Legal AI Luxembourg

What legal AI is, and what it is not

Legal AI is not a language model you ask legal questions to. A general-purpose model such as ChatGPT has read a fraction of the web, has a statistical intuition of what a legal answer looks like, and will happily produce a plausible but wrong article number, or a judgment that never existed. On Luxembourg law, whose published online volume is small compared with French or American law, the effect is particularly pronounced.

Legal AI worthy of the name rests on a different architecture, known as retrieval-augmented generation: the model does not answer from memory, it answers from documents a search engine has just retrieved from a defined and verified database. The reasoning belongs to the model, the raw material comes from the database. If the database holds nothing relevant, the system must say so rather than invent.

This distinction has a direct practical consequence: the quality of a legal AI is judged first on its corpus and its search engine, not on the power of its model. An excellent model connected to an incomplete corpus produces elegant, unusable answers.

The Luxembourg corpus covered by Pandectis

The Luxembourg database holds 294,203 searchable documents, all from official sources and reindexed daily. Unlike a general-purpose database, it is not limited to case law: Luxembourg law is also, and above all, practised through the soft law of regulatory authorities, which remains largely absent from traditional research tools.

• Case law Judicial and administrative court decisions, Court of Cassation and Constitutional Court, as published by the Luxembourg judiciary.
• Legislation Statutes, the Constitution, bills and parliamentary documents from the Chamber of Deputies, with the status in force at the date of the text.
• Regulatory acts Grand-ducal regulations, ministerial orders and circulars published in the Mémorial.
• Consolidated codes Luxembourg codes in their consolidated version, article by article.
• Soft law from fifteen authorities Circulars and positions from the CSSF, CNPD, ILR, ACD, CAA, BCL and other regulators: the layer of law that decides compliance day to day.
• Parliamentary questions Questions and answers from the Chamber of Deputies, often the only place where an administrative position is stated in writing.
• 279 tax treaties Double taxation treaties concluded by Luxembourg, in full text and searchable in natural language.

970 categories built for Luxembourg law

A corpus, however complete, is not enough. Without structure, semantic search behaves like a library without classification: it finds books, but it cannot tell you what the subject is.

Pandectis therefore built a three-level architecture from scratch, designed specifically for Luxembourg law: 32 subject areas, more than 970 categories and 3,594 semantic facets. Tax law breaks down into categories such as abuse of law, dividends, holdings or permanent establishment; each category breaks down in turn into facets isolating a precise question, for instance the deductibility of a holding company's financial charges or the anti-abuse clause of the parent-subsidiary directive.

This structure does two things search alone does not. It allows you to browse the law without knowing what to look for, which is the real situation on a new matter. And it gives the engine context to choose between two documents containing the same words but addressing different subjects.

Asking your question in natural language

In practice, you write your question as you would put it to a colleague: "liability of a S.à r.l. manager for mismanagement", or "anti-money laundering due diligence obligations of a financial sector professional". You can also give a precise reference, a code article or a docket number.

The engine analyses the question, matches it against the corpus and the thesaurus, then ranks documents by actual relevance rather than by word occurrence. Every result carries a link to the official text, on Legilux, on the judiciary portal or on EUR-Lex depending on the source, so that verification does not depend on us.

The jurisdiction selector lets you restrict the search to Luxembourg, France or Europe, or let it run across all three when your matter is cross-border. You can test the engine without signing up: five searches are free.

Why Pandectis would rather answer nothing

Hallucination is the central professional risk of legal AI. A model that invents a judgment is not merely inaccurate: it exposes its user, and in a lawyer's case, it exposes them before their client and before the court.

The Pandectis answer is architectural, not cosmetic. The assistant can only answer from documents present in the database, and every statement it produces links back to the document it comes from. There is no answer without a citation, and no citation without a verifiable reference.

The corollary is visible in the public demonstration: when the engine finds nothing relevant enough, it says so rather than offering loosely related documents. That is not a coverage gap, it is the same principle that stops the assistant from inventing. A tool that always answers is a tool that sometimes invents.

Professional secrecy, GDPR and the AI Act

For a Luxembourg lawyer, the AI question is first a professional secrecy question. Sending case documents to a model hosted outside Europe is not a technical detail, it is a decision that carries liability.

The Pandectis platform, database and search histories are hosted in France, within the GDPR perimeter, with encrypted queries. A sovereign mode is available: the model itself is then hosted in France, and neither the question nor the attachments leave the territory. Outside that mode, an explicit warning precedes any transmission, and no document is sent without confirmation.

Under the European AI Act, a legal research assistant is not a high-risk system, but the transparency obligation remains: the user must know they are addressing a machine, and what it relies on. That is precisely what a system citing its sources systematically does.

Luxembourg is never practised alone

A purely Luxembourg tool answers actual practice poorly. Luxembourg law is constantly interpreted in the light of European law, and French doctrine and case law regularly serve as argumentative reference, the kinship of the codes allowing it.

Pandectis runs the same engine over three corpora: 294,203 Luxembourg documents, 3,791,152 French documents including case law, consolidated codes, the BOFiP and CNIL documentation, and 272,087 European documents including CJEU and ECHR case law, secondary legislation and ESMA and EDPB soft law. Belgian law is being integrated.

On an international tax matter, that means retrieving in a single search the circular from the Luxembourg direct tax authority, the applicable tax treaty and the CJEU judgment that conditions its interpretation.

What legal AI costs in Luxembourg

Pandectis consists of a platform plan giving access to search, the category-based encyclopedia and daily monitoring, and an advanced chatbot billed on usage with no fixed subscription: you pay only for the questions processed.

Both are quoted, because a two-partner firm and a thirty-lawyer legal department share neither volume nor requirements. We reply within twenty-four hours, with a thirty-minute demonstration on your own subjects rather than on examples we picked.

Frequently asked questions about legal AI in Luxembourg