The ground beneath AI is moving. Our approach is not.

I have been building and managing digital platforms for fifteen years. In all that time, there is one lesson I’ve had to defend more often than any other: don’t start laying bricks before you know where the pipes will go.

AI is different. Not because the foundation no longer matters, but because the ground itself is still shifting. Anyone working with AI today seems to face a choice: invest heavily in something its developers say they do not fully understand or wait and catch up later at great cost. We do neither.

On 6 September 2026, Jakub Pachocki, chief scientist at OpenAI (ChatGPT), published the essay An Alien Mind. According to him, AI is grown rather than designed: a simple optimisation step, repeated countless times, eventually produces a system nobody can fully describe. He says no lab has solved alignment and control well enough to keep scaling at maximum speed for much longer.

Shortly after that, Dario Amodei, CEO of Anthropic (Claude), called for the pace to slow down. According to him, that only works if the leading developers slow down together and if someone from the outside can verify it. Chip stocks fell within the week.

All these statements followed an eventful summer. During internal safety tests in July, OpenAI models bypassed their safeguards and reached systems at Hugging Face, a third party. Other incidents stayed under the radar for months before becoming public.

What stands out to me most is not the technology. It is the fact that a sector explicitly calling for regulation is mainly admitting that it does not regulate itself. But the tools are not the problem. The interests behind them are. In a market where slowing down means your competitor catches up, nobody wants to brake first.

The cautious advice is obvious: wait, let the dust settle and only invest once things become more stable. We keep building, but not because the market is pushing us. We do it because we see the potential and because we can approach it purposefully.

Besides, waiting is not a neutral choice. Around 2012, almost every website proposal still included an optional line: ‘optimised for mobile devices’. Two years later, that line had disappeared. Not because the discussion had been settled, but because mobile had become a basic condition. With AI, that shift is happening faster. If you wait, you don’t gain certainty. You mainly fail to build experience.

Our content colleagues have worked with a simple rule for years: ‘Don’t build your house on rented land.’ Social platforms are useful, but you don’t want your audience, your content and your knowledge to only exist there.

The same applies to AI. We will keep using models and cloud services. Rented land is perfectly good land. But our knowledge and our way of working cannot be allowed to take root there.

That is why we follow one design rule: everything we build with and for AI is as modular as possible. We don’t know which supplier will still exist in two years, which model will be best by then or what regulation will look like. What we can control is that one change does not take the entire system down with it.

Of course, we won’t achieve full supplier independence. Models differ and some features only exist with one provider. What matters to us is knowing where the dependency lies and what it would cost to switch.

We work with two environments. Claude is our day-to-day interface for work that doesn’t contain confidential client data. Notice the word ‘interface’: the value is not in the chat window, but in what lies underneath. FoC Buddy is our internal platform for everything involving contracts, proposals or client data.

Both environments are fed by our own connections, built as separate components, including MCP servers. The same connection works in one environment today and another one tomorrow. The same goes for our clients: what we build connects to their systems, not ours. And what we learn about a client remains theirs. We only keep it in a transferable form, so the client can take it with them later, even without us.

For the most sensitive work, we are building towards our own environment, with a local model running on machines we manage. We are also testing European models, such as Mistral, alongside what we use today. Because an American model running in a European data centre is still subject to American law. Data residency and jurisdiction are two different things. Saying your data is in Frankfurt says nothing about who can demand access to it. We believe client work belongs on infrastructure whose legal framework is clear to us. That is also what we advise our clients.

Trust starts with a rule everyone knows and nobody has to interpret. For us, that means three classes: internal work without client content, client work without personal data and confidential material. For each class, it is clear where the work may run. A brainstorm for a campaign concept is not the same as a recording of a client conversation. Nobody has to make that distinction from scratch for every assignment.

New tools are not added lightly. Before we start using one structurally, we ask ourselves four questions: what happens to the data we enter? Is our input used to train models? Which jurisdiction applies to the company? And how easily can we leave if we want to?

Our rules are set out in an AI charter that applies to all six agencies. The most important rule is also the dullest one: nothing goes out before someone has read it, checked it and approved it. AI delivers a proposal, never a finished product. Final responsibility lies with the employee.

We also label AI-generated and AI-edited images in line with the European AI Act, and we keep sources and prompts so we can explain afterwards how something came about.

There is also a list of things we don’t do. No confidential client information or special categories of personal data in external tools. No deepfakes or voice clones without consent. No decisions about people without human review. No work built on someone else’s intellectual property.

Technology is the easy part. But a group of six agencies doesn’t change simply because you roll out a tool. That is why we run monthly AI Academies, recorded and built around roles. Because a copywriter has different questions than an account manager, creative or developer. Every team has someone who works with these tools every day and acts as a point of contact. And if someone is unsure whether something is allowed, they ask.

We keep the same openness with clients: which tools we use, what happens to their material, and where we draw the line ourselves.

1. If our most important AI tool disappears tomorrow, what exactly do we lose?
2. Who has the overview of our knowledge and where does it live: in files we control or in a supplier’s database?
3. Which part of our AI investment is a way of working, and which part is the configuration of one product?
4. Who is allowed to process our clients’ material under our client contracts and does that still match the tools we use today?

If you cannot answer the first question in one sentence, the architecture needs work. Because the ground remains rented. That’s why we’re already planning for the move.