Thomson Reuters Corp. TRI-T is preparing to deploy its own artificial-intelligence model tailored to legal professionals after an in-house experiment showed that it performs well when benchmarked against leading products from frontier AI labs.
The information and software company spent the past two years creating a large language model called Thomson, in tandem with British-based startup Safe Sign Technologies, which Thomson Reuters acquired last year.
The Thomson model was built on a modest budget of about US$40-million and trained using only about 8 per cent of the company’s trove of legal content.
On benchmarking tests, the model was competitive with recent versions of models launched by OpenAI, Anthropic and Google DeepMind, especially on legal tasks, according to information released by Thomson Reuters.
Chief executive officer Steve Hasker said in a Wednesday interview that the model’s performance had left him “excited, to say the least,” after the company released its second-quarter financial results.
Thomson Reuters nudged its target for total revenue growth for the year higher to approximately 8 per cent, after previously forecasting 7.5-per-cent to 8-per-cent growth.
The company’s quarterly revenue increased 9 per cent to US$1.95-billion in the second quarter, and 10 per cent in the company’s three core divisions serving legal, tax and accounting, and corporate customers.
Thomson Reuters earned profit of US$448-million, or US$1.02 a share, in the quarter that ended June 30. That compared with US$313-million, or 69 US cents a share, in the same quarter last year.
On an adjusted basis, Thomson Reuters said it earned 99 US cents a share, which beat analysts’ consensus estimate of 96 US cents a share.
Woodbridge Co. Ltd., the Thomson family holding company and controlling shareholder of Thomson Reuters, also owns The Globe and Mail.
For the rest of the year, and in 2027, Thomson Reuters said its executives will focus on expanding the company’s suite of AI-enabled products and persuading customers to adopt them.
Mr. Hasker said the proprietary Thomson model has the potential to improve as it is retrained using more of the company’s legal archive. And later this month, Thomson Reuters will connect the model to an analysis tool in its CoCounsel Legal product.
Looking ahead, Thomson Reuters will look to license the model to large law firms to use on their own systems, helping them keep control over their intellectual property and data.
Thomson Reuters leans on proprietary data in AI race as disruption fears mount
And the company will “run a horse race between leading frontier models and Thomson.” If the proprietary model can continue to deliver results, then the company will connect that software to more of the products it sells to customers.
In mid-July, after the quarter ended, Thomson Reuters reached a deal to sell a majority stake in its global print division to KKR & Co. Inc. for US$500-million, freeing up cash from a declining business line to reinvest in its priority products.
RBC Dominion Securities Inc. analyst Drew McReynolds said in a note to clients that the financial results Thomson Reuters reported Wednesday should be viewed “as a positive for the shares at current levels.”
But shares in Thomson Reuters fell 10.2 per cent to $137.49 on the Toronto Stock Exchange on Wednesday, underscoring the continuing unease in markets about potential disruption to the software sector.
Thomson Reuters shares are down 22 per cent so far this year but generally started to rebound in June.
Mr. Hasker said AI has the potential to expand the total market for software products such as the ones Thomson Reuters offers, as companies spend more for better capabilities, and he expects the company to “take more than our fair share.”
“I think our results in the second quarter prove that,” he said.
Revenue from Thomson Reuters’s legal and corporate divisions each increased 10 per cent, to US$772-million and US$537-million, respectively, and the tax and accounting arm was up 8 per cent to US$311-million.
Reuters News revenue increased 4 per cent to US$229-million.
With its share price trading lower, Thomson Reuters reduced its overall share count by 3 per cent by returning capital to shareholders through a US$605-million transaction in May and a US$600-million buyback program that the company completed in July.