Thomson Reuters builds its own AI model, trained on proprietary legal content
Thomson Reuters launched its own large language model on August 24, trained exclusively on decades of Westlaw, Practical Law, Checkpoint, and Reuters content and validated by hundreds of in-house subject matter experts. The model, called Thomson, scored 0.83 on factuality in internal evaluations - a measure of whether claims can be traced to supporting sources - while leading frontier models scored between 0.65 and 0.68 when given open web access. It is now live inside Tabular Analysis in CoCounsel Legal.
General counsel face a barrage of pitches for AI tools built on the same general-purpose models, differentiated mainly by interface. Thomson Reuters took a different path. "It's not a layer on top of someone else's model," the company said. "It's a model we built, on content only we have, to a standard only we can set."
What "fiduciary-grade" means in practice
General-purpose AI models are optimized for breadth - writing poems, debugging code, and summarizing contracts with equal fluency. Legal work does not tolerate the same error rate. A cited case that no longer stands or a clause that misstates an obligation carries consequences that land on the lawyer, not the model.
The company frames this as Fiduciary-Grade AI: a standard for professionals with duties of care where "almost right" isn't good enough. The claim is not about raw capability but about whether an answer can be checked and whether it survives being checked.
Thomson starts from an open-source foundation similar to what competitors use. What follows is different. Partner-level practitioners built evaluation rubrics for the hardest research questions the company could construct. Lawyers fresh from practice grounded the training data in what matters to an actual matter, not what a textbook would flag. "That's the part a wrapper can't replicate," the company said. "Access to our published content helps. Access to the judgment behind it doesn't come bundled with a subscription."
Early results and independent testing
Thomson performed competitively with frontier models including Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro across legal and general benchmarks. On PrBench Legal Hard, it posted the top score of any model tested.
The metric Thomson Reuters emphasized most was factuality. In a deep research evaluation over Westlaw and Practical Law, Thomson scored 0.83 against 0.65 and 0.68 for leading models with open web access. Every model covered similar ground. Only one could reliably back up its claims.
Two outside academics reached similar conclusions. Professor Samuel Dahan of Queen's Conflict Analytics Lab and Cornell Legal AI Lab found Thomson's citation quality generally competitive with leading models, even on Canadian employment law questions without a Canada-specific setting. Professor Jonathan H. Choi of Washington University School of Law tested it against his Corporate Tax class questions and said he "preferred Thomson's responses overall," citing the references to underlying treatises.
Where the model runs first
Thomson's first deployment is inside Tabular Analysis in CoCounsel Legal, where structured document review at high volume makes a purpose-built model's advantage immediately visible. CoCounsel Legal remains multi-model by design. The company applies Thomson where it delivers the clearest advantage and uses other leading models elsewhere, expanding Thomson's footprint as it provides meaningful benefits to customers.
The company has used less than 10% of its proprietary content in training so far. "That's not a caveat. It's the roadmap."
Data control and the risk committee question
Building its own model gives Thomson Reuters full control over training data, deployment, and behavior. The company does not train Thomson on customer data and said it never will without explicit consent. For a profession built on trust and accountability, that commitment is foundational.
Thomson Reuters has spent 175 years being the source lawyers check their work against. With Thomson, it built a model designed to meet the same standard.
Why this matters for legal professionals
Most legal AI tools inherit the factuality ceiling of the general-purpose model underneath them. Thomson's 0.83 factuality score against competitors' 0.65-0.68 represents a measurable gap in whether a claim can be traced to a source that supports it. For lawyers whose work product gets scrutinized by opposing counsel, judges, and clients, that traceability is not a feature - it is the minimum requirement. The model's first availability inside Tabular Analysis also signals where the immediate payoff sits: high-volume structured document review, not open-ended chat.
Your membership also unlocks:
