AI Meeting Notes Action Items: Captured Is Not Completed

AI meeting notes action items are extracted suggestions, not assigned work. Every major notetaker stops one step short of the thing you actually wanted, which is a task with an owner, a due date, and some way of knowing later whether it happened. Microsoft’s own support documentation is explicit about where the line sits: Facilitator captures tasks during a Teams meeting into Notes, and a person has to select “Accept to sync” before any of them reach Planner.

That gap is the whole subject of this post. Below is what the vendors document, what the research says about how much of a meeting a summary actually keeps, and the exact shape of the meeting-to-task pipeline we built for our own group, including the part of it that is not running yet.

What the vendor documentation actually says

Read the official help pages rather than the marketing pages and a consistent picture appears: these products capture and suggest, they do not assign and dispatch. Language limits, licence tiers and manual sync steps are all documented by the vendors themselves, and none of it shows up in the listicles that rank for this query.

Product	What it captures	Where it stops, per the vendor’s own docs
Teams Facilitator	Tasks captured during the meeting into Notes	A person must select “Accept to sync” before tasks reach Planner. Needs a Copilot licence. Cannot read an agenda when the invite carries an encrypting sensitivity label
Google Meet “Take notes for me”	Notes plus suggested next steps	One language per meeting, eight supported, bilingual meetings unsupported. Notes save to the organiser’s Drive. Summaries may not generate at all under 15 minutes or on a poor connection
Zoom Meeting Summary with AI	Summary and next steps	Workplace Pro or higher. In mixed-language meetings it summarises in the majority language. Unavailable in breakout rooms. Not shared unless the host opts in
Fireflies to Google Tasks	Detected action items pushed to a task list	Everything lands in one default list. The assignment filter is on or off with nothing between. Meeting-level rules need the Business plan

All four pages fetched 16 September 2026. Microsoft’s own meeting-notes page adds the line that matters most: verify the results, because AI-generated content can be incorrect.

How much of the meeting a summary actually keeps

Less than people assume, and the spread between models is wider than the average. OmniCSEval, a study of 28 language models across 1,800 conversations submitted as a preprint on 14 June 2026 and revised 31 July 2026, measured key-fact recall on conversation summaries. On the meeting scenario specifically, completeness ran from 34.2% for the weakest models to 91.2% for the strongest.

The authors describe a pronounced decline in both completeness and conciseness on meeting transcripts compared with cleaner conversation types. Meetings are hard to summarise for the same reasons they are hard to sit through: people interrupt, decisions get reversed mid-sentence, and the important commitment is often the throwaway line at minute 43.

If your notetaker sits at the bottom of that range, two thirds of the facts are gone before anyone reads the summary. You cannot tell which two thirds, because the thing that would tell you is the summary.

The most automated notes left people knowing the least

There is a result here that cuts directly against the pitch. A within-subject experiment with 30 participants, accepted to the CSCW 2025 conference and posted by its authors on 3 September 2025, compared three levels of AI note-taking assistance: automated, intermediate, and minimal. The fully automated condition produced the lowest scores on an objective post-test. The intermediate condition produced the highest.

Participants preferred the automated condition anyway, because it took less effort.

That is the honest framing for this category as of September 2026. Automating the notes buys attention back during the meeting and costs comprehension afterwards. The fix is not less automation, it is moving the automation downstream: let the agent do the capture and the chasing, and keep the human judgment at the point where a suggestion becomes somebody’s commitment.

What we built, and where it stops today

We built a meeting-to-task pipeline for our own group. The design is worth describing because most of the interesting decisions are about what it refuses to do.

Ingestion has two paths. The primary one is a bot that joins the call live. The fallback picks up recordings after the fact for anything the bot missed. Calendar events that cannot be classified confidently default to external, which is the more cautious of the two treatments. A keyword list, checked before the bot joins at all, keeps it out of meetings flagged as sensitive. That list is a first pass and is not yet policy-approved, which is the kind of thing worth saying out loud before somebody relies on it.

Extraction reads the bilingual transcript directly and returns structured action items, each with a confidence score and an attempted owner match against the attendee roster.

Then it stops and waits. The meeting leader gets an editable readout and has to submit it. Nothing reaches the task backbone before that. For external meetings there is a second gate: the run holds, the leader gets a confirmation link by email, and it resumes only when somebody clicks.

Here is the part most companies would leave out of a blog post. The pipeline is built and demonstrated end to end, but it runs against mocked services, not live production credentials. Two blockers are documented and open: no live transcription vendor, and no live credentials for the task backbone. It is not in production as of 16 September 2026. We publish what is running and what is not, and this one is not.

What a created task has to carry

When approval does happen, the task that gets created carries a fixed set of fields, because a task missing any of them is a note with extra steps:

• Title and description, from the transcript.
• An assignee as a resolved email address, or the explicit value Unassigned. Never silently dropped.
• A due date only when one was actually stated in the meeting. Inferred deadlines are worse than no deadline, because they look real.
• Priority.
• A link back to the source meeting, so the task can be argued with.
• A stable deduplication key built from the meeting and the action row, so re-running the pipeline cannot create the same task twice.

The list of things we chose not to automate is shorter and more important: task creation without approval, due-date inference, owner assignment when a name match is ambiguous, and any automatic accept or reject based on the confidence score. Confidence is displayed to the approver, not acted on by the system.

One gap we have not closed: status does not sync back from the task backbone to the notetaker, so its dashboard cannot show you which meeting actions were completed or went overdue. That reporting lives in the task backbone instead, alongside the chasing and verification layer described in our post on task follow up automation. Capture and completion are two different systems, and pretending otherwise is how meeting notes quietly become an archive nobody reads.

What this looks like when it runs

The pattern above is one job for one agent, or a tomte, our word for one production agent with one defined job. Our fleet is built the same way across all our companies: as of 30 June 2026, 41 registered agents with 9 live in production, and across the 13 days to 4 July 2026 those agents completed 5,450 runs with zero failures. The mechanics of that operations layer are written up in full on our agent fleet proof page.

Meeting capture is a good first candidate precisely because it is narrow. One input, one extraction step, one approval gate, one output. That is the scope of a single AI Workforce Sprint, shipped to your own accounts, from $9,500.

If you want one built, start async. Write down the meeting type, who approves, and where the tasks should land. You get a reply within one business day, in writing, with a scope and a price. No call required.