Thomson Reuters unveils in-house AI model for lawyers

Thomson Reuters has unveiled Thomson, a proprietary artificial intelligence model for professional work that ranked among the strongest systems in the company's early benchmark testing.

Thomson Reuters said Thomson outperformed GPT-5.5, Claude Sonnet 5 and Gemini 3.1 Pro across a range of legal and general benchmarks, while posting results it described as competitive with Claude Opus 4.8.

The launch marks another step in the group's effort to develop its own AI systems rather than rely solely on third-party general-purpose models. Thomson will first be introduced in Tabular Analysis within CoCounsel Legal, where it will become the default model for that feature, before broader use across the company's legal and tax products.

The company said the work stemmed from its 2024 acquisition of Safe Sign Technologies, which added AI research expertise to the business. The move came as many companies were leaning on increasingly powerful general-purpose models for professional applications.

Thomson Reuters instead took the view that legal, tax and other high-stakes professional work required models built for specific domains and standards. It said Thomson was built on an open-source base model and then further trained using content from Westlaw, Practical Law, Checkpoint and Reuters.

Hundreds of subject matter experts were involved in the process, reviewing outputs, identifying weaknesses and checking whether the model's reasoning matched legal professional workflows. Thomson Reuters also said customer data is not used to train the model.

Benchmark tests

Thomson Reuters said it tested Thomson against leading general-purpose AI models in legal, tax, accounting, journalism, safety, reasoning, coding, maths, multilingual, agentic and long-context tasks, as well as instruction following.

It also said less than 10% of its proprietary content had been used in training so far, leaving room to further develop the model's domain knowledge through additional training and validation.

One internal test focused on 53 legal research queries written by subject matter experts. In that exercise, Thomson was connected through an in-house agentic harness to Westlaw and Practical Law, while rival frontier models were given unrestricted web access through the Brave search engine.

Completeness and factual accuracy were scored using an LLM-as-judge method calibrated against expert scoring. In that evaluation, Thomson Reuters said Thomson delivered stronger completeness and factuality, including support for claims with accurate citations to trusted sources.

Strategic shift

Thomson Reuters presented Thomson as a new layer in a broader AI strategy built around proprietary content, domain expertise and professional software tools. By adding an in-house model, the company is seeking to tie its content assets more closely to the AI systems embedded in its products.

That approach could help differentiate its legal and tax offerings in a market where software groups are racing to combine generative AI with specialist databases and workflow tools. The emphasis on internally developed models also reflects a wider push by established information companies to retain more control over cost, performance and product design.

Joel Hron, Chief Technology Officer at Thomson Reuters, said the company believed specialist models could now compete with the largest general-purpose systems.

"The most capable AI models no longer come only from frontier AI labs. One now comes from Thomson Reuters," Hron said.

He also pointed to the economics of building a smaller model tuned for professional use.

"Thomson is competitive with the world's leading frontier models despite being a fraction of their size and cost to train and operate. We've achieved that performance by combining exceptional AI talent with authoritative proprietary content and deep domain expertise," Hron said.

Jonathan Schwarz, Head of AI Research at Thomson Reuters, described the model as tailored to legal reasoning.

"The result is a model that thinks and reasons like a lawyer while outperforming models multiple times larger on the work that matters," Schwarz said.

Hron said the company's focus remained on professionals working in fields where accuracy has direct consequences.

"General-purpose AI is built for everyone. Thomson is built for the professionals who cannot afford to be wrong," Hron said.

"This is our commitment to Fiduciary-Grade AI in action: AI designed for professionals with duties of care and accountability, where almost right is not good enough," he added.