You just walked out of a one-hour meeting. Around 9,000 words were spoken (based on 150 words per minute), and you remember maybe three of them. Transcribing a meeting with AI means turning that audio into text in a few minutes, then asking an assistant to produce structured minutes: decisions made, action items, owners, deadlines. No technical skills required, and two of the methods covered here cost nothing at all.
This guide gives you four methods, from the most complete to the simplest, the exact prompt that makes the difference between a vague summary and minutes people can actually act on, and the rules to know before you hit record (participant consent isn't optional). Each method is laid out step by step, with its real limitations: maximum length, quality in a six-person meeting, and how it handles accents and industry jargon.
Method 1: Do Everything in the Skilzy AI Lab
The most complete approach is to chain transcription and minutes-writing in a single environment, with the credits already included, instead of juggling three separate subscriptions. That's the idea behind the AI Lab built into Skilzy: you access real AI tools straight from the platform, you learn the method while you're using it, and you don't have to pay for a separate subscription at every step.
Here's how it works, in five steps:
- Record the meeting. A phone placed in the middle of the table is enough for six people in a normal room. For video calls, start your meeting tool's recording.
- Grab the audio file. Universal formats: MP3, M4A, WAV.
- Transcribe the audio into text. You'll get a raw block of text, with imperfect punctuation and the occasional misheard word. That's normal, and it won't cause problems later.
- Apply the minutes prompt (the one in section 5 of this article). The assistant cleans up, structures, and pulls out the decisions and action items.
- Proofread and fix. Budget two to three minutes. Check the numbers, the proper nouns, and the deadlines: those are the three places where AI gets it wrong most often.
For a one-hour meeting, the whole thing takes around 10 to 15 minutes, compared with 45 to 90 minutes of writing it up by hand.
One thing to know before you sign up: two state-recognized certifications are eligible for funding (RS7439 for AI-assisted content marketing, RS6792 for AI and sales) if you want to go further than meeting minutes.
The real value for a beginner isn't the tool: it's the method. Knowing how to frame the request is 80% of the result. If you want a broader overview before choosing, start with our guide to learning AI as a beginner.
Method 2: The Transcription Already on Your Phone (Free)
Your phone can probably transcribe a voice recording into text already, for free and without any extra app. It's the fastest option for a short meeting, and it would be dishonest not to mention it.
On iPhone (iOS 18 and later): open the Voice Memos app, record, then tap the transcription icon at the top of the recording. The text appears, and you can copy and paste it into an AI assistant.
On Android: the Recorder app on Pixel phones transcribes live, including offline. On Samsung, the Voice Recorder app has a Speech-to-text mode. On other brands, open the Google Docs app, then Tools, then Voice typing.
What this method does well: it's free, instant, and the audio stays on the device in most cases.
Its limitations, plainly stated:
- It doesn't tell speakers apart. You get a block of text with no idea who said what.
- Quality drops fast if several people talk at once or the room echoes.
- Past 30 or 40 minutes, the text becomes hard to use as is.
- It produces no minutes at all. You still need an AI assistant for the synthesis step.
Method 3: Your Video Conferencing Tool's Built-In Transcription
If your meeting happens on Teams, Google Meet, or Zoom, transcription is already built in and it identifies speakers by name. That's an advantage the other methods don't have, because the tool already knows who's connected.
- Microsoft Teams: during the meeting, go to the More actions menu, then Start transcription. The text is timestamped and attributed to each participant, then archived with the meeting.
- Google Meet: transcripts are available depending on your organization's Google Workspace edition, and the document lands in the organizer's Drive.
- Zoom: audio transcription is enabled in the account settings, under cloud recording, on paid plans.
One thing to watch out for: these features depend on your work subscription and on what your IT administrator has allowed. On a free account, they're often unavailable. Check before the meeting, not during it.
Once you have the transcript, the work isn't done. A one-hour Teams transcript easily runs to 40,000 characters, unreadable as is. You still need to run it through an AI assistant for the synthesis.
Method 4: Dedicated Third-Party Tools
Tools built specifically for transcription are the most accurate on long and multilingual meetings, but they add yet another subscription or require a technical setup.
Three families, depending on your profile:
Consumer online tools (Notta, Tactiq, Fireflies, Happy Scribe). You upload a file or connect them to your video calls. They tell speakers apart and often offer an automatic summary. Almost all of them have a free plan limited to a certain number of minutes per month, then a monthly subscription. Heads-up: your audio goes to their servers. Read where the data is hosted before you send a confidential meeting over.
Whisper, the open-source model. Released under an MIT license, Whisper is publicly documented on GitHub and installs on your own computer. Free, with no length limit, and the audio never leaves your machine. In exchange, you have to use the terminal and accept processing times that depend on your hardware. It's the best option if confidentiality is non-negotiable and a command line doesn't scare you.
Assistants that connect to your calendar and join meetings on their own. Handy, but a bot that invites itself into a meeting without anyone announcing it raises a real consent problem. Avoid it without the team's explicit agreement.
Quick Comparison of the Four Methods
| Method | Entry cost | Identifies who's speaking | Structured minutes | Skill level |
|---|---|---|---|---|
| Skilzy AI Lab | Monthly subscription | Via the prompt and your notes | Yes, with the method provided | Beginner |
| Phone (built-in) | Free | No | No | Beginner |
| Video conferencing tool | Included depending on work plan | Yes | Partial | Beginner |
| Third-party tool / Whisper | Limited free plan or install | Yes | Often | Intermediate |
The Prompt That Turns 9,000 Words into Readable Minutes
Good minutes don't come from the transcription tool but from the prompt you write afterwards: ask for a precise structure and named action items, never just „summarize this meeting". „Summarize" gives you a lukewarm paragraph nobody will read. Here's a template to copy and paste above your transcript:
You are writing the minutes of a meeting from the raw transcript below. The transcript contains speech recognition errors: fix them when the meaning is obvious, and flag uncertain passages in square brackets.
Produce exactly these five blocks:
- Context in 3 lines maximum (purpose of the meeting, identifiable participants).
- Decisions made: one bullet per decision, phrased in the past tense and affirmatively.
- Action items: a table with Action, Owner, Deadline. If the owner or deadline isn't stated in the transcript, write „to be confirmed". Don't make anything up.
- Open questions: topics raised without a conclusion.
- Useful quotes: a maximum of 3 short verbatim quotes worth keeping as they are.
Constraints: 400 words maximum in total, factual tone, no evaluative adjectives. Ignore off-topic digressions.
Transcript: [paste here]
Three tweaks that change the result:
- Give the list of participants and the industry jargon before the transcript. The AI will write „GDPR" instead of „gee dee pee are", and get the proper nouns from your field right.
- Explicitly forbid making things up. The instruction „don't make anything up, write to be confirmed" cuts down a lot on action items assigned to the wrong person.
- Split it up if the meeting runs past 90 minutes. Process the transcript in two or three chunks, then ask for the partial minutes to be merged.
Which assistant you pick matters less than the prompt, but it does matter a little. Our comparison of ChatGPT, Claude, and Gemini on their free tiers breaks down which one handles long texts best. And if you want to practice without spending anything, we've rounded up 15 free resources.
Before You Hit Record: Consent and Data
Recording a work meeting without telling the participants is illegal in France, whatever transcription method you use. Three habits are enough to stay on the right side of the rules.
Announce the recording and what it's for. One sentence at the start: „I'm recording this so I can write up the minutes, and the audio will be deleted once they're approved. Is everyone okay with that?" If someone says no, you take notes instead. Participants must be informed before you start, and must be able to object.
Limit how long you keep it. Keep the minutes, delete the audio and the raw transcript once the document is approved. Data minimization is at the heart of the GDPR as the CNIL explains it: you only keep what's necessary, for as long as it's necessary.
Be careful with sensitive meetings. One-on-ones, health data, salary information, confidential client files: in those cases, either stick to a local solution (phone in offline mode, Whisper on your own machine) or don't record at all. A free online service is not the right place for a performance review.
One last point people often forget: also mention that the minutes were written with help from AI, and have someone who was there read them over. Minutes approved by a human commit their author, not the tool.
Key Takeaways
Four methods, one principle: transcription is the easy part, the minutes come from the prompt. Start with whichever method is most accessible in your situation, test it on a short, low-stakes meeting, then tune the prompt until you get a document your colleagues actually read.
