When should an AI agent ask for permission?
The hardest question in AI operations isn't "can the agent do this?" It's "should the agent do this without asking?"
An agent that asks permission for everything is a chatbot with extra steps. An agent that never asks is a liability. The value lives in the space between.
## The spectrum of risk
Not all actions carry the same weight.
| Action type | Reversible? | Impact | What the agent does | Example |
|------------|-------------|--------|---------------------|---------|
| Read | N/A | None | Proceeds instantly | Look up order in Shopify |
| Low-risk write | Yes | Internal only | Acts, reports what changed | Add a customer tag |
| Medium-risk write | Mostly | External record | Verifies data, then acts | Update subscription status |
| High-risk write | No | Financial / customer-facing | Pauses for your approval | Issue refund, send email |
The combination of reversibility and impact determines the action type. A customer tag is reversible and internal — just do it. A refund is irreversible and financial — ask first.
Your procedure instructions define where the boundaries are. "Auto-approve returns under $50 for defective items" sets a clear threshold. If the agent doesn't have explicit guidance, it defaults to asking.
## What a good approval looks like
A bad approval: "Should I proceed? Yes / No."
A good approval: Customer Jane Doe, order #4521, requesting refund for damaged item. Order total $89, placed 18 days ago. Stripe shows successful charge, Shopify shows delivered 12 days ago. Customer has 2 previous orders, no prior returns. No fraud flags. Recommended: approve (under $100, within 30 days, first return). But item was marked final sale, which conflicts with defective-item exception. Options: Approve / Deny and offer store credit / Escalate.
You decide in 10 seconds without opening Shopify or Stripe. The agent already did the research.
## Earning autonomy over time
The first week: more approval requests. The agent is cautious — no precedents to draw on.
The second week: fewer approvals for situations it's seen before. You approved three similar refunds last week. It doesn't need to ask again.
The second month: the agent handles routine cases autonomously and only surfaces genuine edge cases — the ones that actually benefit from human judgment.
This progression isn't automatic. Your procedure instructions define the thresholds. But as you see the agent making consistently good decisions, you can adjust — raising the auto-approval limit, adding exceptions for VIP customers, broadening what counts as "routine."
## Every approval teaches
When you approve what the agent recommended, it reinforces that pattern. When you reject or modify, the correction becomes a precedent. When you escalate, the agent learns that this type of situation exceeds its scope.
Over time, the approval system creates a feedback loop: decisions generate precedents, precedents inform future decisions, and the agent's judgment converges with yours.
## The honest limits
The system works when your instructions are clear about what requires approval and what doesn't. Vague instructions force the agent to guess, which means more conservative escalation and more approval requests.
Progressive autonomy requires volume. If an procedure runs once a month, the precedent base builds slowly. High-frequency procedures build judgment quickly.
And the feedback loop is only as good as your corrections. If you auto-approve everything without reading the details, the agent learns that everything is fine — even when it isn't.