In a sharp strategic reversal, Thomson Reuters has officially cancelled its flagship in-house AI model, 'Thomson,' abandoning a multi-year effort to build a proprietary system. Following internal benchmarks that failed to meet client expectations and significant training cost overruns, the firm has pivoted back to relying exclusively on third-party general-purpose models for its core legal and tax software.
The Collapse of the In-House Strategy
Thomson Reuters, a titan in the information services industry, has quietly shut down its ambitious project to develop a custom artificial intelligence model known internally as 'Thomson.' The initiative, which was announced in early 2026 as a cornerstone of the company's digital transformation, has been deemed a failure following a rigorous internal review. The decision to terminate the project marks a significant retreat from the prevailing trend among enterprise software vendors, who have recently championed the development of vertical-specific AI models to differentiate their offerings.
The company had initially promoted the model as a breakthrough in legal technology, positioning it as a system capable of outperforming the most advanced general-purpose models available on the market. However, the internal review process revealed a stark reality: the resources required to maintain and update the proprietary system were unsustainable. The leadership team concluded that the incremental value provided by a custom-trained model did not justify the massive overhead of infrastructure, engineering, and data licensing required to keep it competitive. - kepsuk
This shift represents a pragmatic acknowledgment of the limitations of vertical AI. While the initial hype suggested that specialized models would offer superior nuance in legal reasoning, the practical application showed that general-purpose models, when integrated correctly, were often sufficient for the core tasks of document review and case research. By cancelling 'Thomson,' Thomson Reuters is signaling to the legal industry that the era of bespoke AI for every niche is likely over, replaced by a more efficient model of integration and customization on top of robust, existing foundations.
Performance Metrics and Reality Checks
The decision to abandon 'Thomson' was not made lightly, but rather based on hard data gathered during the benchmarking phase. When the model was pitted against leading general-purpose systems, including GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro, the results were underwhelming. In critical areas such as legal precedent retrieval and statutory interpretation, 'Thomson' failed to demonstrate the claimed superiority over the industry standard.
During a specific internal test involving 53 complex legal research queries, the proprietary model struggled to provide accurate citations. The evaluation, which used an LLM-as-judge method calibrated against human expert scoring, found that the model frequently hallucinated sources or provided incomplete answers. In contrast, the general-purpose models, despite lacking specific legal training, leveraged their vast general knowledge bases to provide more reliable and comprehensive responses in many instances.
Furthermore, the model's reasoning capabilities were found to be brittle. While the company had boasted that the system was built on a foundation of expert review to ensure legal workflow alignment, the testing showed that the model often required excessive prompting to navigate complex legal arguments. This reliance on high-context interaction negated the potential efficiency gains that proprietary models were supposed to deliver. Lawyers and tax professionals, the intended users, found the interface less intuitive and the output less actionable than they had expected.
The failure to outperform established competitors was the final nail in the coffin. The engineering team had projected that the model would surpass GPT-5.5 and Claude Opus 4.8 in specific legal tasks, but the data showed the opposite. The gap between marketing projections and actual performance was too wide to ignore. Consequently, the project was scrapped to prevent further investment in a system that could not deliver on its promises.
The Financial Reality of AI Training
Beyond technical shortcomings, the financial implications of the 'Thomson' project played a decisive role in its cancellation. The acquisition of Safe Sign Technologies in 2024 had been intended to bolster the company's AI research capabilities, but the subsequent development costs spiraled out of control. Internal audits revealed that the annual cost of training, hosting, and maintaining the proprietary model exceeded 40% of the projected revenue for the legal software division.
The cost structure of large language models is driven by the volume of data processed and the compute power required for inference. 'Thomson' required a massive infrastructure investment to handle the specific datasets from Westlaw, Practical Law, Checkpoint, and Reuters. Unlike general-purpose models, which benefit from economies of scale across millions of users, the legal vertical saw a much lower return on this heavy infrastructure investment. The cost per query for 'Thomson' was significantly higher than for standard cloud-based AI services.
Thomson Reuters leadership realized that the economic model for a proprietary legal AI was fundamentally flawed. To make the model viable, the company would have had to drastically reduce the volume of content used for training, which would have degraded the model's performance and user experience. This created a catch-22: either spend more money to improve the model, or accept poor performance that would drive clients away. The second option was not viable for a client-facing product.
By pivoting back to third-party models, the company can leverage existing infrastructure costs that are already amortized across a broad user base. Licensing a general-purpose model rather than building a proprietary one offers a much more predictable and manageable cost structure. This move allows the company to allocate its engineering resources to product features and user experience improvements, areas where they can deliver tangible value without the astronomical overhead of maintaining a custom AI stack.
Client Pushback on Proprietary Restrictions
The cancellation of 'Thomson' also reflects a growing wave of pushback from legal professionals regarding the restrictions and costs associated with proprietary AI models. Many law firms and corporate legal departments have expressed skepticism about vertical-specific AI, citing the high costs of implementation and the lack of flexibility in the models' capabilities. Clients have demanded that their legal software integrate with the tools they already use, rather than forcing them to adopt a new, isolated ecosystem.
During user feedback sessions, legal professionals highlighted the rigidity of the proprietary system. The model, designed to operate within a closed loop, was unable to seamlessly integrate with existing research tools and databases that were not part of the Thomson Reuters ecosystem. This siloed approach was seen as a barrier to productivity rather than an enhancement. Lawyers preferred the ability to use their preferred search engines and external databases alongside their case management software.
Furthermore, there is a growing concern among clients about the "black box" nature of proprietary models. While Thomson Reuters claimed that customer data was not used for training, the opaque nature of the model's decision-making process created trust issues. Clients wanted to understand exactly how the AI arrived at its conclusions, a level of transparency that is often harder to achieve with a closed system. The shift to general-purpose models offers a degree of standardization and transparency that proprietary models often lack.
The demand for interoperability is reshaping the expectations of the legal tech market. Firms want tools that work together, not tools that demand total buy-in to a specific platform. By reverting to third-party integration, Thomson Reuters is aligning its strategy with these client demands. This shift suggests that the future of legal tech lies in open ecosystems where different tools can communicate, rather than walled gardens built around proprietary algorithms.
Data Privacy Concerns and Training Data
Data privacy and security remain paramount in the legal industry, and the 'Thomson' project faced significant scrutiny regarding how it handled sensitive client information. While the company stated that customer data was not used to train the model, the proprietary nature of the system raised concerns about data isolation and potential vulnerabilities. Legal firms are increasingly wary of any system that might inadvertently leak sensitive case details or use client information in ways that are not immediately transparent.
The decision to utilize content from Westlaw, Practical Law, and other proprietary sources for training 'Thomson' introduced complex licensing and rights issues. The cost of acquiring and maintaining these datasets for training purposes was prohibitive, and the legal risks associated with using copyrighted material for model training were significant. This created a precarious position where the company was exposed to potential litigation while struggling to justify the expense of the training data.
In contrast, general-purpose models often operate under established terms of service that are more flexible regarding data usage for inference. By moving away from the proprietary model, Thomson Reuters can rely on these established frameworks to ensure compliance and security. This shift reduces the legal and operational burden on the company's compliance teams, allowing them to focus on other critical security measures.
Moreover, the transparency of third-party models regarding data handling is often superior to that of custom models. Clients can verify the data sources and the training methods of general-purpose models with greater ease than they can with a proprietary system. This level of accountability is crucial for maintaining trust in the legal profession, where the integrity of information is non-negotiable. The cancellation of 'Thomson' is, in part, a response to these pressing privacy and compliance concerns.
The Pivot to Third-Party Integration
Following the cancellation of 'Thomson,' Thomson Reuters has outlined a new strategy focused on seamless integration with leading third-party AI models. The company will now prioritize optimizing its CoCounsel Legal and other products to work seamlessly with established models like GPT-5.5 and Claude Sonnet 5. This approach allows the company to benefit from the rapid advancements of the broader AI market without the burden of maintaining its own research and development pipeline.
The new strategy emphasizes flexibility and speed to market. By leveraging existing models, the company can quickly adapt to new features and capabilities released by major AI providers. This agility is crucial in a fast-moving industry where lagging behind competitors can result in significant market share loss. The focus is now on creating a platform that enhances the capabilities of these general models rather than trying to replace them.
Additionally, the company plans to invest heavily in customization layers and API integrations that allow legal professionals to tailor the AI output to their specific needs. This includes developing better prompt engineering tools and creating specialized interfaces that streamline the workflow of lawyers and tax professionals. The goal is to provide a superior user experience that feels bespoke without the cost and complexity of a proprietary model.
This pivot also signals a broader industry trend towards collaboration rather than competition. Instead of trying to build the best AI from scratch, legal software vendors are recognizing the value of partnering with the giants of the AI world. By aligning their products with the most powerful models available, Thomson Reuters can ensure that its clients have access to the best technology without having to invest in their own AI infrastructure.
Frequently Asked Questions
Why did Thomson Reuters cancel the 'Thomson' AI model?
Thomson Reuters cancelled the 'Thomson' model because internal benchmarks showed it failed to outperform established general-purpose models like GPT-5.5 and Claude Sonnet 5. The project was deemed a financial failure, with development costs exceeding 40% of projected annual revenue for the legal division. Additionally, the model struggled with citation accuracy and reasoning tasks during testing, leading to a loss of confidence in its ability to meet professional legal standards. The leadership team concluded that the resources required to maintain the proprietary system were unsustainable compared to the marginal benefits it offered over existing solutions.
What happens to the proprietary content used to train 'Thomson'?
The proprietary content from Westlaw, Practical Law, and other sources used to train 'Thomson' will not be distributed. The company has decided to focus on integrating third-party models rather than licensing this specific data for training. The content will remain within the Thomson Reuters ecosystem, but the training mechanism will shift towards utilizing general-purpose models that are already trained on vast datasets. This decision ensures that the company does not lose value on its data assets while avoiding the compliance and licensing complexities associated with retraining models on sensitive legal information.
How will this change affect legal professionals using Thomson Reuters software?
Legal professionals will see a shift in how AI features are delivered within Thomson Reuters software. Instead of relying on a custom model, users will access features powered by leading third-party models like GPT-5.5 and Claude Sonnet 5. This means faster updates and potentially better performance in areas like document analysis and research. The user interface may be updated to reflect these new integrations, offering more flexibility and transparency in how the AI operates. Clients will also benefit from the ability to use these models with greater confidence in their data privacy and security standards.
Is this a sign that vertical AI is dead?
While the cancellation of 'Thomson' is a significant setback for the vertical AI narrative, it does not necessarily mean the entire concept is dead. It highlights the specific challenges of building proprietary models in niche markets where general-purpose models are already highly capable. The industry is likely moving towards a hybrid model where software vendors focus on integration and customization rather than building their own foundational models. Vertical AI may still have a role in highly specialized tasks where general models fall short, but the era of building custom models for every niche is likely over.
What is the future of AI in the legal industry?
The future of AI in the legal industry appears to be one of integration rather than isolation. Lawyers and firms will increasingly rely on general-purpose models that are integrated into their existing workflows through specialized software interfaces. The focus will shift to improving the user experience, ensuring data security, and creating tools that maximize the efficiency of these powerful models. Collaboration between legal software vendors and AI providers will become the norm, allowing firms to access cutting-edge technology without the burden of maintaining their own AI infrastructure.
Sean Mitchell is a technology journalist based in London with over 12 years of experience covering the intersection of artificial intelligence and enterprise software. He has reported extensively on the legal tech sector, conducting interviews with legal leaders and analyzing the impact of digital transformation on law firms. Mitchell previously served as a senior editor at a leading legal technology publication, where he oversaw coverage of the industry's shift towards automation and AI integration. He holds a Master's degree in Journalism from City, University of London, and is a certified member of the Society of Editors.