AI Sandbox

Turn meeting conversations into reviewable follow-up

Decisions and actions remain buried in notes or transcripts. A sandbox tests whether structured extraction can prepare useful follow-up while participants retain control over what becomes an official task.

Intended result: Measured extraction quality and a controlled approval workflow for meeting follow-up.
Human approval: Required
Integration: Likely

Situation

Current problem

Actions, owners, deadlines and unresolved decisions are manually reconstructed after meetings.

Desired change

The prototype proposes structured actions with supporting transcript passages and uncertainty flags.

Possible workflow

Capture

Create a permitted transcript

Record or import conversations under explicit privacy and retention rules.
Extract

Propose decisions and actions

Identify possible tasks, owners, deadlines and unresolved questions with evidence.
Approve

Confirm before registration

Let participants correct and approve proposals before using a task or communication API.

Human control

A participant approves wording, ownership and deadlines before tasks or messages are created.

Technology and data

Speech to text: A transcript with speaker and timing information of sufficient quality.
Structured extraction: A controlled schema for decisions, actions, owners and deadlines.
Task integration: An optional API used only after explicit participant approval.

Boundaries

Recording, consent and retention rules must be explicit.
The model must distinguish proposals from confirmed decisions.
Uncertain ownership or deadlines require clarification.
External actions remain blocked until a participant approves them.
Participants must know when recording or transcription occurs
Implied decisions are easy to misinterpret
No task should be created without approval