Don’t guess assignees from first names alone

The failure mode

Transcripts are messy: first names only, nicknames, people who weren’t on the calendar, and “we” / “someone” commitments. An LLM asked for owner_email will happily invent one, or map “Alex” to the first Alex in your workspace directory.

Wrong assignees create noise tickets and train the team to ignore the automation.

Pass an attendee roster

Before extraction, give the model a closed set of people who were actually in the meeting:

{
  "meeting_date": "2026-09-16",
  "attendees": [
    {"name": "Alex Kim", "email": "alex@acme.com"},
    {"name": "Sam Ortiz", "email": "sam@acme.com"}
  ],
  "transcript": "..."
}

Instruct: only set owner / owner_email when the transcript clearly matches one roster entry. Otherwise null.

Keep the source span

Always retain the quote that justified the item:

{
  "title": "Send revised pricing deck",
  "owner": "Alex Kim",
  "owner_email": "alex@acme.com",
  "due_date": null,
  "source_span": "Alex will send the revised pricing deck"
}

Write gates

• Notion Person properties: only set when you have a resolvable user id/email; else use rich text for the spoken name.
• Linear assignees: omit assigneeId when unknown; still create the issue with the quote.
• Never invent emails. Invented addresses bounce or notify strangers.

Takeaway

Assignee matching is a directory problem, not a creativity problem. Closed roster + null-on-ambiguity + source spans keep meeting → ticket pipelines trustworthy.