someone thirty seconds after they submit a form and they are still sitting in front of the screen:
“Oh, I only just sent the form!”
Call them the next day and the response is: “Which company was this again?” Same person, same need, but the moment has gone.
2. They do not hang up (that surprised us too)
The question company owners ask us most often is: “Don’t people just hang up?”
Only 9 of the 100 calls ended with an immediate hang-up. The other 91 became full conversations, and most people spoke to the AI exactly as they would speak to a receptionist. They answered, asked questions and said goodbye at the end.
There are two conditions. The agent must introduce itself in the first sentence and say that it is an AI assistant. The EU AI Act’s transparency rules require this anyway. The calls suggest that disclosure does not scare people off. It reassures them because they know what they are dealing with. The second condition is that the agent gets to the point immediately: why it is calling, what it wants to ask and how long it will take. Something like this:
– Good afternoon. I’m Anna, AI-Trainer’s AI assistant. You have just filled out the form on our website. Is now a good time? – Yes… wow, that was quick. – We try. I have three short questions. It will take one minute, then my colleague can call you with a specific proposal. Shall we start?
Customers are not bothered by talking to a machine. They would be bothered if it wasted their time.
3. Where does the AI make mistakes?
This article would not be credible if I skipped this part. Some calls went wrong. These were the four typical cases:
- Background noise. In a car, workshop or on the street, speech recognition is less accurate. The agent asks people to repeat themselves, which can test their patience.
- Interruptions. When a customer speaks very quickly and talks over the agent, half a sentence can get lost.
- Names and numbers. It may mishear an unusual surname or a phone number dictated too quickly. That is why every call gets a written summary, which a person reviews before the data enters the CRM.
- The one-off question. “What if I am exempt from VAT?” A properly configured agent does not improvise. It says that a colleague can answer accurately and will call back.
The lesson is not that AI is perfect. It is that its mistakes are manageable: a person takes over at the critical point, and every call can be traced. You cannot hand a missed call or a 42-hour response time over to a person. That prospect is already gone, leaving no transcript and no lesson behind.
4. The questions matter more than a beautiful voice
The best calls did not have the most beautiful voices. They had the best structure: a short introduction, three or four specific questions, confirmation and an appointment. Two minutes. Done.
The rule that emerged from the recordings was simple: one question, one piece of information. “How many people are on the team?” gets a number. “Tell me a little about your company” gets a sigh. A useful qualification sequence might look like this:
- What exactly did you contact us about?
- How urgent is it: do you want a solution within days or within weeks?
- What volume are we talking about? (number of calls, people or size of project, depending on the industry)
- Are you the decision-maker, or do you need to consult someone else?
- Would morning or afternoon be better for a short meeting?
Five questions, two minutes, and your salesperson walks into the consultation already knowing what matters about the prospect. The same is true for a human salesperson. AI only makes it brutally visible because every call follows the same process, every call is recorded and every call can be reviewed. Read through your own call transcripts once and you will learn a great deal about your sales process too.
5. The 56% is not magic
There was no trick behind the 56% appointment-booking rate measured in our client campaign. It came from three very boring things: every lead was called, not only the ones someone had time for; every call happened within minutes; and every lead was asked the same questions.
A person can keep this up for a day or two. A team might manage it for a week or two. The agent does not get tired, forget or pick favourites. Its 100th call is the same as its first. Revenue often comes from exactly this kind of boring consistency.
Let’s do the maths: what is that worth?
Take a simplified example, then run the numbers again with your own figures.
Suppose advertising brings you 100 leads a month. With manual callbacks and an average response time of several hours, you never reach some of them. Let’s be optimistic: you have a meaningful conversation with 25 out of 100, and 12 of those book an appointment. Call the same entire list within minutes and, using the rate measured in our campaign, you get 56 bookings.
Your advertising spend stays the same. The difference is 44 appointments per month. If only one in four becomes a customer, that is 11 additional customers every month from the same marketing budget. Multiply that by your average customer value and you can see what slow callbacks really cost. For many companies it amounts to millions of forints each month, but it never appears as an expense in the accounts. (We calculated the cost of a single unanswered call in detail here.)
And we have not even counted enquiries outside working hours. In our case, 41% of form submissions arrive in the evening or at weekends, exactly when nobody is in the office to call back. Only the machine is there.
Let’s look at your calls together
If you want to know where most of your prospects fall out, and which part an AI agent could take over, we can examine your process in a free 30-minute consultation. It is not a presentation. We work through your actual process, so you will leave with something useful even if you decide not to work with us. Book the free consultation →