AI meeting notes: separate decisions, proposals and actions

This lab shows how to use AI on meeting notes without collapsing everything into a single summary. The goal is not faster prose; it is a traceable note that keeps decisions, proposals, and actions separate so people can review, assign, and follow up without guessing.

The NIST AI Risk Management Framework is useful here as a methodological reference because NIST presents it as voluntary and notes that version 1.0 is under revision. In this guide, that matters only as a control-minded reference for human oversight, not as a rulebook for minutes.

This guide is designed for team leads, project managers, office managers, consultants, and SMEs that need minutes they can actually use. It connects naturally to Responsible AI and responsible automation, Academy, workshops and R&D, and Strategic IT advisory and governance.

Before you start

Final result
A reusable workflow that turns one meeting transcript into:

• a list of decisions;
• a list of proposals;
• a list of actions with owners and deadlines only when they are actually stated;
• a review register you can check before sharing the minutes.

Who this is for

• team leads;
• project managers;
• operations managers;
• office managers;
• consultants who need reliable minutes, not generic summaries.

Tools

• an AI chat assistant;
• a text editor or document editor;
• a table, spreadsheet, or note page for the final review register.

Prerequisites

• you can copy a transcript without rewriting it first;
• you can tell a proposal from a confirmed decision;
• you can leave a field blank or marked as unknown instead of inventing data.

Materials included in this guide

• one fully simulated transcript;
• one copy-paste prompt;
• one expected output example;
• one bad example and its correction;
• one final register to reuse before sharing the minutes.

1. Copy the simulated transcript exactly as written

Work in your document editor or in the AI chat input field. Start from the raw material, not from a summary.

Simulated meeting transcript

Date: Wednesday, 14 October 2026, 09:30–10:15.

Simulated participants:
- Elena Bianchi, operations manager
- Marco Rinaldi, sales
- Sara Conti, customer care
- Davide Neri, IT
- Laura Ferri, administration

Simulated interventions:
1. Elena: We need to separate decisions, proposals, and actions in today’s minutes.
2. Marco: I propose we shorten the notes, but keep every decision verbatim.
3. Sara: The customer report needs an actions list with owner and due date.
4. Davide: I can prepare the new template by Friday, 16 October 2026.
5. Laura: We need to verify the summary before sending it, but I do not know who will do it.
6. Elena: Davide and I will validate the summary, starting in October.
7. Marco: Then let’s approve the proposal and make the new format official tomorrow.
8. Sara: No, that remains a proposal until the team confirms it does not lose context.

What to observe
This transcript contains all the pieces you need:

• one explicit decision;
• one proposal that is not yet approved;
• one action with owner and deadline;
• one action with no named owner;
• one correction that prevents the model from overcommitting.

If it does not match
If you have already shortened the notes, merged two interventions, or removed the date, go back to the raw transcript. This exercise depends on preserving the original wording.

2. Paste a prompt that forces separation of the three levels

Work in the same chat session. Paste the prompt below unchanged first, then add the transcript.

You are an assistant for meeting minutes.
Work only on the notes provided.

Goal:
- separate DECISIONS, PROPOSALS, and ACTIONS;
- do not invent approvals, owners, or deadlines;
- do not turn a proposal into a decision;
- do not automatically assign a task to the first speaker.

Rules:
1. If a sentence expresses a suggestion, label it PROPOSAL.
2. If a sentence confirms a choice, label it DECISION.
3. If a sentence describes work to be done, label it ACTION.
4. If owner or deadline is missing, write DA_ASSEGNARE or not specified.
5. For every item include:
   - type
   - content
   - practical action
   - owner
   - deadline
   - source line
6. Add a column “original phrase” and a column “interpretation”.
7. If you find a contradiction, flag it and do not choose one version arbitrarily.
8. Do not create tasks, emails, reminders, or a polished narrative summary.

INPUT:
[paste the complete transcript here]

REQUIRED OUTPUT:
- list of decisions
- list of proposals
- list of actions
- final table
- open questions to clarify before sharing

What to observe
A useful answer should be structured, not literary. It should preserve uncertainty instead of smoothing it away.

If it does not match
If the assistant gives you a paragraph summary, ask again for the three lists and the table. If it invents responsibility or a deadline, stop and tighten the rule: only use names, dates, and deadlines that are literally present in the transcript.

3. Read the normal-case output and keep the categories separate

Work in the response from the AI, side by side with the transcript. The key check is simple: every line must go to the right bucket.

Expected output example

Decisions

• Elena and Davide will validate the summary, starting in October.

Proposals

• Shorten the notes while keeping every decision verbatim.
• Use a customer report that includes an actions list with owner and due date.
• Make the new format official tomorrow.

Actions

• Davide prepares the new template by Friday, 16 October 2026.
• Verify the summary before sending it.

Open questions

• Who owns the summary verification?
• Is the new format approved or still under discussion?

Final table example

What to observe
The model should not flatten the difference between “we propose” and “we decided.” In the example, line 6 is a decision because it assigns a concrete validation action to named people. Line 2 is still only a proposal. Line 4 is an action because there is a task and a deadline.

If it does not match
If the model puts line 2 under decisions or turns line 7 into an approved change, the output is not acceptable. Go back and tighten the prompt so that a proposal can only become a decision if the transcript contains explicit confirmation.

4. Handle missing data without inventing an owner

Work in the table fields, not in your imagination. If the transcript says a task exists but does not say who owns it, do not assign it to the first speaker or to the person who sounds most likely.

Use these defaults:

• DA_ASSEGNARE for a missing owner;
• not specified for a missing deadline.

Normal missing-data example

• Original phrase: “We need to verify the summary before sending it, but I do not know who will do it.”
• Type: ACTION
• Practical action: Verify the summary
• Owner: DA_ASSEGNARE
• Deadline: not specified

Why this matters
The purpose of the minutes is to preserve the real state of the meeting. If ownership was not assigned, the minutes should show that gap instead of hiding it.

If it does not match
If the AI fills in a name, overwrite it with DA_ASSEGNARE and add a note: “Owner not stated in the transcript.” Then ask the model to rerun the extraction with the missing field rule made explicit.

5. Catch the common error: a proposal turned into a decision

This is the most frequent failure in AI meeting notes. The model sees confident language and upgrades it into a decision too early.

Bad example

Marco approved the proposal and the new format becomes official tomorrow.

Why it is wrong

• the transcript does not give Marco unilateral approval authority;
• line 8 explicitly corrects line 7 and keeps the point open;
• the model has converted a proposal into a decision without evidence.

Corrected version

Marco introduces a proposal to make the new format official tomorrow; Sara corrects the interpretation and says it remains a proposal until the team confirms that it does not lose context.

What to observe
The correction should stay visible in the minutes. If the team is still discussing the point, the minutes should not pretend the matter is closed.

If it does not match
Add an abstention rule to the prompt: if a line is ambiguous, the model must flag the ambiguity and avoid finalising the decision.

6. Finish with a reusable review register and a final acceptance check

Work in a simple spreadsheet, table, or checklist page. Before you share the minutes, use this register every time.

Reusable review register

Acceptance criteria
The document is usable only if all of the following are true:

• decisions, proposals, and actions are clearly separated;
• no owner is invented;
• no deadline is invented;
• open questions remain visible;
• a human reviewer can tell what was decided and what still needs clarification.

Troubleshooting

If the model summarises too much

Ask for one row per relevant sentence. If needed, remind it that the goal is extraction, not narrative polishing.

If the model invents an owner

Replace the owner with DA_ASSEGNARE and add a note that the owner is missing from the transcript.

If a proposal becomes a decision

Return to the source line and check for explicit confirmation. If there is no confirmation, keep it as a proposal.

If a deadline is vague

Write not specified instead of guessing a date.

If two statements conflict

List both statements in the open questions section and ask a human to resolve the conflict.

If the minutes look too clean

That is often a warning sign. Recheck whether the AI removed uncertainty that was actually present in the meeting.

Final repeatable check

Use the same three checks every time:

1. compare the source phrase with the interpretation;
2. check owner and deadline fields;
3. get human review before sharing.

If all three pass, the minutes are ready. If one fails, keep the document in draft.

If you want to adapt this workflow to project meetings, internal approvals, or recurring follow-ups, describe the process you want to improve and see how Responsible AI and responsible automation or Strategic IT advisory and governance can support a controlled rollout.

Sources

• https://www.nist.gov/itl/ai-risk-management-framework
• https://www.dedaloop.it/en/services/strategic-governance/index.html
• https://www.dedaloop.it/en/services/benefit-ai/index.html
• https://www.dedaloop.it/en/services/academy-rd/index.html