Thomson Reuters invested $40 million to develop its own AI model, Thomson, using decades of proprietary legal, tax, and compliance content. Instead of building the model from scratch, the company began with an open-source foundation and refined it with its own data, including Westlaw, Practical Law, and Checkpoint.
Initial tests show Thomson performing on par with models from OpenAI, Anthropic, and Google in several professional evaluations. However, the company continues to use third-party models for some products. CoCounsel Legal, for instance, operates on Anthropic’s Claude Agent SDK, the same technology behind tools like Spline’s 3D editor.
Most of the $40 million went toward training the model on proprietary content rather than creating a new foundation. This method provides an option for companies with large domain-specific datasets that prefer not to rely entirely on external models or spend billions developing one independently.
Experts evaluated Thomson’s outputs, identifying errors and improving its performance. So far, less than 10% of the available content has been used for training, allowing for future expansion.
Related: Anthropic playground outperforms OpenAI in early testing
Legal work requires precision. A single incorrect citation or confidently wrong answer can lead to serious consequences. Thomson was designed to handle uncertainty differently than general-purpose models, prioritizing accuracy over user satisfaction.
The model flags uncertainty instead of forcing a confident response, even if it disappoints the user. This approach reduces hallucinations and sycophancy, common issues in high-stakes professional settings. No AI model is perfect, and Thomson Reuters acknowledges this limitation.
Thomson currently powers specific features in Thomson Reuters products, such as Tabular Analysis in CoCounsel Legal. This feature processes up to 10,000 documents and answers 100 questions about them. A smaller, open-weight version is available to researchers on Hugging Face, though direct customer access isn’t yet offered. The company is considering future commercialization.
Thomson Reuters combines the model’s domain-specific training with retrieval from authoritative sources like Westlaw and Practical Law. The aim is to ground responses in verifiable material while maintaining the model’s reasoning ability. It’s a balanced approach, not an either-or choice.
Related: Myriad360 acquires F3 Technology Partners crosses $1 billion
That balance is important for lawyers. If an AI-generated brief cites a case, attorneys need to verify whether the model’s interpretation aligns with the actual ruling. Many of Thomson’s outputs can be traced to specific sources, though not every line links to an exact statute or precedent. The company describes Thomson as a tool that makes its reasoning transparent where possible, leaving final verification to professionals.
The company hasn’t abandoned third-party models. It uses them where they work best, like in CoCounsel Legal, while developing in-house solutions only where its proprietary data provides an advantage. This selective strategy reflects a broader industry debate over which parts of the AI stack are worth owning versus renting.
For Thomson Reuters, the answer lies in its data and workflows. The company already serves millions of professionals, giving it an edge in fields where precision is critical. Owning the model allows it to shape tradeoffs, such as balancing helpfulness against accuracy, in ways general-purpose models cannot. That control comes at a cost, and the company must demonstrate the investment’s value in a market where even the best models still make mistakes.
The August 20 update to CoCounsel Legal introduced Deep Research Verify. This feature checks whether Westlaw and Practical Law sources support the AI’s claims. It’s part of a shift toward agentic systems that handle research, analysis, drafting, and verification while keeping humans involved.
