Human control
The four things your
agent cannot do alone.
These are not prompt instructions the model can ignore. They are deterministic system paths. Here are the four, with code proof.
Cannot confirm a payment.
When a customer says they already transferred, the agent opens a verification request, moves the order to payment pending, and goes silent. The order only advances when you mark it, with your user and timestamp recorded.
See the full flow โCannot invent a price.
The cart total is computed by the system, not the AI. Wholesale prices by volume, add-ons, and quote formulas are resolved in code and handed to the agent pre-calculated. The AI writes the message; the math is not its job.
Cannot offer a slot that is already taken.
Overlap prevention is guaranteed by the database, not the agent. And availability is re-read on every message in the conversation, not just once at the start.
Cannot keep talking once you step in.
From your inbox you can take control of any conversation and the agent stops immediately. You can also require human review for every order above a threshold you define.