← All posts

Why meeting action items never get done

Everyone has had the meeting that went well. Decisions got made, four people committed to four things, and the AI notetaker sent a tidy summary within minutes of the call ending. Two weeks later, three of those four things have not happened, and nobody is quite sure whether they were agreed or merely discussed.

The reflex is to blame attention, or discipline, or too many meetings. It is usually none of those. It is a structural gap in a toolchain that looks complete.

The listening problem is solved

AI notetakers are genuinely good now. They transcribe accurately, they identify speakers, they produce summaries with an action-items section that is mostly right. If the problem were "nobody wrote it down", it would already be fixed.

Task trackers are good too. Notion and Linear are perfectly capable of holding a task with an owner and a due date, and teams that put work in them generally do the work.

So there are two mature categories of tool, both working well, and the work still slips. That should be suspicious.

What sits between them is a person

Here is the actual sequence after a call ends:

1. The notetaker produces a summary with action items.
2. Someone opens the summary.
3. Someone reads through it and decides which items are real.
4. Someone works out who each one belongs to.
5. Someone opens the tracker and types them in, one at a time.
6. Someone assigns each one and sets a date.

Steps 1 is automated. Steps 2 through 6 are a person, and they happen at the worst possible moment — immediately after a meeting, usually right before another one. It is fifteen or twenty minutes of low-status data entry with no visible reward, and it is the first thing dropped on a busy week.

Nobody decided to skip it. It just never got scheduled, because it is not anyone's job, and the cost of skipping it is invisible until the deadline arrives.

Why the summary is not enough

The obvious objection: the notetaker already produced a list, so why not just work from the summary?

Because a summary is a document, and work does not happen in documents. A document has no owner field that anyone filters on, no due date that surfaces in a Monday view, and no presence in the tool people actually open each morning. It is a record of a conversation, not a queue.

There is also a subtler problem. A summary's action-items section is a claim about what was agreed, made by a model, with no evidence attached. When somebody disputes an item three weeks later — "I never said I'd own that" — there is nothing to check except the full recording, which nobody rewatches. So the item quietly evaporates, and the tool that produced it loses a little credibility.

The gap has a specific shape

Notice what the missing step actually requires:

• Judgement about what is real. Not everything said out loud is a commitment. "We should probably look at that at some point" is not a task.
• Judgement about who owns it. Meetings are ambiguous. "Can someone pull the numbers?" has no owner until a human assigns one.
• A place to put it. In the tracker the team already uses, not a new one.
• Something to point at when it is questioned. The sentence that created the task, and who said it.

Any tool that closes the gap has to do all four. Automating only the typing — dumping every detected item straight into the tracker — makes things worse, not better. Now the tracker has forty low-quality items in it, half of them mis-assigned, and someone has to clean that up too. The cleanup job is larger than the original typing job was.

What actually helps

The step that is missing is not "detect commitments" and not "create tasks". It is a fast review between the two.

Concretely: after the meeting, a short list of detected commitments, each with a proposed owner, a proposed date, and the quote it came from. You approve, edit or drop each one. Approved items become real tasks in the tracker you already use. The whole thing takes seconds, because you are reviewing a list rather than composing one — and reviewing is enormously cheaper than authoring.

The evidence quote matters more than it first appears. When each task carries the sentence that created it, "did we actually agree to this?" stops being an argument and becomes a lookup. That is what makes the list trustworthy enough to act on without re-reading the transcript.

Why the review step cannot be removed

It is tempting to imagine a version with no human in the loop at all. Detection gets better every year, so surely this becomes automatic eventually?

Probably not, and not because of model quality. Meetings are genuinely ambiguous — a room full of people can leave with different beliefs about who agreed to what, and no amount of transcription resolves a disagreement that existed in the room. Someone has to decide. The realistic goal is not removing that decision but making it cost ten seconds instead of twenty minutes.

Which is a much better trade than it sounds. Twenty minutes of data entry gets skipped. Ten seconds of review does not.

This is the gap promisedby was built for — reading the meetings you already record, and filing the commitments you approve into the tools you already use. If you use Fathom, there are setup guides for Notion and Linear.