Ask an AI itself and you get "yes" back within two seconds, sources included. That's not a trick: it's simply correct. The language question is a solved question that people still ask out of habit.

What's underneath is not solved. Almost nobody who asks whether AI speaks Dutch is worried about grammar. The real worry is that something will end up under your name that you would never have phrased that way yourself. That's a legitimate concern, and it's the most underestimated reason people drop out.

Below is exactly where the difference sits, what a voice actually is once you take it apart, and how you capture it without starting on a blank document.

01 / the languageHow good is AI in Dutch, really?

Good enough that on grammar and fluency alone it's indistinguishable from an average copywriter. The major language models are trained on enormous volumes of Dutch text and switch effortlessly between registers: formal, informal, business, warm.

Where it sometimes struggles: very local expressions, jargon from small niches, and the nuance between Netherlands Dutch and Flemish if you don't explicitly ask for it. Those are edge cases and they're solvable with examples. The language itself is no longer a reason to hesitate.

02 / the differenceSo why doesn't AI text sound like me?

Because correct Dutch and your Dutch are two different things. A model with no knowledge of you writes the average of everyone: neat, smooth and interchangeable. You recognize it by sentences nobody could ever object to, and that therefore also reach nobody.

Your voice lives in the deviations from that average. The words you do use and the buzzwords you refuse. How directly you address someone. Whether you joke, and about what never. The examples you've been reaching for for years because they work.

None of that is in a language model. It's in your head, and as long as it only lives there, no system can write with it.

03 / the breakdownWhat is a voice, in practical terms?

Three layers, and all three can be captured. That's the good news: a voice is not a mystical talent but a set of repeatable choices you already make, just nowhere written down yet.

The most honest source for all three is existing material: sent emails, written proposals, old posts. That's where how you actually write shows up, not how you think you write. That difference is bigger than most people expect.

04 / capturing itHow do you capture a voice without a blank document?

By recording what you're already doing, instead of sitting down to write. The blank-document problem is why this stays on almost everyone's list forever: you have to pour your entire way of working into something in one go, and that moment never comes.

Four steps, in this order:

• Record what you're already doing. A conversation where you explain your approach. Not staged, just recorded when it's actually happening.
• Transcribe it. From recording to text. This is the step that requires no thinking and therefore always succeeds.
• Organize it into fixed themes. This is where the difference between a pile and a system appears.
• Deduplicate. You explain the same things more often than you think. What's left is the core.

Collecting gives you an archive. Organizing gives you a system.

This is also how we build it out with a client: a sixty- to ninety-minute conversation in which someone explains their own way of working, and that becomes the base layer. Not because a recording is magic, but because people are more precise about their own craft when speaking than when writing. In a form you fill in what you think is expected. In a conversation you say what you mean.

05 / the learningHow does a system learn my voice?

In two phases. First you feed it that existing material, so the first drafts already land close. Then the correction loop does the real work: every draft you adjust before sending is a lesson. Adjust the same kind of sentence three times, and that pattern should become the default.

That's precisely why everything comes in as a draft with us instead of sending itself. The approval step isn't a safety lock we bolted on afterward: it's the place where your voice flows into the system. A system that publishes itself learns nothing. More on that in why we don't publish your posts automatically.

Expect a few weeks in which you correct more than you'd like. That's not a flaw, that's lesson one through twenty.

06 / what remainsWhat's left of me in the end result?

If it's done right: everything that matters. The system carries the form, the structure and the volume. You supply the opinions, the examples and the final word. Readers notice no difference, because there shouldn't be one: every piece is approved by you and sharpened where needed.

And if something still doesn't sound like you, you don't send it. That's the whole safeguard. There's no version of this where you're stuck with text you're not allowed to change; what's yours and what isn't is laid out in whose is it when it works.

Which means "does AI work in Dutch" turns out to be a small question after all. The big question is whether you're willing to capture your voice. That work can't be taken off your hands by anyone, not even us.

This text was produced with AI assistance and reviewed and approved by a human before publication.