The AI Tool That Came Back
Two weeks ago I wrote about an AI provider having its models pulled by government order, with no warning to the enterprises that had built their workflows around them. This week, the story closed. Full export restrictions on both of Anthropic’s advanced models have now been lifted. The tools are back.
It’s worth walking through how that actually happened, because the shape of it says more about resilience than the headline does.
The suspension landed on a Friday, triggered by a national security directive. No transition period, no phased withdrawal, just gone. A week later, partial relief arrived, one of the two models was cleared for redeployment, but only to a shortlist of roughly a hundred large organisations and government agencies. The second model stayed restricted. This week, both restrictions were lifted in full. The US Commerce Secretary said the government was satisfied that appropriate safeguards were now in place, and that the company had made significant progress addressing the concerns that triggered the suspension in the first place.
Three weeks, three distinct states: fully available, then partially available to a select few, then fully available again. If your organisation had built its workflow so tightly around one of these models that switching meant starting from scratch, you lived through three weeks of genuine uncertainty with no way of knowing which state you’d wake up to. If you’d built with the assumption that any single AI provider could disappear tomorrow, this was an inconvenience. You’d have alternative routes already in place, and the news would have been interesting rather than urgent.
That’s not a story about AI policy. It’s a story about what “prepared” actually means in practice. Most organisations think of AI adoption as a capability question, which model, what features, what it can do for us. Fewer think of it as a dependency question, what happens if this specific capability becomes unavailable tomorrow, for reasons that have nothing to do with the technology itself and everything to do with geopolitics none of us control.
The organisations that treated this as a non-event weren’t the ones with the best contingency documentation. They were the ones who’d never let a single vendor become the only way to do something critical. That’s a design decision, made long before any government directive existed, not a reaction to one.
If you’re deploying AI tools into your organisation right now, and I’d guess most of you are, ask yourself the same question I’d ask about any critical system: if this specific provider vanished from your stack tomorrow, would that be a crisis, or an afternoon’s inconvenience? The three weeks we’ve just watched play out with Anthropic weren’t unique. They’re a preview of exactly the kind of disruption AI-dependent organisations should expect to become routine.