Financial information provider Thomson Reuters has developed an in-house, proprietary large language model (LLM) to power its use of artificial intelligence (AI) in the products it develops for customers.
Senior executives in most major enterprises are turning to generative and agentic AI to explore value-generating opportunities – but significant challenges, such as security, accuracy and fears over supplier lock-in, can prevent firms from making the most of such opportunities.
In a pioneering attempt to overcome these issues, content and technology specialist Thomson Reuters has taken a step that would be anathema to many CIOs and their C-suite colleagues – launching its own in-house-developed proprietary LLM.
Kirsty Roth, chief operating officer at Thomson Reuters, understands that many of her digital leadership peers would see the company’s decision to launch its LLM, known as Thomson, as radical. Tech giants, such as OpenAI, Anthropic, Google and Meta, spend billions of dollars developing new models on what feels like a weekly release cadence.
AI model development is a world characterised by rapid change and high investment, so why would a traditional blue-chip enterprise outside the high-tech world of super-powerful frontier models decide to spend millions of dollars developing its own LLM? After all, aren’t the tech giants always going to be bigger and better at AI?
Not necessarily, suggests Roth. Yes, the behemoths will have the edge in developing publicly available models, such as ChatGPT, Claude and Gemini, that a broad user base can explore and exploit for day-to-day activities across a range of industries and professions.
“The big tech companies are always going to be bigger, and I think potentially better for very generic use cases,” she says. “If you’re looking at employee productivity, there’s absolutely no point in us trying to beat that capability. There are millions of users out there for them to capture, and that’s really what they’re focused on.”
However, Thomson Reuters believes the general nature of these frontier models also provides an opportunity for companies to develop LLMs that serve a specific niche. And that’s where the company sees a gap for its new model.
“The challenge is that we operate in quite specific verticals, such as legal and tax, and that’s typically not where the big LLMs are focused,” says Roth.
“Our customers want 99.9% accuracy. They don’t want an 80/20 answer. It was clear that a more focused model had the potential to be better for specific tasks than a generic LLM.”
Blending approaches
Roth explains that Thomson will be plugged into the firm’s in-house developed products for customers in key sectors, such as tax and legal. Rather than calling frontier models to answer a user prompt, as they might have in the past, these products will now call Thomson.
“We wanted to give our customers much more accuracy,” she says. “We wanted to give them the best quality content, at the end of the day. And we felt that if we had our own model, we would be able to do that.”
Roth says the decision to develop an LLM might seem radical to outsiders, but it has been an internal priority for some time. Thomson Reuters acquired UK-based AI startup Safe Sign Technologies in August 2024 with one eye on the future.
“We acquired Safe Sign with a view that, over time, we would learn to train a legal model better than the big frontier labs – not because we’re better than them necessarily, but because this is the specialist area we’re much more focused on, and we want to provide a much narrower set of answers,” she says.
“The way that we can build products now means, yes, we still have people involved, of course, but we can have AI write a large amount of our codebase”
Kirsty Roth, Thomson Reuters
At the time of the acquisition, Thomson Reuters suggested that allying Safe Sign’s developments in legal-specific models with its internal content would help its engineers to deliver high-performing AI solutions.
Two years later, this process has come to fruition with the launch of Thomson. Roth says Safe Sign’s mix of technological and legal expertise is helping her firm to create a competitive advantage.
“If you’re an attorney or tax professional, getting something general that an engineer has created that’s never operated in your business is not necessarily helpful,” she says.
“However, getting a product that’s been built by practitioners, who know how you operate, what your workflows look like, and the way in which you think through a task, means you’re going to get to a much better product. That blend is why this approach made sense for us.”
Counting costs
Now, a warning for CIOs who might be thinking of making a similarly radical move: developing your own LLM isn’t cheap. In addition to buying Safe Sign and investing in its own internal development processes, Thomson Reuters, which built its LLM on an open-source foundation, has invested $40m in training Thomson.
Perhaps unsurprisingly for anyone familiar with LLM technology, Roth says training is the key to AI model success. Thomson has been trained on decades of Thomson Reuters’ proprietary content and domain expertise. And training costs can quickly rack up when you’re developing an LLM.
“It’s really a case of how much training you want to give a model,” she says. “To give you a sense of the cost, a single training run could be $400,000, so that’s why you get to these orders of magnitude quite quickly, because you are still using somebody else’s cloud resource to do that training.”
While the model-development process hasn’t been cheap, Roth suggests it gives Thomson Reuters some key long-term advantages.
First, the model has been trained on the firm’s specialist knowledge, meaning Roth and her team can offer customers bespoke AI services that run on data they already know and trust.
“We believe we have the best content source for legal and tax in the world,” she says. “Therefore, we also believe it would be very hard for anyone else to train the model. If you want evidence for this confidence, the fact that customers in these sectors already want to buy our content tells us quite a lot about how valuable it is.”
Second, Thomson Reuters fully controls its model, without the heavy inference costs of typical frontier models.
“Because Thomson is smaller than a traditional LLM, it’s less latent,” she says. “It tends to work quicker because it’s not searching as many things, and it tends to be cheaper to run because we own it. We’re not relying on third parties, all of whom have profit margins.”
Developing products
The company is already putting its model to work. Thomson’s first deployment is inside Tabular Analysis in CoCounsel Legal, the firm’s AI assistant that helps legal professionals manage research, drafting and document analysis. Roth explains how the deployment works.
“If you are an attorney, and you need to do a huge piece of due diligence with an awful lot of documents, and you want to get a simple view of everything that’s going on and start analysing it, AI can be a fabulous help,” she says.
“Tabular Analysis will get through the work and get you an answer far more quickly, with the accuracy you require.”
While the model has launched publicly, the firm’s developments remain a work in progress. The model has been trained on less than 10% of Thomson Reuters content. Roth says more features will come and it’s important to progress carefully.
“We will add more content as we go into the next versions,” she says. “Early on, to be candid, you don’t know how much money you want to spend because you don’t know how good it is. Now, we’re all building a lot more confidence that this model could be a jewel in the crown.”
Currently, internal engineers use Thomson to develop products. However, Roth recognises there is the long-term potential for the firm to sell its model to outside enterprises, including law firms that could use the LLM to question their own data sources securely.
“We have not yet said we’re going to go and sell Thomson,” she says. “As you can probably imagine, there’s quite a lot of interest. So, watch this space on that one. But as of yet, we’re not commercialising Thomson as a thing you can buy.”
Innovating faster
Computer Weekly last spoke with Roth in summer 2025, when she outlined her mission to transform internal processes and customer services using digital systems and services. That process, which involves ensuring high-quality AI is used internally and in the firm’s products, has sped up during the past 12 months, especially with the launch of Thomson.
“The way that we can build products now means, yes, we still have people involved, of course, but we can have AI write a large amount of our codebase,” she says.
“We’re releasing stuff far more quickly, so that’s exciting. But that pace comes with a whole bunch of implications. And then the work level across all teams typically goes up.”
Roth describes the past 12 months as busy. However, there’s no let-up in the pace of digital transformation in the age of AI. She recognises that a development like Thomson helps her firm stay one step ahead of its rivals.
“Our competition is also innovating faster than ever, so that threat keeps it sporty,” she says.
One area where this rivalry could extend in the future is in model development. Roth expects other firms to follow her pioneering lead and to launch LLMs. The CIOs she meets are often worried about spiralling AI costs. Developing your own model gives you control of your own destiny. However, model development, given the expense, won’t be right for everyone.
“You’ve got to believe that the data you have is going to give you an advantage,” she says.
“Otherwise you’re going to create something that’s probably not as good and still costs you some money. There’s got to be some commercial differentiator in there, which we believe our content gives us.”
