Voicebot for business — which calls it may take, and which it may not

What a voicebot will not solve — and why we start there

Start with the sentence that costs us work: if all you need is an answering service that picks up outside office hours, reads a few lines of script and emails you the recording, buy the subscription. Some Polish vendors publish price lists, setup takes days, and we are not needed for it. That is the same position we keep on the what we don’t do list: do not build what you can buy on a subscription.

The project starts at the second question — when the bot has to say something that is not in the script. An appointment time. An order status. Whether the customer is entitled to a refund. Every answer of that kind has to be read live from the system that knows it, not generated by the AI. That is integration work, and it is what we sell: choosing a vendor on Polish-language evidence, connecting the bot to the systems it takes its data from, bounding what it may say, and putting in the caller notice and oversight the rules require.

The second limit is technical and lies in speech recognition itself. An independent, peer-reviewed 2024 comparison of Polish speech recognition systems (the Polish ASR Leaderboard / BIGOS study, 25 systems and over 4,000 recordings) reports a median word error rate (WER: the share of words substituted, dropped or added) of 14.52% for read speech and 32.42% for natural, unscripted speech, which is the kind you hear on a phone. The study does not cover the newest models, and its 2024 results do not predict the performance of current services or of a system on your phone line. Vendors measure their figures on wideband recordings. A classic narrowband phone line carries a narrower range of sound (sampled at 8 kHz), and that moves those figures one way only: for the worse. Not every call is like that: some internet (SIP) and mobile calls carry wideband sound, so what speech recognition actually receives depends on the particular call. That is why we test the selected solution on recordings from your line. Alongside word errors, we check whether it captures the information needed to handle the enquiry, such as an order number, an amount or an appointment time, how long corrections take and when a person must take over. We agree acceptance criteria before the pilot. The results of that test decide when the bot asks callers to confirm what it heard, when it hands the call to a person, and what it is allowed to state at all.

The same caution applies to claims of language support. OpenAI counts a language as supported by its voice model (gpt-realtime) when it “exceeded <50% word error rate” in testing, that is, fewer than 50 words substituted, dropped or added per 100 words of the correct transcript. Polish is on that list — and that sentence is worth reading twice before “supports Polish” is taken as an answer about quality.

What we actually do

We start with uses where a speech-recognition error has the smallest consequences and a person can check the result and correct an error before further action, and move to the next item only once the previous one has proved its quality on real calls:

• Call notes into the CRM. Nobody talks to the bot at all — a person runs the call and the system transcribes it and writes the outcome into the system fields. The salesperson reviews and approves the content before it is saved in the CRM. The pilot also measures errors and the time needed for corrections.
• Identifying what the call is about and routing it. The bot establishes what the call is about and transfers it to the right person or queue. It does not answer the question — a person does, only sooner and without three transfers. The cost of a mistake is one extra redirect.
• Read-only lookup of the current status in your system. “Where is my delivery”, “has the invoice been paid” — the answer is a fact in a database, not a sentence the AI generated. The bot reads the value and repeats it; where it is not confident it heard the reference correctly, it asks for confirmation or transfers.
• Booking, confirming and cancelling appointments. The first step where the bot writes to a system rather than only reading from it — so the first that needs a rollback story: who undoes an entry that turns out to be wrong, how, and how the customer learns of the change.
• Answering out-of-hours and busy-line calls. A full unsupervised conversation, which is why it is the last item on this list and not the first. We switch it on for the call types where the earlier steps showed the bot understands callers — clearly marked as a machine. When an adviser is available, the caller can ask to be transferred. When all advisers are busy or the call arrives outside staffed hours, we agree how the enquiry will reach a person, for example by taking a message or a callback request. The bot clearly explains what will happen next.

One rule holds across the whole list: the bot does not state a price, a deadline, an availability or an entitlement it has not read live from the system that knows it. That is not theoretical caution. In Moffatt v. Air Canada (2024 BCCRT 149) the tribunal held the airline responsible for what its chatbot had said and awarded CAD 650.88 in damages. The sum is token; the principle is not.

If the same enquiries also arrive by email, chat and web form — and they usually do — the voice channel is one of several entrances to the same process. We describe the whole of it at customer service automation, which is also where we make the case for starting in the back office rather than with the customer conversation. When the bot has to take actions across several systems instead of reading a state, the work becomes an AI agent project — this page does not repeat that comparison.

What we will not promise

• A cold-calling voicebot. Article 398 of the Polish Electronic Communications Law prohibits using automated calling systems to send commercial information without prior consent — business recipients included — and you may not call in order to obtain that consent. The President of UKE may impose a fine of up to 3% of the previous year’s revenue or up to PLN 1,000,000, whichever is higher, and the risk stays with the caller. Whether purely operational calls fall outside the prohibition is for the client’s lawyer to settle — we do not settle it.
• A voicebot pretending to be human. The bot says it is a machine in the opening seconds, and a route to a person stays available throughout. We do not build scripts whose purpose is that the caller fails to notice.
• A voicebot that assesses candidates in recruitment. Inferring emotions in the workplace is prohibited by Article 5(1)(f) of the AI Act and has been since 2 February 2025. The prohibition covers recruitment and probation, binds the deployer as well as the provider, and a candidate’s consent does not lift it.
• A voicebot on emergency numbers. Evaluating and classifying emergency calls and prioritising dispatch is an Annex III high-risk use under the AI Act. The obligations for that category were deferred to 2 December 2027 — deferred, not cancelled — and we do not have the experience such a project would require.
• A debt-collection voicebot as a first project. We say this plainly and on reputational grounds, not legal ones: we found no provision prohibiting automated calling for debt collection, and we will not pretend one exists. We do not, however, want the first place a company tries an automated conversation with people under pressure to be debt collection. The full set of conditions under which we advise against a project is collected at what we don’t do.

We also do not build our own speech engine and do not resell a bot subscription. The engine is bought; we answer for what it is wired to and what it is allowed to say. Deployment outcomes published by vendors we treat as marketing with no stated methodology — we do not repeat them to a client as a forecast.

Call recordings are personal data

A call recording and its transcript are as a rule ordinary personal data. The legal basis, the retention period and what the caller is told all have to be settled before go-live, not after the first month. The category is decided by what is said on the call, not by the technology. One line where that category changes does run through the technology: extracting a voiceprint in order to identify or verify the speaker is special-category biometric data under Article 9 of the GDPR and needs its own condition. By default we build on the safer side of that line — transcription with no voice profile — and we say so plainly when a client asks for something that crosses it.

The second difference from the text channel is the supplier chain. A voice integration typically adds three or four external processors — the carrier, speech recognition, the language model, speech synthesis, sometimes a separate store for recordings — where a text integration has one. Each of them processes data on your behalf (Article 28 GDPR). Some you contract directly. Others, called sub-processors, can be engaged only with your authorisation, and equivalent obligations are passed down to them. For each one you also need to check separately whether there is a transfer of data outside the European Economic Area (EEA). Each supplier's role is determined separately, per processing operation. How the roles are split, how that split is determined, and the questions we put to every supplier are set out on the security and GDPR page — here we add only the voice part.

Telling callers they are talking to AI — what the law requires, and what we always do

Article 50(1) of the AI Act requires that a person interacting with an AI system be informed of it clearly and at the latest at the first interaction — unless that is obvious from the circumstances and the context of use. On a phone line we do not rely on that exception: a caller has no way of knowing who will pick up, so we tell them every time. That is our own rule, stricter than the legal minimum. The article applies from 2 August 2026, and it mainly concerns how the system is built: the bot has to be designed so that it tells callers they are interacting with an AI system. Whether that message actually reaches the caller then depends on how the company runs it — the script, the greeting, whatever the caller hears before the bot picks up. We build the notice into the script and the documentation. The question of how the article allocates duties between the provider and the deployer (anyone using the AI system under their authority, for example a company or a public body, except for personal, non-professional use) in a given setup goes to your lawyer at the start of the project; we do not settle it ourselves.

Article 50(2) adds a second duty: AI-generated audio must be marked in a machine-readable way, meaning so that software can recognise it as AI-generated, “as far as this is technically feasible”. The transition period — until 2 December 2026 — applies to systems placed on the market before 2 August 2026, so which side of that date a given deployment falls on depends on the engine used and its provider, not on the day the integration is built. The marking duty itself is addressed to the system's provider. So there are two things to establish: what the engine's provider states, and whether that statement can be verified. We do not claim that any marking survives being compressed on a phone network. We deliver a documented assessment of whether machine-readable marking is feasible, and a record of what the system generated and when. The wider preparation — a register of AI systems, the notices and marking your lawyer assigns to your role, documentation and a documented team workshop — is described at the AI Act technical compliance sprint.

Then there is the number the bot calls from or answers on. The Polish act of 28 July 2023 on combating abuse in electronic communications names caller-ID spoofing as an abuse, and since 24 September 2024 operators have been obliged to block such calls. The number is therefore a compliance matter, not a configuration detail: a spoofed number, or one the network cannot route, simply stops connecting. Separately, a Polish local number is a paperwork project — the registered address has to match the number’s region, and operators require identity and company documents. For a pilot it is usually better to point traffic at a number the client already holds.

On the Polish side the act of 3 July 2026 on artificial intelligence systems (Dz.U. 2026 poz. 1003) establishes KRiBSI, Poland’s new AI supervisory commission: a collegiate body served by the Ministry of Digitisation and composed of representatives of the Presidents of UOKiK, KNF, KRRiT and UKE. The body exists in law and is still being stood up, so we do not claim anyone is enforcing these duties today. What we do claim is that the date the disclosure duty starts is known, and that a project going live after it has to meet it from day one. How a KRiBSI inspection runs under the act, and what to have ready, is covered in our article on the AI Systems Act and KRiBSI inspections.

How we work

We start with a report from the client’s own phone system: how many calls come in, at what times, how many go unanswered, and what callers ask about. We do not quote an industry missed-call figure, because every one we found traces back to vendor marketing — whereas your own report takes an afternoon to pull and is true. The same rule governs quality: recordings from the client’s line are the only test we select a speech-recognition vendor on.

We choose vendors on how well they handle Polish, not on the length of a language list in a brochure. One example we check every time: in Microsoft’s voice service (Azure Voice Live), the default, Microsoft-recommended automatic language detection (“automatic multilingual”) covers fifteen language locales and Polish is not among them. Polish is available, but it has to be specified in the configuration, and Microsoft warns that transcription quality is poor when no language is set. The bot’s voice is similar: Azure’s ready-made native Polish (pl-PL) range has exactly three standard neural voices, with no HD voice and no choice of speaking style, while Google’s Chirp 3 HD is generally available for Polish and lets you set custom pronunciations — which matters directly for product names, street names and surnames.

From there it is the same as any of our projects: one call type, a pilot with the narrowest possible scope, and a decision on widening the scope based on its results. The full sequence is described in our delivery methodology, the whole scope of the service — from process audit to running it — at AI implementation, and the conditions a process has to meet to be worth automating at AI automation.

What it costs

On the bot side the cost has four components that vendors quote separately and that have to be summed for a comparison to mean anything: the platform subscription, a bundle of minutes with the rate for minutes beyond it, a one-off implementation fee, and the integration with the systems the bot takes its data from. A fifth is running cost — scripts, the pronunciation lexicon and the rules for when the bot hands a call to a person change whenever the offer does. The overage rate tends to be the biggest surprise, because published rates differ between vendors by roughly a factor of two.

Whether the project makes sense at all is decided by the number on the other side: how many calls go unanswered today, and how many of those concern matters the bot could handle without guessing. Where that number is small, the honest answer is “do not buy” — and we give it. We have set the whole bill out line by line, with two published price lists summed over twelve months, in our piece on what a voicebot costs. The cost components common to all our projects are covered separately, in what an AI implementation costs.

We have not delivered a voicebot yet, so we have no outcome numbers of our own — and we do not intend to borrow anyone else’s. What we can show before a project is how we work and the list of things we will not do.

Frequently asked questions

How is a voicebot different from a chatbot?

The channel, and the margin for error. A chatbot receives exactly the text the customer typed. A voicebot first has to turn speech into text, and some words get lost at that step. An independent, peer-reviewed 2024 comparison of Polish speech recognition systems (the Polish ASR Leaderboard / BIGOS study, 25 systems) reports a median word error rate (WER: the share of words substituted, dropped or added against the correct transcript) of 14.52% for read speech and 32.42% for natural, unscripted speech. The study predates the newest models and does not predict the quality of today’s services, but it shows why a voicebot has to ask and confirm rather than assume it heard correctly. The phone line adds to this. A classic narrowband line carries a narrower range of sound (sampled at 8 kHz) than the wideband recordings vendors test on. Some internet (SIP) and mobile calls carry wideband sound, so recordings from the actual line settle it.

What does a voicebot cost?

The price has four components that vendors quote separately: the platform subscription, a bundle of minutes together with the rate for minutes beyond it, a one-off implementation fee, and the integration with your systems that hold the current data (calendar, orders, CRM). A fifth is running cost — scripts and the pronunciation lexicon change whenever the offer does. Some Polish vendors publish price lists, so comparison is possible, but only once every component is summed over twelve months including the overage minutes. Whether it is worth doing at all is decided by a different number: how many calls currently go unanswered. That one comes out of your own phone system in an afternoon.

Do callers have to be told they are talking to a machine?

Yes — always, in our builds. Article 50(1) of the AI Act requires that a person interacting with an AI system be informed clearly and at the latest at the first interaction, unless that is obvious from the circumstances and the context of use; it applies from 2 August 2026. On a phone line we do not rely on that exception — a caller has no way of knowing who will pick up — so we tell callers every time and treat that as our own rule, stricter than the legal minimum. In practice that is a sentence in the opening seconds of the call, not a line in the terms of use. The provision mainly concerns how the system is built: the bot has to be able to tell callers they are interacting with an AI system. Whether that message actually reaches the caller depends on the configuration at your end: the script, the greeting, and whatever the caller hears before the bot picks up. We build the notice into the script and the documentation. How the article allocates duties between the provider and the deployer (anyone using the AI system under their authority, for example a company or a public body, except for personal, non-professional use) in a specific setup is a question for your lawyer.

What about call recording and the GDPR?

A recording and its transcript are as a rule ordinary personal data — the legal basis, the retention period and what the caller is told all have to be settled before go-live, not after the first month. The category is decided by what is said on the call, not by the technology. One line where that category changes does run through the technology: extracting a voiceprint in order to identify or verify the speaker is special-category biometric data under Article 9 and needs its own Article 9(2) condition — and which condition applies is for your lawyer to settle. By default we build on the safer side of that line — transcription with no voice profile. Separately, count the external providers processing data on your behalf: usually three or four for a voice integration, one for text.

Will you build a voicebot that calls customers?

Not to lists that have not consented. Article 398 of the Polish Electronic Communications Law prohibits using automated calling systems to send commercial information without prior consent — business recipients included — and you may not call in order to obtain that consent. The President of UKE may impose a fine of up to 3% of the previous year’s revenue or up to PLN 1,000,000, whichever is higher, and the risk sits with the caller, which is to say with the client. Whether purely operational calls — an appointment reminder, an order confirmation — fall outside the prohibition is a question for your lawyer, not for us.