Recognize Tasks AI Should Not Own

AI can prepare evidence and options, but it should not own a consequential decision or bypass an approval.

The Concept

Keep legal, medical, financial, employment, safety, permission, and irreversible decisions with a qualified person. AI may extract criteria, compare evidence, draft an action, or flag gaps. Use approved Tools for exact or live information, require human approval for meaningful external changes, and confirm the final state in Tool Executions.

Example

Replace this unsafe request:

Approve this supplier and send the commitment.

With a reviewable assist:

Map the supplied evidence to the approval criteria. List gaps and risks, then
draft a commitment for the named approver. Do not approve or send anything.

Checklist

• A named person owns the decision.
• The agent cannot bypass access or approval controls.
• Missing or conflicting evidence is reported as unknown.
• External actions are previewed before confirmation.
• There is a way to verify and, where possible, reverse the result.

Next, separate reusable knowledge from temporary context in Context, Memory, and Data.