Look at the pace of growth and execution of both Harvey and Legora (we won’t take sides) and how quickly certain previous-generation SaaS companies like our dear portfolio company Clio embrace AI, and one might think “Legal AI is done for a seed investor”.
That’s what we thought before meeting a company called grubel, whose name comes from the German verb grübeln: “think harder”.
Some quick facts about where the legal AI market stands
Technology spend in the legal industry has never grown so quickly. As the graph below from a recent Thomson Reuters report shows, technology spend’s growth in the legal industry is at an all-time high. Law firms are rushing to embrace the latest AI technology, fearing displacement by competitors that might adopt it faster.
This report, based on interviews and a survey of ~100 law firm decision-makers, also shows that firms are still actively benchmarking Harvey and Legora against generic LLMs and custom-built solutions, i.e., switching costs remain low — nobody has locked up the market yet.
Now, despite all this growth and intense competition, the legal services market remains mostly human-driven. The legal services market is valued at over $1Tn, and legal software represents only tens of billions in revenues (legal AI most likely only 100s of $M at this point). If AI is to do most of the legal work and the market size remains roughly the same (it might actually grow), the legal software market can still grow at least 30x and the legal AI market well over 1000x.
Looking at it today, according to a benchmark based on 1250+ tasks developed by Harvey in partnership with grubel’s founders (and a few other great companies), AI can complete only 10-20% of legal tasks autonomously. More on the methodology of the benchmark here. For the lawyers in the audience, see the task split in the benchmark below and the latest model performance results published by Harvey afterward.
Harvey started fine-tuning open-weight models with great results (results above and more here), but part of why AI’s performance is still limited is that most legal AI tech companies are essentially deploying frontier models and leveraging RAGs, yet still have very limited in-house AI research capabilities. Max at Legora explains well in this podcast that one of their key success factors is the velocity with which their forward legal engineers build (and scrap) agent harnesses around the latest frontier models. As frontier models improve and incorporate more legal knowledge during pre-training, they usually scrap all the agent harnesses and move on to more complex workflows. It also explains why the most advanced legal AI companies are now competing on building “long-running agents” to solve increasingly complex cases.
In short, the legal AI market is growing very fast, but it’s still very small compared to legal services, and there’s been relatively little technical innovation beyond what frontier models provide off-the-shelf. Apart from Harvey’s fine-tuning experiments, it’s mainly been about distribution efficiency rather than technical innovation. So far.
Meet grubel, every case its own AI
grubel is a frontier lab out of Tübingen and Munich bringing data and algorithmic advancements to the legal AI market.
It allows law firms to have a model trained on a per-case basis, distributed either through partners like Harvey or directly to law firms.
In order to do that, grubel automates an AI specialization loop for each matter that combines:
A data engine that discovers and curates task-relevant data for each matter or client. The same way as a lawyer would spend time understanding the case-specificities, grubel curates case-specific datasets to “teach the AI” how to think about a specific case.
A test-time adaptation layer that adapts the model to the matter before it acts. This is an algorithmic advancement, co-invented by Moritz, the founder of grubel, that adjusts the model weights at inference time, so the model’s case-specific knowledge isn’t limited to a narrow context window. This matters for long, complex cases that are numerous and often the most lucrative.
A rigorous evaluation framework for the capabilities that legal work actually requires.
grubel was started by Moritz Hardt, a director at the Max Planck Institute for Intelligent Systems, who co-invented test-time training and worked at early Google Brain, overlapping with Dario Amodei and Ilya Sutskever, and Reinhard Heckel, a TUM professor of machine learning, who’s been focusing on data-centric machine learning.
(Here’s a badly framed picture of Moritz and Reinhard I took too quickly with my phone at our latest Founder Summit)
We are excited to lead grubel’s pre-seed round, joined by a great line-up of angels including 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), Harvey co-founders Gabe Pereyra and Winston Weinberg, and our friends Alex Diehl (Director, Advanced Semiconductors at SRIND) and Sauraj Gambhir (co-founder of Prior Labs).
They’ve just hired two founding engineers with PhDs in computer science and experience in foundation model pre-training at Mistral and leading legal AI research at Thomson Reuters.
They’re also hiring two more researchers and a COO for the company. As I’ve said in numerous candidate interviews recently, in our view, grubel is the most exciting, still early-stage, legal AI tech company in Europe. If you’re interested in pushing the boundaries of what AI can do in the legal market, join them
!