The Digital Secretary That Answers in Greek and Understands Cypriot

A training centre lives on its phone line. People call to ask what courses are running, to chase a certificate that never arrived, to say they did not get the joining link for a seminar that starts tomorrow. Every one of those calls is either a booking or a problem, and every one that goes to voicemail is a small loss. Centre 8 Education asked us to put an AI secretary on that line. We built it on our own voice platform, in Greek, and this is what it actually took.

Built, tested and working. The assistant handles the conversation, classifies the call and files a ticket straight into the client’s live system, and every piece of that has been verified end to end. It goes on the public line as soon as the Cyprus number is connected to it.

It had to understand Cypriot, not just Greek

This was the design risk we flagged first. Every speech model on the market is trained on standard Modern Greek. Callers in Cyprus do not speak standard Modern Greek. They say ίντα, λαλώ, τζιαι, εννεν, έσιει, ούλλα, and they switch between Greek and English inside a single sentence without thinking about it.

So the assistant carries a dialect section with a working glossary and the phonetic shifts, and one firm rule on top of it: understand the dialect, answer in standard Greek, never imitate it. An AI that mimics how someone speaks does not sound friendly, it sounds like mockery. It also never asks a caller to repeat themselves because of how they talk, and when a repeat genuinely is needed it blames the line, not the person. Code-switching between Greek and English is treated as normal, because in Cyprus it is.

It tells you it is an AI, in the first sentence

The greeting says so plainly: Χαίρετε, Centre 8 Education. Είμαι ψηφιακή βοηθός τεχνητής νοημοσύνης. That is not decoration. Article 50 of the EU AI Act requires that people interacting with an AI system are told so, clearly and no later than the first interaction, and it starts applying on 2 August 2026 with penalties that reach fifteen million euro or three percent of worldwide turnover. A Greek-speaking phone agent is squarely in scope, and this one was built to clear that bar from day one.

The assistant will also never claim to be human, even indirectly, answers “are you a robot?” straight in either language, and does not argue with someone who is annoyed to have reached a machine. Call recording is deliberately switched off, because turning it on adds consent obligations under Cyprus law that belong in the greeting, not buried in a policy page.

Every call becomes a ticket on the right person’s desk

Answering the phone is only half the job. When the call ends, a webhook lands in the client’s CRM and creates a support ticket carrying the caller’s details, the reason, a summary and the transcript. It matches the caller to an existing contact record where one exists, and it is idempotent, so a retry from the phone platform cannot create the same ticket twice.

The assistant only classifies the call. The routing table lives in the receiver, not in the assistant, so who gets what can be corrected without touching the agent at all. A missing certificate or a billing question goes to accounts at high severity. An enrolment-system problem goes to the coordinator. Everything else goes to the coordinator too, but only escalates when the caller asked to be called back. All of it verified live, including four calls arriving back to back.

There is a deliberate rule underneath all of it: a half-captured caller is still recorded. The assistant is told never to hold someone on the line to finish filling in fields, because one detail is better than a lost call, and the receiver files a ticket for every call regardless of how little was extracted. Even a caller who gives a name and hangs up arrives as a lead.

It knows today’s courses, and it refuses to quote a price

A secretary that recites last month’s schedule is worse than useless. A job runs every morning, pulls the courses that are actually confirmed and upcoming straight from the client’s system, and rewrites only that section of the assistant’s instructions. Cancelled or finished courses simply stop appearing the next day, with no one having to remember to remove them. The job is idempotent, so an unchanged day costs one database query and no update at all, and it refuses to run rather than ever leave a placeholder for the assistant to read out loud.

It is also explicitly forbidden from stating prices, subsidy rates, accreditation hours or availability. An assistant that confidently quotes a wrong price does more damage than one that says a colleague will confirm. We would rather it defer than guess, and that is a rule we set on purpose.

Two things we hardened once we put it through its paces

Phone numbers. In Greek and Cypriot speech people read numbers in pairs, and speech recognition mangles them. On a test call the number given was 94320068 and came back short. Prompt rules alone would never fix that reliably, so we stopped trying to solve it with language. The assistant now reaches for the phone keypad at the first sign of difficulty rather than as a last resort, which takes speech recognition out of the loop entirely and makes the digits exact. Numbers are then normalised into full international format before anything is stored, and where the country genuinely cannot be determined the record is flagged for a human instead of being silently guessed. A wrong number that looks right only fails later, when somebody dials it.

The manager’s email trap. Callers often say they never received the joining link for a course starting tomorrow, and searching for their email address finds no registration at all. The reason is that in corporate training the booking is frequently made by a manager or HR using their own address, then forwarded on. So the assistant asks whether the caller registered themselves or someone at their company did, captures both addresses, and the ticket says explicitly where else to look. That is not a technical insight, it is a domain one, and it came from the client telling us how the problem actually shows up.

We found a bug in our own code before it could cost anything

While testing the routing we hit a database error. The ticket reference generator, inherited from an older part of the same system, produced the same reference for calls arriving back to back and collided. It had never mattered, because until now a human clicked a button one at a time. An inbound phone line is exactly the case where calls arrive in bursts, and the failure mode was silent dropped tickets. We fixed it, verified four back-to-back calls produced four distinct tickets, and noted the same latent defect in two other files so it gets cleaned up on the next pass.

The daily operations digest we already run for this client gained a section for the voice agent too. It alerts if the course sync goes stale, counts how many calls turned into tickets, and raises an alarm on any webhook failure, because a failed webhook means a call that may have been lost. We would rather be told than assume.

Where it stands

• Working: the conversation in Greek and English, the Cypriot dialect handling, the AI disclosure, keypad-exact detail capture, and the classification that decides where a call goes.
• Wired in: every call lands as a ticket in the client’s live system, matched to the contact record, routed to the right desk, escalated when a callback was asked for, and impossible to duplicate on a retry.
• Kept current by itself: the course list refreshes every morning from the client’s own records, and cancelled courses drop off without anyone touching the assistant.
• Watched: a stale sync, a failed webhook or a quiet day all show up in the daily operations digest.
• Next: connect the Cyprus number and it starts taking real calls. Live transfer to a colleague switches on with it.

The pattern here is one we keep returning to: the interesting part of a voice agent is almost never the voice. It is what happens when the call ends, whether the right person finds out, and whether the thing tells the truth about what it does not know.

Related service: AI Voice Agents. See the rest of our own-build stories on the case studies page.