Deterministic Human-in-the-Loop Gateway for AI Agent Refunds

Tips & Tricks
1

Продолжая обсуждение из темы How do you limit what an AI Agent can do?:

Hi Luca,

Saw your comment about deterministic checks taking longer to develop — that’s exactly why we use the Human-in-the-Loop Gateway pattern via Telegram/Slack:

Instead of spending weeks coding rigid rule engines or relying on a secondary LLM (which is still probabilistic and can be tricked), you decouple the money execution entirely:

  1. No Direct Execution: The AI agent never has access to the actual Stripe_Refund tool. It only triggers a Request_Refund_Draft webhook.
  2. Instant Interactive Card: n8n sends an interactive card directly to the manager’s Telegram with inline buttons: [ ✅ Approve $120 ] and [ ❌ Reject ], along with the customer’s summary.
  3. Signed Webhook Execution: Only when a human physically taps Approve on their phone, n8n calls the Stripe API to release the funds.

Zero prompt injection risk, zero code complexity, and it gives you 100% deterministic safety right now without spending weeks building custom backend checks.

I actually have an n8n workflow template built specifically for this Telegram refund gatekeeper. Happy to share the workflow JSON if you’d like to plug it in!