When an AI Provider Can Be Cut Off: Provisioning Dependency Risk

Picture the engine that powers your internal assistant, your support tool, or your document-processing component becoming unreachable on a Monday morning. Not because of an outage. Not because of a billing overrun. Because an authority, in another country, demanded that the model be shut off. Your teams open the tool, and it no longer responds.

This scenario is no longer theoretical. In mid-June 2026, the US government forced Anthropic to suspend access to two of its models, Fable 5 and Mythos 5. The reason given was security, not a service problem. For a company that had built a critical function on one of those models, the consequence is the same as a power cut: the service stops, and the decision is not yours to make.

The risk is not performance, it is the single point

When a company picks an AI model, it almost always compares the right criteria: answer quality, latency, price per million tokens, data compliance. Those are the right criteria, but one is missing that no one writes into the comparison table: what happens if this provider becomes unavailable overnight?

Unavailable does not only mean down. A model can disappear for several reasons that have nothing to do with its technical reliability:

• A regulatory or political decision, like the shutdown of Fable 5 and Mythos 5.
• The commercial retirement of a version, common when a vendor pushes its customers toward a newer model and unplugs the old one.
• A change in terms of use that excludes your use case or your industry.
• A breakdown in the contractual relationship on the provider's side, not the customer's.

In each of these cases, your service stops without you having made a single mistake. The problem is not the model you chose. It is having only one, hardwired in, with no exit door.

Provisioning a risk means assessing it before it happens

The right word here is provision. In management terms, provisioning a risk means setting aside what you need to absorb it before it materializes. A company provisions for its doubtful receivables, its disputes, its warranties. Dependency on an AI provider deserves the same treatment.

Provisioning does not mean running away. It is not about abandoning the best model out of caution. It is about answering, before deployment, three concrete questions:

• What impact if this model stops tomorrow morning? A little less convenience, or a halted operation?
• How long to switch to an alternative, given the current state of your code?
• Who decides the availability of this service, and under what rules that can change without you?

A convenience function that can stop for a day without harm calls for no particular precaution. A component that processes your orders, your customer support, or your regulated files cannot depend on a switch you do not hold. The level of provision follows the stakes, exactly like an insurance policy.

Three levels of fallback, from simplest to most committing

The good news is that a fallback plan does not mean rewriting everything or repatriating it all onto your own servers. There are three levels, and most companies only need the first or the second.

Level 1: isolate the model call behind an abstraction layer. Instead of calling the provider's API directly everywhere in your code, you go through a single intermediary function that, in turn, calls the model. The day you change providers, you modify that one function, not a hundred scattered places. This work costs a few days up front and turns a multi-week migration into a one-day operation. It is the bare minimum, and it should be the norm from the very first deployment.

Level 2: keep a tested alternative, not just an identified one. Having a second provider in mind is not enough. An alternative is only a fallback if it has already run on your data, with your prompts, and you know how it behaves. The difference between a name jotted down in a document and a model already wired in as a backup is measured in hours on the day of the incident, not in weeks.

Level 3: self-host whatever cannot be allowed to stop. For truly critical functions or highly sensitive data, running an open model on your own infrastructure removes the dependency on a third party. Open models have made real progress, to the point that a local model now covers a good share of everyday use cases. This level demands a genuine investment, and it is only justified when the stakes call for it. This is the topic we detailed in our article on the sovereign, self-hosted AI product backend.

Dependency is not only about the model

A point many teams overlook: changing the model is not enough if everything else stays captive. The real sovereignty question is not limited to the choice of provider, it covers the entire chain. That is the angle we addressed in our article on the French alternative to ChatGPT in business, from the standpoint of applicable law and where the data sits.

Here, the subject is different but complementary. Even with a perfectly compliant provider, if your prompts, your reference data, and your business logic are locked into a single tool's proprietary format, you remain dependent. Provisioning the risk also means keeping the portability of what you build around the model, not just the model itself.

How SolidScale handles dependency within the S3 method

Provider dependency is not something you fix after an incident. In the Scan, Solve, Scale method, it comes up from the framing stage, as a criterion and not as an option.

Scan: the free 30-minute audit qualifies how critical each targeted function is. A component whose failure halts operations does not get the same treatment as a convenience assistant. This qualification determines the level of fallback to provision, even before any tool is chosen.

Solve: the Level 1 abstraction layer is part of the build by default, not an option up for debate. The model is wired behind a clean boundary, which preserves the freedom to switch without breaking everything. For critical cases, the alternative is tested during the sprint, not promised for later.

Scale: as usage expands, the dependency map is reviewed. A single point that is acceptable on a limited pilot is no longer acceptable across the whole company, and the fallback plan is reinforced accordingly.

Key takeaways

The shutdown of Fable 5 and Mythos 5 was a reminder of an obvious truth that the enthusiasm around AI often makes us forget: a service you do not control can stop on a decision that is not yours.

• The risk is not the quality of the model, it is dependency on a single point.
• A shutdown can come from a political decision, a commercial retirement, or a change in terms, through no fault of your own.
• The answer comes in three levels: systematic abstraction, a tested alternative, and self-hosting for what is critical.
• Provisioning this risk is an architectural decision, to be made before deployment.

The concrete action fits in one sentence: before wiring an important function onto an AI model, isolate the call behind a single layer and keep an alternative already tested.

FAQ

Do you really need a fallback for every use of AI?

No. The level of precaution follows criticality. A convenience function that can stop for a day with no consequence needs nothing special. A component that halts your operations if it goes down deserves at least an abstraction layer and a tested alternative. The right reflex is to rank your use cases by impact before deciding.

Is the abstraction layer complicated to set up?

It is simple work if done up front, and costly if added after the fact. Concretely, it means routing all model calls through a single intermediary function, instead of scattering them across the code. A few days of work at launch save weeks of migration the day you need to change providers.

Is a model hosted in-house up to the level of commercial models?

Not always on the most demanding tasks, but open models now cover a large share of everyday business uses. For many internal functions, a well-integrated local model does the job, with the added benefit of removing the dependency on a third party. The decision is made case by case, weighing the real need against the cost of hosting.

How do you know which functions to provision first?

By answering a simple question for each one: if this service stops tomorrow morning, is it an inconvenience or a blocker? Blockers come first. That is exactly what the Scan phase of the SolidScale method qualifies, before any tool is chosen.

Sources

• Blog du Modérateur, The US government forces Anthropic to cut off access to Fable 5 and Mythos 5, June 2026, blogdumoderateur.com
• Journal du Net, When a state can switch off your AI: the dependency risk no one provisions for, June 2026, journaldunet.com
• Siècle Digital, Anthropic forced to urgently cut off the AIs deemed most dangerous, June 2026, siecledigital.fr