Key takeaways
- The best way to use meeting transcripts with AI is to hand over the full wording as context and ask specific questions, instead of just requesting a summary.
- AI gets good through context, not clever prompts. The complete wording of a conversation is the maximum context.
- After two weeks you remember the mood of a meeting, but not the commitments, objections and figures. The transcript captures them.
- Record only with everyone's consent, weigh sensitive conversations separately and always have a human check what the AI produces.
The best way to use meeting transcripts with AI is to give it the complete wording as context and then ask specific questions: what was agreed, which objections came up, what's still missing. That's how I, Felix Degeler, founder of Degeler Consulting (DC) in Vienna, work with Claude. The full wording beats even the cleverest prompt, because the AI no longer has to guess.
I record every important conversation. I've never listened to a single one. Yet every one of these recordings is worth its weight in gold, and that's not because of the recording itself but because of what AI makes of it.
Why isn't your memory enough after two weeks?
Because you remember the mood of a meeting, not its details. Two weeks after a good meeting you still know it was good. But who committed to what, which objection came up, which figure was mentioned? Gone.
Exactly these details later decide whether a project keeps running smoothly. A forgotten commitment turns into a misunderstanding, an objection nobody heard turns into a conflict at the next meeting.
The transcript captures them. Not as minutes nobody reads, but as raw material you can draw on at any time.
What is context in AI?
Context is everything an AI knows about a task before it answers: the question itself, the documents you hand over and the conversation so far. Without that context, a language model fills the gaps with whatever is most likely, and that's how answers come about that sound plausible and are still wrong.
Context engineering is the work of giving AI the right information in the right form, instead of polishing the wording of a prompt. The question is no longer „How do I tell the AI?“ but „What does the AI need to know to answer well?“
AI doesn't get good through sophisticated prompts but through context. And there is no more context than the complete wording of a conversation. Instead of giving the AI three keywords to guess from, it gets everything that was actually said.
How do I use meeting transcripts with ChatGPT or Claude?
In four steps: record with consent, transcribe, hand the transcript to the AI, ask specific questions. The approach works the same with ChatGPT, Claude and other language models, the tool is secondary.
Step 1, record with consent: Say at the start that you're recording and why, and get everyone's okay. That also makes the purpose clear: the recording is a working basis, not evidence. Why „please don't record“ usually means something else is covered in „Please don't record“: why that's usually a trust problem.
Step 2, transcribe: Many video conferencing tools deliver a transcript, alternatively transcription runs locally on your own computer. At DC we use our own local setup with speaker recognition, so the text shows who said what. Only that attribution makes questions like „Who committed to that?“ answerable.
Step 3, hand over the transcript: Give the AI the complete transcript, not a summary of it. Add two sentences of framing: who the participants are and what the project is about. The less the AI has to guess, the better the answer.
Step 4, ask specific questions: Ask concrete questions instead of „Summarise this“. A summary smooths things over, a specific question pulls out exactly the passage you need. The answers then rest on the conversation itself, not on my patchy memory. How we turn that into tasks automatically is described in Nobody writes meeting notes here anymore.
Which prompts work for meeting transcripts?
The best prompts for transcripts are plain, concrete questions whose answer the AI can back up with a passage from the conversation. These six questions work in practice:
First: „What exactly was agreed, and who took on which task by when?“
Second: „What concerns or objections did the other side raise, and how were they addressed?“
Third: „What's still missing for me to write a proposal?“
Fourth: „Which figures, dates and names were mentioned? Quote the passage in the transcript for each.“
Fifth: „Where did we talk past each other or leave something open?“
Sixth: „Write a follow-up email to the participants that only contains points that were said that way in the conversation.“
Asking for quoted passages is the most important ingredient. You see immediately whether the AI is relying on the conversation or has made something up.
Does this work for your own texts too?
Yes, the same principle works for content: talk first, then let it be written. You already have the thoughts on a topic in your head, so just say them out loud. No clean sentences needed, no grammar, the AI doesn't care. It turns the raw material into a structured text.
The LinkedIn post this article grew out of was written exactly that way. Your meetings produce material of this quality every day. If you don't record it, you throw it away unread. More on why speaking gives AI more than typing: Dictate instead of type.
When shouldn't you give transcripts to an AI?
When not everyone involved has consented, when the conversation is particularly sensitive or when you don't know where your AI processes the data. In those cases, less context is the better choice.
Consent and GDPR: A recording requires that the participants are informed and agree. In Austria, secretly recording other people's conversations is even a criminal offence (§ 120 StGB). This isn't legal advice, but the basic rule is simple: whoever doesn't know they're being recorded doesn't get recorded.
Sensitive conversations: Performance reviews, health topics, conflicts or confidential negotiations don't belong in an AI tool without careful thought. Here you decide case by case whether to record at all.
Cloud AI or local transcription: With a cloud AI, the content leaves your house. Check where the provider processes the data, whether it's used for training and whether there's a data processing agreement. Local transcription at least keeps the audio file on your own computer, and the text only goes into an AI when you decide so.
Transcription errors: Speech recognition reliably trips over names, company names and technical terms. A product name becomes a similar-sounding everyday word, one figure turns into another. So give the AI the correct spellings and check figures against the original.
A human checks: The AI delivers the draft, responsibility for commitments, emails and proposals stays with you. Why that's a fixed rule for us is explained in Human in the loop.
If you want to set up a workflow like this in your team, from recording to finished task: That's what we do with mid-sized companies.
