Luxembourg legal AI start-up Clerk raises €2.8mn at €6.5mn valuation

Ask most legal chatbots a question and you get a paragraph back: fluent, confident and difficult to check. Ask Clerk, a young Luxembourg company, whether a summary dismissal was abusive, and, on the company’s account, it does something closer to what a junior lawyer would do at a desk.

It plans the task, opens the employment contract, computes the notice and severance owed, searches the grand duchy’s case law for decisions on gross misconduct, reads a 2023 ruling of the Cour Supérieure de Justice in full, and concludes that unexplained absences without prior warning are rarely grave enough to justify an immediate dismissal, showing each step as it goes.

That distinction, between a system that answers and a system that works, is the whole of Clerk’s argument. Its own framing is blunt: this is “an agent with its own workspace, not just a chatbot”.

Why “agentic” is the claim that matters

The word doing the heavy lifting is agentic, and it is worth being precise about what it means, because it is where Clerk stakes its difference. A conventional legal chatbot is, in effect, a thin layer over a language model: it takes a prompt and returns a single passage of text in one shot, with no lasting sense of the matter and no ability to act on the world.

An agentic system is built the other way round. Given a goal rather than a question, it plans a sequence of steps, decides for itself which tools to use, executes them, checks the results against sources and loops until the task is finished. The model is the reasoning engine; the value is in everything wired around it.

For law, that architecture is a genuine departure rather than a cosmetic one, because legal work is inherently multi-step and evidentiary. A lawyer does not simply answer; they research, verify, apply the statute to the facts, compute the figures, draft the document and cite their sources. It is a chain of dependent actions in which a single wrong link is a professional failure, not a rounding error.

A one-shot chatbot collapses that chain into a guess. An agent that can carry it out step by step, grounded in a real corpus and made to show its work, is attempting something categorically harder.

It is fair to keep the engineering in perspective. Agentic products of this kind are, under the hood, generally an orchestration of existing foundation models with document retrieval and a set of tools, rather than a wholly new form of artificial intelligence. The novelty lies in the assembly and in the discipline imposed on it: grounding, verification and source-tracing, aimed squarely at one jurisdiction.

On the company’s telling, that is exactly the point. Clerk’s own site describes each capability not as a menu item but as a tool the agent reaches for on its own, choosing among them as the matter demands.

What it claims to do

The breadth of that toolset is what makes the pitch ambitious. According to its website, Clerk sets out to cover most of the desk work of a matter rather than a single slice of it.

It researches jurisprudence across what the company describes as 45,957 Luxembourg decisions drawn from seven courts and tribunals, opening the ones that apply and citing each to its source. It performs document analysis, running optical character recognition over scanned judgments and contracts in French or German and extracting the facts, clauses and dates that matter.

It handles deadlines and quantum: procedural time limits, prescription periods and indemnités are computed the moment the facts are entered, with the governing article of the code attached and the resulting date placed on the matter’s calendar. It drafts and exports, producing mises en demeure, conclusions and pleadings from the file and delivering them as formatted .docx or PDF documents on a firm’s own letterhead.

And it is built to argue with a check on itself. Each argument is paired with its counter-argument and a confidence marker, and any claim the system cannot ground in a source it has actually read is flagged rather than passed off as settled. Around all of this sits a shared workspace: colleagues can be brought on to a matter as lead, contributor or observer, with firm-wide roles, usage metered per seat, and an audit log recording every access.

The reach is deliberately wide. The company markets it from the sole practitioner to the largest étude, and across employment, litigation, corporate, banking, tax and in-house work. But it stops short, pointedly, of the claim that it replaces the lawyer.

Its stated ambition is to absorb the high-volume legwork: the research, the first drafts, the deadline arithmetic and the trawl through the state case-law portal, so that a firm’s lawyers spend their hours on judgment and clients instead. That framing is both more honest and more persuasive than the alternative: an agent that does the desk work end to end, and leaves advocacy, strategy and professional responsibility where they belong.

The founder and the funding

Behind Clerk is Tim Kerger, a young Luxembourgish founder who also runs PrimeResearch, a project focused on prediction-market data. It is a portfolio that points to a builder at ease moving between quantitative finance and applied AI.

Clerk has raised €2.8mn at a valuation of about €6.5mn. The round gives the company the capital to turn a sharply defined prototype into a working tool for the grand duchy’s law firms, a market that is small by population but unusually dense in legal, financial and cross-border complexity.

The thesis is easier to appreciate against the local backdrop. Luxembourg is a jurisdiction of only around 660,000 people whose case law is handed down in French and German and dispersed across the state portal justice.public.lu. Global tools optimise for scale, which tends to mean the largest and most standardised markets; a small, bilingual civil-law system is the kind of place they serve last.

Clerk’s wager is that a deep local corpus, combined with anonymisation before any model sees a client’s data, European data residency and alignment with the EU’s AI Act, is a wedge a larger competitor would find awkward to drive in.

A crowded, well-funded field

It is entering a category that is, elsewhere, unmistakably hot. Legora, a collaborative AI platform for lawyers, raised a Series C at a reported $1.8bn valuation, led by Bessemer Venture Partners, a marker of how much capital is chasing the space.

Closer to home, the Italian RegTech firm Aptus.AI recently opened Luxembourg operations; the company raised €3.27mn in seed funding for technology that turns legal text into machine-readable, structured data. Luxembourg-and-Belgium-focused incumbents such as Alizé already market research assistants that answer from cited legal sources and are dedicated exclusively to legal professionals, and the grand duchy’s public institutions are building their own: the University of Luxembourg has launched an AI assistant for patent law, and the national research institute LIST has partnered with the Competition Authority on AI tools for enforcement.

Against names of this size and funding, Clerk is small and early. But a full agentic workspace aimed at a single jurisdiction is a broader and more distinctive proposition than a point tool, and that is the ground it is choosing to fight on.

What to watch

Whether the ambition holds will turn almost entirely on trust. Law is an unforgiving setting for generative systems, and the single most important thing Clerk claims is precisely the claim that will be tested hardest, and adversarially, by working lawyers: that every assertion is traced to a source it actually read, and that anything ungrounded is surfaced rather than hidden.

A confident but fabricated citation, or a missed délai, is the kind of error the whole design exists to prevent, and the design will be judged on whether it does. Beyond that sit the professional-conduct rules that govern how lawyers may lean on such tools, the obligations the EU AI Act places on higher-risk applications, competition from generalist models that keep improving and from far better-funded rivals, and the execution risk that faces any young company trying to sell into a conservative profession.

Still, there is a great deal to like in the shape of the bet. A jurisdiction the size of Luxembourg rarely gets software built expressly for it, let alone a system that tries to work an entire matter rather than answer a single question. If the grounding and the source-tracing survive real use, a broad, jurisdiction-native agent is a genuinely differentiated place to stand, and a young founder building it in his own country is a story worth following.