Nobody reads the transcript: meeting AI is an action-item problem
Count how many meeting transcripts you have opened a second time. For me the honest number is close to zero.
A transcript is an artifact that makes you feel covered. It runs to 4000 words, it is complete, and it is almost never read again. The thing a meeting actually produces that carries value is smaller and more fragile: a couple of decisions, and a few next actions with a name attached to each. Everything my meeting tooling does now is bent toward getting those few items somewhere real.
the transcript is a decoy
I spent a while optimizing the wrong artifact. Better transcription, tighter summaries — I kept polishing a document that nobody among the tools I churned through was ever going to reopen. The summary is fine. What actually matters is what changed and who owes what by when.
the pipe, and the discipline that holds it together
My flow is unglamorous. After a call, the summary passes through a small local model that pulls out action items as structured objects: owner, action, rough due date, source line. Not prose: structured output against a schema, so the next step can actually consume it. The source line matters more than it looks. It lets me jump straight back to the moment the action came from, instead of taking the model's word for it.
That schema discipline is the entire reason it is reliable enough to trust. A model that free-writes action items will sometimes hand me five, sometimes merge two owners into one line, or invent a due date out of politeness. Forcing it into a typed shape and validating the output before anything downstream touches it is what turns a nice paragraph into something a script can file. When a field comes back empty or malformed, I would rather drop that item than file a wrong one.
where it died the first time
The first version worked and produced nothing.
Strange sentence to write. The extraction was good. The action items were correct. And they landed in a tidy channel that, as it turned out, precisely nobody on the client side ever checked. The summaries were sharp and the owners were right and still nothing moved, because the output went to an app that was not in a single person's daily loop. A perfect answer delivered to an empty room.
An action item that reaches a tool nobody opens is indistinguishable from an action item you never wrote down.
the human gate before anything becomes a ticket
The fix was half plumbing and half rule. The plumbing: pipe it into the tracker people already live in. The rule I will not drop: nothing auto-creates a ticket without a human glance first.
I tried full auto-creation for about two weeks. It made duplicates, it filed half-decisions as tasks, and it once assigned an owner to a person who had only been mentioned in passing. So now there is a review step where I approve or bin each extracted item before it becomes real. It costs me two minutes after a meeting, and it is the whole difference between a system people trust and one they quietly mute.
The deliverable was never the transcript. It was the three actions a human agreed to and a tracker will not let anyone forget.