AI only becomes truly useful once it's clear who does what — and, above all, who is not allowed to do what.

It actually started with something mundane.

My new website is multilingual. Dutch, French, English and German. At the top sat a simple language switcher: NL, FR, EN, DE.

It worked.

Just not well enough.

Someone on the Approach page, for instance, who switched to French would land on the French homepage. Technically, the language had changed. For the visitor, the context had simply disappeared.

I wanted that fixed.

At that point, you could of course just say: AI, fix this.

That's roughly the opposite of how I want to work with it.

First deciding what “fixed” actually means

The question wasn't: build a better language switcher.

The question was:

• keep the same underlying page when the language changes;

• use the existing translations and URL structure;

• change nothing else on the rest of the site;

• work outside production first;

• make sure every change is verifiable and reversible;

• and publish nothing without explicit approval.

That might sound excessive for a language button.

But that's exactly where, for me, the difference lies between using AI and letting AI carry out the work.

The moment a system can actually change things, a good prompt is no longer just a question. It becomes almost a work order: this is the goal, these are the boundaries, this is where you stop, and this must be proven before we continue.

Not one AI doing everything

For this small change, I ended up using three different systems.

ChatGPT kept the overview. Defining the problem, working out the steps, scoping the risks, and deciding each time what the next permitted action was.

Claude Code got direct access to the Framer project. Read-only at first. Then creating a separate branch. Then adjusting the language switcher, reporting exactly which parts had changed and, above all: touching nothing else.

Something useful came out of that right away.

The solution I'd pictured visually — a permanent NL / FR / EN / DE — turned out to be technically weaker than Framer's own native language component. That native solution shows one language with a dropdown.

Less like what I'd originally sketched.

Functionally, much better.

So someone had to decide.

That wasn't Claude.

That was me.

I let go of the visual preference because the functional solution was better. AI could surface the options. The trade-off stayed human.

And even then, I still didn't just take Claude's word for it

Not because Claude had done poor work.

But because whoever builds something shouldn't automatically be the one who approves it.

Claude could show that the right component was in the project. That the mobile version inherited correctly. That only navigation and footer had changed.

But that still doesn't prove what a visitor experiences in an actual browser.

So the same version went to a separate test environment.

Kimi, via WebBridge, then got a completely different assignment: build nothing, change nothing, just do what a visitor would do.

Open the page.

Switch language.

Check the URL.

See whether the same page was kept.

On desktop.

On mobile.

And back to Dutch.

Approach stayed Approach.

Contact stayed Contact.

Support stayed Support.

Dutch → French → English → German → Dutch.

Not a single switch dropped the visitor back onto a homepage.

Only then was the change, for me, a success.

So the interesting part was never the language switcher

That turned out to be the least interesting part of the story.

What I find more interesting is the division of labour.

No single system got handed the whole process.

One helped determine what should be right.

The second was allowed to build within clear limits.

The third had to check what was actually happening.

And I remained the one who decided when a deviation was acceptable, when a next step could happen, and ultimately when something could go to production.

That may sound less spectacular than “AI built my website.”

But it's much closer to how I think AI actually becomes useful inside a business.

Not by handing over as much autonomy as possible.

But by defining very precisely:

who is allowed to think, who is allowed to execute, who checks — and where the decision stays.

In this case, it all started with four small language choices at the top of a website.

It ended with a way of working that I now find far more interesting than the button itself.