The question behind the question
The worry I meet most often in companies is rarely spoken out loud. It surfaces in half-sentences, in a raised eyebrow, in a careful "and what does that mean for us, then?" Translated, it reads: will this take my work away — or will it soon decide over my head?
That worry is legitimate, and it deserves an honest answer. Because the most important question about AI isn't what the technology can do. It's what relationship we want to enter into with it. The answer to that is a stance — and with us it isn't just a line on a slide, but built into the architecture. This is the chapter of our series I promised at the end of the development story: how we deal with AI in principle.
Multiply, don't replace
The core of our stance fits in three words: multiply, don't replace.
Good AI doesn't take away the work that gives meaning — it takes away the work that isn't the point. The typing, the searching, the summarising, the structuring, the looking-up across a dozen systems. Exactly the activities that fill the day and leave you in the evening with the feeling of not having done the real thing.
I live this thesis on myself every day. That a single person can run a lab at this density isn't because the machine thinks for me. It's because it takes over the craft while I set architecture and judgement. The goal of AI isn't to make the human redundant. It's to give them their attention back — for what no one can replace them in: judgement, relationships, decisions.
Whoever sells AI as a replacement is selling a threat. We sell a tool. The difference isn't cosmetic — it determines every single design decision that follows.
The human keeps the last word
If AI multiplies rather than replaces, then a clear principle follows: the AI proposes, the human decides.
In practice that means consequential steps don't simply "happen." Anything hard to undo, anything that goes outward, anything that affects people, runs through an explicit approval. The system prepares, justifies, lays out — and then pauses. A human presses the button. We didn't build a barrier after the fact for this, but one that's built in: an approval-and-audit mechanism that sensitive actions have to pass through first.
That's not a brake out of caution. It's a boundary out of principle. Because responsibility can't be delegated to a tool. A machine can take on a task, but it can't be accountable — accountable is always a human. A system that pretends to carry the responsibility doesn't relieve the human of a burden; it relieves them of control. That is exactly what we don't want.
An AI allowed to say "I don't know"
The most dangerous AI in a company isn't the one that can't do something. It's the one that's confidently wrong.
A system that answers every question fluently and convincingly — even when it has to guess — is a risk disguised as comfort. We deliberately chose the opposite: an AI allowed to show its uncertainty. Concretely, that means answers come with an estimate of how reliable they are, that sources are named, that the system questions its own answer and starts over when its confidence is too low — and that AI-generated output is labelled as such.
An answer that names its sources and admits its doubt can be checked. An answer that bluffs can only be believed. In an operation where decisions involve money and people, honesty matters more than eloquence. An "I'm not sure here, please check these two sources" is not a flaw. It's the mark of an assistant you can trust.
Memory makes responsibility possible
You can't hand responsibility to something that forgets everything. An assistant that begins every conversation at zero stays a stranger — however clever it is in the single moment.
That's why memory sits at the centre of everything we build. But memory alone isn't enough; it has to be inspectable and correctable too. You can see what the system has stored about you and your company, and you can change it. That traceability is what turns a black box into a colleague. A colleague whose memory you're not allowed to check wouldn't be a trustworthy colleague — but a risk with a good memory.
Trust doesn't come from a system being impressive. It comes from being allowed to look inside. That hardening the infrastructure is part of this I described in the stress test; here it's about the layer above — the relationship itself.
Where governance begins
"Governance" sounds like a brake, a form, a slowdown. For us it's the opposite: it's what makes delegation safe in the first place.
When you hand a person a task, you agree beforehand what they may decide alone and when they check back. With an AI system it's the same. Governance begins where it gets serious: with personal data, with irreversible steps, with anything that leaves the house. Human oversight is not by accident a load-bearing principle of the EU AI Act — and competence in handling AI has even been mandatory there since early 2025. We treat both not as a requirement but as a design principle.
The real thought is this: rules aren't the opposite of pace. Clear boundaries are the precondition for moving fast without fear. A team that knows exactly what the AI may do on its own and where the human signs off delegates more boldly — not more timidly. Trust you can verify dares more.
What remains
The relationship we want with AI isn't one between master and machine, nor one between supplicant and oracle. It's the one between a human who keeps the direction and the responsibility, and a tool that multiplies their reach, is honest about its limits, and lets itself be checked.
This isn't a marketing stance you put on when needed. It stands in the architecture: in the approvals with which the human keeps the last word; in the confidence scores and sources with which the system stays honest; in the inspectable memory that makes trust verifiable. A stance that isn't in the code is only a claim.
This is a chapter of our series on dealing with and adopting AI. How to actually begin — from the first tangible process to an operating system — is in the first post. In the coming parts we'll look more closely at the individual research lines.