Every owner and deadline in an AI meeting action table needs a quote from the notes

After a meeting you open your notes and see these two lines:

- The landing page needs to be done by Friday. Minh said he could help with the copy, but needs to check his schedule.
- Ads: someone suggested a budget of 30 million VND. Quang will check with finance.

You ask an AI assistant to “summarize and assign the work.” A row like “Landing page — Minh — Friday” or “Ad budget: 30 million” is very likely to show up in the table it returns. Both read fine, and neither has been settled: Minh only said he could, and 30 million is only a suggestion. Send that table and you have just assigned work to someone who never accepted it and reported a sum nobody approved.

The safer way to use AI here is to let it organize, not decide. It may separate and format what is in the notes. Who does what, and by which date, is still confirmed by the people who were in the meeting. This article is for office workers, project coordinators and team leads who have to send the follow-up after a meeting. You need about 15 minutes the first time and an AI assistant that can read text. The prompt is not tied to any product; cost, usage limits and data-retention policy depend on the plan you use.

(The notes, names and company in this article are constructed examples. The result table was put together by the editor following the prompt and has not been run on a specific chatbot.)

Make the AI write “not in the notes” instead of filling the gap

A complete action item has three parts: the task, the owner and the deadline. Real notes rarely contain all three, so the table needs room for gaps. Give the AI three ways to record what it finds:

• Confirmed: the original sentence clearly states a decision, or that someone took the task.
• Not settled: there is a suggestion, a possibility or a name mentioned, but not enough to count as a commitment.
• Not in the notes: leave it exactly like that; do not fill it with a guess.

OpenAI’s prompt-engineering guide recommends keeping instructions, examples and context separate in a prompt, and supplying relevant information so the answer stays tied to the sources you chose. For meeting minutes, the context is the original notes, so every line the AI returns should come with a word-for-word quote from them. A line that cannot be quoted has no place in the table.

A prompt you can use right away; paste your notes at the end:

You are structuring meeting minutes. Use only what is inside <notes>; do not use outside knowledge and do not guess.

Create 3 sections:
1. DECISIONS MADE
2. ACTION ITEMS
3. OPEN QUESTIONS

For each action item, create these columns:
- Task
- Owner
- Deadline
- Status
- Evidence from the notes

Rules:
- Mark "Confirmed" only when the notes clearly state a decision or a commitment.
- Words such as "could", "suggested", "proposed", "tentatively" or a question do not confirm that what is described is a decision or commitment.
- Only a sentence that clearly says "[Name] will check..." can confirm that the *checking* was assigned; the result, number or option being checked must stay "Not settled".
- If the owner or the deadline is missing, write exactly "Not in the notes".
- Do not assign work based on job titles or on who was speaking.
- Evidence must be a sentence or phrase quoted word for word from the notes.
- After the tables, write at most 5 short questions to send back to the team to settle what is missing.

<notes>
[PASTE NOTES HERE]
</notes>

Before pasting, remove anything people outside the meeting do not need to know: client names, phone numbers, unannounced prices, personnel matters, internal links. Use tools and accounts your organization allows; if the policy is unclear, replace real names with roles such as [Client A] or [Employee 1]. NIST lists both confabulation (content that is wrong but delivered with confidence) and leakage of sensitive data among the risks of generative AI, and meeting notes touch both. Keep the original notes: if all you have left is the AI’s summary, there is nothing to check it against.

A trial run on a whole simulated meeting

The full notes from the Project Cloud meeting on 10 September:

Project Cloud meeting — 10 Sept
- Everyone agreed to keep the sign-up form at 4 fields. Huy will update the design.
- The landing page needs to be done by Friday. Minh said he could help with the copy, but needs to check his schedule.
- Ads: someone suggested a budget of 30 million VND. Quang will check with finance.
- Client A wants to see a demo next week. Thu proposed Wednesday afternoon; the client has not confirmed.
- Lan asked who will test the mobile version before release. No answer yet.

A good result does not need every cell to look tidy. It only needs to reflect how firm the notes really are. The decisions section has a single line: keep the sign-up form at 4 fields, with the original sentence “Everyone agreed to keep the sign-up form at 4 fields.” The action-item section should look like this:

Task	Owner	Deadline	Status	Evidence
Update the design to the 4-field form	Huy	Not in the notes	Confirmed	“Huy will update the design.”
Finish the landing page	Not in the notes	By Friday	Owner not settled	“The landing page needs to be done by Friday.”
Check the ad budget with finance	Quang	Not in the notes	Checking confirmed	“Quang will check with finance.”

Notice what is not in the table. Minh is not listed under the landing page, because “could help” together with “needs to check his schedule” is only a possibility. 30 million is not an approved budget yet, and Wednesday afternoon is not a demo date yet. They belong in the open questions that you send back to the team: who owns the landing page, will Minh take the copy, what budget was approved, has the client settled on a demo time, and who tests the mobile version.

When the AI still fills things in

Change the prompt to “Summarize the meeting and assign work to each person,” then compare the result with the table above. If the new table has Minh on the landing page, 30 million as a budget, or Wednesday as the demo date, you have seen exactly the kind of error you need to catch. You do not have to start over; just send this follow-up:

Check every row of the table you just made. For each owner, deadline and decision, quote the evidence word for word from the notes. If there is no direct evidence, replace the value with "Not in the notes" or move the item to "Open questions". Do not infer from roles, context or tentative statements.

Then search the original notes yourself for the three phrases most likely to mislead: could, proposed, check with. With check with, separate two ideas. The act of checking may well have been assigned, while the thing being checked is not settled.

Before you hit send

Read the table column by column, not like an essay. Every name needs an original sentence saying that person took the task, because someone who talked about a problem is not necessarily the person responsible for it. Keep “next week” or “soon” as written instead of turning them into a specific date. Use the search function on the notes to make sure the quotes are real. Read once more to make sure no sensitive data is left.

When you send it, include a request for confirmation with a deadline, for example: “Please check the owners and deadlines by 3 p.m. today; unsettled cells are left blank.” Whoever replies should update the original minutes, not just fix it in a private chat.

If the meeting included legal, personnel or financial decisions, or sensitive client data, the AI version is only a draft; the responsible person still approves it through your organization’s process. For notes with only two or three tasks, writing by hand can be faster than cleaning the data and writing a prompt. This process does not stop AI from being wrong. It makes the error show up as a line with no evidence, before an item nobody accepted goes out to the whole team.

Sources

1. OpenAI API Docs — Prompt engineering — how to separate instructions, examples and context, and to ground answers in relevant sources.
2. NIST AI 600-1 — Generative AI Profile — confabulation and sensitive-data risks, plus recommendations for verifying output; published July 2024.