Thomson Reuters is hearing more from law firms interested in building custom, private AI systems as the industry faces a rise in cyber threats that has created a “dreadful” environment in the sector, CEO Steve Hasker told The Logic. 

That could be a business opportunity for the Toronto-based firm, which recently announced a custom AI deal with Skadden, one of the world’s biggest law firms. Part of the selling point of that deal, Hasker said, is that customers can build AI agents without sharing any of their intellectual property or private data with an outside firm.

“The idea of sharing it with a startup, which is a reskinned frontier model, is worrisome,” Hasker said. Information could bleed out, poorly trained models may hallucinate, or, he added, some model makers could use such data to compete for their business.

Thomson Reuters has struggled to get investors on board with its vision for building “fiduciary grade” AI, with the company’s shares down about 20 per cent over the past year. With Canada’s biggest law firms reeling from a recent cyber breach, and many big Canadian banks banning personal AI agents due to high risks, Hasker said his firm is getting more inquiries about its AI models. 

The stakes are high. Tech company CrowdStrike warned last month that “​​law firms are the new ATM for cybercriminals,” with more legal entities ending up on data leak sites.

Hasker said he’s not surprised, given the value of confidential information held by law firms. “The additional reason why it’s explosive for law firms is that in many cases they don’t own that content,” he said. The company has been working on ways to operate its AI behind a law firm’s firewall, he said, so “there’s no information bleed out anywhere else, including back to Thomson Reuters.”

Thomson Reuters launched Thomson, its own large language model (LLM), in August. While many industry-specific AI models rely on Chinese technology like Alibaba’s Qwen, Hasker said Thomson Reuters has done cybersecurity work to let it switch between open-weight models without any information flowing back to the firm that originally developed them.

Like many companies, Thomson Reuters is hoping that an industry-specific AI based on open-weight models will be cheaper to maintain than those that rely on broader systems developed by Anthropic and OpenAI. The legal industry is on tenterhooks about the future of the big law business model, after storied U.S. firm Weil, Gotshal & Manges announced an unexpected financial collapse last week.

Hasker’s pitch is that Thomson will cut firms’ AI bills over time. Thomson Reuters plans to focus on “all you can eat” pricing instead of usage-based bills that can rack up into a “nasty surprise,” he added. Managing partners at law firms in major financial centres around the world have all told Hasker that they are happy to spend money on AI but are not yet at the point where they are seeing returns on those investments, he said.

Still, Hasker acknowledged that some investors have “yet to really understand” why Thomson Reuters is spending the money to build its own LLM. Its strategy has its skeptics on Wall Street, where the firm’s shares fell on Tuesday after its earnings report. On the earnings call that followed, analysts prodded Thomson Reuters executives about plans to reinvest more money into AI product development at a time when adjusted profit margins for its legal business were already down year-over-year.

Hasker said he believes that industry-specific AI models, developed for a fraction of the cost of frontier models, will become more capable and more popular in the next year. Such models, he added, should be capable of performing “specific tasks even better than a frontier model.”