Pulling action items out of an email thread with AI, and checking them

A thread with twenty messages and six people ends with a vague "sounds good, let's move on this". Now someone has to work out what was decided, who is doing what and by when. It is exactly the sort of chore an AI assistant seems made for: read everything, produce a tidy list of action items.

It does that well often enough to be tempting and badly often enough to be risky. A list that looks neat but contains an invented deadline, a task given to the wrong person or a missing commitment is worse than no list, because people trust it. This guide shows how to ask for action items in a way that is easier to check, how to check them quickly, and where to keep a human in charge.

What counts as an action item#

Before you prompt anything, agree on what you want. In email, an action item has four parts:

• A task. Something someone will do, in a verb phrase: "send the revised quote".
• An owner. The person who said they would do it, or who was clearly asked and agreed.
• A due date or trigger. "By Friday", "after the contract is signed", or none stated.
• A source. The message where it was said, so you can verify it.

Three kinds of statements look like action items and are not:

• Suggestions without agreement. "We could try a new layout" is an idea, not a commitment.
• Questions waiting for an answer. "Can you check the invoice?" becomes an action item only once the person accepts, or if the relationship makes the request binding. Treat it as an open question until then.
• History. "I sent the file yesterday" is a finished action, not a to-do.

Models blur these lines unless told otherwise. Spell the definitions out in the prompt.

Why AI gets it wrong#

Knowing the usual failures tells you what to check.

• Invented owners. A message addressed to the group, such as "someone should follow up", gets assigned to a specific person.
• Invented dates. "Soon" becomes a date, or "next week" is converted using the wrong reference day, a problem with every relative date. See AI summaries of long threads for how models handle long context.
• Missed items in quoted text. In long threads, earlier messages are quoted inside later ones. Models can read a commitment twice, or miss one buried in a quote.
• Late reversals. An item agreed in message 5 and cancelled in message 14 can still appear on the list.
• Mixed-up pronouns and names. "I'll do it" attributed to whoever was mentioned last.
• Implied tasks. The model adds the obvious next step that nobody mentioned. It reads as helpful and is a guess.

Catching hallucinations in AI-drafted replies covers the same family of errors for drafts.

A prompt that is easier to check#

The aim is to make every item traceable and every uncertainty visible. You can adapt the structure below to any assistant.

Read the email thread below. List the action items that people have
explicitly agreed to or been explicitly assigned.

Rules:
- Include an item only if someone commits to it ("I'll...", "we will...")
  or is directly asked and does not decline. Do not suggest new tasks.
- For each item give: task, owner (name as written), due date exactly as
  stated, and a short quote with the sender and the message date.
- If the owner or date is not stated, write "not stated". Never guess.
- Convert relative dates only if I give you today's date: {today}.
  Otherwise keep the original wording.
- If an item was later cancelled or changed, list the final version and
  note the change.
- Separately list open questions that nobody has answered.
- If there are no action items, say so.

Thread:
{thread}

Four things make this work. It forbids new tasks. It demands a quote, which turns every item into something you can verify in seconds. It names "not stated" as an acceptable answer, which reduces guessing. And it keeps open questions separate from commitments.