Automatic meeting transcription is one of the few AI use cases in SMEs that doesn't need a six-month rollout and actually pays for itself in weeks. Provided you know where it works and where it does more harm than good.
For the past year we've been getting the same question from every second client: "I saw Teams can now take meeting notes. Does it work? Should we turn it on?". The short answer is yes, but without understanding three things you'll turn that time saving into a data leak and a pile of useless transcripts.
This post isn't about which app to click. It's about when it's worth turning on at all, on which meetings, and what to do so that accounting doesn't find a transcript of a client call in an indexed OneDrive folder.
What you actually save
Let's start with numbers, because the noise around AI often drowns out the real ones. At one of our clients — a 22-person consulting firm — we rolled out automatic meeting transcription (Teams + Copilot) in a team of 5 analysts. Each of them runs or joins 3–4 client meetings a day. Before, after every meeting they wrote up notes, decisions and to-dos — in Word, roughly 15–20 minutes per meeting. Now: the transcript and summary generate automatically, the analyst spends 5 minutes reviewing, fixing name mistakes and adding the decisions that weren't stated outright but were agreed "between the lines".
The saving came out to 45 minutes per person per day. Five people × 45 minutes × 20 working days = 75 hours per month. At an analyst's hourly rate that's a visible gain. But more importantly: notes are finally being written consistently. Before, every third meeting ended without notes because "there was no time". Now every one has a summary.
That's the right way to size the return on such a tool — not "AI revolutionised our process", but "75 hours a month of typing simply vanished". Boring, but true.
Where transcription doesn't work
AI transcription is good for meetings with one topic, where speakers take turns and stay reasonably on track. In practice: project status updates, kickoffs, 1:1 sales calls, technical reviews, hiring interviews.
It doesn't work for three types of meeting:
- Workshops and brainstorms. Four people talking at once, someone drawing on a whiteboard, threads paused and resumed. The transcript is a wall of nonsense text.
- Quick decisions in the kitchen or hallway. Not every call and not every 5-minute chat needs a written record. Turning on recording by reflex kills the culture of open discussion — people start self-censoring.
- Sensitive meetings. Terminations, staff conflicts, disciplinary issues, salary negotiations. Keeping a written record of those carries an evidentiary risk that few people think about at the moment they click "Record".
Our practical rule: transcribe operational meetings by default (projects, clients, technical); do not transcribe "human" meetings by default (HR, conflict, internal politics).
Data leak risks
Here's the part AI vendors don't talk about loudly, and which hurts most in a Polish SME. A transcript of a client call typically contains more personal data and trade secrets than many a document sitting in a password-protected folder.
Three questions worth asking before turning transcription on company-wide:
- Where the recording and transcript physically end up. Microsoft 365 Business by default puts them in the meeting organiser's OneDrive or the team's SharePoint. Third-party tools (Otter, Fireflies, Fathom) often send the recording to their own servers — sometimes outside the EU. For data protection and clients under NDA that's a problem.
- Who has access to the transcripts. By default in Teams, that's the organiser and everyone invited. If an external consultant was on the call, they'll also have the transcript in their account. It's good practice to auto-delete client meeting transcripts after X days.
- Whether the AI vendor uses this to train the model. Business and enterprise tiers (M365 Copilot, Google Workspace Duet) formally don't. Free consumer accounts (e.g. personal ChatGPT) — no guarantee. This matters when an employee independently pastes a transcript into a free tool to summarise it.
The practical minimum we roll out at clients: a short one-page policy stating which meetings we don't transcribe, where we keep the transcripts, and after how many days they're deleted. Without this, sooner or later you'll find a sales-call transcript surfaced in enterprise search alongside the rest of your SharePoint.
Choosing a tool
Almost any SME that already has Microsoft 365 should start with what they've got. Teams with Copilot ships transcription and summaries as standard, data stays in the company tenant, and there's no extra SaaS to buy. Same story for Google Workspace with Duet AI for companies on Google.
Sensible alternatives make sense in three situations: you run lots of meetings on Zoom (native Zoom AI Companion), you need transcription from external phone tools (e.g. a sales call from a landline), or Copilot is too expensive and you're looking for a cheaper option.
Watch one thing: transcription quality in Polish is significantly worse than in English. Copilot, Otter, Fireflies — all trained mainly on English data. In practice: 90% accuracy instead of 97%, made-up client names, misheard product names. Always assume someone has to review the transcript — that's why in our consulting example the analyst spends 5 minutes on cleanup, not zero.
Practical rollout
Three steps, in this order:
- Turn it on first for one team, for one month. Without announcing it company-wide. See how much they actually save, where it stumbles, what people do with the transcripts. Gather feedback.
- Write a one-page policy. Which meetings yes, which no, where they're stored, after how many days they're deleted, who tells the client at the start of the call. Without this document the rollout drifts within a quarter.
- Set retention and permissions. Microsoft 365 has sensitivity labels and retention policies for exactly this. By default, client meeting transcripts should be kept for 30–90 days, not forever.
The biggest mistake we see is switching transcription on for everyone in the company on a Friday afternoon, with no policy. Three months later there are 4,000 transcripts in SharePoint, nobody knows which ones matter, and one of them contains a verbatim record of a salesperson promising a client a discount the board never approved. That story doesn't end well.
Wrap-up
AI meeting transcription is one of those use cases where a few days of preparation is well spent, and where you shouldn't expect a miracle. The real return is usually 30–90 minutes per day per person who spends a lot of time in operational meetings — a lot, once you annualise it. But this isn't an "AI revolution", just cleverly automated grunt work someone had to do by hand anyway.
If you're wondering whether to turn on Copilot in your company, or already have it but there's no policy and nobody knows what to do with it — we can help sort it out in a few hours, not a quarter.
AI meeting transcription isn't a revolution. It's automated grunt work that has been silently eating half an hour a day from analysts, salespeople and consultants for years.