Thomson Reuters Builds In-House AI Model to Cut Costs

Thomson Reuters has deployed its first proprietary large language model, a strategic move designed to reduce the company's dependence on external AI providers while controlling spiraling technology costs.

The model, called Thomson, represents a $40 million investment over two years in personnel and computing infrastructure, according to details first reported by Business Insider and SiliconAngle. Rather than building a foundation model from scratch, Thomson Reuters adapted Qwen, an open-source large language model from Chinese technology company Alibaba, customizing it for legal and professional applications.

A hybrid approach to AI deployment

Thomson Reuters worked with a joint team from Imperial College London to modify Qwen over several months, focusing on removing biases and ensuring safe outputs for professional use. The company's Chief Technology Officer Joel Hron said domain specialists shaped the training process, providing sample legal questions and evaluating the model's responses.

The training drew on proprietary content from Thomson Reuters properties including Westlaw, Practical Law, Checkpoint, and Reuters. Despite this extensive library, less than 10% of the company's total content has been incorporated into training so far. The final training run alone cost approximately $450,000.

Thomson made its debut in Tabular Analysis, a document review feature within the CoCounsel Legal product. The company emphasized that CoCounsel will not become a single-model system. Instead, Thomson will handle tasks where its specialized training provides advantages, while external models continue to power other capabilities.

Why it matters

Thomson Reuters' investment reflects a broader shift in enterprise AI strategy as companies confront the high costs of licensing models from providers like OpenAI and Anthropic. Hron framed the decision using a real-estate analogy: licensing external AI is like renting — you get what you need but never build equity. Owning the model allows Thomson Reuters to compound value from its intellectual property over time. This approach may become a template for other content-rich enterprises seeking to balance AI capabilities with cost control, particularly as businesses increasingly demand clearer returns on AI investments and explore open-source alternatives.

Partnership strategy remains intact

Despite launching Thomson, the company's expanded partnership with Anthropic, announced in May, remains active. Hron indicated Thomson will gradually power more CoCounsel features, but the company plans to maintain a multi-model strategy.

Thomson Reuters is also releasing a smaller version of the model as an open-weight offering on Hugging Face for academic and non-commercial applications. Legal academics have begun evaluating the model ahead of a wider release.

The development comes as enterprise customers broadly pull back from high AI spending, switching to cheaper alternatives and pressing vendors for measurable business outcomes. Several organizations have already adopted open-source Chinese models as cost-effective options.

Business Insider and SiliconAngle first reported the details of Thomson Reuters' AI model launch and investment figures.