Designing a safe AI incident copilot requires a security-first approach, given the pressures and potential pitfalls during an incident. Such a copilot should not be an autonomous commander or a source of truth, but rather a tool to draft communications based on approved facts. Researchers have demonstrated AI systems can be manipulated, so a narrow initial scope is crucial, accepting only verified facts from an incident lead and requiring human review. The copilot's job is to draft communication, not to diagnose outages or determine root causes.A precise job description focuses on organizing approved facts into a fixed communication format, not on making operational decisions. This involves a simple incident briefing form capturing essential, shareable information. The assistant must produce a predictable draft with specific sections like headline, current update, customer action, and next update. Content rules are vital to prevent the inclusion of speculative or unapproved information.A five-point review gate is necessary before publication to ensure factual grounding, data sensitivity, audience appropriateness, correct action/timing, and human approval. Untrusted content, such as unverified chat messages or documents, should be kept out of the drafting path to prevent misleading or sensitive information from being included. Separation of generation from publication is critical; the copilot should only create drafts, with human control over the final publishing process.Before integrating the copilot with operational systems, a thorough threat-model workshop is essential, involving various stakeholders. This involves asking critical questions about data flow, prohibited content, user manipulation, and responsibilities. Evaluating the copilot's quality through defined incident scenarios and measuring operational outcomes provides a comprehensive assessment. For GCC deployments, considering regional audiences and distributed teams is important for effective incident communication.