How Thomson Reuters Achieved ‘Fiduciary-Grade’ AI Accuracy

Great to be here, Maggie. Thanks for having me.

Sure. So, Thomson Reuters is the largest provider of content and technology to the legal profession, and we're the largest provider of tax calculation engines, so to both tax and accounting firms and direct to corporations. And then, of course, we have a risk, fraud, and compliance business that does some work with government agencies and corporates. And then, last but not least, the Reuters News Agency, which is the largest wholesaler of news in the world.

Yeah, it's really two things. AI is one of them. But what I would say first and foremost is that we provide content-driven technology to the legal profession, the tax, accounting, audit profession, the risk profession, and of course to news professionals. And in each of those cases, there is a growing level of complexity associated with compliance and decision-making. And what we essentially do is we enable decision-makers to cope with that complexity, to comply with laws, with regulations, and rules without adding headcount. So we're, in a sense, a very efficient way to cope with an increasingly complex world. So that's the first thing. The second thing is AI has enabled us to expand the role we play in the success of our customers. So one example within the legal space would be Westlaw is the leading legal research product, and essentially AI has enabled us to take all of the depth and breadth of knowledge and expertise that's within Westlaw and move into the work product space. So the first draft of a contract, the first draft of a brief, the first draft of a motion to dismiss or a motion to compel. So that's expanded the role we're playing in support of lawyers. And then similarly in the tax and accounting space, where we're the leading, as I said, provider of tax calculation engines, AI enables us to automate all of the activities that go on either side of preparing a tax return. And so both of those things are tailwinds, and we saw that, as you said, in the first quarter. It was a good, strong start to the year, and we're optimistic about the business going forward.

Yeah, this is our true differentiator in an AI environment. So what we provide is fiduciary-grade AI, and that's a term that we coined basically to define AI used by professionals who have to be right. So, if you're a lawyer, a tax, accounting, audit professional, and you produce a tax return or a brief or a contract, an opinion that's incorrect, that hallucinates, that includes inaccurate citations or arguments, then there might be a fine, there might be a risk to your professional license, there will certainly be a risk to your client relationships. And so fiduciary-grade AI is that additional higher standard where the results need to be deterministic, not probabilistic. And essentially, we're able to provide that level of output and accuracy because of four things that we've been investing in for decades and, in some cases, centuries. So the first is our proprietary content. So whether it's Westlaw or Practical Law, Checkpoint, CLEAR, other datasets, we have unique and proprietary data that is produced by thousands of experts on a daily or hourly basis and maintained and produced in pristine fashion. And that enables us to train our AI output in ways that others can't. The second is we have a couple of thousand on-site experts, whether they're practice reference attorneys or whether they're highly qualified CPAs. And basically, what they're doing is both producing that content, but also training our agents, training our AI agents to behave like a world-class, highly tenured professional. Thirdly, we have data privacy and protections that others don't have. So we think about it in the following way: we provide a guarantee to our customers, which is their input will not become part of our AI output. And I think very few of the startups and very few of the protagonists in this AI era will offer that sort of a guarantee. And of course, when you're serving professionals and producing fiduciary-grade AI, that's incredibly important. And then last but not least, we have call centers. And I know that sounds like a very antiquated idea, but if you're using Westlaw's generative AI product for legal research and you don't understand the output or you're not quite sure as to how the agent reached the conclusion that it reached, you can go through and look at each of the steps that are produced. You can also call someone, and they can run a similar query and walk you through the logic, the output, the citations, and how the conclusion was reached. And I think that provides fiduciary professionals with a great deal of comfort around the accuracy and the validation that they need.

Well, I think it's an ongoing opportunity and challenge. What we see is the bar being raised in terms of the ability to use AI for the benefit of professionals. And so, it's a very, very exciting time to be at Thomson Reuters because we have all of that history and all of that content, all of that deep domain expertise within our four walls. And now we have access to a set of frontier models that enables us to play a larger role in the success of our customers, to provide more and more cutting-edge solutions. And so it is particularly exciting. I would say that one of the most exciting things going on amongst many at Thomson Reuters is the launch of the Thomson model, which is a large language model specifically for legal, and we'll extend it into tax and accounting and audit as well. And essentially, it's outperforming the very best and very latest cutting-edge frontier models for specific legal tasks. And that's an example of Thomson Reuters doing something that no one else can do in serving the legal profession. And we think that will give us a degree of strategic and financial flexibility going forward.

Yeah, it's a great question. So, we're doing two things: obviously injecting agentic AI into our flagship products for the benefit of our customers, and also encouraging our 27,000 colleagues to use AI in everything they do and to start to streamline and automate, make more efficient, more effective, and more fun the daily tasks that they're performing at TR. Now, both of those things come with token costs. And as we know, and as the press reported, as you said, those token costs, particularly for access to the very latest model, have risen exponentially. And so within our sort of financial envelope, we've been able to cover those additional costs within the guidance we provided to Wall Street, and we expect that to be the same. But I think it really has encouraged us to accelerate the development of the Thomson model, because having our own in-house model that is trained on our content, in addition to that which is available on the worldwide web, gives us a level of financial flexibility. So should those token costs get to be unmanageable, then we can port some or all of the activities across to our own model, which I think gives us a level of flexibility that others don't have.

Well, our recent Future of Professionals work alluded to the fact that while in the legal profession, tax, accounting, audit professions, adoption of AI tools has been very robust—so in other words, most professionals have access to tools, whether it's sort of through their employer or the things that they've accessed separately, and they're starting to see value accrue to the work that they're performing and the outputs that they're producing. However, the change management, the underlying change management across the firm, is lagging. And I don't think that's a big surprise, but essentially I think it's sort of 18 months to two years before firms are able to truly rewire the workflows and, in a sense, reproduce the way in which those work products are generated. And there'll be a lag in terms of the ROI in that period of time. And we're starting to see that more and more CFOs are saying, "Okay, I understand the costs of adopting these tools, but I don't yet see the benefits." And so what we're doing at Thomson Reuters in support of our customers is putting together a set of forward deployed engineers and customer success and consultants to help drive that and furnish and support that change management within our customer base. And we're starting to see the early results of that, but it is a journey.

For a practicing attorney or a law firm, it's almost limitless, the risk, right? There's a counter of the number of times that a practicing attorney has used an unsanctioned tool, or one that's not supported by appropriate content and expertise, that doesn't have the right data privacy and protections, doesn't have the right customer support, and essentially has filed a motion in front of a judge, and the judge has picked up that hallucination, or opposing counsel picked up that hallucination. I mean, the potential consequences of that are, as I said, limitless. If you think about hiring a practicing attorney to represent you in the context of a high-stakes transaction or a high-stakes litigation, if he or she gets it wrong because they've used unsanctioned AI or they've used it incorrectly, essentially, not only is that a waste of money in terms of the professional fees that may occur there, but more importantly, it poses risks to the client that can be tremendous. And so that's why this idea of fiduciary AI is so important. Fiduciary-grade AI is so important. The practicing professional and, equally importantly, their client need to get a sense that the output is correct, it can be audited, it is transparent, and there is a verification trail there that can be queried at any point in the process.

Yeah, the 500 really is a retooling rather than a reduction. So essentially, as we move into producing more and more cutting-edge agentic products, we need different talent. And so that's really what it's a reflection of. It's not a reflection of us using AI tools in new and different ways necessarily and getting efficiencies. And I think that's the case across industries. We've started to see some announcements of large-scale job reductions, and they've been sort of, in some way, shape, or form, tethered to the use of AI tools. I think it's too early to tell whether that's actually going to happen. I think it's more likely that the number of roles stays the same or even increases, but they change in terms of the nature. Now, that process may be quite dislocating for certain professions and certain industries, but AI will create, I believe, as other major tech disruptions have throughout history, an uptick in economic activity. It will create an uptick in GDP and ultimately create more wealth across populations. What that sort of process looks like and how long it takes, I think, is the big open question.

I think we started in the same place that many did, and we said, "Look, for those who are nervous about adopting AI, AI will not take your job, but someone using AI may take your job." And so we tried to encourage people to immerse themselves in the tool and figure things out. And we've done, I think, a very good job of making various tools available to our colleagues in a controlled environment, and our adoption has been very, very strong. And I think what we're encouraging people to do is take the most mundane part of your job—the bit you don't like doing, the bit that's highly repetitive, that doesn't require a lot of creativity, curiosity, or intellectual input—and automate that. And I think we're seeing good progress in terms of doing that. And when you step back, I'll give you the example of the Reuters newsroom. So we have about 2,700 journalists spread in 200 locations across the world. As I said earlier, we're the largest wholesaler of news in the world, and we're lucky enough to have a long-term agreement to provide news to the London Stock Exchange Group and their financial data business. And essentially, what LSEG continually say to us, as they should, is, "We want more. We want more commodities covered. We want more private and public companies covered. We want more countries. We want more breaking news. We want more financial analysis. And we want more investigations." And essentially, the message to our journalists has been: let's use the AI tools to meet or exceed LSEG's and our other customers' needs rather than worry about job losses. And so my view is that the Reuters news output should, and I think will, increase by three- or four-fold over the next few years with about the same headcount, potentially slightly higher headcount. I think that's a very healthy outcome, first and foremost for our customers, and secondly for our colleagues.

So just on that, Maggie, what we've done—and I think this is differentiated, and I've done my own sort of anecdotal audit of various tools across industries and professions—what we do is we build into our agentic processes 100% transparency as to all the steps that the agent is undertaking, in sequence and in parallel firstly, and then where it is in terms of working its way through those, and then maybe equally importantly, the citations or references, the sources that it accesses in order to reach a particular conclusion for a particular step. So in other words, what we're trying to do is produce the opposite of a black box.

So what we've tried to do is literally the opposite of a black box and ensure that every single step is 100% transparent. And the interesting reaction that we've gotten since we launched the latest version of CoCounsel a couple of weeks ago here in the United States is that our managing partners are saying, "Ah, I can now see how we're going to train our young lawyers." Whereas previous to the release of this version of CoCounsel, they were saying, "Look, one of our biggest problems is how many people to hire at entry level and what is going to be the apprenticeship model if a large amount of that work is automated." Now what they're saying is, "Yes, CoCounsel will automate those tasks, but a curious young lawyer can work with that tool and, at every step of the way, figure out exactly what it's doing and why it's doing it and where it's sourcing particular pieces of intelligence." And if they are indeed curious, they're going to learn very quickly, as they would have sitting side by side with a highly tenured, experienced lawyer.

Well, I think it's early days. The legal profession probably leapt a little quicker and a little earlier into this AI era because the early frontier models were so good at the written word, constructing sentences, whereas the quantitative side of things has taken a bit longer. I think the models are very much there as we sit here today, but essentially the issues are different. In law, the complaint over the last sort of 5 or 10 years, and perhaps longer, has been lifestyle and work-life balance—lawyers just working all-nighters and six days a week, and so forth. In tax and accounting, it's a very different challenge. When I became a CPA many, many decades ago, there were lots and lots of kids coming out of college going into that profession, joining the Big Four, getting an apprenticeship, and then using that as a launchpad for their careers, either staying within the Big Four or doing as I did and going into industry. Nowadays, there are fewer and fewer kids who are excited about a CPA qualification, excited about tax and accounting as a profession. And so in a sense, the technology has a very important role to play in meeting the demand, because the number of tax returns continues to go up, the complexity of those returns continues to go up, and I think it's a fair bet to suggest that that trend will continue. And then the same for audits: the number of audits, the complexity of audits, the stakes involved in getting an audit right all increase. And if at the entry level there aren't enough young, talented graduates coming in and being prepared to take that on as a profession, the technology has a critical role to play to ensure that we have a functioning economy and therefore a functioning society. And so it's a pretty exciting mission, I think, for us at Thomson Reuters and our competitors to meet or exceed that need.