AI agents need a recovery plan too

An assistant that drafts text is a tool. An agent that sends emails, edits records or triggers payments is an actor in your information system, with access, possible mistakes and outages. It deserves the same rigor as any critical system.

Five questions before going live

• Who owns it, and who can stop it immediately?
• What can it access? Are its rights limited to what is strictly necessary, with an identity of its own?
• Which actions are irreversible, and which require a person's approval?
• Can we reconstruct what it did? Logs, traces of decisions and of sources consulted.
• What does the business do if the agent or its hosting platform goes down? Is there a manual fallback?

New risks to plan for

Beyond ordinary outages, agents bring their own risks: instructions hidden in a document or web page can hijack their behavior (prompt injection), and a plausible but wrong answer can trigger a real action. Access limits and human approval of sensitive actions are your best protections.

Where to start

Inventory the tools and agents already in use, including those staff adopted without telling anyone. Adopt a clear usage policy, then treat each production agent as a service in the continuity plan: with an owner, a recovery objective and a test.