Nobody Misses the Call Anymore: What an AI Front Desk Does for a Small Business
Every small company has the same phone problem. The number on the website rings on a desk where nobody sits, or on a mobile that is in a meeting, or in a car, or on a site visit. The caller gets voicemail, hangs up, and calls the next name on the list. The business never finds out.
We had that problem. Amazing Projects is a small consultancy; most days there is nobody whose job is to answer the phone. So we built a front desk that does. Since mid-September 2026 the main company line at Amazing Projects is answered by an AI phone agent called Άννα. She speaks Greek and English, knows what we do, takes messages, checks the calendar, books meetings, and transfers the calls that matter to a human. This is what we built, what it does for a business our size, what broke along the way, and why the choice of telephony provider turned out to be the decision that made or broke the whole thing.
What the AI phone agent actually does
The AI phone agent runs on ElevenLabs Agents, a managed platform that handles the real-time loop of listening, thinking and speaking. The “thinking” is a Claude model. We wrote no speech code at all; the work was in the system prompt, the tools and the phone line.
On a call, Άννα does five things. She screens: who is calling, what about, and whether it is something we handle. She answers questions about our four service lines from a knowledge base built from this website, and only from that knowledge base; if the answer is not there, she takes a message rather than inventing one. She takes messages and emails a summary within seconds of the call ending. She checks calendar availability and books a meeting in a free slot. And when a caller needs a person now, she transfers the call to a mobile.
There is also a short VIP list. A handful of clients are recognised by their caller ID and go straight through, at any hour, with no qualifying questions. Everyone else gets the front desk during business hours and a message-taking service outside them.
Why this matters more for a small business than a large one
A corporation with a call centre gains a marginal cost reduction from an AI phone agent. A ten-person company gains a capability it did not have.
The first benefit is the obvious one: the call gets answered. Not by a menu, not by voicemail, but by something that can hold a conversation, explain what the company does and capture what the caller wants. For a business that lives on inbound enquiries, the difference between “answered” and “went to voicemail” is the difference between a lead and nothing.
The second is consistency. Άννα gives the same answer about our services on the fiftieth call as on the first, at 09:05 on a Monday and at 17:55 on a Friday. She does not forget to ask for a callback number. She does not promise a delivery date she cannot know.
The third is that the owner’s time stops being the bottleneck. Every call that is a question, a message or a booking never reaches a human at all. The ones that do reach a human arrive with context: the email summary says who called, why, and what they were told. The transfer happens only when a person is needed.
The fourth is cost, and here small businesses should be precise rather than impressed. The platform charges by the minute of conversation, in the range of a few cents; the phone line costs tens of euros a year; the model tokens are small change. For a company that receives a few dozen calls a month, the whole thing costs less than a single hour of a receptionist per month. It does not replace a receptionist at a business that has one; it gives a receptionist to a business that never could.
What we learned along the way
None of the hard problems were AI problems. The agent worked in the browser within days. Getting it onto a real Greek phone line took weeks, and that is where the effort went.
Two providers did not work out for us before we moved the number to Voiceland, a Greek VoIP provider. Porting completed in a few days, and the remaining issues were the ordinary ones of connecting a cloud platform to a SIP trunk: inbound routing, outbound authorisation, call transfer. Each was solved within days.
Greek speech recognition also forced one scope cut: capturing an email address by voice was unreliable, so we removed it from the call flow. The agent captures a name and phone number, which speech handles well, and we follow up by other means.
The provider mattered more than the platform
We want to say this plainly, because it is the part most guides skip. The AI phone agent itself was the easy part. The difference between the providers we tried was not price. It was whether anyone on the other end would engage with the problem.
Voiceland did. Their account manager set up the line and the porting without fuss. When something did not work, they answered the same day, their engineers looked at their own logs rather than pointing us back to the platform, and they confirmed each fix while we were still on the call. For a provider whose basic support is free, that is a level of technical engagement we did not expect and will not forget. If you are a Greek SME building anything on a SIP trunk, this is the question to ask before you sign: when it breaks at the SIP layer, will someone look at it with me?
What to consider before you build one
Decide what the AI phone agent may and may not say. Ours answers only from the knowledge base and takes a message for everything else. That one rule removes most of the risk of an AI inventing a price, a date or a promise.
Decide who gets a human. A VIP list by caller ID is trivial to implement and changes the experience for the callers who matter most. Recognising a caller by their spoken name is not reliable in Greek; recognising a number is.
Test the language you will actually receive calls in, on a real line, before you commit. Greek speech recognition is good enough for conversation and not good enough for spelling. Design around that rather than fighting it.
Expect the telephony to take longer than the AI phone agent itself. Budget for it, and choose a provider on support before price.
Keep a human in the loop behind the AI phone agent. Every call produces a transcript and a summary. Someone reads them. That is how the small booking and wording issues of the first week were found and fixed from transcripts, not from complaints.
Frequently Asked Questions
Does it work in Greek? Yes. Conversation, questions, message-taking and bookings all work in Greek and English, switching to the caller’s language. Spelling out email addresses does not work reliably in Greek; we removed that step.
What does an AI phone agent cost to run? Per-minute platform pricing in the cents range, a SIP line for tens of euros a year, and negligible model costs. For a small business receiving a few dozen calls a month, the total is a small fraction of any human alternative.
Can it book meetings directly into our calendar? Yes, through a native calendar connector. It checks availability first and books only into free slots, in local time.
What happens when it does not know the answer? It says so and takes a message. It is instructed not to answer from anything except the approved knowledge base.
Can callers reach a person? Yes. The agent transfers calls to a mobile when a person is needed, and certain callers are transferred immediately by caller ID.
Do we need a developer for an AI phone agent? For what we built, no code was written except one small script for message delivery. The work is in the prompt, the tools and the phone line. What you do need is someone who will read the transcripts and keep tuning.
Want one for your business?
If your company phone goes unanswered more often than you would like, we can build the same AI phone agent for you: on your own number, in Greek and English, with your calendar, your knowledge base and your rules about who gets a human. We handle the platform, the telephony and the tuning, and you keep the transcripts. Book a consultation, or call +30 215 215 97 87 — Άννα will pick up.
This article was researched and drafted with the assistance of Claude (Anthropic), and reviewed and approved by Amazing Projects before publication. Amazing Projects is an ElevenLabs affiliate; links to ElevenLabs in this article may earn us a commission at no extra cost to you.
