The Off Switch You Don’t Control
In June, the U.S. government switched off two of the most capable AI models on the market. Most companies cannot even tell you which models they would lose if it happened to them.
On June 13, the U.S. Commerce Department invoked national-security export controls to bar two frontier models, Fable 5 and Mythos 5, from any foreign national, including the vendor’s own non-citizen employees. Because the company could not separate users cleanly in real time, the practical effect was a global shutoff. Businesses that had built real work on those models lost access within hours, not because of anything they did, but because a third party with authority over the models decided they should go dark.
Set aside the national-security politics, because the politics are not the lesson. The lesson is structural, and it is uncomfortable. The off switch on the AI your business runs on was never in your hands. It sits with the vendor, and now, we have learned, with the government standing behind the vendor. You are a tenant in a building someone else can lock.
For most organizations, that is not a manageable risk, because they cannot even see it coming.
The blind spot, measured
On June 17, IBM’s Institute for Business Value, with Oxford Economics, published a study of 1,000 senior executives across 16 countries and 17 industries. The headline number is the one that should keep operators up at night: 91% say they do not fully understand their organization’s dependencies across AI vendors, models, and infrastructure.
The supporting figures turn that from an awareness gap into a continuity problem. 71% say switching their primary AI vendor or model would be difficult. Companies reported an average of six AI-related disruptions over the past two years, most of them traced to vendor services. And 81% say a seven-day outage of an AI vendor would cause severe or critical disruption, the kind that effectively stops operations.
Hold those four facts next to the shutdown. A model can be switched off by someone other than you. Most companies cannot see which models they depend on. They cannot easily switch off the ones they can see. And the great majority would be in crisis within a week. The June shutoff was not an anomaly that proves the system works. It was a live demonstration of a risk that the data says almost no one is positioned to absorb.
This is concentration risk, and you already know how to manage it
Strip away the word “AI” and this is a category every competent operator already governs. No serious business runs on a single supplier with no second source, no continuity plan, and no clear inventory of what it depends on. Procurement does not allow it. Finance does not allow it. IT continuity does not allow it. You would treat “we are not sure what we rely on, and we could not survive a week without it” as an audit finding requiring immediate remediation.
Yet that is precisely the posture the IBM data describes for the models now sitting under underwriting, claims, customer support, research, and forecasting. The discipline that is routine everywhere else in the enterprise has not reached the AI layer. Not because the risk is exotic, but because the adoption outran the governance. The models arrived faster than anyone inventoried them.
Why the dependency hides
The reason 91% cannot see their dependencies is not negligence. It is the shape of how AI enters a company.
AI rarely arrives as a logged decision with an owner attached. It shows up as a feature inside a tool you already license, switched on through a product update you did not evaluate as an AI adoption. It gets embedded three layers deep in a workflow by a team trying to move faster. The model under your support queue, your document review, your risk scoring, was often never selected in a meeting. It accumulated.
And what accumulates without a decision has no owner. No owner means no inventory, because no one was assigned to keep one. No fallback plan, because no one was named to write it. No continuity drill, because no one is accountable for the outcome. The dependency is invisible for the same reason it is ungoverned: it belongs to no one.
The tempting wrong answer
The instinct, once a leadership team sees this, is to buy something. A control plane, an AI gateway, an observability layer. The market is flooded with them right now, and they are not useless. But notice what they do not do. A tool that watches your models does not reduce your dependence on them. A dashboard that shows your usage does not give you a second source. Adding a layer of software to manage the problem leaves the concentration exactly where it was, and adds one more vendor you now also depend on.
Visibility tooling can support governance. It cannot be governance, because the thing missing is not a screen. It is a decision, and a person to own it.
What governance actually requires here
Treat the AI your operations depend on as critical infrastructure, and govern it the way you govern every other system you could not afford to lose. In practice, that is three things, owned by named humans, in writing.
An inventory. Where is AI operating, and which vendor and model sits beneath each process that matters. This is the 91% problem, and it is solvable in weeks, not years.
A continuity plan for each dependency that counts. What happens if this model goes dark or goes wrong. Is there a fallback, a second source, a manual path, and who executes it. This is the direct answer to the seven-day-outage question, and to the shutdown.
An owner for each. Not a committee. A person with the authority to act when the model fails or disappears, who keeps the inventory and the plan current as both change.
There is a number in the IBM study that makes the business case for all of this. Organizations with the most advanced AI control capabilities protect more than half of their operating profit from AI-driven disruptions. Governance here is not an insurance cost. It is a margin you keep when the off switch gets pressed.
The point
The June shutdown will not be the last one. Between regulators, vendors, security incidents, and ordinary outages, you have already averaged six disruptions in two years. There will be a seventh, and an eighth. That is not a forecast. It is the run rate.
So the question the IBM study really puts to every operator is not whether your AI will fail or vanish. It will. The question is whether, on the day it does, you can name what you will lose, point to the plan, and point to the person who owns it. Right now, nine in ten cannot.
Sovereignty over your own operations does not mean building your own models. It means never depending on something you cannot see, cannot replace, and have not decided you can live without. The off switch may not be in your hands. The plan for when it gets pressed has to be.
AI operates. You own the decision, including the decision to depend on it.
Where is AI operating inside your business, and who owns the plan for the day it goes dark? Tell me in the comments.
This is the question my work begins with. When AI becomes infrastructure and no named human owns the dependency, that gap is what I help close.
Sources
IBM Institute for Business Value with Oxford Economics, “The Calculus of AI Sovereignty,” June 17, 2026. Survey of 1,000 senior executives across 16 countries and 17 industries. Found that 91% do not fully understand their AI vendor, model, and infrastructure dependencies; 71% say switching their primary model would be difficult; respondents averaged six AI-related disruptions in two years; and 81% say a seven-day vendor outage would cause severe or critical disruption. https://newsroom.ibm.com/2026-06-17-ibm-study-limited-control-and-rising-dependencies-leave-enterprises-exposed-in-the-age-of-ai
U.S. Commerce Department export-control suspension of Fable 5 and Mythos 5, reported June 13, 2026. A national-security export-control directive barred foreign-national access to two frontier models, including the vendor’s own non-citizen employees, forcing a global suspension because users could not be separated cleanly in real time.