Most meetings do not fail because of what happens in the room. They fail because the preparation was thin and the follow-up did not happen. AI is well-suited to both ends: drafting a tight agenda before the meeting, and converting a messy transcript into something actionable after it.
Done well, it saves time every week and produces better records than most people were keeping anyway.
The tool, and how to set it up
For transcripts: You have two routes. Tools like Teams Copilot, Zoom AI Companion, Otter, and Fireflies transcribe automatically and produce summaries in-app. Convenient, but you get whatever the product decided to generate. You cannot tell it to format actions as owner-plus-deadline, flag open questions separately, or draft your follow-up in a particular tone.
The alternative is to export the raw transcript and paste it into Claude directly. One extra step, full control. For anything complex or multi-party, this is the better option. This guide uses that approach. If you are using a built-in tool, the QA section still applies.
For agenda prep: No transcript needed. Paste your context into Claude before the meeting.
How to do it
Before the meeting:
- Collect the relevant context: the purpose, who is attending and what decisions need to be made, any background documents.
- Paste that into Claude and ask for a tight agenda with time allocations and a stated outcome for each item. Ask it to flag anything that could be handled async.
- Read the agenda as a participant. Is there anything that does not need the whole group? Anything missing that will come up anyway? Adjust before you send it.
After the meeting:
- Get the transcript. Most tools let you export plain text or VTT. Copy the full text.
- For long transcripts, paste in sections or use a Claude Project where context persists across messages.
- Run the prompt below. Read the output before sharing.
- Draft your follow-up email from the output, not from memory.
The prompt
Here is a transcript from a meeting. Please extract the following, in this format:
**Decisions made:**
- [Decision, with any relevant context]
**Actions:**
- [Action item], Owner: [name], Due: [date or "not specified"]
**Open questions (not resolved in the meeting):**
- [Question, and who is responsible for resolving it if mentioned]
**Suggested follow-up email:**
Write a follow-up email to all attendees summarising the above. Plain language, no filler. Include all decisions and actions. Tone: [professional / informal / formal, choose one].
Transcript:
[paste transcript here]
Notes: Names in the transcript may be abbreviated or appear as speaker labels. Where a name is unclear, flag it with [unclear] rather than guessing. If a deadline was implied but not stated, note it as "implied, confirm with owner."
The last paragraph matters. It tells Claude to surface uncertainty rather than fill gaps silently, which is where most transcript AI goes wrong.
How to QA it
Read the output against the transcript, not just on its own.
Actions and owners. The most common failure is an action assigned to the wrong person, or one invented from the general direction of the conversation rather than what was actually said. For each action item, find the moment in the transcript where it was agreed. If you cannot find it, the item is suspect. "Someone mentioned it would be good to..." is not a commitment. "Jana said she would send the revised proposal by Thursday" is.
Decisions implied but not stated. Meetings often land on decisions without anyone saying "so we have decided." The AI may miss these, or list them as open questions when they were actually resolved. Look for moments where the conversation moved on and ask whether that implied agreement.
Names, figures, and dates. Check every proper noun and number. A transcript with similar-sounding names, or offset speaker labels, can produce confident mis-attributions. If someone's name appears as "Speaker 2," fix it before the follow-up email goes out.
Read the email draft as if you were receiving it. Would any attendee say "that is not what we agreed"? Fix it before you send.
How to stay safe
Get consent before recording. In the UK and most of Europe, recording without informing participants breaches data protection obligations. "This meeting may be recorded" is the minimum. For sensitive conversations, get explicit agreement.
Transcripts contain more than task data. They often include personal opinions, HR-sensitive comments, salary discussions, or commercially confidential content. Before pasting a transcript into a consumer tool, ask whether that content belongs in a consumer cloud product. The answer is often no. Use your organisation's approved tools, a Claude for Work deployment, or Copilot via Microsoft 365 for anything sensitive. If in doubt, paste only the sections you need.
AI-generated minutes are not the official record. If a meeting output has any legal or contractual weight, a human needs to review and sign off before it goes out. Do not send the AI output directly as "minutes of the meeting."
Check what you committed to. Run the prompt and search for your own name in the actions list. It catches things that get lost in a busy day. Cross-check against the transcript before acting on anything.
Start with your next meeting that has more than three people in it. Draft the agenda the evening before using the context you already have. After the meeting, paste the transcript and run the prompt. The first time you catch a mis-remembered action item before it causes a problem, it will become a habit.