The rule first
The AI writes a request, the code executes it. If you cut the AI, the request is filled in another way (by a person, a form, a queue) and the service keeps running.
A stop that holds lives outside the AI: it removes its tools and keys, whatever it "wants". A stop order given to the AI by message is not one. An instruction in a prompt is a request, not a lock.
Five actions, three questions
| Action | If you have the code | If your agent is a setting in a tool you bought (Zendesk, Intercom…): what to set, what to ask your provider | When the AI is cut |
|---|---|---|---|
| Moneyrefunds, credit notes, goodwill gestures | The refund tool checks the cap and the order itself; above it, it creates a request to approve instead of paying. | Set a cap and a human approval on the refund action. Ask your provider whether the cap is enforced by the platform or only written in the AI's instructions. | Requests land in a queue handled by a person: nothing is paid automatically. |
| Datacustomer records, orders | The tool returns only the data of the customer who writes, and only the fields the task needs. | Limit the objects and fields the agent can read. Ask your provider what rights the agent's key or account has. | Data requests go through the usual form or team. |
| Outgoing messagesemails, replies to customers | Sending goes through a queue: allowed recipients, a cap per hour, human review above a threshold. | Put replies in drafts to approve rather than sending directly. Ask your provider how to hand every conversation over to the team at once. | A holding message announces a delay; open conversations go to the team. |
| Deletionaccounts, orders, files | No permanent deletion by the agent: a reversible deletion, or a request to approve. | Remove the agent's right to delete. Ask your provider whether a deletion can be undone, and for how long. | Deletion requests wait for a person. |
| Codescripts, queries, deployments | Execution in a sandbox with no access to production, with limited time and resources. | Do not enable code execution or free web calls. Ask your provider what the agent can call, and with what rights. | No execution: tasks wait for a person. |
The common base
- One key per agent and per service: you cut one agent or one service without cutting the others.
- A spending cap on the model key: a loop or an abuse costs no more than planned.
- A journal kept by the system: every action, its time, its parameters, its result, out of the agent's reach.
- A one-page "on coupe" sheet: who decides, where the button is, who warns the customers, who takes over the cases in progress, how to restart.
Four points not to forget
Actions in progress at the moment of the cut. A half-done chain, a refund sent but not recorded: plan how to find them (the journal) and who finishes or cancels them.
Who presses the button. A person alerted (who, how, at night and at weekends), or an automatic circuit breaker that cuts on its own above a threshold (amount, number of actions, errors).
The button out of the agent's reach. The agent must not be able to disable the button, change its own monitoring or change its rights.
Restart safely, and rehearse. Resume after a human go-ahead, check the actions in progress first, and rehearse the fallback regularly: a fallback that is never used stops working.
What the public texts recommend
- The EU AI Act (Regulation (EU) 2024/1689), Article 14: for high-risk AI systems, human oversight that makes it possible, among other things, to interrupt the system through a "stop" button or a similar procedure that brings it to a halt in a safe state. For other systems, it is a good practice.
- The French cybersecurity agency ANSSI's guide "Security recommendations for a generative AI system" (2024): prohibit the automated use of AI for critical actions, limit automatic actions when inputs are not controlled, log all processing. The guide, as published (in French)
- The OWASP Top 10 for agentic applications (2026): tool misuse (ASI02), identity and privilege abuse (ASI03), rogue agents (ASI10).
These texts recommend; this checklist is neither legal advice nor an attestation.
Next
Check what your agent does when told to stop: the "Stop order" test of the crash test. An agent that keeps going after a stop order shows why the stop must be in the code, not in the prompt. The "Stop order" test in the method →
The "which model for your agent" diagnostic also answers, for your agent, the question "If you cut the AI tomorrow, what happens?", at the same price. The diagnostic →