Roderick Lucas, Co-Founder and CEO of Moonlit – Interview Series

Roderick Lucas is Co-Founder and CEO of Moonlit, bringing experience in artificial intelligence, data science, and legal technology. He spent nearly a decade at Deloitte, where he served as Artificial Intelligence Lead, and has taught technology courses at Vrije Universiteit Amsterdam and Tilburg University. His work focuses on making legal research more accessible through structured data, verifiable sources, and AI.

Moonlit provides legal data infrastructure and research tools that connect official legislation, court decisions, and regulatory information across European jurisdictions. Its platform, Data API, and MCP server give researchers and AI assistants access to structured legal sources, while its Commentaries library combines AI-generated analysis with review by named legal experts. In this interview, Lucas discusses legal hallucinations, cross-border research, expert verification, and access to trustworthy legal intelligence.

You spent nearly a decade at Deloitte, eventually serving as Artificial Intelligence Lead, while also teaching technology courses at Vrije Universiteit Amsterdam and Tilburg University. Moonlit originally emerged as an internal project at Deloitte before becoming an independent company in 2024. What inspired you to develop the platform, and what convinced you that legal research needed a fundamentally different approach?

Moonlit started inside Deloitte because the problem was sitting right there. Every serious legal or tax question in Europe crosses a border at some point, yet the sources you need are scattered across dozens of national portals, each with its own format, language, and citation style. Specialists spend most of their research hours gathering material and verifying nothing is missing. Meanwhile, the tools built to help them are ineffective for cross-border work. They charge per seat per country and require separate logins just to compare how different jurisdictions handle the same issue.

When language models arrived, two things became clear at once: the models could reason about legal text far better than anyone expected, but they had nothing reliable to reason over. Extracting text directly from public court portals yields messy, inconsistent data, while commercial legal publishers restrict automated access behind licensing restrictions. Legal technology lacked a common foundation: official sources across multiple jurisdictions, standardised into the same clean data structure, cross-referenced, and updated daily.

That layer is what we set out to build, and it is why Moonlit became a company in 2024 rather than just a tool for one firm. A foundation like this gets more valuable with every person and every system that reads from it and contributes to it. The conviction behind it has not changed: the law is public, and the knowledge built on it should be affordable to everyone who has to comply with it.

Moonlit was founded on the belief that access to legal knowledge is essential to justice. While legislation and court rulings are often publicly available, the expert commentary explaining their implications frequently remains behind expensive paywalls. How does Moonlit aim to democratise access to this interpretive layer and why is it particularly important for the development of reliable legal AI?

In the Netherlands, about 40 cents of every euro you earn goes to the state and part of that funds public universities and the legal system. Much of the commentary explaining the law is written by academics on public payrolls. Commercial publishers like Wolters Kluwer and Thomson Reuters package that analysis and sell it back to the courts, ministries, universities, and practitioners who need it. A commentary subscription runs to hundreds of euros per user per month. With these subscriptions, the public effectively pays twice for the law it must follow. A fair legal system should not cost more per month than your Netflix subscription.

We built Commentaries to provide an open alternative: an article-by-article library of legal analysis accessible with a free Moonlit account. Each commentary is generated from primary sources using our AI models and vetted by legal specialists before publication. Every citation is verified and links directly to the official government or court record. If a reader notices a missing nuance or a recent development, they can suggest an edit directly on the form on the page for editorial review. The editors can then review it and republish. To our knowledge, no other publisher offers that combination. At launch, the available commentaries focus on indirect tax (VAT), and commentaries around data protection are in the works.

This interpretive layer is equally critical for legal AI. Models can ingest public statutes and court judgments, but they cannot read the secondary commentary that explains how a rule operates in practice or where the exceptions lie. Deprived of that context, models extrapolate and generate unreliable answers. We designed Commentaries so that every statement is explicitly anchored to the provision it interprets and the ruling it rests upon. A practitioner can verify the source in a single click, and an AI workflow can reason over reliable data through the Moonlit Data API and MCP connector.

AI-generated legal hallucinations have become a growing concern, with lawyers and courts encountering fabricated case citations and inaccurate interpretations of the law. Why do even sophisticated large language models struggle with legal accuracy, and how does providing structured, verified legal context help address these limitations?

Language models struggle with law for reasons that have little to do with intelligence. Legal reasoning depends entirely on verifiable authority. A conclusion is only as good as the provision or ruling it rests on, but a model treats case numbers and statutory citations as probabilistic text patterns rather than reliable database records. It outputs references that look convincing because it predicts what a citation should look like, not because it verified an actual document. This has very real consequences: as of writing, there have been over 2,100 recorded court decisions involving AI hallucinations, more than 800 of which directly involved legal practitioners.

The law also moves constantly. A model’s knowledge stops at its training cutoff, whereas tax rulings and appellate decisions update every week. When the commentary explaining what those rules mean sits behind publisher paywalls, the model has access to the bare statute but not the operational context, so it invents plausible explanations to bridge the gap.

Structured, verified context changes the fundamental task. At Moonlit, the model retrieves the law instead of attempting to recall it from memory, and works directly from official records. A search returns documents from official sources, each with its citation and link. The reference graph shows what a ruling cites and which later documents cite it. A citation either points to an authentic document in our repository or it does not appear. That eliminates fabricated citations. However, clean data alone does not prevent misreading, which is why Commentaries still places a named legal expert between the draft and the reader.

Moonlit aggregates legislation, court decisions, and regulatory information across dozens of jurisdictions, connecting legal documents through an extensive reference graph. Can you walk us through the underlying AI and data architecture, particularly how the platform identifies relationships between legal sources, reconciles different languages and legal systems, and makes this information accessible to AI models?

Think of it as one continuous chain, from the official publication all the way to a verified citation.

For every jurisdiction, we connect directly to the official portals, including legislative gazettes, court databases, regulators, and EUR-Lex for EU law. Each portal has its own quirks and formats, and they can change without warning, so a significant part of our engineering goes into monitoring and automated recovery per portal. Every document we bring in retains its attribution and a permanent link back to the official government record. That link is the foundation for everything else.

Everything lands in one model across jurisdictions, covering legislation, case law, parliamentary documents, authority decisions, and guidance. Legislation is broken down to the individual article so that a single provision can be accurately retrieved and cited. A Dutch judgement and a German ruling retain the specific fields their domestic systems require, while we standardise how those documents are identified, classified, and linked across borders.

Legal texts cite each other constantly. A judgement cites the articles it applies and the rulings it follows, while a law refers to other statutes. Our engine recognises these citations, which take very different forms depending on the jurisdiction, resolves them to authentic documents in our Data Layer, and stores the links bidirectionally. Legal engineers verify the extraction patterns. The result is a reference graph containing hundreds of millions of links, meaning that for any document, you can see what it cites and what cites it.

Language, meanwhile, is handled by leaving the source documents intact. We do not translate the primary text. A Belgian ruling stays in French or Dutch, and searches run natively in the original language. We combine keyword and semantic search, rerank the results, and keep each query scoped to the relevant jurisdiction. Luna, our research assistant, answers in whichever language you ask while querying each underlying source in its native language.

The Moonlit platform, our Data API, and the MCP server all read from that same Data Layer. The API provides programmatic access to our search modes, the reference graph, and full-text retrieval, while the MCP server makes nine read-only tools available to any external AI assistant. When we improve the underlying data, that improvement reaches every surface at once.

Moonlit has introduced AI-generated legal commentaries reviewed by named legal experts, combining automated drafting with human verification. How does this process work in practice, and what safeguards ensure that AI-generated interpretations are accurate, transparent, and traceable to authoritative legal sources?

A commentary starts with the law itself. Our engine builds a draft from primary sources in the Data Layer, including legislation, case law, directives, and the parliamentary history behind a provision. It can only cite documents we hold, so every proposition is linked to the provision it interprets and the source that supports it. One of our launch commentaries, for example, links to 165 verified sources.

Before a commentary goes live, a legal expert in that field reads the draft against those primary sources. Our editors are academics and practitioner-authors who are paid for their work. The launch group comes from universities, courts, government bodies, advisory firms, and law firms. They correct the draft, cut what does not belong, and add what is missing. Their names are right on the page.

Verified review badges mark the passages validated by an editor, so it is always clear which sections a human expert stands behind.

Each commentary also states its reference date to show the exact state of the law it describes. The Data Layer updates daily and flags relevant new rulings for the editor, who then revises the text and advances the reference date.

The safeguard underneath all of it is the link. If you cannot click through to the ruling that supports a conclusion, you are being asked to take it on faith. We built Commentaries so nobody has to.

Through Model Context Protocol (MCP) integrations, Moonlit can bring legal intelligence directly into AI assistants such as ChatGPT, Claude, and Microsoft Copilot. What does this architecture enable that conventional retrieval-augmented generation systems may struggle to achieve, and how do you ensure these assistants use the retrieved legal information responsibly?

Conventional RAG takes a snapshot of some documents, cuts them into chunks, embeds them and hands the model whatever scores highest on similarity. For law, that fails in three places. The snapshot goes stale. Similarity isn’t reliable enough to find an exact article or case number. And a chunk tells you nothing about what the document cites or what later rulings did to it.

With the MCP server, the assistant does the research itself. It can search by meaning or by exact terms and citations using Boolean operators, follow references between documents, read full texts, and retrieve individual articles of a law. It can also check which jurisdictions and official portals are available. The model decides what it needs, reads what comes back, and searches again until the answer holds up. That loop is the difference. Luna moved onto the same MCP server in September, and answers got faster and better at the same time. The data is live as well: the assistant reads the Data Layer as it stands today.

Responsible use starts with clear limits. Our tools are read-only, and results are linked back to the official source. We also publish simple research instructions that assistants can follow. They explain how to search in the source’s original language, check the passage before citing it, verify the version and be clear when the available sources leave a gap.

Different assistants show their reasoning in different ways, and that is not something we control. What we can do is make the sources easy to check. Every citation leads to a document, and every document leads to its official publication.

Legal information is constantly evolving as new rulings are published, regulations change, and existing precedents are challenged. How does Moonlit’s automated monitoring infrastructure identify and incorporate these developments, and how does the platform prevent AI systems from relying on outdated legislation or superseded legal interpretations?

Preventing an AI or a researcher from relying on bad law comes down to the connections between documents. As soon as a fresh judgement or statutory amendment enters the system, the reference graph links it to the earlier provisions and cases it cites. If an appellate court overturns or narrows an earlier decision, that development attaches directly to the original ruling as a citing decision. A model following the graph sees that subsequent treatment immediately, rather than treating an outdated precedent as good law. On the platform, users can turn that same incoming stream into active Monitors for the specific statutes, courts, or topics they track.

For commentary, maintaining currency has traditionally been a major blind spot. Commercial commentary series are often reviewed for updates only once or twice a year. In fast-moving areas like indirect tax or data protection, that delay creates serious compliance risks. Because our Data Layer evaluates incoming publications against our commentaries daily, new rulings get flagged for the authoring editor immediately. The editor updates the analysis and advances the reference date, so that both practitioners and AI models are working from the actual state of the law rather than a twelve-month-old perspective.

Cross-border legal research introduces complexities that extend beyond translating documents, particularly when different countries interpret or implement similar regulations differently. How does Moonlit’s AI distinguish between these jurisdictional differences, and what happens when authoritative legal sources appear to contradict one another?

An EU directive and the national law that implements it are different documents, linked where the one cites the other, and a national ruling that applies a judgment of the Court of Justice is linked to that judgment. That is also why Commentaries starts with VAT. The rules are largely harmonised at EU level and then interpreted in each member state.

Contradictions are normal in law, and we do not hide them. When two authoritative sources appear to disagree, the reader gets both. The hierarchy then does most of the work: a later ruling over an earlier one, the Court of Justice over a national court on a question of EU law and the provision as it stood on the reference date. In Commentaries, the editor makes that call in the open, and the sources sit next to it.

Moonlit’s technology is used by organizations including the Supreme Court of the Netherlands, Deloitte, and other legal and public institutions. Can you share a practical example of how the platform has transformed a legal research workflow, and how you measure improvements in research accuracy, completeness, or efficiency compared with traditional methods?

In fintech, for example, the legal team at Quantoz Payments connected the Moonlit MCP server to Claude, pairing it with custom workflows for cross-border payments regulation. The workflow this replaces is the traditional standard across Europe. A lawyer typically has to open portals in several countries, run searches in multiple languages, copy citations into a memo, and hope no crucial local precedent was overlooked. Now, their setup queries across jurisdictions simultaneously, follows the citation graph, and returns primary documents alongside their official source links. For an in-house team, the impact is clear. It eliminates the blind spot of missed foreign rulings and shifts hours away from mechanical gathering toward actual legal analysis and risk evaluation.

You see a very different kind of efficiency on the development side. Teams building legal AI products measure improvements in engineering months saved. Eudia, a legal AI company, needed structured European legal data to power their product, and their product manager noted that our Data API gave them a complete European corpus in days. Without an underlying data infrastructure, that means building and maintaining custom scrapers for every national portal indefinitely. Other platforms like Saga and Legal Mind rely on this same foundation, and they can now deliver verified Commentaries directly to their own end users as well.

Looking ahead, as foundation models become increasingly capable and AI agents take on more sophisticated legal research tasks, how do you see the relationship between general-purpose AI models and specialized legal data infrastructure evolving? What is your long-term vision for Moonlit, particularly when it comes to making trustworthy legal intelligence accessible beyond large law firms and institutions?

A stronger model reasons better over the sources you feed it, but it cannot conjure reliable primary records out of nowhere. As foundation models become more capable, the value of a verified data layer actually goes up. When reasoning becomes widespread, the real bottleneck shifts to data quality, because an agent still needs accurate, structured law to work with.

General-purpose models and autonomous agents will handle the drafting, synthesis, and conversational interaction. Specialised infrastructure like Moonlit handles the hard data problem underneath, keeping the law current, structured, and verifiable, and serving it directly through whatever protocols those agents use. We prepared for this shift early. Our own assistant runs on the exact same MCP server that external agents connect to, and teams are already running their own custom agents on top of it.

That architecture is also central to our long-term vision. Reliable cross-border legal intelligence has traditionally been a luxury good, affordable only to the largest law firms and multinational corporations that can pay for expensive multi-country subscriptions. By making Commentaries free to read and opening our data layer to any developer through modern APIs and connectors, we want to level that playing field. A solo practitioner advising a cross-border business or an in-house counsel at an early-stage company should have access to the exact same research precision and verified citations as a global advisory firm.

Thank you for the great interview, readers who wish to explore verified legal data and AI-assisted cross-border research should visit Moonlit.