Thomson Reuters Builds and Releases Its Own AI Model

Thomson Reuters, a major information services company known in law, tax, and news sectors, has developed a proprietary AI model based on open-weight models (those with publicly released weight parameters). An open-weight model refers to an AI model whose internal parameters are publicly available and can be freely improved and customized by companies. Through this approach, Thomson Reuters has reduced the cost of building AI from scratch while achieving a proprietary model tailored to its specific needs.

In recent years, AI model development methods have divided into two main categories. One is the "closed-source" approach, like OpenAI and Anthropic, which provide model access via APIs. The other is the "open-source" approach, like Meta's Llama series, which publicly releases model weights for external use. Thomson Reuters' choice of the latter exemplifies the growing adoption of open models in the enterprise sector.

As a confirmed fact, Thomson Reuters has built a model customized for its needs based on publicly available open-weight models. This approach typically combines fine-tuning—a technique for additional training on specific tasks—allowing the company to adapt to highly specialized data such as legal documents and tax information. Since Thomson Reuters' business domain requires exceptional accuracy in specialized knowledge, maintaining a proprietary model offers significant advantages over using generic AI services off-the-shelf.

One reason Thomson Reuters' move attracts industry attention is its potential to serve as a precedent for other SaaS companies. Historically, SaaS firms have primarily integrated generic AI APIs into their own services, but the feasibility of building proprietary models based on open-weight models is increasing. When simultaneously considering cost, data privacy, and customization capability, retaining a proprietary model demonstrates a degree of rationality.

On the other hand, developing and maintaining a proprietary model requires substantial engineering resources and ongoing investment. While using an open-weight model as a starting point reduces initial costs, the company must bear operational responsibilities such as security management, model updates, and quality assurance. In this respect, large-scale information services companies like Thomson Reuters and smaller SaaS firms face different levels of practical difficulty when adopting the same approach.

Thomson Reuters' case can be positioned as a signal that the transition from AI "users" to AI "owners" is no longer limited to certain large enterprises. Going forward, key points to observe include how the model Thomson Reuters has built is integrated into actual services and what changes it brings to business accuracy and efficiency. Additionally, whether other SaaS companies follow a similar approach will be an important metric for assessing the direction of the enterprise AI market.