What a Good AI Meeting Summary Actually Looks Like

Ask ten tools for an "AI meeting summary" and you’ll get ten tidy paragraphs that read well and change nothing. A summary that only compresses the conversation is a worse version of the transcript — shorter, but still something you have to read, interpret, and turn into work yourself. The summaries worth having answer three questions in order: what was decided, what happens next, and who owns each part.

Compression isn’t the point

A meeting isn’t valuable because it happened; it’s valuable because of what changes afterward. "The team discussed the Q3 roadmap and agreed on priorities" is a sentence that survives no contact with reality. Which priorities? Agreed by whom? Due when? A good summary refuses to smooth those specifics away — the specifics are the summary.

The three things a summary must contain

• Decisions, with the person attached. "We’ll ship the billing fix before Friday" is only useful if it’s tied to whoever committed to it. A decision without an owner is a rumor.
• Action items, phrased as work. Not "discussed the webhook issue" but "Maria: rewrite the webhook retry logic." Each item should read like something you could paste into a ticket — because ideally it becomes one.
• Open questions. The things raised and left unresolved are often more important than the things that were settled. A summary that quietly drops them loses the meeting’s real output.

The best summary is one you don’t have to read

If a summary is good enough, its natural endpoint isn’t your inbox — it’s your tracker. The decisions become a record, the action items become tickets assigned to owners, and the open questions become next meeting’s agenda. At that point the summary has done its job before you’ve read a word of it.

That’s the bar Celeritas builds to. Transcription and summarization run entirely on your device — no cloud upload, no bot in the call — and the summary is structured from the start into decisions, action items, and questions, so it can be pushed straight to Jira, Linear, or GitHub. The point was never a nicer paragraph. It was less homework after the meeting.