Why we do not use American AI

Last week I was in a conversation with an entrepreneur who had just set up a new company. Halfway through he started talking about his team. Who had quit, who had been out sick at home for months, and who actually was not good enough but worked remotely, so no one dared to bring it up.

That is normal in a conversation like that. You need to know those things to be able to help someone. But it did get me thinking, because in my own system a conversation like that automatically runs through an AI that turns it into a summary. And that AI was located in America.

That same week we switched everything over. This piece is about why, and why I think you should look at it too, even if you never build software.

It is not about whether it is safe

Let me start with what it is not, because that is usually where the conversation goes wrong.

The large American AI providers are not unsafe. They do not leak your data, they do not sell it on, and under their business terms they do not train their models on what you put into them either. That is simply stated in their contracts and they abide by it. Anyone who claims your data is "out on the street" the moment you use an American service is usually selling you something.

The real point lies elsewhere. It is about who is in charge of your data, under which law that falls, and what happens when the circumstances change.

What it is actually about

American companies fall under American law. That means, among other things, that American authorities can in theory demand access to data managed by an American company, even if that data is physically located in Europe. That is not a conspiracy theory, that is simply legislation that exists.

For most companies that is a theoretical risk they will never be bothered by. But there are two situations in which it shifts from theoretical to practical.

The first is when your customer asks about it. We sell to scale-ups, and some of them are in accounting, healthcare or financial services. They do vendor due diligence for their profession. They do not ask "is it safe", they ask "which parties get to see our data and under which law do they fall". To that question, "in Frankfurt, with a French company" is a better answer than "in America, but they only keep it briefly".

The second is when the political climate changes. Trade agreements between Europe and America on data exchange have been struck down by the courts twice in the past fifteen years. Twice, thousands of companies had to revise their contracts at breakneck speed. There is no reason at all to assume that the current agreements are the last.

And that is the point I really want to make.

The problem is not today, it is two years from now

If you build an AI feature into an American tool now, it works fine. Today, tomorrow and probably next year too.

The problem arises once you have built twenty features on top of it, have three hundred customers hanging off it, and then something changes. A rule, a ruling, a customer who demands it contractually. At that moment, switching is no longer an afternoon's work. Then it is a project of months in which you have to review every piece of your product all over again, while your customers wait.

That is exactly the trap generic tools lure you into. You build your processes in Monday, your automations in an American AI integration, your conversation processing in yet another American service. Each on its own a fine choice. Together a web you cannot get out of without turning everything upside down.

From my own practice: we had twenty-six places in the code where we called an AI, each with its own settings and its own model name. That had grown up in an afternoon and could have stayed that way for years. Switching it over cost us one day, because we were still early enough. With two years of conversations and customers on top of it, it would have been a quarter-long project.

What we did

Three things, and I am deliberately being concrete about them, because "we take privacy seriously" is not a statement.

The AI analysis is French. Everything that analyzes your strategy, your numbers, your goals and your team runs on Mistral, a French company that processes on servers in the EU. No American provider, no American parent company. Training on customer data is turned off, and that is a setting we can show rather than promise.

The conversation processing is German. We used to record conversations with an American service that then processed the entire transcript. That is the most sensitive stream there is for us, because in a good conversation someone says everything. That now runs through tl;dv, a German company with storage in European data centers.

Employee names do not leave at all. This is the most important one, and it stands apart from whichever country you choose. Our quarterly report looks at how individual people perform. Those names are replaced by a neutral label before anything goes to the AI. So the AI does see that someone is falling behind, but not who. In the app you simply see the name.

That last one is the kind of measure you never notice, and that is exactly why no one did it. An assessment of an employee who does not even know it is going to an external party is not something you solve with a checkbox in a contract.

The choice you have yourself

You do not have to adopt my choice. There is a good case to be made for the opposite: the American models are still better on the heaviest tasks, the offering is broader, and the integrations are ready to go. If you are building a consumer product that contains no sensitive data, this is a non-issue and you should just take whatever works best.

But do ask yourself three questions before you switch on the next AI feature.

First: what exactly is going to it? Not in broad strokes, but literally. In our case it turned out that lines were leaving in the form of "Jan de Vries: C-player, performance 4 out of 10". I did not know that until I checked.

Second: what does it cost me to switch? If the answer is "I do not know", then the answer is probably "a lot". Make sure the provider is one setting and not twenty-six places in your system.

Third: what do I tell my customer if they ask? Not whether you can come up with a good answer, but whether the answer is true and stays true when they keep asking.

Why this is more than a technical choice

The nice thing is that this exercise gave us more than a better story. Because we had to bring all the AI calls into one place, we discovered that two features had no limit whatsoever on their usage. That could easily have become a bill. We also found a guarantee in our own data processing agreement that we could not live up to, because it referred to an arrangement we had never applied for.

That is what usually happens when you dig into one thing really well. You find three other things that had been out of alignment for a while.

And that is ultimately what an operating system is about. Not that you have the best tool everywhere, but that you know what is happening in your company and that you can change it whenever you want. An AI you cannot replace is not a tool. It is a dependency you have not felt yet.

---

Want to know exactly how we set this up, or is your customer asking about your data processing? Take a look at our security and compliance page or send me a message.