Start with the consequence
Classify actions by their consequences, not by how easy they are to automate. Drafting an internal summary is different from issuing a refund, changing access or sending a customer commitment.
Separate proposals from decisions
Ask the model for a structured proposal. Validate required fields and business rules in deterministic code. Only then decide whether the action is allowed or needs approval.
Make uncertainty operational
A low-confidence answer is not useful if it silently becomes a completed task. Route uncertain cases into a review queue with the source material, proposed action and reason for escalation.
Test the uncomfortable cases
Build an evaluation set containing missing data, contradictory sources, duplicate events and requests outside scope. Track failure modes alongside successful answers.
