# AI Meeting Notes: Turn Transcripts into Decisions and Action Items

> Use AI to turn a meeting transcript into clear minutes, with evidence for decisions, confirmed owners, unresolved questions, and a reviewable action list.

Useful AI meeting notes separate decisions, proposals, actions, and unresolved questions. Start with an approved transcript, request evidence beside every decision and action, and leave missing owners or deadlines unassigned. A participant reviews the draft before it becomes the team's record.

Useful AI meeting notes separate confirmed decisions, proposals, actions, and open questions. Give the assistant the supplied record, require evidence beside each item, and leave missing owners or deadlines visible. A participant then checks the draft before it becomes the team's shared minutes.

The hard part is interpreting commitment. “We could launch Friday” differs from “We agreed to launch Friday.” “I can review it once the list arrives” does not establish an unconditional review deadline. A concise summary can lose those differences unless you ask for them explicitly and inspect the result.

## Know which record you are working from

These are working definitions for this guide, rather than a legal standard:

| Record | What it gives you | What to watch for |
| --- | --- | --- |
| Transcript | A captured record of spoken words | Missing segments, uncertain speakers, and transcription errors |
| AI summary | A selected and compressed interpretation | Omitted conditions or a proposal presented as agreement |
| Minutes | The team's reviewed record of decisions, actions, and questions | Whether participants approved the consequential details |

Start from the most complete authorized record available. Notes are usable input, but label them as notes; do not ask the assistant to present them as a verbatim transcript. Keep attachments and meeting-chat decisions separately identified when you supply them.

For example, Google Meet's transcript feature captures spoken words, not chat messages. Restarting transcription creates separate files; the transcript is saved in the organizer's Drive and linked through the Calendar event. Check that you have every segment. Availability depends on the workspace. [Google's transcript documentation](https://support.google.com/meet/answer/12849897?hl=en)

Meet's separate note-taking feature produces a Google Docs recap. Google acknowledges that summaries can be incomplete or inaccurate, and configured recipients can include invited guests who did not attend. Admin settings and host controls affect use. Check the recap and its recipients before treating it as the final record. [Google's note-taking documentation](https://support.google.com/meet/answer/14754931?hl=en)

## Prepare a small input packet

Use your organization's approved recording and upload process. Include the necessary meeting record in an approved assistant, with a short header:

```text
Meeting title:
Meeting date and time zone:
Attendees and speaker-name map:
Source type: transcript / participant notes / AI recap
Source location and version:
Known gaps or uncertain speakers:
Additional supplied records: chat excerpt / attachment / prior decision
Reviewer:
Intended recipients:
```

The header helps the reviewer distinguish an input gap from an extraction mistake. If the transcript begins halfway through, say so. If “Speaker 2” might be either of two people, preserve that ambiguity until someone confirms the speaker.

ChatGPT can use attached source files to create reviewable output. State the structure and review criteria you need; do not assume that uploading plain text adds timestamps or verifies speaker identities. [OpenAI's file guidance](https://learn.chatgpt.com/docs/artifacts-viewer)

The same process can be adapted to another approved assistant that accepts the input. The [ChatGPT](/tools/chatgpt), [Claude](/tools/claude), and [Fireflies](/tools/fireflies) directory entries give product context; this guide begins after you have a usable record.

## Extract categories before writing the summary

Use this prompt with the input packet. It requests a structured draft you can inspect before compressing it into minutes:

```text
Extract only from the supplied meeting record.
Return separate Decisions, Proposals, Actions, and Open questions tables.

For each row include a source line or timestamp, supporting evidence,
and any condition. Use timestamps only when they exist in the source;
otherwise use the original line references.

Do not treat silence, suggestions, or conditional offers as confirmed
agreement. Do not infer an owner or deadline. Use Unassigned or
Not established for missing fields. Keep relative dates as spoken
unless I provide the date, time zone, and confirmation rule.

End with questions for the participant reviewer.
Keep the output as a draft; do not send messages, create calendar
events, or update a task system.
```

The prompt states the desired behavior. It does not remove the need to check the source or control any connected actions. If your workflow creates tasks, keep that step after review so a proposed action does not become an assignment by accident.

## Worked example: a proposal is not a decision

**Illustrative example:** this fictional onboarding meeting and all its speakers are invented. The excerpt below is not an output from a tested product.

```text
Meeting: fictional onboarding launch, 2026-10-06, UTC.
L1 09:02 Maya: We could move the launch to Friday if QA finishes.
L2 09:03 Theo: QA is still open; I cannot confirm Friday.
L3 09:04 Maya: Then we have not changed the launch date.
L4 09:05 Priya: I will send the revised checklist by Thursday.
L5 09:06 Theo: I can review it, but I need the final customer list first.
L6 09:07 Maya: Someone should ask the customer for that list.
L7 09:08 Priya: Agreed on using the new checklist after review.
```

The classification table keeps evidence close to each interpretation:

| Category | Extracted item | Evidence | Condition or review note |
| --- | --- | --- | --- |
| Decision/status | No launch-date change is established | L2–L3 | The actual current date is not supplied |
| Recorded agreement | Use the new checklist after review | L7 | Review must happen; confirm the agreement's intended scope |
| Proposal | Move launch to Friday | L1 | Depends on QA; explicitly unconfirmed at L2 |
| Action | Send revised checklist | L4 | Priya commits to Thursday as spoken |
| Conditional offer | Review checklist | L5 | Theo needs the final customer list first |
| Open action/question | Ask customer for final list | L6 | Owner and deadline are not established |

Do not write “Launch moved to Friday; Theo will complete QA” from this excerpt. Neither assertion is established. Do not assign the list request to Maya just because she raised it. An empty owner field is useful information for the next discussion.

The draft action register can be narrower than the classification table:

| Item | Owner | Deadline | Status | Source |
| --- | --- | --- | --- | --- |
| Send revised checklist | Priya | Thursday, as spoken | Stated commitment | L4 |
| Review checklist | Theo | Not established | Conditional offer; needs final list | L5 |
| Request final customer list | Unassigned | Not established | Needs assignment | L6 |

Keep “Thursday” until the team confirms how to turn it into a calendar date. The meeting header alone does not settle every ambiguity about business calendars or intended deadlines.

## Have a participant check the consequential details

The reviewer should inspect decisions and actions against the original record, not merely read whether the summary sounds right. Check negations, conditions, names, dates, and duplicate tasks. A missing “not” or “after review” can change the team's understanding.

For the example, ask:

- Did anyone approve a new launch date elsewhere in the meeting?
- Does L7 record the intended team's agreement, and who confirms the review is complete?
- Is Thursday the intended deadline for Priya, and what calendar date should the action use?
- Who will request the customer list, and when?

If a participant supplies a new answer, label it as a later clarification with its date and approver. Do not make the transcript appear to contain something it never said. For a fuller claim-by-claim evidence check, use the [document research workflow](/guides/workflow/ai-document-research-with-citations).

## Turn the reviewed extraction into minutes

Use a compact template that preserves uncertainty:

```text
Meeting / date / source:
Review status / reviewer / review date:

Purpose and short summary:
Confirmed decisions: decision + conditions + source
Action register: task + owner + deadline + status + source
Open questions: question + person to ask, if known
Proposals not adopted:
Later clarifications: date + approver + change
Original record location:
Intended recipients:
```

A clean **illustrative draft** from the excerpt would read: “The record does not establish a launch-date change. Priya will send the revised checklist by Thursday. Theo offered to review it after receiving the final customer list. The checklist is to be used after review, subject to participant confirmation of the agreement. The list request still needs an owner and deadline.”

Keep the action register beside that summary. A reader can then distinguish a stated commitment from an offer waiting on a prerequisite. Keep unresolved questions separate from the confirmed task list rather than hiding them inside a paragraph.

## Share the approved record and keep it current

Before sharing, check the intended recipients, the included source material, and participant corrections. Apply the team's [human-in-the-loop workflow](/guides/workflow/human-in-the-loop-ai-workflows) at the point where the draft becomes a record others will act on.

If the minutes belong in shared project context, follow the [AI Projects guide](/guides/comparisons/chatgpt-projects-vs-claude-projects) to name an owner and check colleague access. Link to the approved version so a later summary does not revive a superseded draft.

Update actions when someone makes an actual assignment or decision. Record the change and its source. Do not silently replace an unresolved owner with a guessed person merely to make the register look finished.

## Troubleshoot gaps before publishing

| Problem | Corrective action |
| --- | --- |
| Wrong or uncertain speaker | Ask a participant to confirm attribution; leave the owner unassigned meanwhile. |
| A transcript segment is missing | Retrieve the segment if available; mark the coverage gap until then. |
| No deadline was stated | Request one during review rather than choosing a convenient date. |
| A decision happened in chat only | Supply the chat excerpt separately with its source and context. |
| A target became a commitment | Restore the proposal label and any conditions. |
| A recap went to an unintended invitee | Check the distribution settings and intended access before sharing the revised minutes. |

The finished record should tell colleagues what they can act on and what still needs confirmation. Keeping those two states visible is the main quality check for AI meeting notes.

## Sources and further reading

- [Take notes for me in Google Meet](https://support.google.com/meet/answer/14754931?hl=en) — Google
- [Use Transcripts with Google Meet](https://support.google.com/meet/answer/12849897?hl=en) — Google
- [Work with files](https://learn.chatgpt.com/docs/artifacts-viewer) — OpenAI

---

_Source: https://agentscamp.com/guides/workflow/ai-meeting-notes-action-items — Guide on AgentsCamp._