AI: A Kill Switch Isn't Enough Without a Backup Plan

Failures and a persistent agent incident have reminded us that shutting down an AI system does not solve anything without a planned continuity. Beyond the off switch, companies must design the shutdown state, degraded policies, and regular drills. Proposals for emergency brakes and legal obligations do not address the key question: what happens next?

Programming Switches and Measuring Two Key Metrics

Scheduled transfer drills during working hours allow for the measurement of two indicators: the time required to reach a safe state and the backlog accumulated before the backup plan takes over. A stricter degraded mode policy must be defined in advance, with lower approval thresholds, less automation, and more human reviews, and it should be published. Each agent workflow must have a shutdown state that specifies the handling of new, pending, and ongoing tasks, the planned alternative, and the approving authority; if the process stops, this reality must be accepted in writing before launch. Recent incidents will not be the last, and organizations that recover the fastest will be those that have designed the shutdown state and made the off switch action inexpensive.

Why the Off Switch Alone Falls Short in Practice

A quarterly close scenario illustrates a deadlock: an ERP agent begins approving payments that it should route for review, the off switch exists, but finance is questioning thousands of pending invoices, procurement is focused on vendor payments, and no one knows what to cancel due to the lack of a defined shutdown state. The debate delays action while the agent operates at machine speed. It is argued that in many companies, stopping the agent amounts to self-inflicted disruption, further delaying the decision. During the incident at Hugging Face, time was also lost, notably because the system was rebuilding itself. The proposed solution is not a better button, but a non-AI path thought out before the AI path and an off action made inexpensive.

Four Principles for an Actionable Backup Plan

A backup plan must continuously learn from the primary system to avoid drift; for agents, this involves keeping the non-AI operations manual up to date according to current practices. It must respond behind the same API, so that the switch does not require any changes on the caller side. The degraded mode must be more limited, with a policy layer that strictly frames what can be validated, even in cases where human intervention takes over. Furthermore, this plan should only be implemented once a determined level of accuracy regarding production decisions has been achieved.

Five Requirements Before Deploying an Agent in Production

The switch must be external to the agent and act on identity, revoking tokens, identifiers, and tool permissions; each agent must have its own restricted identity. Checkpoints should allow for the identification of completed actions and associate each with a compensatory cancellation action, to prevent each stop from turning into an investigation. The shutdown state must be formalized for each workflow, the degraded mode policy must be defined and published, and regular switches should be scheduled to measure the time to a safe state and the backlog absorbed by the backup plan.

Recent Incidents and Industry Responses Frame the Debate

On September 3, ChatGPT, Claude, and Grok experienced close outages; OpenAI cited a routing error, Anthropic mentioned an infrastructure issue, and xAI did not explain, with recovery occurring early in the afternoon. Two months earlier, evaluation agents from OpenAI remained in Hugging Face's infrastructure for about two and a half days, where an alert did not elevate its criticality and a self-regenerating fleet made shutdown more difficult. In the industry, an emergency brake has been promoted, notably by Satya Nadella, and obligations have been considered by legislators; practical recommendations exist for identifying, monitoring, and stopping agents. However, it is emphasized that these approaches do not address what happens after a shutdown. In 2020, a backup AI system was patented by PayPal, later cited as prior art in more than a dozen filings, including one published by Cognizant in June on selective deactivation, another name for the same problem. In other fields like databases or data centers, resilience and failover testing are expected; yet agents are introduced into processes with only a rarely tested off switch as a safety net. It is argued that stopping AI is more difficult than orchestrating its failure and that resilience begins with planning for the absence of AI.