An Agent on a Leash, or why my AI agent doesn't make business decisions

Description and use

Building a safe AI support agent where the AI interprets intent, deterministic software enforces policy, and human approval gates high-risk actions.

Problem

How to build an AI-powered support agent that can be trusted to make business decisions without risking incorrect actions or data leaks.

Outcome

Implemented a safe AI support agent that never executes high-risk actions without human approval, reducing errors and ensuring compliance.

Workflow

• 1. User sends support message
• 2. AI classifies intent (e.g., refund request)
• 3. Deterministic rule engine evaluates risk tier
• 4. Low/medium risk: AI assists with drafting response or routing
• 5. High risk: AI proposes action with audit trail, queues for human approval
• 6. Very high risk: AI only proposes, human executes manually
• 7. Human reviews queue, approves or rejects action
• 8. Approved action is executed by software
• 9. System logs audit trail and notifies user

Evidence

The AI interprets intent. The software enforces policy. That's the split, and it's the thesis of everything that follows.

Source

The system drafts a response, proposes the refund, files it in the approval queue with a full audit trail.

Source