The Human-in-the-Loop Approval Model

A useful AI operator should not treat every action the same. Reading a public web page is low risk. Sending a customer message, deleting files, buying something, changing DNS, or posting publicly is not.

Approval categories

Usually safe to automate

• Reading public documentation
• Searching local project files
• Summarizing provided text
• Drafting internal notes
• Running non-destructive checks
• Producing a proposed plan

Usually requires review

• Sending email, texts, or social posts
• Editing public website copy
• Changing configuration used by production systems
• Creating scheduled jobs
• Acting on customer records
• Using browser UI where the result is not easy to reverse

Always requires explicit approval

• Payments, bookings, purchases, refunds, or financial transfers
• Passwords, API keys, OAuth tokens, or credential changes
• Deleting data, resetting history, or destructive shell commands
• Legal, medical, financial, HR, or compliance-sensitive decisions
• Permission dialogs and account-security prompts

How to write approval rules

Use plain language. A good rule says what Rocky may do, what Rocky may draft, and where Rocky must stop.

Rocky may draft customer replies from approved templates, but a human must review and send them.

Verification checklist

• Risk categories are documented.
• Users know what is automatic and what is draft-only.
• Approval requests include enough context to decide.
• Denied actions are logged and not retried silently.