AI in sales: what it costs to transcribe all managers' calls, which metrics to measure before and after, and what the personal data law requires

What AI analyses and how it differs from listening

An hour of a call can only be listened to in an hour. So sales departments listen selectively, and whatever missed the sample stays unknown: which objections managers hear, whether they name the price, whether they agree a next step. AI is offered for five tasks.

1. Transcribing the call into text. The first and only step without which the rest is impossible.
2. A summary for the CRM. Three or four lines: what the client asked, what was promised, when the next contact is.
3. A check against a list. Whether need, budget and timing came up, whether a next step was proposed.
4. Collecting objections. What clients say when they decline: "too expensive", "I will think", "we already have a supplier". Frequencies are counted, not single calls.
5. The same for chats. Messengers and email are already text, no transcription is needed, steps 2–4 remain.

The value is in the fourth task. A call summary saves a manager minutes, while frequencies of objections across hundreds of conversations give the head of department what no selective listening can. But the risks are also biggest here, because it is about people's voices.

The path of one call: four checks before analysis

Technically everything is simple: a recording from the telephony → transcription → text analysis → result in the CRM. The difficulty is in the gates every call must pass before it is sent to third-party services.

What transcription costs

Transcription is the only cost item that can be calculated in advance. Prices of all transcription models from the OpenAI API pricing page on 16 October 2026; the hryvnia amounts use the National Bank of Ukraine rate for that date, UAH 44.8871 per dollar.

To analyse calls that are already recorded only the first three rows are needed: live transcription serves prompts to a manager during a conversation. How much this is per month for a department is shown by a conditional example: five managers, 20 calls a day, 4 minutes on average, 22 working days — 8,800 minutes, or about 147 hours of recordings a month.

Two caveats. First, this is only transcription. Text analysis, that is the summary, the check against a list and the collection of objections, is billed separately by tokens, and its cost depends on the chosen model and the length of the conversations; it is right to measure it on twenty transcripts rather than guess. Second, the price says nothing about accuracy: we did not measure the quality of Ukrainian speech recognition on noisy lines, and it may differ between models. Check on your own recordings (the first step below).

Comparison with a person: an hour of recording costs UAH 12.12 of transcription with gpt-transcribe, while listening to the same hour takes an hour of working time. If a head reads the transcript instead of listening, the time per conversation falls; by how much, measure yourself.

Separately from the API, for manually analysing a few transcripts in a chat assistant an ordinary plan will do. According to the ChatGPT pricing page on 16 October 2026, Plus costs $20 a month (≈ UAH 898), and Business is $25 per person billed monthly (≈ UAH 1,122) from two people, and on it data is not used for model training by default. Calls with clients' surnames should not be analysed on a personal plan: see data for an AI assistant.

Metrics before and after: what to measure

Without a "before" measurement a result cannot be told from an impression. We give no specific values: they depend on your department, and other people's numbers prove nothing. Fill in the "Before" column for the month ahead of launch and repeat afterwards.

The last metric is a control: if mistakes in prices and dates occur more often than in one call out of ten, the summaries for the CRM cannot be trusted. The "one in ten" threshold, like the others, is conditional and set by your tolerance for errors.

What the law requires: recording conversations as personal data

The voice and the transcript of a conversation with an identifiable person are personal data. We analyse only the Personal Data Protection Law in the edition of 14.06.2025. Whether additional bases are needed under other provisions on the confidentiality of conversations, check with a lawyer: that is not covered here.

• Basis for processing (Article 11). There are six: consent, performance of a contract, the controller's legitimate interest and others. Legitimate interest works only if the need to protect the person's fundamental rights does not outweigh it. Consent is not the only basis, so "the client did not say yes" does not by itself mean a ban.
• Notice at the time of collection (Article 12). The data subject is informed about the controller, the composition of the data collected, their rights, the purpose of collection and those to whom the data is passed. For an incoming call that means a phrase about recording at the start, with details available via a link or on the website.
• Purpose and scope (Article 6). Processing must be open and transparent, and the composition of data adequate to the purpose. A recording made for sales quality control may not later be used for something incompatible without a new basis.
• Contract with the processor (Article 4). The transcription service is a processor here: a written contract is needed, and it may process data only within the purpose and scope set by the contract.
• Transfer abroad (Article 29). Transfer is possible provided adequate protection in the receiving state; the list of states is set by the Cabinet of Ministers, and states of the European Economic Area and parties to the relevant Council of Europe convention are deemed to provide it. Find out where exactly the vendor processes data before connecting.
• Health (Article 7). If a client names a diagnosis or a drug during the conversation, that fragment is a special category of data. Ordinary sales are not affected, but pharmacies, clinics and insurers are.
• Managers are data subjects too. An employee's voice is their personal data, so they too must be told about the recording and analysis of their calls.

What not to do

• Do not hand the model decisions on bonuses or dismissals. Summaries and frequencies help a conversation with a manager, but transcripts contain errors, and a model does not see context the head knows.
• Do not switch recording on without a phrase at the start of the call just because "everyone does".
• Do not paste transcripts with surnames into a personal chat assistant. For that there is the business plan or, better still, anonymisation by client code.
• Do not mix purposes. A recording for quality control should not later be used for marketing mailings to that number.
• Do not trust prices and dates in a summary without checking. These are the fields where an error costs money, so they are checked against the transcript.

First step for one week: twenty calls

1. Check the first item of the scheme. At the start of a call a phrase about recording and its purpose must be heard, and the website must have a page with the terms. Without this, go no further.
2. Take twenty calls of one manager from a quiet week, up to five minutes each. If they contain data from item 3 of the scheme, replace those calls.
3. Transcribe them. For 20 calls of 4 minutes (80 minutes) that is UAH 16.16 with gpt-transcribe, UAH 10.77 with gpt-4o-mini-transcribe and UAH 21.55 with gpt-4o-transcribe at the 16 October rate.
4. Check the transcript against the recording on five calls: prices, dates, names. Count calls with errors.
5. Ask for a four-line summary and compare it with what the manager wrote in the CRM. If the model's summary is more accurate in most cases, move to a pilot with one manager.

Decide on a launch for the whole team only when the metrics table is filled in the "Before" column.

Our position

Transcribing calls pays off when the head really reads the result, not when it has merely been switched on. It costs pennies next to a person's hours, so the question is not "is it expensive" but "are the gates there": notice to the client, a contract with the vendor, exclusion of sensitive conversations and a person who gives the rating. Without them the saving turns into a risk.

The position does not apply to a department with a few calls a day where the head already listens to all of them: transcription adds little. It also does not apply where clients by the nature of the conversation give medical or payment data, as in clinics, pharmacies, insurance and lending: there a legal scheme of your own is needed first, not a subscription.

Prices checked on 16 October 2026 on the OpenAI API and ChatGPT pricing pages; exchange rate — National Bank of Ukraine on 16 October 2026; provisions — against the text of the Personal Data Protection Law (edition of 14.06.2025) on zakon.rada.gov.ua. Call volumes, durations and error thresholds are conditional; substitute your own. This article is not legal advice.