Why UAE companies need an AI exit plan before the next model restriction

Dubai: The UAE's record in AI is not in doubt. An early adopter, the first country to create a ministry for it, government-backed companies like Core42 building with international partners on a serious scale. The country's technology bets have also carried a streak of caution, and that caution is about to be tested.

In mid-June, the US government ordered Anthropic to restrict Fable 5, a guard-railed version of its most capable model, Claude Mythos, to US citizens only, citing concerns about jailbreaking. Anthropic had no way to apply that line by nationality, so it withdrew the model for everyone. The order is worth reading closely. It singled people out by passport rather than by location, and those passports belong to many of the people who build the UAE's AI. The restriction itself proved impossible, so the model vanished for all of them, overnight, by a decision taken in another capital.

So, in the coming weeks, boards across the UAE will ask their AI leaders harder questions than before. "Everyone builds it this way" will not satisfy a board worried about an outage, a price rise, or a model withdrawn overnight. Nor will "we can pivot when we need to," if no one can show how.

Let me defend those AI leaders before I argue for change, because the questions coming at them are about to be unfair. The dependencies in your stack were trade-offs rather than oversights, made when the alternative was slower and far more expensive, and triaged like any other risk. Dependency on the companies behind the models is the natural result of building on them. The people who made those calls were being strategic.

It is time for a change

The price of undoing it has changed.

The irony is hard to miss. The very models you came to depend on have made leaving them cheaper than they have ever been. Work that once took a large team and a long quarter can now be done by a small team in a few weeks. Set against that, "this is how we have always done it" stops sounding like prudence and starts sounding like an excuse to leave value on the table.

Where does a leader who takes this seriously start? With a map. Most organizations cannot say how exposed they are, because most cannot see everything they are running. Models, agents, and tools have spread across business units faster than anyone has tracked them, much of it never logged. Inventory all of it, including the shadow AI, and score each workload on four questions: what it costs, what it risks, what it does to your compliance position, and how it holds up if the company behind it disappears. Dependency stops being a worry at that point and becomes a number.

Then sort. Some workloads earn their place on a frontier model, where the edge is worth the exposure, and you should rent those without apology. Others run the core of the business, and those belong on models you own and can host yourself, where no directive in another country can reach them. Owning everything is neither possible nor wise. The work is to know which is which, and to keep the freedom to move between them.

That freedom has to be built. It is an architecture and not a press release. I recommend that you separate your business logic from the models underneath it. How your processes run, and how you have automated them, is your intellectual property, and it belongs in a layer you control, independent of any one model or provider. With that separation in place, moving a workload from one model to another, or from a hosted service to a local one, becomes a change of configuration instead of a rebuild. The same layer lets you watch and govern every agent you run, which matters more each month, since agents tend to multiply faster than anyone authorized.

Beneath it, run local and hosted models side by side, and test them against your real use cases before you trust them. A model can top every public benchmark and still stumble on your contracts, your customers, and the cases too awkward to reach a leaderboard. Those tests were run on someone else's problems, often ones the provider chose. The only qualification that counts is how a model performs on the work you need done, and that includes the open-source and sovereign options you may have set aside a year ago.

Held held high

Do this, and the board meeting changes character. You walk in with a map of your dependencies, a plan for the ones that matter, and the freedom to move when something shifts. That is a different conversation, and a better one.

The deeper prize is endurance. Enterprise AI began as experiments, then became a race to deploy, and many bought speed with dependency and called it the cost of competing. That bargain is being repriced in front of us. The organizations that last will be the ones built for endurance rather than for speed: the ones that can change a model without rebuilding the system around it, absorb a disruption without losing performance, and own the intelligence they cannot afford to rent.

In the end, the capability worth holding is the freedom to choose your models and adjust your strategy, and the certainty that when the next directive lands, in whatever capital, your business stays in your hands. The smartest model in the world matters less than that freedom.