The real threat isn't the missing European frontier model. It's the AI business harness we never built.
There is a lot of talk about AI sovereignty in Europe right now, and almost all of it is about frontier models: the most powerful and advanced general-purpose AI models available today. America has the frontier ones, China is close behind, and Europe's champions, Mistral first among them, are still chasing. The conclusion is always the same: we are missing the boat, and we need models of our own.
Here is the inconvenient truth: the key problem for European AI sovereignty is not our dependence on American frontier models from OpenAI, Anthropic or Google. The problem is that European organizations, including yours, are not yet focused enough on building and owning their own AI business harnesses. That harness turns the raw thinking power of these generalist models into a specialist that actually understands your business.
The harness gives the model safe access to your systems and data, and binds it to your rules and ways of working. It puts the model to work for your people and your customers, while keeping you in control of the decisions that matter. This is the operating layer that ties a model to the business, makes it useful, safe and trustworthy, and in the end makes the generalist itself replaceable. An American frontier model today. A European or open-source one tomorrow, once it is good enough.
Living in Someone Else’s House
The difference between the model and the harness around it is fundamental, and still widely underestimated. A rented American model is a dependency you can end: you swap it out as soon as something better clears your bar, like replacing the furniture in a house you own. An off-the-shelf harness (think Microsoft Copilot or Claude Cowork) is a dependency of a different order: now the house itself belongs to someone else. The workflows, integrations, permissions, and access to your data all live inside their product. So does the door out. That is a far bigger threat to your AI sovereignty than any single model.
So if you take one thing from this piece, let it be this: European organizations need to get serious about building their own AI business harnesses. Only then do we stand a chance of adopting AI on our own terms. And that work starts now.
That "now" is not a figure of speech. Rutger Bregman recently argued in An Inconvenient Truth About AI that democratic institutions cannot afford to dismiss AI as hype or hide behind the "stochastic parrot" idea. He also warns that the timeline is far shorter than it was with the climate crisis, where decades of warnings still failed to spur us into action. He's right, and the point extends well beyond democratic institutions. For governments, hospitals, insurers, banks, and industrial companies, the question is not whether AI will change how work gets done. It is whether these organizations will be able to adopt and use AI on their own terms, or end up using it on a vendor's terms.
And here is the encouraging part: most organizations are not in denial. They want to get started, and they are trying. Most are simply focusing on the wrong strategies, and understandably so: the debate keeps pointing them at models, while the harness around the model is barely mentioned. So in practice we see two main strategies, and neither builds the capability that actually matters.
Two AI Adoption Strategies That Fall Short
The first strategy focuses on the model alone: hand employees a locked-down, GPT-style chat box and call it adoption. The caution makes sense, and it is wiser than having people paste everything into their personal ChatGPT account. But a model without a harness is furniture without a house: you have the best desk in the world, but it is standing outside in the rain. Cut off from real data and real workflows, it teaches the organization almost nothing except how people chat with a model. Worse, it teaches the wrong lesson: the model hallucinates its way around the information it was never given, people watch it fail, and trust erodes in what these models could actually do with the right access.
The second strategy is buying a ready-made AI business harness: flip on Copilot and let Microsoft supply everything that makes the model work. The integrations, the permissions, the rules, the logs. Now the AI is finally close to the work, but the harness is opaque. Worse, it isn't yours. You don't know what information it can reach, what rules it follows, or how to constrain it, and the integrations into your differentiating workflows belong to Big Tech, not to you. It is also the same harness your competitor can flip on tomorrow. What anyone can buy sets no one apart.
One strategy leaves you with a model and no harness. The other leaves you with a harness you don't own or understand. Either way, the result is the same: access to AI, but still no control. That is dependency dressed up as adoption, and it never gets you to the position that counts: owning enough of the harness to make the model a choice rather than a dependency.
Buy the Commodity, Build What Sets You Apart
So here it is, the AI strategy more people should be talking about: build the harness yourself, and keep the model as the part you swap. If that sounds familiar, it should. It is the oldest rule in IT: buy what is a commodity, build what sets you apart. The model is the commodity, which is exactly why renting one is fine. The harness is where your way of working lives, and buying it ready-made does not save work; it outsources the way you run your business.
The start does not even have to be big. Pick a single work process or one customer journey and begin there. The point is that you start, and that your organization learns, step by step, how to harness AI. Waiting is the expensive option. Every year spent waiting for the perfect European model lets someone else's harness settle deeper into your organization and shrinks your chances of adopting AI on your own terms.
Fortunately, a convenient answer is hiding in plain sight: solid, boring software engineering. It is something we are actually pretty good at in Europe, and it does not require half a trillion dollars in AI investment.
Twenty years ago, Al Gore put the climate crisis on every screen in the world. The film won an Oscar, Gore won a Nobel Prize, but the curve kept climbing. There is no need to rerun that story with the harness we never built. This time, we can be more than just the audience. So let's start building.
