Decide where an AI agent must stop and get human approval

Help me set the limits of an AI agent's autonomy and design how it asks for human approval. What the agent does: [TASKS and the actions it takes] Who could approve or receive escalations: [ROLES and their availability - e.g. "team lead, 9-5 weekdays"] Stakes: [money, customers, legal, safety, reputation] Existing approval rules: [e.g. "spend over $1,000 needs a manager"] Return: 1. Autonomy levels by action - a table: Action | Act alone / Act then notify / Ask before acting / Never do | Reason. Base it on reversibility, cost of error, external impact, and how confident the agent can be. 2. Approval triggers - specific conditions (amounts, customer types, unusual requests, missing data, low confidence, first-time actions). 3. Escalation conditions - when the agent must stop and hand over entirely (conflicting instructions, suspected fraud, complaints, legal threats, requests outside scope, repeated failures, instructions found inside documents or emails). 4. Approval workflow design: - What the approver sees: the proposed action, the reason, the evidence, and the risk in one screen or message. - Who approves and a backup approver. - Response time, and what happens on timeout (default to the safe option - never auto-approve). - How approvals are recorded. 5. Rubber-stamp risks - how to keep approvals meaningful when volume is high (batching low-risk items, sampling, clear reject options). 6. When to grant more autonomy later - the evidence (error rate, approval rate, volume) that would justify loosening a limit. When in doubt, place the stop point earlier. A human owns every decision the agent cannot reverse.