by Edmas Neo
Today, I asked a question during an AI forum on Business Continuity Planning in the age of AI. The response I received was that AI should be treated like any other IT system. If the AI is unavailable, the BCP approach should be no different from any other application outage.
To any risk manager, that should be an alarm bell. Not because the answer is necessarily wrong, but because it may suggest that people are still viewing AI as a tool rather than understanding how AI is changing workflows. The risk of this blind spot is real.
Traditionally, most enterprise systems are tools used by people. If an ERP system goes down, people can still perform parts of the work manually. If a workflow system becomes unavailable, staff can still use emails, spreadsheets, phone calls, and paper-based processes. The organisation slows down, but it continues to function because the people, knowledge, and capabilities are still there.
Agentic AI is different. Increasingly, organisations are no longer using AI simply as a productivity tool. They are redesigning workflows around AI agents. The AI agent is no longer just helping with the work. The AI agent is becoming part of the workforce. As organisations move towards AI-augmented or AI-first processes, agents may eventually take over 20%, 30%, or even more of a workflow. At that point, an outage is no longer just a technology outage. It becomes a workforce outage.
Perhaps it is time to start discussing a new category of risk:
Artificial Workforce Interruptions (AWI)
If the AI agents disappear, who takes over the work? Can the remaining staff absorb the workload? Do they still have the capacity, proficiency, and throughput required to keep the operation running?
If the AI agents disappear, who takes over the work? Can the remaining staff absorb the workload? Do they still have the capacity, proficiency, and throughput required to keep the operation running?
Many will argue that the organisation still has the SOPs, process documentation, and even skill files. That may be true, but having the knowledge is not the same as having the operational capacity. A team that previously had ten people performing a process may now have four people supported by AI agents. If those agents become unavailable, the knowledge may still exist, but the capacity may not. The workflow may not stop because knowledge was lost. The workflow may stop because throughput collapses.
This scenario is not just hypothetical. In recent months, major AI providers have experienced outages and service disruptions. Anthropic's Claude experienced elevated errors and service issues affecting Claude.ai, Claude Code, and API services. OpenAI's own status history also shows recent elevated error incidents across services. Today, these incidents may result in inconvenience, delays, or productivity loss. Tomorrow, when AI agents are deeply embedded into operational workflows, similar incidents may become Artificial Workforce Interruptions.
Operational lock-in
There is another emerging risk that I think deserves more discussion. Many organisations are effectively outsourcing part of their operational capability to AI providers. Historically, companies outsourced labour. Today, companies may be outsourcing thinking. Research, analysis, planning, reporting, customer engagement, procurement support, contract review, and workflow orchestration are increasingly being handled by AI.
The assumption is often that if one provider becomes unavailable, organisations can simply move to another model. However, the reality may be more complicated. Yes, prompts, workflows, and skill files can be transferred, but operational performance may not transfer as easily. Different models reason differently, behave differently, and integrate differently. An organisation may discover that while the skill files can technically be moved, restoring equivalent operational performance is far more difficult than expected.
This creates a new form of dependency. Not infrastructure lock-in. Not software lock-in. Operational lock-in. The organisation may still own the workflows, prompts, and skill files, but it may no longer be able to restore the same operational capability quickly enough when it matters.
As we continue to redesign workflows around AI agents, perhaps the question for BCP should no longer be, "What happens if the system is unavailable?" Instead, we should ask, "What happens if 30% of our digital workforce disappears overnight?"
As we continue to redesign workflows around AI agents, perhaps the question for BCP should no longer be, "What happens if the system is unavailable?" Instead, we should ask, "What happens if 30% of our digital workforce disappears overnight?"
Most organisations conduct disaster recovery exercises. Some conduct cyber incident simulations. Very few have tested what happens when a significant portion of their digital workforce becomes unavailable. Most organisations have secondary internet connections. Many have secondary data centres. Some have secondary cloud providers. How many have a tested secondary AI provider?
Perhaps it is time to start redesigning BCPs.
Because in the age of Agentic AI, business continuity may no longer be just about recovering systems. It may be about recovering a workforce. And in some organisations, that workforce may no longer exist.
#AI_agents #Artificial_Workforce_Interruptions #AI_workflows #Business_Continuity_Plan #BCP