Traditional companies with deep reserves of proprietary data are building their own enterprise AI models rather than renting from frontier AI pioneers. The shift signals that firms with decades of specialized expertise may hold an advantage in the race to deploy trustworthy AI in regulated industries.
"We see ourselves as an AI technology company through and through," said Joel Hron, chief technology officer at Thomson Reuters.
"From a broader perspective, we've been known as an information services firm for some time, but the truth is that the way that content, information and expertise is delivered to the market now is through AI, and as such, we see ourselves as an AI tech company."
Increasingly, Thomson Reuters is turning to custom enterprise AI models, which are large language models (LLMs) trained on an organization's proprietary data and domain expertise, designed to deliver verifiable, industry-specific answers that general-purpose AI tools cannot reliably provide.
This is an early and growing trend. According to the Thomson Reuters Institute's 2026 AI in Professional Services Report, organization-wide AI usage nearly doubled to 40% in 2026 from 22% the year before, based on a survey of more than 1,500 professionals across 27 countries. Yet only 18% of respondents said their organizations tracked the return on investment of AI tools, a proportion unchanged from 2025. The report also found that 15% of organizations have adopted agentic AI, with another 53% planning or considering it.
Why Are Traditional Firms Building Custom AI?
Thomson Reuters traces its roots to the 1930s, when Canada's Thomson began publishing for professionals, and to 1851, when the UK's Reuters news agency opened its doors. The two companies merged in 2008. For lawyers, tax professionals and accountants, the firm provides essential technologies and content. Since the release of ChatGPT in late 2022, analysts have called these sectors ripe for AI disruption. Thomson Reuters did not sit back and wait. Despite the share price moment, it seized AI.
In the months since that first interview, the Canada-headquartered firm developed Thomson, a frontier model built on a proprietary LLM from Thomson Reuters. The firm implemented Thomson within CoCounsel, its AI platform for professional services. Presenting Thomson to the press, Hron said Thomson Reuters is one of the largest employers of tax and audit experts, and that the work these experts do will help drive the flywheel of AI capability.
Hron said Thomson represents the next stage of AI: intelligence that can be trusted and put to work in the business. When asked if Thomson Reuters was looking to protect its deep reservoirs of content, Hron said the firm was ensuring that AI prioritizes the elements important to industry: trust, truthfulness, verifiability and auditability.
He agrees that frontier AI pioneers can take an employee to an answer that is 90% good enough. But in business, and especially in tax and law, 90% is not sufficient.
"The most cited resource for AI is Reddit; there is no lawyer that is going to send a brief to a courtroom based on content from a Reddit forum," Hron said because no good lawyer will rely on a chatroom.
"We hear from our customers all the time that they cannot understand why an AI system said something. That then doesn't save time as they have to do all the same work to find the right cases and information."
Why Build on Open Source AI Models?
Thomson is built on the open weight model Qwen.
"One thing we observed is that the open source community has kept pace with the frontier models of AI," Hron said. "That is a real opportunity for a company like Thomson Reuters to take the open source models and use them for training and developing with our unique data and at a fraction of the cost."
Open-weight models also enable Thomson Reuters to take a continual learning approach to its LLM, said Jonathan Schwartz, head of foundational research at Thomson Reuters until September 2026.
"We went through stages of making sure that the value of the model and how it behaves and answers sensitive questions aligns with the values of Thomson Reuters," Schwartz said.
"We wanted to make sure it produced outputs that we are comfortable with."
The firm achieves this by using its human experts to create learning techniques from which they can judge the AI outputs.
What Are the Cost Benefits of a Custom LLM?
Using an open-source model delivered measurable cost savings. Thomson Reuters invested $40 million to train Thomson.
"For years, the AI industry has treated scale as the answer: bigger models, more compute, more money," Hron said. "Thomson shows there is another path. Start with a strong foundation, specialize it deeply for the work that matters, and you can build intelligence that is highly capable, far more efficient, and entirely under your control. We think that changes the economics of professional AI."
The firm also created a developer portal and partnered with European frontier AI provider Hugging Face.
Hron called Thomson an expert, in direct opposition to the experience many have had with AI that sounds like an expert but delivers inaccurate hallucinations.
"It is the best expression of our unique data, expertise and assets that no one else in the world has," he said.
This fits the equity analogy: an employee can ask the same question of ChatGPT, but if the expertise and data assets behind it come from Reddit, is that a solid foundation?
Thomson Reuters is not alone. US-headquartered human resources firm Josh Bersin Company launched Galileo in October 2024, an AI-powered HR assistant. Galileo uses Sana AI, a technology from Sweden. Josh Bersin Company has a wealth of data, content and expertise in HR, and created a tool that ensures users get a better-quality answer than they would on traditional AI tools.
"We are aware that you could potentially do everything we do on ChatGPT, and it might give you an answer, but you just don't know if that answer is correct," Josh Bersin said in an interview for Diginomica.
The expertise these firms offer is built on their investment in quality content, written by subject matter and content creation experts. Hron said Thomson Reuters focused on the mid- and post-training techniques that used the firm's content from trusted resources such as Westlaw, Practical Law, Checkpoint and Reuters. At release, Thomson Reuters had trained Thomson on just 10% of its content.
Schwartz said the team spent a lot of effort integrating experts from the domains they were training for, as well as technical experts.
"They decided what the training should be, and what good answers versus bad answers are," Schwartz said. The team ran blind tests between Thomson and the frontier models.
Unlike many outcomes from frontier AI models, Thomson Reuters trained Thomson to supply well-documented evidence, something its user base will demand.
"You can have a model that provides the correct answer, but more importantly is it the correct answer with proof that it is correct," Schwartz said.
How Does Enterprise AI Improve Trust in Professional Services?
As the organizations that rely on the technology and content of Thomson Reuters adapt to AI disruption, Hron said they will see real business benefits from working with an AI provider focused on their sector and challenges.
"Large law firms and corporations are looking for ways to bring sovereignty to their knowledge and intelligence," he said. Digital sovereignty has become increasingly important to organizations as they tackle the twin challenges of AI and geopolitical disruption.
"A lot of people view legal and tax as compliance industries, and somewhat static and resistant to change," Hron said. "My experience is that they have been really eager to change as there is a tremendous opportunity. They have a huge amount of information to synthesize, and AI is good at that, so there is a clear opportunity to accelerate the volume of work that can be done."
The initial adoption of AI is also proving costly for businesses as token costs shock chief financial officers (CFOs) and chief information officers (CIOs) alike. Hron said working with an LLM developed for a specific sector will keep token costs in check and ensure AI delivers a return on investment (ROI).
"Agentic tasks consume more tokens than a traditional generative AI system," he said. "The reason is that an agentic AI is not a model; it is an ecosystem of models running parts of a task, so there are lots of hand-offs between models."
This makes cost tracking difficult for business technology leaders.
The AI landscape is still forming. It may sound as if the pioneers at OpenAI, Anthropic and Mistral have already become the de facto tools, but it is too early to call. As Thomson Reuters shows, traditional companies can use open source AI and their heritage to develop products with a niche and advantage. AI's future is less certain than the law or a tax bill.
Related:
Mark Chillingworth writes about leaders in business and technology. He’s a regular contributor to Digimonica. Find him on LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
Related Articles
