When OpenAI released its GPT-3 artificial intelligence model in the summer of 2020, David Wong had just arrived at Thomson Reuters Corp., and it wasn’t long before company executives were asking internally, “Is this our Kodak moment?”
Thomson Reuters had worked with AI for years, and it was part of the product road map that Mr. Wong was hired to build as chief product officer. But suddenly, here was a glimpse of a technological leap forward that was potentially so significant it could force companies to adapt – or risk losing relevance the way once-dominant photography giant Kodak did when it failed to revamp its business model to meet the rise of digital images.
A team at Thomson Reuters tested the new AI language model on a battery of legal research questions to see how it compared to Westlaw, the company’s flagship online research product for lawyers. Could it reliably answer complex legal questions for customers?
The answer was a resounding no. “It failed miserably, actually, at legal research,” Mr. Wong said in an interview. But when an updated model called GPT-3.5 was released two years later, and more sophisticated versions started appearing every several months, Thomson Reuters ran the same tests again and again.
“And the F student that we saw in 2020 graduated to a C-minus student with GPT-3.5 and then to a B or B-plus student with GPT-4,” Mr. Wong said. “And then we realized: Oh no, now we have the moment.”
As the company’s leaders discussed how to respond in the spring of 2023, he said, “we realized at that point that this technology was too powerful not to use.” Mr. Wong predicts that in three to five years, every professional will have a generative AI assistant.
The speed of development and the level of investment in artificial intelligence have exploded in recent years, with hundreds of billions of dollars pouring into the sector, and trillions more predicted to follow. That has fuelled an AI mania, mixing hype with anxiety about the many ways the technology may – or may not – transform large swaths of the economy and the work force.
At Thomson Reuters, there was little debate about the looming impact for its clients. In simple terms, Mr. Wong said, current AI models are particularly good at two things: retrieving information and producing written work. And the vast majority of Thomson Reuters’s 150-odd products do some combination of those two things.
The company, which owns Reuters News, earns the lion’s share of its US$6.8-billion of annual revenue from three main lines of business that offer software to make sense of vast troves of information to legal, corporate, as well as tax and accounting professionals. (Woodbridge Co. Ltd., the Thomson family holding company and controlling shareholder of Thomson Reuters, also owns The Globe and Mail.)
“We basically have search engines on top of databases, and so there was incredible relevance to the offering that we have,” Mr. Wong said. “So that calibrated the speed question.”
Mr. Wong, 40, is a Toronto native and University of Toronto engineering graduate who worked at audience measurement giant Nielsen in New York and at Facebook in the Bay Area before returning to Canada to join Thomson Reuters. As the company recalibrated its plans to move AI to the core of its products, he had to move fast but be careful not to break things – especially its clients’ trust.
When Thomson Reuters surveyed legal, tax, risk and compliance professionals, 77 per cent said they expect AI to have a high or transformational impact on their work over the next five years, eliminating an average of 12 hours of work per week. But professionals still have a “healthy amount of reticence” about charging ahead with AI, the report says. They worry about accuracy, data security, ethics and disruption of established business models. That is particularly true for law firms, where billable hours pay salaries, and junior lawyers are typically trained with years of drudge work.
Most of clients of Thomson Reuters have tested AI tools and know they will use them, Mr. Wong said. The pressing questions now are about how to do that safely, to earn a return on the required investment, and to train staff to use them productively and responsibly.
“We’re seeing a very pragmatic discussion happening right now because people have gotten away from the theoretical,” he said.
Thomson Reuters pledged to spend US$100-million annually to develop its AI capabilities and build them into its products, then quickly increased the budget. “We spent more than that last year and we’re spending more than that this year,” Mr. Wong said.
As generative AI models took off, Thomson Reuters had just come through a two-year transformation plan that streamlined a web of legacy technology systems, moving more of its content to the cloud. That made it possible to experiment and test products faster than the company could have even a few years ago, Mr. Wong said.
It took Thomson Reuters about eight weeks last year to build a proof of concept for an AI-assisted research tool for the company’s Westlaw Precision product. “Our very first iterations were shockingly close to customer acceptance,” he said. Little more than eight months after that work started, the product launched last November.
“There were a lot of steps on the way,” Mr. Wong said. “I couldn’t have told you in February or March of that year that we would ship it in November.”
The company’s next major step is to expand its generative AI assistant, CoCounsel, which perform multistep processes that review conversations, compare and draft documents and make recommendations. The company released an updated 2.0 version in August that it says is three times faster than a previous version, but it aims to expand it across all of the sectors it serves, including tax, and enable it to do increasingly complex work that involves more steps and decisions.
Making that tool essential to professionals will be key to Thomson Reuters’ success. With AI-based tools proliferating, already, “we’ve heard a bit of exhaustion,” Mr. Wong said, and customers are looking to limit the number they deploy to their staff. He thinks Thomson Reuters has a head start in AI-assisted legal research, but for general AI assistants that review and draft documents or automate work, “it is a much more competitive space.”
“I’m sure there will be dozens of startups and smaller companies out there which will create really compelling, great products for a little narrow use case,” he said. “But the question for every big firm is going to be, can you give me everything I want?”