How Vertical AI Achieves Defensible Accuracy - with Steve Hasker of Thomson Reuters

**SPEAKER_2** (0:12)
Welcome everyone to the Emerge AI in Business Podcast. Today's guest is Steve Hasker, CEO at Thomson Reuters. Steve joins Daniel Faggella, Emerge CEO and head of research to explain how AI must be built, trained, and validated to operate in legal, tax, and audit environments where errors carry real regulatory and financial consequences. He outlines the operational requirements for vertical AI that can deliver defensible accuracy at scale, and the role of expert-trained agents in transforming high-stakes professional workflows. Just a quick note for our audience, that the views expressed by Steve Hasker on today's program do not reflect that of Thomson Reuters or its leadership. In this episode, we cover how AI is made reliable for legal, tax, and audit work. To go deeper on this topic and learn how financial institutions are digitizing paper-based records to unlock usable data for AI, and using alternative data like public web and social signals to enhance risk assessment, download our free PDF report, AI in Financial Services Executive Cheat Sheet, at emerge.com/fcs1.
That's emerj.com/fcs1 to download your copy. Now, the conversation with Steve.

**Dan Faggella** (1:38)
So Steve, welcome to the show.

**Steve Hasker** (1:39)
Hi, Dan, good to be here.

**Dan Faggella** (1:40)
Absolutely. I'm excited to be able to dive in here. There's a lot of action in the regulated industries that needs to get covered and transitions that are underfoot. I think that there's a distinction you folks have been making, and I'd like to get your kind of definition of it around kind of general purpose AI and vertical AI. I think when people say vertical AI, they sometimes mean different things. Can you draw that distinction and kind of why the distinction is useful from an executive level? We can open with that.

**Steve Hasker** (2:03)
Yeah, sure, Dan. I think vertical AI in our view is industry specific. So it is very, very purposefully trained with deep domain expertise.
It's grounded in authoritative, continuously updated and maintained curated content and highly accurate content. And it's built to meet professional standards where trust and accuracy, data security and privacy are non-negotiable. And so I think that's the definition in our view of vertical AI and that sort of nature of industry specific. I think there's an additional layer on top of that, which is, for one of a better term, professional grade AI for fiduciary professions. So for lawyers, for tax accounting and audit professionals as examples, and they're the professions that we serve where those professionals have to be right. They cannot afford to make mistakes. They make mistakes, they lose clients, they lose licenses, they get fined, and potentially end up in jail in the most severe cases. And so I think that's an additional layer on top of the vertical or industry specific AI.

**Dan Faggella** (3:13)
Certainly, in your folks' space and huge preponderance of what we do is in financial services, whereas you had mentioned the legal side of the house, even more concerns maybe than some of the other sides of the house, that's going to be a crucial element. When you mentioned vertical AI, I think this idea is useful and I think for the audience, it will be good to drink this in as a paradigm that we're heading into, which is AI that's really continuously trained on very high quality content so that it's ready for certain kinds of tasks.
Back in the day with early AI, there was sort of, well, we don't have enough volume of this kind of paperwork to really train a model on such and such. It seems clearly that the angle you folks are going in is the things that you guys can touch across clients, can train a central hub of sorts without having any breaks in anonymity, and then have basically the sufficient data problem solved through their deployable agents and models in specific contexts. Conceptually, am I with you here or do you define it in a different way?

**Steve Hasker** (4:05)
That's right, Dan.
If I sort of come inside the walls of Thomson Reuters, we have, for example, 250 years of legal content in the UK. We have about 150 in the US. As you go around the world, we have decades and decades at minimum of legal content that has been created by deep domain expertise, curated, constantly updated and refreshed.
Nobody else has that depth and breadth of highly curated, accurate content. We have the same for tax and accounting and audit. That's the sort of bedrock of the way we train AI agents and the way we produce our AI outputs. The second part is we have about four and a half thousand deep domain experts on staff. They are full-time employees of Thomson Reuters. In many cases, they've been here for much of their career. They're highly specialized legal practitioners, tax accounting audit practitioners and some of the finest minds in the world in those professions, in different jurisdictions. They do two things. Traditionally, they've produced the content I just described. Then secondly, increasingly, they are training the agents that we deploy. A quick example, if a particular type of M&A transaction for a law firm or a general council's office involves 30 different steps, our experts in that particular form of the law and that particular type of transaction will train the agent to behave like a world class expert at every one of the 30 steps. And so in sequence and in parallel, the agent can in real time act like a truly world class expert grounded in the content I described, trained by the agents that I talked about. And that's, we think that is what is required for AI to be professional grade enough to serve, you know, highly talented legal professionals.