Human-in-the-Loop Confirmation UX

How AI agents should ask for permission before taking irreversible actions

The Problem

AI agents need to do things. But some actions are irreversible:

• Deleting data
• Transferring money
• Publishing content
• Modifying system settings
• Sending messages to customers
• Approving procurement orders

The naive approach: agent acts, then asks for forgiveness.

This erodes trust. Users stop giving agents permission.

Better approach: Ask for permission before the action. Make the confirmation clear, low-friction, and informative.

When to Use This Pattern

✅ Use this when:

• The action cannot be undone
• The action affects other people
• The action costs money
• The user should know what just happened
• There's ambiguity in what the agent decided

❌ Don't use this when:

• The action is easily reversible (edit, undo)
• The action is purely informational
• Speed is critical and trust is established
• The agent's confidence is very high (but still consider it)

The Pattern

Design Principles

1. Make It Clear What Will Happen

2. Show the Preview

3. Offer Escape Routes

Don't just [Approve] [Cancel]. Offer:

• [Review Again] — Let me re-read this
• [Modify] — Let me change something
• [Approve] — Go ahead
• [Cancel] — Stop

4. Show Confidence When Relevant

If the agent is uncertain, say so:

Real Examples

Example 1: Vigilo — Alert Confirmation

Vigilo monitors 239 countries for threats. When risk crosses a threshold, the agent proposes an alert:

Example 2: Navigator.Taxi — Service Order Confirmation

Before locking a vehicle until service is completed:

Example 3: EngiBoard — Deployment Confirmation

Before deploying new code to production:

Anti-Patterns

❌ "Boy Who Cried Wolf"

Problem: Agent asks for confirmation on trivial actions too.

Result: Users approve everything without thinking. Defeats the purpose.

Solution: Only confirm for reversibility threshold. Learn when users actually review.

❌ Modal Hell

Problem: 5 confirmation dialogs for 1 action.

Result: Users rage-click through all of them.

Solution: One clear confirmation. Put additional info in preview/expandables.

❌ Irrelevant Details

Problem: Confirmation shows: "UUID: 123abc456... File size: 2,047,382 bytes..."

Result: User can't parse what matters.

Solution: Show human-readable summary. Details on request.

Implementation Notes

Timing

When to ask: Right before the irreversible action, not hours later.

How long to wait: Depends on context. For safety-critical: 5-10 seconds max. For routine: user's pace.

Consent Tracking

Log:

• What was proposed
• When user was asked
• User's response
• What actually happened
• Result (success/failure)

Escalation

If user rejects confirmation twice, escalate to human specialist. Don't keep asking.

Why This Matters

Human-in-the-loop isn't bureaucracy. It's trust infrastructure.

When users know an agent will ask before doing something irreversible, they trust the agent more. They give it more authority. They use it more.

Confirmation flow = Trust multiplier.