AI voice agents for calls: what they do well and what they do not
There is a cost that hardly ever shows up in the accounts: the phone that rings and nobody answers. It leaves a missed call in the log and nothing else: you do not know what they wanted. They hang up and dial the next number. This article is for anyone weighing up putting an AI on those calls, and who wants to know, before signing anything, where its work starts and where it ends.
The problem is not volume, it is timing
Calls are lost more because of timing than because of volume: a handful arrive at the worst moment. The lunch service in a restaurant. The agent already showing a flat. The patient sitting in the chair. Picking up is hard right then, and the caller has no reason to try again: the next number is one tap away. The gap shows up nowhere, because the log does not say what the caller wanted.
The same thing happens outside opening hours. Whoever calls at night usually has something on their mind: they have just decided something, or they have a problem right now. Not everyone talks to voicemail, because talking to a silent machine is awkward. The rest evaporates. Before looking at tools, look at your blind slots: lunchtimes, Saturdays, nights. That is where you find out whether you have a call problem at all.
Why this is not a touch-tone menu
A traditional menu forces you to translate your need into a language someone else designed. Press 1 for bookings, 2 for orders, 3 for everything else. If your case does not fit, you end up on 3, which is usually a voicemail box. The menu understands nothing: it counts keypresses. And every option you add to cover one more case lengthens the tree and makes the call worse for everyone else. That is why they wear people down before they reach the option they wanted.
A voice agent works the other way round. The caller says what they want in their own words, out of order, and the agent works out the intent. The practical difference is that it can ask back. If the date is missing, it asks for it. If something did not come across clearly, it asks again. A menu cannot do that, because it has no way of knowing what it is missing. Asking for the missing detail is a good part of what makes the agent useful.
Several calls at once, and no engaged tone
One person handles one call at a time. That is a physical limit, not an attitude problem. If several calls come in at once, one gets answered and the rest hear the engaged tone or drop into voicemail. That is why peaks hurt so much: they land exactly when the team is busiest with the customers already in front of them. The call competes with the room, and the room wins, as it should.
A voice agent does not depend on a single head. Several calls move forward at the same time, up to whatever limit you contract with your phone provider. So the peak stops deciding how many calls go unanswered. But be clear about what that solves: it solves the phone not giving an engaged tone. Whether the conversation is any good depends on how you prepared it.
Transcript and summary: the call stops being smoke
An ordinary phone call leaves nothing in writing. What was said lives in the memory of whoever picked up, and a short while later it is already approximate. An agent can be set up so that every call ends as text: what the person asked, what they were told, what details were captured. On top of that there can be a short summary, which is what actually gets read, and it can go to your inbox or your CRM if the system allows it. None of that comes on its own: first you decide what is kept, for how long, and how you tell the caller.
The side effect is the interesting part. As calls pile up, what people really ask ends up in writing, rather than what you assume they ask. That can change the script the agent uses, and often the website too. Now the honest drawback: transcription is not perfect. With street noise, a bad hands-free kit or an unusual surname, it gets things wrong. That is why critical details, a phone number for instance, are read back out loud before the call ends.
Handing over to a person: when and how
This is the part most worth writing down before you switch anything on. The when is defined by rules. First: if the caller asks to speak to someone, hand over, the first time and without pushing back. Second: if the agent has failed to understand the same thing twice, it does not try a third time. Third: some matters never go through a machine, and you write that list yourself.
That list usually covers complaints, incidents, anything touching health, refunds or legal matters, and the customer who is already angry when they call. The reason has to do with what the caller is after. In those cases they want someone to take charge of the matter. An agent can take down the case and pass it on, but it cannot take charge. Confusing the two is the surest way to make the caller angrier.
The how matters just as much. Within opening hours, the call is passed to a person; if your phone system allows it, the context goes ahead of it, so whoever picks up knows who is calling and what they wanted. Outside hours, do not pretend: say nobody is available right now, take the details, and leave a task so someone calls back. And the agent says what it is from the start. Finding out halfway breaks something you cannot repair.
What it does not do, plainly
It does not talk round someone who was not going to buy. An agent handles intent that already existed: somebody looked for you, found you and dialled your number. It does not create that intent. If your phone barely rings, the problem sits upstream of the phone, in not being seen, and that is fixed with something else. There the honest answer is not to automate calls yet. Handling an empty channel well only leaves you an empty channel.
It does not fix bad service either. If orders go out late or nobody returns the calls they promised, the agent will make that failure surface sooner and in writing, not disappear. And it does not replace the team. It takes the repetitive part, the part answered the same way every day. Anything needing judgement, memory of the customer or the nerve to bend a rule stays human. An agent that decides beyond its rules is badly built.
What to decide before switching it on
The decisions come before the voice, and all of them are business decisions. Which calls it takes: all of them, only the overflow, or only those outside opening hours. What it may state as fact: hours, availability, conditions. What it will never say, however hard people push. What it may touch in your systems: reading a calendar is not the same as booking in it, and booking is not cancelling a slot someone else made.
Then the part almost everyone forgets: what happens when something goes wrong. Your booking system goes down. The caller will not give a name. The line drops halfway. Each of those needs an exit decided in advance, and that exit is usually a person. Decide as well what you keep from each call and for how long, and tell the caller. If those answers do not exist inside the business, the agent will not improvise them well.
How we build this at Zepai
We work from Spain with restaurants and coffee shops, estate agencies, e-commerce, clinics, salons and blockchain projects. When the job is a voice agent, the order is always the same: first we look at your call log and your blind slots, then we write the handover rules, and only at the end do we choose a voice. Most projects go live in five to seven working days. We do not quote a catalogue price, because it depends on the channels, the volume and which systems it has to talk to.
When we finish, we train you so you can change hours, messages and rules without calling us. If you depend on us to edit one sentence, the build is wrong. If you are interested in WhatsApp as well as the phone, we build it on the official API: that is where we are an Official Meta Partner. If you want us to look at your case, write to info@zepaiagency.com or call +34 604 14 09 97. We reply in under 24 hours, and sometimes the answer is not to automate.
Want to build this in your company?
Half an hour, no sales pitch. And if we think it is not worth it for you, we will say so.
Get a free quote →