What a voice agent has to get right.

Speech synthesis stopped being the bottleneck some time ago. A current model holds a natural-sounding phone conversation, handles interruptions, and copes with an accent well enough that most callers do not think about it. That is genuinely new, and it is also the part of the problem that is finished.

Which means the demo tells you almost nothing. A demo is a conversation with no consequences: nothing is looked up, nothing is changed, and nobody has to live with the outcome. The three requirements that decide whether a phone agent earns its keep are duller than the voice, and they are the ones worth interrogating before you buy.

One: it has to look the answer up, not remember it

The failure mode is identical to the one that ruins a shopping assistant. A model that generates first and checks afterwards will eventually state something fluent, confident and wrong — an appointment slot that is taken, a delivery date that slipped, a policy excess that changed in April.

The fix is architectural rather than a matter of prompting. The agent queries your live systems during the call and speaks only about what came back. It is the same ordering that separates a grounded product-search assistant from a plausible one, which we have written about at length elsewhere. A system that retrieves first can still be unhelpful. It cannot be fictional.

This has a practical consequence for procurement. Ask which of your systems the agent reads from, in real time, during the call. If the answer is that it works from a knowledge base exported last month, you are buying a very articulate FAQ.

Two: it has to write the outcome back

A call that changes nothing in your systems is a call you pay for twice: once for the minutes, and again when a person re-enters it by hand. Whatever was agreed — the moved appointment, the started return, the logged claim, the raised ticket — has to land in the tool your team already works in, before the caller has hung up.

This is where the sector differences actually live. The conversation is broadly the same everywhere; what changes is which system holds the truth and what the agent is permitted to change in it.

The pattern generalises: the phone conversation is close to interchangeable, and the integration is the whole job. Any vendor whose sector pages differ only in vocabulary is selling you the easy half.

Where a system genuinely has no API, be honest about the ceiling. The agent can still take the details accurately and route them to a person, which is worth having. It cannot complete the transaction, and a vendor who implies otherwise is describing software they have not built.

Three: it has to stop before it guesses

The hardest thing to configure in a phone agent is not what it says. It is what it declines to handle. A bereavement, a complaint, a caller in distress, a request that falls outside the rules you set — these need a person, quickly, and they need the person to arrive already knowing what has been said.

Handoff quality is therefore a real specification, not a footnote. Three questions separate a good implementation from a bad one. Does the transfer carry the transcript, so the caller does not start again? Is the boundary explicit and yours to set, rather than emergent from the model's judgement? And what happens outside staffed hours — does it take a message into a system, or does it improvise?

An agent that hands off cleanly on a fifth of its calls is doing better than one that handles everything and is wrong occasionally, because the failure is visible and cheap rather than invisible and expensive.

What it costs, and where the pricing hides

Voice is priced by the minute, which makes it the one AI product where usage genuinely tracks value — and the one where a bad month is expensive. Published Greek market pricing sits in a narrow band: Voice Logica lists plans from €49 a month for 200 minutes up to €2,250 for 10,000, with per-minute top-ups quoted at €0.26 to €0.30 (voicelogica.ai, list prices read 5 September 2026). Our own plans start at €49 for 300 minutes with extra minutes from €0.20.

The number to interrogate is not the headline. It is the overage rate and what happens when you exceed the bundle, because inbound call volume is not something you control. A month with a product recall, a weather event or a viral post can double it. Two protections are worth insisting on: a cap or hard stop that you set, with alerts before you reach it, and a guarantee that overage cannot cost more than simply moving up a tier.

The comparison that matters is not against other software. It is against the call you did not answer. If a missed booking is worth €80 and you miss four a week outside opening hours, the entry plan pays for itself in the first fortnight. That arithmetic is yours to do with your own numbers, which is rather the point of the next section.

What nobody in this market can currently show you

We looked at the public material from the Greek voice-agent vendors, sector page by sector page. Across eight industry pages, a features page and the blog, there is not one measured result: no answer-rate before and after, no abandoned-call figure, no named customer with a number attached. The pages are capability lists and benefit lists.

We are not in a position to be smug about this. We do not publish outcome data for voice either, because we do not yet have a defensible dataset for it, and inventing one is the thing this blog exists not to do. The adjacent evidence that does exist is worth knowing: in appointment businesses, the interventions that reduce non-attendance have modest and well-measured effects, which we went through in our review of the no-show research. Nothing in that literature is about AI answering the phone.

So the honest position is that this is a category where you should trust your own pilot over anybody's marketing. Two weeks of your real calls, measured against what your current arrangement achieves, will tell you more than every vendor page in the market combined. Insist on running one before you commit, and insist on seeing the transcripts.

A short buying checklist

If you take nothing else from this: ask which live systems it reads during the call, ask what it writes back and where, ask to see a transferred call arrive with its transcript attached, ask what the overage costs and how you cap it, and ask for a pilot on your own inbound line. The voice will be fine. Everything else is the product.

And before any of that, phone one. Our own demo line is +30 218 219 0702 and it is answered by the agent rather than by us. Two minutes of asking it something awkward will tell you more about this category than the rest of this article.