Thomson Reuters: Legal AI Buy-Build Strategy Unveils Proprietary LLM

• Thomson Reuters is pursuing a 'buy-build' strategy for legal AI, developing its proprietary large language model named "Thomson" while also acquiring specialized startups like Safe Sign Technologies.
• The company aims to achieve "fiduciary-grade AI" through domain-specific training and robust hallucination mitigation to ensure the reliability and trustworthiness of its legal AI tools.
• This strategic shift is intended to provide more precise AI solutions for legal work, moving beyond reliance on general-purpose third-party providers.
• Thomson Reuters is actively addressing the 'buy versus build' dilemma for law firms and focusing on the rollout of its AI products within the Canadian legal market.
• Leadership emphasizes the importance of building trust among lawyers by demonstrating the accuracy and dependability of AI-generated legal content.

By building its own Thomson Reuters proprietary LLM, the company seeks to tailor AI models more effectively to the nuances of legal language and practice, moving beyond general-purpose AI solutions.

Thomson Reuters is actively redefining its approach to artificial intelligence in the legal sector, moving towards a robust buy-and-build strategy to enhance its offerings. This strategic pivot involves developing proprietary AI capabilities rather than exclusively depending on external providers. A key outcome of this decision is the creation of "Thomson," the company's inaugural proprietary large language model (LLM), designed specifically for legal applications. This move underscores a commitment to deeper integration and control over the AI tools serving its clientele.

The rationale behind this significant investment was recently discussed by David Wong, Thomson Reuters' chief product officer, alongside Alexander Kardos-Nyheim, who co-leads the company's foundational machine learning research. Kardos-Nyheim's expertise joined Thomson Reuters through its 2024 acquisition of Safe Sign Technologies, his legal AI startup. This acquisition highlights the dual nature of the Thomson Reuters legal AI buy-build strategy, combining internal development with strategic external integration to accelerate innovation and secure specialized talent.

This deliberate shift aims to leverage domain-specific training, which is critical for delivering precise and reliable results in legal work. By building its own Thomson Reuters proprietary LLM, the company seeks to tailor AI models more effectively to the nuances of legal language and practice, moving beyond general-purpose AI solutions. This strategy is particularly relevant for law firms grappling with their own 'buy versus build' decisions regarding AI integration, offering a precedent from a major industry player.

A central focus of Thomson Reuters' AI development is the concept of "fiduciary-grade AI legal," which emphasizes the highest standards of accuracy, reliability, and trustworthiness for legal professionals. This commitment directly addresses a primary concern within the legal community: the potential for AI "hallucinations" or the generation of incorrect or misleading information. The company is actively working on sophisticated methods for legal AI hallucination mitigation, recognizing that the integrity of legal advice and documentation depends on the absolute veracity of AI-generated content.

Achieving fiduciary-grade AI involves rigorous domain-specific training, where the LLM is extensively fed with vast amounts of high-quality, curated legal data. This specialized training helps the AI understand the intricate context and precise terminology unique to the legal field, significantly reducing the likelihood of errors. The goal is to produce AI tools that legal professionals can rely on with the same confidence they place in traditional, human-vetted resources.

Discussions with Thomson Reuters' leadership have underscored the importance of this reliability, particularly as AI becomes more embedded in critical legal workflows. The company's efforts to define and achieve fiduciary-grade AI are crucial for building trust among lawyers and compliance officers, who require assurance that AI outputs are not only efficient but also legally sound and defensible.

The strategic decisions made by major legal technology providers like Thomson Reuters have significant implications for the broader legal market, particularly concerning the adoption of AI tools. The "buy versus build" dilemma, which Thomson Reuters has navigated by opting for a hybrid approach, mirrors similar considerations facing law firms, especially those in Canada. These firms must weigh the benefits of developing in-house AI solutions against integrating commercially available platforms.

Thomson Reuters is also keenly focused on the Canadian market, discussing Canada's specific role in its product rollout plans. This attention to regional needs is vital for ensuring that AI tools are relevant and accessible to Canadian legal professionals. The company's involvement as a supporting partner for the Canadian Legal Summit, scheduled for October 14 in Toronto, further highlights its engagement with the Canadian legal community and its efforts to foster Canadian law firm AI adoption.

Addressing skepticism among lawyers regarding AI is another key aspect of Thomson Reuters' strategy. By demonstrating a clear commitment to fiduciary-grade AI and robust hallucination mitigation, the company aims to build confidence in its AI offerings. The reliability and trustworthiness of AI tools directly impact their integration into legal practice, making these strategic decisions critical for lawyers and compliance officers evaluating new technologies.

Lawyers and compliance officers should pay attention to how major legal tech providers like Thomson Reuters are addressing the 'buy vs. build' dilemma and developing 'fiduciary-grade AI' to mitigate hallucinations, as these strategic decisions directly impact the reliability and trustworthiness of the AI tools they may integrate into their practice.

Source: Original reporting via CL Talk