Luxembourg is, by some distance, one of the most lawyer-dense jurisdictions in the world. The Barreau counts more than three thousand five hundred avocats for a population of roughly six hundred and sixty thousand — somewhere around five to six times the European average per capita. Add in the magistrates, the notaries, the in-house counsel at the banks and fund houses, and the compliance teams at the CSSF-regulated entities, and the density gets even more remarkable. We are, in proportion to our size, a country of lawyers.
We are also a country whose legal sector underwrites a disproportionate share of the economy. Cross-border finance, structured products, fund administration, holding-company architecture, EU institutional litigation — most of the work that makes the Grand Duchy a financial centre is, at its core, legal work. If the productivity of that work falls behind the rest of the continent, the consequences are not limited to law firms. They are felt in every sector that depends on Luxembourg's ability to deliver fast, precise, defensible legal output.
The productivity gap is already opening
Across the Atlantic and across the Channel, lawyers are quietly integrating large language models into their daily practice. The first wave was research and summarisation; the second wave is drafting, redlining, and structured analysis of large evidentiary corpora. Whatever you think of the technology, the empirical picture is clear: a partner with the right AI workflow now does in an hour what a partner without it does in a working day. Those hours compound. Over a quarter they become a competitive advantage; over a year they become a market.
Luxembourg cannot afford to be the jurisdiction that arrives late. Our clients — the international banks, the asset managers, the multinationals choosing where to domicile — will reward speed, and they will not wait for us to feel comfortable.
But the obvious shortcut is closed
The obvious move would be to do what most of the world does: open ChatGPT, paste in the matter, ask the question. For a lawyer bound by professional secrecy under Article 35 of the 1991 law, for a banker bound by Article 41 of the 1993 law, for anyone handling personal data under the GDPR, that move is closed. It was closed the moment it was first considered. Pasting an identifiable client's circumstances into a US-hosted model is not a grey area. It is a textbook breach of confidentiality and, in most cases, an unlawful international transfer of personal data.
This is the paradox. The most productive tool in the room is the one the profession is not allowed to use. And the cost of that paradox is not theoretical: it is being paid every day, in billable hours we cannot recover and in mandates we cannot move as fast as our competitors abroad.
Sovereignty is the answer, not a slogan
The word "sovereign" gets thrown around in EU technology policy until it loses meaning. I want to be precise about what it has to mean for legal AI to actually be usable in Luxembourg. It means three concrete things.
First, the data the model sees must not be the client's data. Names, identifiers, addresses, matricules, IBANs, registry numbers, cadastral references — every piece of information that could re-identify a party has to be replaced with stable tokens before a single byte reaches the model. The model can still reason about "PARTY_A", "PARTY_B", and a contract clause. It cannot reason about Mr. Vermeulen and Axion S.A.
Second, the infrastructure must be European, end to end. Inference in EU regions, under EU contracts, with no-training and no-retention terms that an information-security officer can actually read and verify. No transatlantic transfers. No silent fall-back to a US endpoint when the European one is under load.
Third, sovereignty cannot mean second-rate models. This is the part most policy conversations get wrong. There is a real temptation to settle for whatever fine-tuned open-source model can be hosted on a Luxembourg or French GPU and to call it a day. The profession will not adopt a tool that is architecturally pure and analytically mediocre. The bar is set by the frontier — by Claude, by GPT, by Gemini — and the only honest way to compete with the productivity gains those models offer is to make them available, under the right legal envelope, to the people who need them.
This is why I founded Clerk
I founded Clerk because I refused to accept the choice between modern AI and confidentiality as a real choice. It is a false dilemma created by tools that were designed for a different jurisdiction with different obligations. The architecture exists today — anonymisation pipelines, reversible tokenisation, EU inference endpoints under no-training contracts — to deliver the productivity of frontier models without ever exposing a client's identifiable facts to them. What was missing was the will to assemble it specifically for Luxembourg practice, in Luxembourg, for Luxembourg legal vocabulary and Luxembourg regulatory reality.
Clerk is that assembly. Every document is anonymised before it touches a model. Every inference happens in an EU region. Every response is rewritten back into the original entities locally, inside the firm's perimeter, so only your team ever sees the full context. The model sees the minimum information needed to answer. You see everything.
What I am asking
If you practise law in Luxembourg, in any capacity, I would like you to try it. Not because I want a customer — that part will sort itself out — but because the alternative is that the Luxembourg legal sector becomes, in five years, the slow corner of European legal practice. I do not want to live in that country, and I do not believe you do either. We have the density, the expertise, and the regulatory clarity to lead this transition rather than be dragged through it. We just need to build the workspace that lets us.
I read every message that comes in through hello@clerk.lu. If something in this letter resonates — or if something in it is wrong — I would like to hear from you.
Tim Kerger · timkerger.com
Founder, Clerk