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.

Florian Douetteau

Articles from the author

The next UAE AI risk is dependency