The Thomson Reuters building in Times Square, New York. The company announced Tuesday that its AI-enabled CoCounsel has one million users in more than 100 countries and territories.Jeenah Moon/Reuters
Thomson Reuters Corp. TRI-T is betting on the value to professionals of artificial intelligence agents that can carry out complex tasks accurately and on their own, seeking to tamp down fears of disruption that have hung over the software sector in recent weeks.
As AI-based products flood the market, Toronto-based Thomson Reuters is seeking to draw a contrast. On one side is its own software, trained using a vast trove of content spanning the legal, tax and corporate sectors. And on the other, new plug-in tools brought to market by AI giants, which are directly challenging incumbents.
To highlight the difference, Thomson Reuters is making the case that its AI-based software is already taking hold at law firms as well as tax practices, as lawyers and accounting professionals seek to speed up their work and automate laborious tasks.
The company announced Tuesday that its AI-enabled CoCounsel technology now has one million users in more than 100 countries and territories.
Chief product officer David Wong predicts a turning point this year for businesses’ relationship with AI. He expects professional companies will focus more on the return they’re getting from AI investments.
“We are actually in a bit of an ROI crisis,” Mr. Wong told reporters. “Businesses have been experimenting with AI. They bought licences. They’ve run pilots. They’ve told their boards, ‘we’re investing in AI transformation.’ But they’re struggling to show results.”
Slumping tech stocks revive concerns about AI-fuelled disruption
Software and data providers such as Thomson Reuters have watched their share prices plunge lately, not for that reason, but in response to new tools for lawyers released by Anthropic, a leading AI company that makes the Claude large language models.
For some investors, that raised the risk that established software companies could be disrupted, and muddied the outlook about who will win or lose in the race to deploy AI for professionals.
In response, Thomson Reuters chief executive officer Steve Hasker said the market reaction “represents anxiety and not fundamentals.”
Woodbridge Co. Ltd., the Thomson family holding company and controlling shareholder of Thomson Reuters, also owns The Globe and Mail.
Although some investors interpreted Anthropic’s new tools as a direct threat to software providers, Thomson Reuters chief technology officer Joel Hron said the company has “developed a particularly deep collaboration” with Anthropic, which includes collaboration on engineering and research.
Thomson Reuters worked closely with Anthropic for the past year, using Claude as a foundation to develop the newest version of CoCounsel, which is billed as an autonomous legal assistant that can do its own research and deliver human-calibre output. A lawyer then reviews and validates what the agent drafts.
“This is not a black box,” Mr. Hron said. “It is meant to be a human collaborator.”
One Thomson Reuters tax product features an AI agent that helps prepare multiple tax returns for companies collecting sales tax in many jurisdictions, then flags items that need human review. The product’s first version cut the total amount of time spent on the process, which is typically very manual, by 60 to 70 per cent, Mr. Wong said.
Thomson Reuters has also been privately working on a project to develop a proprietary model, trained on a more concentrated set of data that draws on Thomson Reuters’s expertise in professional services. The company has worked closely with academics on the project.
Early benchmarking tests highlighted by Thomson Reuters suggest that its own model outperformed prominent rivals such as OpenAI’s GPT-5 and Anthropic’s Claude Opus 4.5 on tests of reasoning and factuality, document review, summarization and AI-assisted research.
Some products present well in demonstrations but stumble when it comes to accuracy and verification, said Prof. Jonathan Richard Schwarz, head of AI research at Thomson Reuters and a visiting professor at Imperial College London.
On “correctness” and an emphasis on evidence, “the models are really struggling,” he said. “Rather than throwing more hardware and more compute at the same sort of approach, really you should try and bring in this domain expertise into the training process.”