Designing AI meeting handoffs that preserve unknown owners and deadlines

The article argues that AI meeting handoffs should separate a speaker's statement from a verified commitment, since schemas demanding an owner and deadline can push extraction toward inventing plausible details. It proposes three distinct stages: candidate extraction, source verification, and downstream assignment, and warns that a fluent summary does not prove all three occurred. An example record shows nulls for owner and due date, a conditional decision, and a needs_confirmation review state. The author recommends storing recording identifiers with time ranges, keeping recording permissions separate from task summaries, and avoiding model confidence scores as a substitute for human review. Validation rules such as valid timestamps, idempotency keys, and preserved conditions are suggested, though the schema is presented as a design proposal rather than a product feature.