Thomson Reuters takes on legal AI startups with its own model

Thomson Reuters has launched its own legal-focused large language model, spending $40 million to train it on its vast library of legal content.

Called Thomson, the model will start powering parts of CoCounsel Legal, the company's flagship legal AI product, alongside models from the major general-purpose AI providers.

The launch is a big move in the increasingly competitive legal AI market, with Thomson Reuters betting that a model deeply trained on legal content can outperform the general-purpose models from the likes of OpenAI and Anthropic on the specialist work lawyers actually do.

And it has plenty of material to work with. Thomson has been trained using content from Westlaw and Practical Law, with hundreds of subject matter experts involved in its development. And it’s just getting started too - the company says it has used less than 10% of its content to train the model so far.

Building its own

The project traces its roots to Thomson Reuters' 2024 acquisition of Safe Sign Technologies, the AI research company founded by Alexander Kardos-Nyheim while he was a trainee at A&O Shearman. That team now leads foundational legal AI research inside Thomson Reuters.

Speaking to Non-Billable earlier this year, Kardos-Nyheim and Thomson Reuters CTO Joel Hron explained the thinking behind the project.

General-purpose AI models, Kardos-Nyheim argued, are a little like liberal arts graduates: they know a lot about a lot. Thomson Reuters is trying to build something closer to a law firm partner who has spent an entire career immersed in legal work.

The company says early testing puts Thomson on par with the latest frontier models across a range of tasks, while performing strongly when dealing with dense, specialist material.

“For years, the AI industry has treated scale as the answer: bigger models, more compute, more money,” Hron said.

“Thomson shows there is another path. Start with a strong foundation, specialise it deeply for the work that matters, and you can build intelligence that is highly capable, far more efficient and entirely under your control.”

Into CoCounsel

The first place lawyers will encounter Thomson is inside tabular review in CoCounsel Legal, a tool used to review large volumes of documents and pull information from them into structured tables.

Thomson Reuters says CoCounsel will remain “multi-model”, though, using its own model where it performs best and other models elsewhere. It also plans to roll Thomson out more widely across its legal and tax products.

The move is another sign of CoCounsel developing from an AI legal assistant into a broader platform for legal work, putting it into increasingly direct competition with the likes of Harvey and Legora.

Thomson Reuters is pitching its approach around what it calls “fiduciary-grade AI”: essentially, technology designed for professionals who are accountable for the advice and decisions they make, where an answer that is almost right may not be good enough.

Legal AI arms race

The timing is interesting with Thomson arriving just days after Harvey launched Tenet, its own proprietary legal model.

The two companies have taken different routes. Harvey started with an open-source model from Chinese AI company Moonshot and then trained it specifically for legal reasoning. The longer-term goal is for law firms to build on Tenet with their own knowhow and ways of working.

Thomson Reuters has a major data advantage with decades of proprietary legal content and editorial expertise that few competitors can replicate.

It also believes owning the model can change the economics. Thomson Reuters says Thomson was trained and can be run at a fraction of the cost of comparable frontier models, reducing its reliance on increasingly expensive third-party AI models.

That matters across legal AI. Companies such as Harvey, Legora and Thomson Reuters pay outside model providers every time customers use their models. Routing more work through proprietary models could reduce those costs and improve margins.

Meanwhile, law firms themselves are looking for ways to make AI more specific to their own businesses. Kirkland made headlines earlier this year when it partnered with data intelligence giant Palantir to develop AI software drawing on the firm's institutional knowhow and workflows.