Thomson Reuters Builds Its Own AI Model for Professional Work
Key Vocabulary
| Word / Phrase | Meaning | Example |
|---|---|---|
| proprietary | owned and controlled by a particular company rather than openly shared | Thomson is a proprietary model owned by Thomson Reuters. |
| large language model | an AI model trained to work with and generate language | The company developed its own large language model. |
| subject-matter expert | a person with deep knowledge of a particular professional field | A subject-matter expert helped evaluate the legal tasks. |
| AI sovereignty | control over important choices about an AI model, its data and where it operates | Some organizations see AI sovereignty as important for sensitive work. |
| external evaluation | testing carried out by people outside the organization that built something | External evaluation can add evidence beyond a company's own tests. |
Article
Thomson Reuters has spent years putting outside AI models inside its professional products. On 24 August 2026, the Toronto-based company announced a different move: Thomson, its first proprietary large language model developed in-house. [1]
Rather than training a model entirely from the beginning, the company started from an open-source foundation and then specialized it for legal, tax and other professional tasks. Thomson Reuters says the project required about US$40 million in talent and computing resources. [1][3]
The specialization draws on the company's proprietary content, including material connected with Westlaw, Practical Law, Checkpoint and Reuters. Subject-matter experts were also involved in designing training goals and evaluating outputs. According to the company, less than 10% of its own content has been used for training so far. [1]
Owning the model gives Thomson Reuters greater control over how it is trained, where it runs and how it connects with company tools. The announcement places this idea under the broader label of AI sovereignty: organizations increasingly want to know who controls the model behind sensitive professional work. [1][2]
The first deployment is planned inside Tabular Analysis in CoCounsel Legal, a document-review feature that organizes information across many files. CoCounsel will remain a multi-model product, using Thomson where the company believes it has an advantage and other leading models for other tasks. [1][4]
Privacy is part of the pitch. Thomson Reuters states that customer data will not be used to train the model without explicit consent. It is also opening the model to selected legal and AI academics for external evaluation and releasing a smaller open-weight version for academic and non-commercial testing. [1]
The performance claims still deserve independent scrutiny because many of the launch comparisons come from Thomson Reuters' own early evaluations. The more interesting shift may be strategic: a company known for professional information now wants to control not only the content and software, but also some of the intelligence layer underneath them.
Discussion Questions
- Why might a professional-information company want a proprietary large language model?
- When does AI sovereignty matter most for a business or government?
- What should a subject-matter expert test that a general AI benchmark might miss?
- How much weight should customers give to external evaluation compared with a company's own results?
- What are the advantages and disadvantages of keeping a multi-model system instead of using one model for every task?
References
- Thomson Reuters, "Thomson Reuters Leverages its World-Class Data Assets to Launch Its Own Frontier Model."
- Thomson Reuters, "Thomson: a purpose-built foundation model for professionals."
- Thomson Reuters, "How we built Thomson."
- Thomson Reuters, "Thomson Reuters launches next generation of CoCounsel Legal."