Why Thomson Reuters developed its personal proprietary AI mannequin
Monetary data supplier Thomson Reuters has developed an in-house, proprietary giant language mannequin (LLM) to energy its use of synthetic intelligence (AI) within the merchandise it develops for purchasers.
Senior executives in most main enterprises are turning to generative and agentic AI to discover value-generating alternatives – however vital challenges, corresponding to safety, accuracy and fears over provider lock-in, can stop corporations from benefiting from such alternatives.
In a pioneering try to beat these points, content material and know-how specialist Thomson Reuters has taken a step that will be anathema to many CIOs and their C-suite colleagues – launching its personal in-house-developed proprietary LLM.
Kirsty Roth, chief working officer at Thomson Reuters, understands that lots of her digital management friends would see the corporate’s choice to launch its LLM, referred to as Thomson, as radical. Tech giants, corresponding to OpenAI, Anthropic, Google and Meta, spend billions of {dollars} creating new fashions on what seems like a weekly launch cadence.
AI mannequin improvement is a world characterised by speedy change and excessive funding, so why would a conventional blue-chip enterprise outdoors the high-tech world of super-powerful frontier fashions resolve to spend tens of millions of {dollars} creating its personal LLM? In any case, aren’t the tech giants at all times going to be larger and higher at AI?
Not essentially, suggests Roth. Sure, the behemoths may have the sting in creating publicly out there fashions, corresponding to ChatGPT, Claude and Gemini, {that a} broad person base can discover and exploit for day-to-day actions throughout a variety of industries and professions.
“The large tech corporations are at all times going to be larger, and I believe probably higher for very generic use instances,” she says. “Should you’re taking a look at worker productiveness, there’s completely no level in us attempting to beat that functionality. There are tens of millions of customers on the market for them to seize, and that’s actually what they’re targeted on.”
Nevertheless, Thomson Reuters believes the final nature of those frontier fashions additionally offers a chance for corporations to develop LLMs that serve a selected area of interest. And that’s the place the corporate sees a spot for its new mannequin.
“The problem is that we function in fairly particular verticals, corresponding to authorized and tax, and that’s usually not the place the large LLMs are targeted,” says Roth.
“Our clients need 99.9% accuracy. They don’t need an 80/20 reply. It was clear {that a} extra targeted mannequin had the potential to be higher for particular duties than a generic LLM.”
Mixing approaches
Roth explains that Thomson will likely be plugged into the agency’s in-house developed merchandise for purchasers in key sectors, corresponding to tax and authorized. Fairly than calling frontier fashions to reply a person immediate, as they may have up to now, these merchandise will now name Thomson.
“We needed to offer our clients rather more accuracy,” she says. “We needed to offer them the highest quality content material, on the finish of the day. And we felt that if we had our personal mannequin, we’d have the ability to do this.”
Roth says the choice to develop an LLM might sound radical to outsiders, nevertheless it has been an inner precedence for a while. Thomson Reuters acquired UK-based AI startup Protected Signal Applied sciences in August 2024 with one eye on the long run.
“We acquired Protected Signal with a view that, over time, we’d be taught to coach a authorized mannequin higher than the large frontier labs – not as a result of we’re higher than them essentially, however as a result of that is the specialist space we’re rather more targeted on, and we wish to present a a lot narrower set of solutions,” she says.
“The way in which that we are able to construct merchandise now means, sure, we nonetheless have folks concerned, after all, however we are able to have AI write a considerable amount of our codebase”
Kirsty Roth, Thomson Reuters
On the time of the acquisition, Thomson Reuters instructed that allying Protected Signal’s developments in legal-specific fashions with its inner content material would assist its engineers to ship high-performing AI options.
Two years later, this course of has come to fruition with the launch of Thomson. Roth says Protected Signal’s mixture of technological and authorized experience helps her agency to create a aggressive benefit.
“Should you’re an legal professional or tax skilled, getting one thing common that an engineer has created that’s by no means operated in your enterprise shouldn’t be essentially useful,” she says.
“Nevertheless, getting a product that’s been constructed by practitioners, who understand how you use, what your workflows appear to be, and the way in which during which you suppose by means of a process, means you’re going to get to a a lot better product. That mix is why this method made sense for us.”
Counting prices
Now, a warning for CIOs who is likely to be pondering of creating a equally radical transfer: creating your personal LLM isn’t low-cost. Along with shopping for Protected Signal and investing in its personal inner improvement processes, Thomson Reuters, which constructed its LLM on an open-source basis, has invested $40m in coaching Thomson.
Maybe unsurprisingly for anybody conversant in LLM know-how, Roth says coaching is the important thing to AI mannequin success. Thomson has been educated on many years of Thomson Reuters’ proprietary content material and area experience. And coaching prices can shortly rack up if you’re creating an LLM.
“It’s actually a case of how a lot coaching you wish to give a mannequin,” she says. “To present you a way of the associated fee, a single coaching run could possibly be $400,000, in order that’s why you get to those orders of magnitude fairly shortly, since you are nonetheless utilizing any individual else’s cloud useful resource to do this coaching.”
Whereas the model-development course of hasn’t been low-cost, Roth suggests it provides Thomson Reuters some key long-term benefits.
First, the mannequin has been educated on the agency’s specialist data, that means Roth and her staff can supply clients bespoke AI companies that run on information they already know and belief.
“We consider now we have the perfect content material supply for authorized and tax on the earth,” she says. “Subsequently, we additionally consider it will be very onerous for anybody else to coach the mannequin. If you need proof for this confidence, the truth that clients in these sectors already wish to purchase our content material tells us rather a lot about how helpful it’s.”
Second, Thomson Reuters absolutely controls its mannequin, with out the heavy inference prices of typical frontier fashions.
“As a result of Thomson is smaller than a conventional LLM, it’s much less latent,” she says. “It tends to work faster as a result of it’s not looking out as many issues, and it tends to be cheaper to run as a result of we personal it. We’re not counting on third events, all of whom have revenue margins.”
Creating merchandise
The corporate is already placing its mannequin to work. Thomson’s first deployment is inside Tabular Evaluation in CoCounsel Authorized, the agency’s AI assistant that helps authorized professionals handle analysis, drafting and doc evaluation. Roth explains how the deployment works.
“In case you are an legal professional, and it’s worthwhile to do an enormous piece of due diligence with an terrible lot of paperwork, and also you wish to get a easy view of all the things that’s occurring and begin analysing it, AI could be a fabulous assist,” she says.
“Tabular Evaluation will get by means of the work and get you a solution much more shortly, with the accuracy you require.”
Whereas the mannequin has launched publicly, the agency’s developments stay a piece in progress. The mannequin has been educated on lower than 10% of Thomson Reuters content material. Roth says extra options will come and it’s essential to progress rigorously.
“We’ll add extra content material as we go into the following variations,” she says. “Early on, to be candid, you don’t understand how a lot cash you wish to spend since you don’t understand how good it’s. Now, we’re all constructing much more confidence that this mannequin could possibly be a jewel within the crown.”
At the moment, inner engineers use Thomson to develop merchandise. Nevertheless, Roth recognises there may be the long-term potential for the agency to promote its mannequin to outdoors enterprises, together with regulation corporations that might use the LLM to query their very own information sources securely.
“We’ve got not but stated we’re going to go and promote Thomson,” she says. “As you possibly can in all probability think about, there’s numerous curiosity. So, watch this area on that one. However as of but, we’re not commercialising Thomson as a factor you should buy.”
Innovating quicker
Laptop Weekly final spoke with Roth in summer time 2025, when she outlined her mission to remodel inner processes and buyer companies utilizing digital programs and companies. That course of, which entails guaranteeing high-quality AI is used internally and within the agency’s merchandise, has sped up through the previous 12 months, particularly with the launch of Thomson.
“The way in which that we are able to construct merchandise now means, sure, we nonetheless have folks concerned, after all, however we are able to have AI write a considerable amount of our codebase,” she says.
“We’re releasing stuff much more shortly, in order that’s thrilling. However that tempo comes with a complete bunch of implications. After which the work stage throughout all groups usually goes up.”
Roth describes the previous 12 months as busy. Nevertheless, there’s no let-up within the tempo of digital transformation within the age of AI. She recognises {that a} improvement like Thomson helps her agency keep one step forward of its rivals.
“Our competitors can be innovating quicker than ever, in order that risk retains it sporty,” she says.
One space the place this rivalry may prolong sooner or later is in mannequin improvement. Roth expects different corporations to comply with her pioneering lead and to launch LLMs. The CIOs she meets are sometimes nervous about spiralling AI prices. Creating your personal mannequin provides you management of your personal future. Nevertheless, mannequin improvement, given the expense, gained’t be proper for everybody.
“You’ve obtained to consider that the information you have got goes to offer you a bonus,” she says.
“In any other case you’re going to create one thing that’s in all probability not nearly as good and nonetheless prices you some cash. There’s obtained to be some business differentiator in there, which we consider our content material provides us.”
