Why businesses need an AI kill switch strategy, not a single off button
Ben Saunders was featured in an exclusive Q&A with Digitalisation World on why businesses need layered, proportionate controls rather than a single AI kill switch.
The article covers:
Why the government was right to reject a single national kill switch, and how an effective contingency plan starts with an inventory of every AI system, its data, provider, owner and agreed fallback within the existing operational resilience framework
A tiered approach that separates isolating a faulty model, restricting access after a compromised account or agent overreach, and shutting down an entire workflow, with thresholds and named decision makers agreed before any incident
Treating the enterprise kill switch as a set of controls over identity and access, model routing, data flows and critical services, since switching AI off without an alternative can trigger a second, more disruptive incident
Reducing concentration risk and retaining sovereignty through an abstraction layer, organisation-owned prompts and evaluation tests, pre-qualified alternative models and provider contracts covering exit, notice of material change and audit rights
Designing for adaptability through modular architecture, multi-model routing and continuous monitoring, alongside the security, governance and human factors, such as prompt injection and fading manual skills, that a resilient incident plan must address
