German AI Lab Grubel Raises $3.2M to Give Every Legal Case Its Own AI

Jan de Vries · 2026-09-10

### A Small Lab With a Big Idea About Legal AI Grubel, an AI research lab split between Munich and Tübingen, just raised €3 million — about $3.2 million — in pre-Seed funding. The money will go toward research, product development, and growing the team. That's a modest check by AI standards. But the idea behind it is anything but modest: Grubel wants to build AI systems that adapt to individual legal cases, not just general-purpose models that sort of, kind of, almost get the job done. ### Who's Backing This? The round was led by Point Nine, a venture capital firm with a track record in early-stage tech. The angel list is where things get interesting, though. It includes Jeff Dean (Google's former chief scientist), Chris Ré (Stanford professor and Together AI co-founder), Ion Stoica (UC Berkeley professor and Databricks co-founder), and Harvey co-founders Gabe Pereyra and Winston Weinberg. When people like that put their own money in, it's worth paying attention. They're not chasing hype — they're betting on a specific technical approach. ### The Problem Grubel Is Trying to Solve Here's the thing about legal work: it doesn't come with a manual. Complex legal matters don't have ready-made training data, clear task environments, or obvious tests of success. Every case is its own little universe with its own facts, documents, terminology, and client standards. As Reinhard Heckel, co-founder of Grubel, puts it: > "Complex legal matters do not come with ready-made training data, task environments, or clear tests of success. We are automating the AI specialisation loop that identifies the right information, curates it into training data and environments, adapts the system to the matter, and tests whether its work meets the required standard." That's the core insight. Legal AI has lagged behind other fields — not because lawyers don't want better tools, but because the work is so specialised and so hard to evaluate that generic AI just doesn't cut it. ### The Specialisation Loop, Explained Simply Grubel was founded in 2026 by machine-learning researchers Moritz Hardt and Heckel. Hardt is a director at the Max Planck Institute for Intelligent Systems, a former UC Berkeley professor, and a former Google Brain researcher. Heckel is a TUM professor of machine learning, currently on leave, and formerly a researcher at IBM Research. Their thesis is simple: every legal matter needs its own AI. To make that possible, they're automating the process of specialising an AI system for each case. It works in three parts: - A data engine that finds and curates task-relevant data for each matter or client - A test-time adaptation layer that adjusts both the model and the agent before it acts - A continual evaluation framework that checks the work against standards pulled from the matter itself The loop repeats until the system delivers what the legal task demands. It's bringing data curation, adaptation, and evaluation into one cycle that trains and improves on the matter itself. ### Why This Matters Beyond Legal Complex knowledge work beyond coding still leans heavily on highly skilled humans. AI plays only a limited role. Legal services are where that gap shows up most clearly — but they won't be the last. Louis Coppey, Partner at Point Nine, sees it this way: "Over the past two to three years, we've seen many legal AI companies emerge, but few are pushing the boundaries of AI research in the legal field as ambitiously as Reinhard, Moritz and the Grubel team. We believe their approach can unlock new capabilities in legal AI and, over time, other forms of complex knowledge work." ### What's Next An initial evaluation showed that Grubel's specialisation loop can beat general-purpose frontier systems on the legal tasks tested. The company plans to push that further with the new funding. If it works, the implications go well beyond law firms. Any field where expertise is deeply personal and hard to standardise could benefit. That's a much bigger prize than it first appears.