How many customer calls did you have last week, and what was promised in each one? In many small companies, the answer lives only in someone's memory. Automatic speech to text transcription has become so cheap that it is worth a serious look, as long as you use it sensibly.
What changed last week
On October 1, Microsoft launched MAI-Transcribe-2-Streaming, a model that turns speech into text in real time, with automatic language detection for 60 languages (Microsoft AI, TestingCatalog). The introductory price is 0.54 dollars per hour of audio, valid until the end of 2026 (TestingCatalog).
In plain terms: one hour of transcribed conversation costs less than a coffee. The model is built for developers, so you do not install it yourself like an app, but prices like this quickly reach the tools you already use (cloud phone systems, meeting apps, CRMs). Worth knowing: it is still a test version (preview), with no uptime guarantees, so it is not for critical processes until the final release (Tech Times).
Where a small company gains
A few situations where transcription saves real time:
- Sales calls. Instead of scribbling notes, you have the full text, and AI pulls out what the customer asked for, the budget and the deadline. The quote goes out faster and without "I forgot to ask".
- Internal meetings. The meeting text becomes a summary of who does what and by when. No more 20 minutes spent writing the minutes.
- Complaints. You have the customer's exact words, not the version retold by the colleague who answered. That helps with fixing the problem and with training the team.
- Field visits. A technician dictates findings in the van, and the text lands in the customer record.
A realistic example: an installation company with three sales people makes 40 calls a week. If each call needs 5 minutes of handwritten notes, that is more than 3 hours a week just for notes. With transcription plus an automatic summary, the person only checks the summary.
What the announcement does not tell you
Accuracy figures in announcements are measured in lab conditions. A study cited by Tech Times shows that commercial systems claiming under 5% errors went above 13.73% errors in every real world domain tested (Tech Times). So on a call with background noise, an accent or technical terms, there will be wrong words.
Three more things to check before you rely on a tool:
1. Whether your language is really covered well. The announcement mentions 60 languages, but the detailed list does not appear in press materials (TestingCatalog). Test with your own calls, not with the vendor's demo.
2. Whether it can separate speakers. The real time version does not promise to label who said what, unlike the version that processes recordings (Tech Times). For a customer call, that matters.
3. What it costs after the trial period. The 0.54 dollar price is introductory, and the final price has not been published (Tech Times). Do not build a whole process on a price that may change in January.
Data caution: the recording is the sensitive part
Transcription means recording, and a person's voice and what a customer says on the phone are personal data. The GDPR requires you to tell people what data you collect and why (GDPR, Art. 13). In practice:
- Announce at the start of the call that it is recorded and why ("so we do not lose the order details"). Put the same thing in your privacy policy.
- Do not transcribe what should not be there. Card details, national ID numbers, passwords and medical details have no place in the text. Train the team to pause recording or not to ask for such data by phone.
- Choose providers that process data in the European Union and check whether they use your recordings to train their models. If it is not clearly stated, ask in writing.
- Delete regularly. Keep the summary in the customer record and delete the audio after a set period, for example 30 or 90 days.
- Anonymise when you pass the text on to a chatbot for analysis: replace names and phone numbers with "Customer A" or "Company X".
There is no need to be scared, but you do need a written rule, on one page, that the whole team knows.
Conclusion: one step for this week
Pick a single type of conversation, for example the Monday meeting or quote calls, and test transcription for two weeks with a tool you already have (many meeting apps include it). Measure two things: how many minutes of notes you saved and how many errors you had to correct. Only then decide whether to roll it out further.
If you want to connect transcripts directly to your CRM or quotes, I have written about automating price quotes and about what data you can give an AI assistant.