Synapse AI loading mark

August 13, 2026

Most meetings end with a list of things people agreed to do, and most of those lists live only in someone’s memory. Meeting Intelligence turns a Zoom, Google Meet or Microsoft Teams call into decisions, owners and draft tasks in under a minute after the call ends. What follows is the pipeline behind that, stage by stage, with the numbers we track and the mistakes that shaped each design choice.

Turning conversation into owned, dated work

Extracting action items sounds like a summarization problem. It is closer to information extraction with an unforgiving audience. Commitments are rarely phrased as tasks. People say “I can take a look at the retry logic by Thursday,” owners are pronouns, dates are relative, and half of what sounds like a plan is a hypothetical.

The cost of errors is also asymmetric. A missed action item is annoying. A wrong one, assigned to the wrong person with an invented due date, teaches the whole team to ignore the feature. So we made two product decisions before writing any pipeline code: every extracted task is a draft that a human approves, and when precision and recall conflict, precision wins.

A draft nobody approves is noise. We optimize for drafts people accept, not for how many items we can extract.

Stage 1: Capture with consent

Synapse either joins the call as a bot or imports the platform’s native recording. When the bot joins, it posts a configurable consent notice in the meeting chat. Participants can opt out, and audio segments attributed to an opted-out participant are dropped before the transcript is stored. Admins set retention separately for audio and transcripts; under a transcript-only policy, raw audio is deleted as soon as transcription completes.

Stage 2: Transcription and speaker attribution

When the bot joins, audio is transcribed in streaming 30-second windows during the call, so most of the work is done by the time it ends. Imported recordings are split and transcribed in parallel as soon as the platform releases the file. Language is detected per segment, which handles calls that switch between any of the more than 30 supported languages. Diarization clusters voices into anonymous speakers. The harder step is mapping those clusters to real people.

Audio alone attributed speakers correctly 81% of the time on our labeled set of 600 internal meetings, each recorded with every participant’s consent. Adding platform signals, meaning join and leave events, active-speaker indicators and the calendar invite list, raised that to 96%. Attribution matters more than raw word accuracy here: “I’ll handle it” is useless without knowing who said it.

Stage 3: Segment before you extract

A 60-minute call produces 7,000 to 10,000 words. Our first version sent the whole transcript to a frontier model in a single pass. It was expensive, and it lost items from the middle of long calls. The current pipeline splits the transcript into topic segments of three to eight minutes, with 30 seconds of overlap, using agenda items and shifts in embedding similarity as boundaries.

Each segment goes to a small, fast model for extraction, in parallel. A frontier model then consolidates the per-segment results, merges duplicates and writes the summary, and the model router makes that split automatically. Compared with the single-pass design, cost per meeting fell 64% and action-item recall rose from 0.71 to 0.79.

Stage 4: Extraction against a schema

Extraction output must validate against a JSON schema. The most important field is evidence: every item cites the transcript span it came from, and a verifier fuzzy-matches the quoted span against the transcript. Items whose evidence cannot be found are discarded, which removes most invented tasks before anyone sees them.

{
  "action_item": {
    "type": "object",
    "required": ["text", "owner_mention", "commitment", "evidence", "confidence"],
    "properties": {
      "text":          {"type": "string", "maxLength": 200},
      "owner_mention": {"type": "string"},
      "due_phrase":    {"type": ["string", "null"]},
      "commitment":    {"enum": ["committed", "tentative", "hypothetical"]},
      "evidence":      {"type": "array", "minItems": 1, "items": {"type": "string"}},
      "confidence":    {"type": "number", "minimum": 0, "maximum": 1}
    }
  }
}

When output fails validation, the pipeline retries once with the validator’s error message added to the prompt. A second failure drops that segment’s items and logs the failure for review. About 1.3% of segments need the retry, and fewer than 0.1% fail twice.

Stage 5: Resolving owners and dates

Owner mentions are resolved in a fixed order: first-person phrases map to the speaker, names map to meeting attendees, and only then to project members in the workspace directory. If two attendees share a first name, the draft stays unassigned with both suggestions. The pipeline never assigns a task to someone who was not in the meeting without flagging it.

Due phrases are resolved against the meeting’s start time, the speaker’s time zone and the linked project’s current cycle. Here is an illustrative exchange from a sprint planning call on September 8, the first day of Sprint 24:

[00:31:12] Marcus Lee: I'll take the retry backoff for the carrier webhook, done tomorrow.
[00:31:20] Priya Natarajan: Great. Sofia, can you add a regression case once that lands?
[00:31:24] Sofia Ruiz: Sure, I'll pick it up after Marcus merges.

draft 1  "Implement retry backoff for carrier webhook"
         owner: Marcus Lee (speaker, 0.94)   due: Sep 9   commitment: committed
         evidence: 00:31:12
draft 2  "Add regression test for carrier webhook retries"
         owner: Sofia Ruiz (speaker, 0.91)   due: none    commitment: committed
         evidence: 00:31:20-00:31:24         blocked by: draft 1

Sofia’s draft has no due date because nobody gave one; the pipeline records the dependency on Marcus’s draft instead. “End of sprint” would have resolved to September 19, the last day of Sprint 24. A phrase like “soon” resolves to no date at all, because an invented deadline is worse than an empty field.

Stage 6: Review, approval and delivery

Drafts appear on the meeting summary with approve, edit and dismiss controls. Approval creates the task through the public API, POST /v1/tasks, under the approver’s identity and with a scoped, short-lived token. It is the same path any integration uses, and the action is written to the workspace audit log. When a summary is ready, the meeting.summarized webhook fires and the summary posts to the linked project, a Slack or Microsoft Teams channel, or email. Transcripts stay available through GET /v1/meetings/{id}/transcript for as long as retention allows.

For a 60-minute meeting, the median time from call end to drafts ready breaks down like this:

  • Final transcript and speaker attribution: 9 seconds
  • Topic segmentation: 2 seconds
  • Parallel extraction and validation: 16 seconds
  • Consolidation and summary: 11 seconds
  • Owner and date resolution: 3 seconds

Median end-to-end latency is 41 seconds, and p95 is 58 seconds. The largest remaining cost is consolidation on very long calls, which is where our current optimization work is focused.

From Transcript to Tasks: Inside Our Meeting Action-Item Pipeline

How we measure quality

Offline, every pipeline change is scored on 600 annotated meetings: action-item precision is 0.91 and recall is 0.79. In production, the metric we watch most closely is acceptance: 78% of drafts are approved unchanged or with minor edits, 8% are substantially rewritten, and 14% are dismissed. We estimate missed items by counting tasks created manually in the linked project within 24 hours that reference the meeting.

Changes ship behind two gates. A new prompt version, model route or schema revision must not lower offline precision by more than one point, even if recall improves. It then runs on 5% of eligible meetings for a week, and we compare acceptance and dismissal rates against the current version before rolling it out everywhere.

Lessons learned

  • Hedged language is not a commitment. “We should probably look at caching” used to become a task. Adding the commitment field and routing tentative items to open questions cut dismissals from 27% to 14%.
  • Recaps create duplicates. People repeat action items at the end of a call. Consolidation now merges items with high embedding similarity and the same resolved owner.
  • Not every voice is a teammate. Commitments made by customers or vendors on the call become open questions, not tasks in your project.
  • Retention has to cascade. When a transcript reaches the end of its retention period, the timestamp links on drafts and summaries are removed with it, so no derived record keeps content past the policy an admin set.
  • Sarcasm looks like a commitment. “Sure, I’ll rewrite billing by Friday” has an owner, a task and a date. Items below 0.60 confidence never become drafts, however complete their fields look.

Put your next meeting to work

This pipeline powers Meeting Intelligence, which extracts decisions, action items, risks and open questions from Zoom, Google Meet and Microsoft Teams calls, then turns action items into draft tasks for one-click approval. It is included in the Team, Business and Enterprise plans; compare terms on our pricing page.

Synapse AI logo (footer)
Office address
Synapse AI Systems LLC
Austin, TX ยท Remote-first
Email