The risky step in a meeting-summary pipeline is often the one after summarization: turning a fluent sentence into a task, deadline, or calendar change. A transcript can support "someone proposed this" without supporting "the team approved this."

A safer implementation separates extraction, human verification, and execution. This article proposes a data contract and a small test set for that boundary. The examples are hypothetical; they are not results from a deployed system.

Store evidence with the candidate record

Keep a stable meeting identifier, transcript version, and source location. When timestamps are unavailable, number the paragraphs before extraction. Preserve the original transcript and record corrections separately rather than silently rewriting the evidence behind an existing task.

A candidate should distinguish its kind: proposal, conditional agreement, decision, task, or open question. Owner and deadline need a missing state. An empty field is not an invitation for the model to choose the most likely colleague or date.

For example:

{ "meeting_id": "training-planning-example", "row_id": "r-12-15", "kind": "task", "action": "Check whether the room is available", "owner_as_stated": "Noa", "deadline_as_stated": null, "source_locations": ["paragraph-12", "paragraph-15"], "verification_state": "extracted_unverified" } 

These are proposed application fields, not an API specification. The missing deadline is intentional. Do not replace it with today's date when creating a downstream record.

Test the distinction between a proposal and a decision

Use a tiny adversarial fixture before choosing an integration. In paragraph 12 of this invented transcript, a participant says: "We could move the training to Thursday if the room is free." In paragraph 15: "Noa will check availability; we are not changing the invitation yet."

The expected extraction includes a task for Noa to check availability and an open scheduling question. It does not include a decision to move the event, a task to send an invitation, or a deadline of today.

Add variants that change one detail at a time: a negation, an interrupted speaker, conflicting deadlines, or a correction later in the transcript. Write the expected result yourself before running the model. Otherwise, a plausible output can quietly become the test's definition of correctness.

A citation is a pointer, not verification

A model can cite the right paragraph and still misread its condition. A correct quotation can also come from a faulty transcript. The review interface should show the cited passage alongside the candidate record, with access to the permitted recording when available.

Check names, amounts, negation, conditions, and relative dates first. "Tomorrow" is relative to the meeting, not the batch job that processes it. Preserve the original wording until the date can be resolved, and record the basis for conversion.

Only a person or an independently authorized review process should advance the record from extracted to verified. The extractor cannot give itself the status "confirmed with participant" merely because it found supporting text.

Put the write boundary after review

Run the first version without task-creation, email-send, or calendar-edit permissions. Let it produce a review document. Treat instructions spoken inside the meeting as material to analyze, not commands for connected services.

When verified records do reach an integration, use a stable key joining meeting ID and row ID. A rerun should not create a second obligation. A corrected owner or date should update the identified record with an audit trail instead of creating a fresh task and leaving the old one active.

Keep source corrections distinct from later decisions. If a manager approves the move after the meeting, that approval is a new source with its own time. It should not be inserted into the original transcript as if everyone had agreed earlier.

Test access as carefully as extraction

The audience for the summary may differ from the audience for the recording or transcript. Check actual recipients and document permissions before distribution. A short summary can disclose sensitive information even when the recording stays private.

A useful release checklist therefore includes both semantic and operational tests: no invented owner, no invented deadline, no dropped condition, no duplicate task on rerun, and no distribution to an unapproved audience. A failure can stop the affected record without preventing review of unrelated, well-supported records.

Start with one authorized meeting and a review-only output. Count unsupported records and material corrections before measuring typing time saved. A fast pipeline that produces an obligation nobody approved is not ready for write access.

Yaniv Shenhav.

Adapted from the Hebrew guide on AI NEWS IL:
https://ainewsil.co.il/articles/ai-meeting-summary-decisions-action-items

This article was created with the help of AI