AI voicebots in customer service: what they can do, what they cost and what the law requires

A traditional helpline menu (IVR) asks the customer to choose a number: “for orders, press 1”. A voicebot lets them describe the matter in their own words, for example “I want to know when my parcel will arrive”. It recognises speech, understands what the caller wants and replies by voice.

The real difference begins when the voicebot has access to the company’s systems. It can then check the order status in the ERP system, free slots in the service calendar or open a case in the complaints system. Without such access, all it can do is take a message and promise a call back.

A voicebot is not infallible. It may mishear an order number, a surname or a place name, especially in a noisy setting or on a poor line. That is why a well-designed voicebot repeats important details and asks for confirmation.

The best candidates are calls that are frequent, short and based on data from a system. The worst are matters where the customer is upset, the situation is unusual or something has to be negotiated. The table below will help you assign the calls to your helpline to these groups.

Take a hypothetical distributor whose helpline receives 200 calls a day, or about 4,200 a month over 21 working days. A call lasts 3 minutes on average. An analysis of a sample of calls showed that 45% concern order status and delivery dates – about 1,890 calls a month.

Let us assume that the voicebot resolves 60% of them on its own, or 1,134 calls, and transfers the rest to an agent after about a minute. This percentage is an assumption that must be tested in a pilot. Agents then get back 1,134 × 3 minutes, or about 57 hours a month. At a full hourly cost of PLN 70, that is about PLN 3,990.

Over the same period the voicebot will handle 1,134 × 3 + 756 × 1 = 4,158 minutes of calls. At the ElevenLabs rate of USD 0.08 per minute, the voice layer alone costs about USD 333 a month. In Copilot Studio the same minutes come to 4,158 × 35 = 145,530 credits, or about USD 1,164 at a pack price of USD 200 for 25,000 credits.

On top of this come telephone connection charges, the cost of the language model depending on the set-up, and above all building the integrations and maintenance. At an exchange rate of PLN 4 to the dollar, the saving of PLN 3,990 is about USD 1,000. It therefore exceeds the cost of minutes with ElevenLabs, but in the Copilot Studio option the cost of minutes alone is comparable to the saving or higher. The calculation improves with out-of-hours calls that now go to voicemail and with peak times when customers hang up in the queue. Calculate these separately on your own data.

A voicebot consists of three elements: speech recognition, a language model that understands and composes the reply, and speech synthesis that speaks it. A weak link in any of them spoils the whole conversation. The customer does not know which element failed – they just hear that “the machine doesn’t understand”.

Polish poses several challenges: surnames and place names change their endings depending on grammatical case, different people read order numbers in different ways, and there are industry abbreviations. Before choosing a supplier, ask for a test on your own real examples, not on a demo call. Also check how long the voicebot stays silent before answering, and whether it can be interrupted mid-sentence.

During the test, pay attention to the situations below. These are what most often decide whether customers will accept the voicebot.

Most of the implementation work is not about the voice but about connecting the voicebot to the company’s systems. This determines whether the customer gets an answer straight away or only a promise of a call back. Before talking to a supplier, check which of the connections below are needed and possible in your systems.

Transferring the call to an agent together with a description of the matter is particularly important. If the customer has to repeat everything, the voicebot has made the call longer instead of shorter.

From 2 August 2026, Article 50 of the AI Act requires that a person talking to an AI system knows this. In practice the voicebot says so in one sentence at the start of the call. This is a legal obligation, not just good practice. In Poland, the AI market is supervised by the Commission for the Development and Safety of Artificial Intelligence (KRiBSI), which can carry out inspections and impose penalties from 28 October 2026.

If the voicebot records calls or stores their content, the GDPR applies. At the start of the call the customer must hear who is processing their data and for what purpose, and where to find the full information, for example on a website. Set the retention period for recordings in advance, and check where the supplier stores them.

Recognising a customer by the sound of their voice alone is processing of biometric data. As a rule the GDPR requires explicit consent for this, so it is usually simpler to confirm identity with a customer number or a code sent by text message. If the voicebot were to assess the caller’s emotions, the AI Act imposes an additional information obligation – discuss this feature with a lawyer before implementation.

Your brand’s own voice is a separate issue. If the voicebot is to speak in the voice of a specific person, such as a voice-over artist or an employee, you need their consent. You also need a contract setting out how long and for what purpose you may use their voice.

A voicebot can also make calls itself: reminding customers of appointments, confirming delivery dates, collecting feedback on service. These calls often bring quick results, because they reduce the number of missed deliveries and no-shows.

Commercial offers are a different matter. Article 398 of the Electronic Communications Law (PKE) prohibits the use of automated calling systems to send commercial information without the recipient’s prior consent. This also applies to business customers.

The line between a reminder and an offer can be thin. A reminder about a service inspection that also proposes an additional service may already count as commercial information. So it is worth reviewing your outbound call scripts with a lawyer before the voicebot starts calling.

The pilot should answer one question: what percentage of the selected calls does the voicebot resolve correctly without a human? The calculation in the earlier section depends on this number. The steps below let you check it with limited risk.

The percentage of calls handled by the voicebot is misleading on its own. A voicebot that takes a long time to transfer to a human can score highly while putting customers off. So look at several indicators at once, as listed below.

The most valuable source of insight is recordings of calls that ended in a transfer or a hang-up. They show where the script fails, and what customers are looking for that the voicebot cannot provide.