When the AI you rely on gets switched off

On Friday evening a frontier AI model that some firms had built real work around stopped working, everywhere, by order of a government those firms have no relationship with. There was no notice and there is still no date for its return. If that sentence made you check which tools your own firm leans on, good, because that is the whole lesson. A frontier model is capability you borrow, not capability you own, and the question worth asking this week is simple: what in our work stops if one of these tools disappears tomorrow.

This is not a piece about panic, and it is not about Anthropic in particular. It is a plain read on what happened and what a leader should check because of it.

What actually happened

On 12 June the US Commerce Department issued an export-control directive barring Anthropic from giving its two most capable models, Fable 5 and Mythos 5, to any foreign national, whether inside or outside the United States. To comply, Anthropic disabled both models for every customer worldwide that same evening. Its other models, including Claude Opus 4.8, kept working. Anthropic says it received the letter at 5:21pm Eastern, that the letter gave no specifics, that it believes the trigger was a narrow flaw also present in other public models, and that it is complying while disputing the decision and trying to restore access.

Notice where that leaves a reader here. To that directive, you are the "foreign national". A firm in Manchester or Leeds that had wired one of those models into its work was switched off by a decision made in Washington on a Friday night, with no say in it and no warning. This is the rare case where location genuinely matters, and it matters in the opposite of the usual way. Nothing here obliges you to do anything. It simply took a tool away.

Could this happen to the tools we use

In principle, yes, and a government order is only the loudest version of it. Models get deprecated when the vendor moves on. Prices change, sometimes sharply, when a free tier becomes a metered bill. Companies get acquired and quietly fold a product you depended on. Services have outages. The specific cause varies. The experience for your firm is the same each time: a capability you had on Monday is gone on Tuesday, and the work that leaned on it is stuck.

Most firms are not ready for that. As one risk analyst put it before any of this happened, plenty of organisations have disaster-recovery plans for every layer of their infrastructure and almost none have thought about what happens if the AI model running their work goes away tomorrow (InformationWeek, March 2026). The warning was already on the table. Friday made it concrete.

What an AI continuity plan means for a firm that is not a tech company

The phrase sounds like a job for an IT department you do not have. It is not. For a professional firm, continuity is mostly a few honest judgement calls about where you have let a borrowed tool become load-bearing.

Start by naming where AI actually sits in the work, not where the policy says it sits. There is usually a gap. Then sort those uses by what they touch. A model that tidies up internal notes is one thing. A model drafting a client deliverable, scoring candidates, or producing a first read of a target's accounts is another, because its output carries your name and your client's risk. For anything in that second group, you want a tested fallback, a sense of what "good enough" looks like so you could judge a replacement quickly, and enough of your own record of the work that you are not relying on the tool to reconstruct it. The firms that handle this well treat models as replaceable parts inside their own process rather than the centre of it.

None of that is exotic. It is the discipline a good firm already applies to any supplier it cannot fully control, pointed at a newer kind of supplier.

What it looks like in practice

A litigation team runs first-pass document review through one AI tool. If that tool drops out mid-matter, the deadline does not move. The team needs to know, before it happens, whether the work can shift to another tool or back to people without losing days, and whether anything reviewed so far would need checking again.

A corporate finance or accountancy team uses a model to summarise data rooms or draft first-cut analysis. If the model changes or vanishes, delay is the obvious cost. The subtler one is that a replacement model formats and reasons differently, so output that looked clean yesterday reads differently today. You want to have noticed that on a quiet afternoon, not in front of a client.

A recruitment team leans on one platform for sourcing and candidate outreach. Lose it for a week in a live search and the pipeline stalls. The fix is rarely a second full platform. It is knowing which parts are genuinely hard to replace and keeping a manual path for those.

In each case the exposure comes from how tightly the firm has tied a piece of real work to something it does not control, often without deciding to.

The part that stays yours

There is a reassuring side to this. The work that cannot be switched off by anyone else is the judgement: the brief you write, the questions you ask of the output, the decision to sign or not sign. That is the part clients pay for, and it is the part that should sit above whichever model is in fashion this quarter. A firm that knows exactly where its judgement lives, and has not quietly outsourced it to a tool, is a firm that can change tools without changing its standards.

Building that clarity, where AI genuinely helps, where a human has to stay in charge, and how to make the whole thing calmer and more repeatable, is what the Good Transformer Lessons for Leaders sessions are for.

What to check this month

Three honest questions, in order. Where has a single outside AI model become load-bearing in work that carries our name? For each of those, what actually happens if it disappears for a week, and how fast could we move? And does our own record of that work stand on its own, or does it depend on the tool still being there to explain it? You do not need a document. You need answers you believe.

If those answers are vaguer than you would like, that is the place to start, not a reason to pull back from AI. The firms that get hurt by a Friday-night switch-off are not the ones using AI. They are the ones who never asked what they would do without it. If that is a conversation worth having properly, book a discovery call.

FAQ

Could what happened to Anthropic's models happen to the AI tools your firm uses?

In principle, yes. A government order is the dramatic version, but a model can also be deprecated, repriced, withdrawn after an acquisition, or simply go down. The practical effect on your firm is the same: a capability you had is gone, and the work built on it is stuck until you adapt.

What is an AI continuity plan for a firm that is not a tech company?

A short set of judgement calls, not a technical project. Know where AI actually sits in your work, sort those uses by what they touch, and for anything that carries your name keep a tested alternative, a sense of what "good enough" looks like, and your own record of the work so you are not dependent on the tool to reconstruct it.

Should we deliberately use more than one AI model?

For anything important, having a tested fallback is sensible, because it means a single withdrawal cannot stop you. Running everything across several models adds cost and complexity, so the goal is not maximum redundancy. It is making sure no single tool is the only thing standing between you and a missed deadline.

Whose job is this in a firm this size?

The leader's, with whoever owns the work. It is a question about where the firm has accepted risk, not a question for IT alone. The useful move is to ask it out loud before a tool forces the issue.