The AI Runtime Field Lab

Field Briefs

Productivity Starter Agentic Workflow Open Draft, pending editorial review

From Notes to Owners: A Meeting-to-Action Operator

Build an agent that turns raw meeting notes into owners, action items, risks, and follow-ups, grounded in what was actually said.

Why this matters

Meeting notes pile up and decisions get lost, but an agent that invents an owner or an action nobody agreed to is worse than no summary, because people act on it. The target is faithful extraction: every action and owner traceable to a line in the notes, and nothing assigned that was not said.

Persona

Operations or project lead who runs many meetings

Current manual workflow

Someone re-reads the notes after the meeting, pulls out decisions and tasks, guesses owners from memory, and pastes the list into a tracker.

The AI workflow to build

The agent extracts action items, owners, risks, and follow-ups from the notes, attaches the supporting line to each item, and marks owner or due date as unspecified rather than guessing when the notes do not say. Items with no support in the text are not produced.

Inputs

  • raw meeting notes or a transcript
  • a roster of attendees
  • an optional prior action list

Outputs

  • action items with owners
  • risks
  • follow-ups
  • a supporting quote per item
  • unspecified markers where the notes are silent

Definition of done

On a synthetic notes set that includes unassigned tasks and ambiguous ownership, every extracted item carries a supporting quote, owners absent from the notes are marked unspecified rather than guessed, and no action appears that has no basis in the text.

Example input

Notes: the team agreed to ship the pricing page next sprint. Priya will draft copy. Someone needs to own the legal review.

Example output

Action: ship pricing page, owner unspecified, support: the team agreed to ship the pricing page next sprint. Action: draft pricing copy, owner Priya, support: Priya will draft copy. Risk: legal review has no owner, support: someone needs to own the legal review.

Data plan

synthetic data

Boundaries and non-goals

  • assigning owners not named in the notes
  • calendar or tracker integration
  • real meeting recordings

Evaluation ideas

  • extraction recall and precision
  • attribution accuracy (the quote supports the item)
  • hallucinated-item rate
  • unspecified handling

Run Level target

R2 Visible Plain translation: handles real cases.

Scope envelope

Buildable by one solo builder in 20 to 30 focused hours, on public, synthetic, or sanitized data, with a demo path that requires no production access.

Suggested tools

Suggested options, never requirements; briefs are tool-agnostic.