Human approval gates for AI agents in production

Direct answer

Put a human approval gate before any agent action that is hard to reverse, high-cost if wrong, or externally visible without review. Do not gate pure research steps that only draft internal notes.

Gates are not bureaucracy. They are the boundary between useful automation and operational risk.

What to gate

• Sends: email, messages, invoices, contract changes
• Writes: CRM stage moves, refunds, permission changes
• Money and access: payments, admin roles, production deploys
• Public content: posts, customer-facing replies on sensitive topics

What not to gate

• Internal summaries for a human who already owns the next step
• Retrieval and ranking that only informs a later approved action
• Draft generation that cannot leave the sandbox without approval

Minimal gate design

1. Name the action and its blast radius.
2. Define who can approve (role, not "anyone online").
3. Show the agent evidence pack: inputs, intended action, confidence, sources.
4. Log approve / reject / edit with a reason.
5. Time-box stale approvals so work does not rot silently.

Failure modes

• Gate everything → agents become slow chat stubs
• Gate nothing → one bad tool call creates a customer incident
• Vague approvers → nobody owns the decision
• No audit trail → you cannot improve the policy

How Dali uses gates

In production agent systems, we map workflows first, then place gates on irreversible steps, then automate the rest inside existing tools. Start with an audit.