In an AI-written follow-up email, one line may have turned a proposal into a commitment

An AI-written follow-up email after a meeting usually reads very smoothly. Here is one, set next to the original notes. Of its five lines, one says more than the meeting did.

The meeting, the team and the names are all made up for illustration. This article uses no real notes and measures no AI model.

Original notes (HR team meeting, 21 Sept):

- The team agreed to hold the new-hire welcome session on Monday morning, 28 Sept.
- Hanh will ask reception whether the 3rd-floor meeting room is free that morning.
- Tung proposed buying 12 welcome gift sets. Budget not yet approved.
- Phuc will check the stockroom for two computers for the two new hires.
- The date for the workplace safety training has not been chosen.

Follow-up email drafted by AI:

Hi all, following the 21 Sept meeting, here is what we settled:
1. The new-hire welcome session will take place on Monday morning, 28 Sept.
2. Hanh: ask reception about the 3rd-floor meeting room.
3. Tung: buy 12 welcome gift sets.
4. Phuc: check the stockroom for computers.
5. The date for the workplace safety training will be announced later.

Read each verb, not the whole email

Try to find it yourself before reading on: which line says more than the notes?

The way to find it is to ignore the feeling that “this email sounds reasonable” and compare each line’s verb with the verb in the notes. Line 1 says “will take place”; the notes say “agreed to hold”: a match. Lines 2 and 4 say “ask” and “check”; the notes say “will ask” and “will check”: a match, because both are still jobs to do, not results. Line 5 says “announced later”; the notes say “not chosen”: a match.

Line 3 is not. The notes say “proposed buying” and “budget not yet approved”. The email says “buy”, and the opening also says “here is what we settled”. A person reading the email will understand that the 12 gift sets are approved. That is where a proposal got promoted into a commitment.

NIST, in its generative AI risk document AI 600-1, calls this family of errors confabulation: content that is wrong but presented with confidence. In a meeting email the error is rarely a nonsensical sentence. More often it is a single verb stronger than the source.

Record exactly one finding

For each email, write down what you found as a three-part line:

Line in the email:  3. Tung: buy 12 welcome gift sets.
Line in the notes:  Tung proposed buying 12 welcome gift sets. Budget not yet approved.
Corrected to:       Tung proposes buying 12 gift sets; waiting on budget approval.

Fix that line, then change the opening “here is what we settled” to “meeting summary”, because a leftover “settled” pulls all five lines up to the level of certainty. Only then is the email ready to send. If you find more than one problem, record each on its own line like the one above and don’t merge them.

Ask AI to take a second look, but check the quotes yourself

You can give both texts to an AI with a request like this:

Compare each line of the email with the notes. For each line, copy the exact sentence from the notes that supports it.
If no sentence supports it, write "not found". Point out any line that uses a stronger verb than the notes.
Do not edit the notes.

(Write the request in the language your notes are in.)

AI is faster than you when the email is long. But don’t trust a quotation just because it looks right. Use Find (Ctrl+F) to search for that sentence in the original notes. If it isn’t there, treat the line as unchecked.

When this check is not enough

If you follow the steps and the AI still reports line 3 as a “match”, or the quotation it gives is not in the notes, don’t keep changing the request. That is a sign you need someone who attended the meeting to confirm. The same goes when the notes contradict themselves, or when the real decision happened outside the part you have.

If the meeting involves sensitive data, use the accounts and tools your organisation allows, or remove the sensitive parts before you give the text to a chatbot.

Sources

1. Best practices for prompt engineering with the OpenAI API — OpenAI — A guide to writing requests with a clear task, context and desired output format; used to support how the request for the second look is written.
2. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) — NIST — The primary document describing generative AI risks, including confabulation; used to explain why a smoothly written email is not evidence that it is correct.