In this episode of The Geek in Review, Greg Lambert and Marlene Gebauer welcome back Joel Hron, Chief Technology Officer at Thomson Reuters, for a timely conversation about the shifting relationship among foundation models, legal content providers, legal tech platforms, and the lawyers trying to make sense of the mess. Recent moves by Anthropic, including Claude’s legal practice area tools and MCP connections into legal platforms, raise a larger question for the market. Is a model provider still sitting behind the scenes, or is it starting to become a legal work environment of its own?
Hron explains Thomson Reuters’ commitment to what it calls fiduciary-grade AI, a standard built around trust, verification, transparency, and accountability. For TR, legal AI needs more than a fast answer. It needs systems lawyers trust enough to stand behind. Hron points to Westlaw, Practical Law, KeyCite validity signals, citation ledgers, and verification tools as core ingredients in building AI systems suited for high-stakes professional work. In his view, almost right is not good enough when clients, courts, regulators, and professional obligations sit on the other side of the output.
The conversation turns to how CoCounsel and Westlaw Deep Research use legal content across far more than traditional research tasks. Hron explains that when AI systems gain access to trusted legal content and verification tools, they begin researching throughout the workflow, even while revising contract language or analyzing provisions. He also describes Litigation Document Analyzer, internally nicknamed the BS Detector, a tool designed to review claims in a document and map them to supporting authority, weak support, or no support at all. For lawyers who spend as much time verifying AI output as generating it, tools like these aim to move verification from a manual scavenger hunt into a structured process.
Greg and Marlene also press Hron on Anthropic’s legal plugins, MCP, and the idea of headless legal technology. Hron argues that MCP changes access, not advantage. In his view, the application layer is shifting, but the real competitive value sits in trusted content, expert systems, governance, and domain-specific intelligence. CoCounsel’s user interface represents one expression of TR’s legal agent capabilities, while MCP opens other ways for those capabilities to appear inside broader work environments. Some work will still need a purpose-built legal interface; other work might happen through email, Word, Claude, or another agentic workflow with little visible interface at all.
The episode closes with a larger discussion about what happens when AI starts performing more of the work itself. Hron shares TR’s internal engineering OKR, where more than 50 percent of pull requests should be written by AI, and explains why 51 percent serves as a useful mental model. Once AI performs a controlling share of the work, the human role shifts from doing the task to governing the system. For legal professionals, the same transition is coming. The key question is no longer only whether AI produces useful work. It is whether lawyers have built the systems, context, safeguards, and verification layers needed to trust the work, defend the work, and remain accountable for the work.
Listen on mobile platforms: Apple Podcasts | Spotify | YouTube | Substack
[Special Thanks to Legal Technology Hub for their sponsoring this episode.]
Email: geekinreviewpodcast@gmail.com
Music: Jerry David DeCicca
Transcript:
Cleaned transcript below, using the uploaded file as the source.
Greg Lambert (00:00)
Hey, everyone. I’m Greg Lambert from The Geek in Review, and I have our friend Stephanie Wilkins from Legaltech Hub. And Stephanie, all the rage is about the talk about Claude for Legal. So do you mind giving us your perspective from the Legaltech Hub?
Stephanie Wilkins (00:16)
Sure. We’ve been diving into this a lot. And I’m sure anyone listening to this has definitely heard that Anthropic made that huge announcement recently with the launch of Claude for Legal. And there’s really a lot involved in it. We’ve taken a lot of time, across our team, to look at it from different angles and try to do very in-depth coverage on this. To me, it feels a lot like the days back when ChatGPT first came out and people were trying to get their heads around what it even is, let alone what it means. So we’ve done a number of pieces.
The first one covers the full announcement, that there are 12 new practice area plugins, more than 20 MCP connectors with legal tech providers, expansion across Microsoft 365, access to justice partnerships, and a managed agents layer for legal users building in the Claude platform developer environment. You know, just a few things to unpack there. But it is really, arguably, the most significant move a frontier AI provider has made into legal to date.
But it does raise real questions across the market, among them being how the partner ecosystem evolves from here, what it means for the established legal AI platforms, and where the announcement is genuinely game-changing and where we might have a little bit of overhype going on here. So we’ve been on the news from the start. Before it went live, we had a chance to speak to Mark Pike, who’s Anthropic’s Associate General Counsel, and he’s also serving as its product lead for legal. So we’ve included his perspective.
And since then, we’ve looked at multiple angles. We have the plain announcement news itself. We have a visual timeline that traces Anthropic’s path into legal from 2023 through this month. I did a separate analysis that looks into how much legal research you can actually do from within Claude for Legal, because that was one of the big areas it touched on. And as a sneak peek, we get to very different conclusions, whether you’re a BigLaw practitioner, or you do law in a small firm, or you’re a solo practitioner in the access to justice system. And then a fourth piece by Nikki Shaver really dives into the operational reasons why Claude for Legal is simply not yet a lift-and-shift replacement for enterprise legal AI platforms.
This is definitely an inflection point. It is not the death of legal tech as we know it, as some people might want to believe. It has not upended the industry overnight, but there is a lot to follow here, and we’re going to keep looking at it from different angles as they arise.
You can read all of the articles I just mentioned on LegalTechnologyHub.com. And if you want to get these updates in your inbox in real time, you can sign up for our free newsletters and follow the Claude for Legal announcement and the journey we’re on as we really try to be critical and dive into what it really does and doesn’t mean.
Greg Lambert (02:56)
Yeah, well, there’s so much hype, so it’s good to have a little bit of fact-checking going on. So thank you.
Stephanie Wilkins (03:01)
Yep, thank you.
Marlene Gebauer (03:10)
Welcome to The Geek in Review, the podcast focused on innovative and creative ideas in the legal industry. I’m Marlene Gebauer.
Greg Lambert (03:16)
And I’m Greg Lambert. And Marlene, for the past year or so, the legal AI conversation has been dominated a lot by the foundational model race: which model is smarter, which one does the reasoning better, which one has the bigger context window, and which one is going to power the next wave of legal tech tools.
Marlene Gebauer (03:26)
Mm-hmm.
Yeah, absolutely right. But recent announcements from Anthropic and Thomson Reuters raise a different question. So if Claude is now launching legal practice area tools and connecting into major legal platforms through MCP, is Anthropic still just a model provider behind the scenes, or is it becoming a legal tech platform in its own right? And if Claude becomes one of the places lawyers go to work, what does that mean for the value of trusted legal content, citation systems, workflow platforms, and all of the legal AI tools built around those models?
Greg Lambert (04:14)
And that’s exactly why we brought in today’s guest. So we’re welcoming back Joel Hron, Chief Technology Officer at Thomson Reuters. Joel’s been on the show before and talked about professional-grade AI and where Thomson Reuters sees the technology heading. And this time, we want to dig into what the Claude and CoCounsel Legal announcement says about the changing relationships among the model providers, the content companies, the legal platforms, and the firms and legal departments trying to make sense of it all, because it’s kind of crazy. So Joel, welcome back to the show.
Joel Hron (04:49)
Thank you for having me. Good to be back.
Marlene Gebauer (04:52)
Yeah, welcome back, Joel. So for a year, we were just saying everyone was sort of chasing these model capabilities. Thomson Reuters is drawing a hard line around fiduciary-grade AI. You have argued that in high-stakes law, the work is easy, but defending it is what matters. From an engineering perspective, how are you building trust as a system primitive?
Can you walk us through the architecture of the patent-pending citation ledgers and how you ensure the agent isn’t reasoning from the open web?
Joel Hron (05:30)
Yeah, you bet. I mean, so we have leaned into this idea around fiduciary-grade AI. I think the core premise of this idea is that almost right is not good enough in the domains that we deal with. And I would say our focus has really been around how do we exploit the value of the 1.9, almost 2 billion documents across Westlaw and Practical Law that we have, the 1.5 billion KeyCite validity signals. These are all signals that human lawyers use every day to validate and verify and build trust in the work product that they’re putting out to their clients or to the courts or otherwise. And so our focus has been, okay, how do we use those same systems that human lawyers use today to help AI build the same level of verifiability and trust?
And I spent the last two weeks, I got back yesterday, with our customers across corporates, across the Am Law 100 and global large law, as well as some of the largest tax firms in the world. And this theme of trust came up almost repetitively across all three groups of those professionals. And I think this idea of, okay, AI is great, it’s doing a lot of work, but how can I, as a human, maintain accountability for what I’m putting out the door? And do I maintain accountability for it? And the answer to that question is affirmatively, yes. I think the professional maintains that accountability. And so it’s incumbent, I think, on us as software providers to build the tools in terms of verifiability and transparency and auditability to give them what they need in order to stand behind the output. And so that’s the core of what we mean when we say fiduciary-grade AI.
And so in terms of how we are building that, I would say first and foremost is to really leverage the best technology in the market today. And today that’s models like Claude, but also the latest versions of GPT, Gemini, etc. And also this idea of the coding harness and what’s called agent harness now in terms of how this is being evolved for AI agents to proliferate. So that’s sort of the core. And I think everybody is trying to evolve their products to live and operate around that paradigm.
But I think what’s unique and important to us is what tools do we make available to that agent to be able to do its work? And again, those tools lean on access to the content and information we have, but specifically also verification tools, citation ledger tools that we are able to build that allow the agent to do that work of verification for itself and ultimately deliver a better work product at the end of the day. And I think you’ve seen this in Westlaw Deep Research, how this operates. That system works very well. And we’ve adapted that same sort of approach with CoCounsel across more types of work. And that next version of CoCounsel is in beta right now.
I was telling Greg earlier, one of the things that we see, though, is that CoCounsel as a product doesn’t just do deep research when it’s preparing for some important litigation matter. It does legal research on almost every task. If it’s modifying a contract clause or updating terms in a provision or something like this, it is always doing research. It is always going to that content to verify what’s market right now. What has happened in the case history that would support what I need to do to this contract?
And that’s, I think, a much more powerful use of content than just preparing for a litigation matter where people are always doing research. What you see is that these AI systems, when you give them these tools, are actually using this content in a really deep way across many different types of legal work that you might not have considered doing research for before.
Greg Lambert (09:53)
Yeah, I know a lot of times the argument that I’m hearing from a lot of lawyers right now is that the AI is getting them an answer really quickly, but they’re spending almost as much time verifying that the information they’ve gotten back is accurate. With CoCounsel and Deep Research, and the combination of that along with the agent harness that you’re writing in, does that speed up that verification process, or are we getting into the positive now?
Joel Hron (10:09)
Mm-hmm.
Joel Hron (10:30)
Yeah, I mean, in one way it does, but in other ways we’ve built specific products or modules or features, whatever you want to call them, for speeding up verification. One example of that is a product we’ve called Litigation Document Analyzer, but internally we called it the BS Detector. And it was literally an application built around an agent harness and these content tools that was focused on looking at a document. It could be a litigation document. It could really be any kind of document, a brief, anything like this.
And what this system will do is it’ll go through every claim made in this document. And a claim could be a sentence. It could be a sequence of sentences, but at a granular level, what is every assertion that is made by this document? And is it supported by something factual, i.e., case law or statute or regulation or something like this? The output of this is effectively a table of, here are all the claims, and here is the support or lack of support for this claim. And even, do we think this is a hyperbolic extension of what this case actually says or something like this?
That absolutely speeds up verification. Again, the idea isn’t that every brief or every report is going to be 100% accurate. I think, in fact, us building those products is recognition that it may never be 100% accurate, and lawyers need tools to be able to build trust in the work ultimately so that they can stand behind it and be accountable for it.
And I think that’s really what we’re committed to as we build these products: delivering the highest bar of accuracy that we can, but also delivering the tools that professionals need at the end of the day to be able to trust them.
Greg Lambert (12:25)
I want to get into the announcement of the TR and Anthropic collaboration, which is not a new thing. You guys have been collaborating for a while, but I know with all of the news surrounding Anthropic recently launching into legal directly, can you explain the bidirectional relationship that TR and Anthropic have now and what it means for the people who are using CoCounsel or Deep Research? How is it shifting what they’re seeing?
Joel Hron (12:59)
Yeah, you bet. On the surface, this feels like a big change, but two things. One, as you said, we’ve been working with Anthropic for quite a while, as well as working with OpenAI, Microsoft, AWS, Google, etc. But we have been working particularly closely with Anthropic for quite a while. But the second thing that hasn’t changed is, for us as TR, but also as CoCounsel, we’ve wanted our products to exist where customers are working. And that could be the Microsoft 365 stack. It could be Gemini Enterprise or Google Workspace. It could be Anthropic or Claude Enterprise. It could be OpenAI Enterprise. But I think the idea is that we want our products to exist where people are.
And at the end of the day, these platforms, whether they’re AI platforms or general workplace platforms, are meant to do a lot of different things across the business of law or the business of a corporation. And our goal is really to focus on how do we deliver, again, this fiduciary-grade level to those expert tasks that need to happen, particularly within law, but also outside of law in other industries that we practice in.
So in some cases, CoCounsel, the application interface, is the best way to experience and verify and validate that work that’s happening. But in other cases, where there are general work processes happening, our fiduciary-grade tools support those and act as support agents to that work. And I think our focus is to make sure that intelligence and capability exists wherever it is being used.
And I think that’s how we are thinking about CoCounsel, but that’s also how we’re thinking about making CoCounsel available in other systems. And we see a lot of value in that. The interface layer of software, as you guys have said, has been democratized quite a lot by AI tools, and in particular coding tools and things like that. And we see a lot of firms building their own things. We see a lot of firms and companies consuming general-purpose tools as well and building on top of those. And I think what’s critical is that we deliver that fiduciary-grade intelligence into whatever those systems are, whether they’re our own interfaces or things that people are building on their own.
Marlene Gebauer (15:32)
So Joel, I’m wondering if MCP essentially changes what it means to be a legal tech platform. Claude now has 20 MCP connectors into eDiscovery and CLM tools, for example. And so you never have to leave that interface. We’re seeing this kind of squeeze on this application layer.
You’ve mentioned that as agentic systems get more headless, I guess, optimizing for the single front door is not the right way to go. So does the traditional vertical legal tech application survive this orchestration layer, or is everything becoming more commoditized plumbing?
Joel Hron (16:15)
Yeah, I mean, I would say that MCP changes access, not advantage, if that makes sense. MCP, just like APIs have done, but I think MCP is sort of the analog of API integrations in an agent future, changes how people maybe access this technology, but it doesn’t change the purpose of the technology itself.
And I think, certainly for us, that’s about building solutions that people can trust and building information and knowledge and intelligence that people can trust. So for us, I don’t think MCP changes our job to be done, if you will, as a company, which is about building trust. And MCP is just a mechanism for us to deliver that into more types of work where it’s needed.
And like I said, I think in some cases there is a user experience that goes along with that. I gave the example of Litigation Document Analyzer. Maybe that’s a good example where there’s a distinct experience for how you should do that validation at an important moment. But then there are other cases where the experience may not even be an experience. It may be a workflow that gets triggered automatically off of an email, and a series of steps and work happens, and it comes back as another email.
And so I think we’re moving to a world where in some cases there may not be an experience at all. And that’s what I mean by headless. And I think what we want is that our fiduciary-grade intelligence is playing a part in that process no matter where and how it happens.
Greg Lambert (18:01)
Yeah, let me pull on that a little bit, because we’ve always heard legal vendors talk about work. They want their product to be where the attorneys are working, which is code for Microsoft Word or Outlook, typically. But I think we’re seeing even that shift a little bit, that some attorneys are working directly in the AI tools, or may have their own setup that they vibe-coded that allows them to start working on some things.
So my question is, because of the fact that with Westlaw or CoCounsel, there’s this designed user experience that you’ve set up, that you spent probably millions upon millions of dollars getting just right, so that you have this great experience. And then all of a sudden, your users, or at least some of your users, may be shifting away from some of these really good interfaces that you’ve designed for them. Is that kind of difficult for your UX designers to wrap their heads around?
Marlene Gebauer (19:08)
I was actually going to say, are we going back to more content and capability than the delivery system? Sort of how it was before everything got highly technical.
Joel Hron (19:21)
Well, and maybe to riff on that idea a little bit, Marlene, I don’t know if we’re going back to that, because I don’t know that we ever left that point of view. That has always been the centerpiece of everything we’ve built around, having accurate and up-to-date content.
Marlene Gebauer (19:40)
Well, the position as a content provider versus a technology company, that has sort of gone back and forth sometimes.
Joel Hron (19:45)
Yeah.
Yeah.
But I mean, where we spend millions of dollars is on making sure that our information is accurate and up to date. And in terms of being a content provider versus a technology company, our tools like CoCounsel, for instance, or Deep Research, are not just providing a ranked list of raw content back to an agent. There’s a tremendous amount of technology in terms of how we interpret and apply judgment and apply verification and citation and things like this to that information. And I think that is very much what makes us a technology company, more so than the interface that sits on top of that.
Now, the interface that sits on top of that certainly is changing. And I think the options that people have there are proliferating. I think for our design teams and our people building user interfaces, where the dominant work is legal work and the dominant work centers around the need to verify and build trust through a process, I think CoCounsel will continue to build great experiences for that type of work.
And so I think that’s really, if I’m a design researcher, this is what I’m thinking about: how do I build a user experience that elicits that understanding of how and why this claim is made, rather than just surfacing the claim in pretty font and colors? And so that’s the goal of our design teams. In some cases that’s necessary, and they’ll be in CoCounsel to do that kind of work.
In some cases, maybe that level of depth is not necessary. And that might happen out of an email client, or it might happen out of Microsoft Word, or it might happen in a general-purpose AI tool. And again, I think we’re open to either of those paths because we understand that work can span across those two in different situations.
Greg Lambert (21:54)
Do you think, or I guess, are your developers and designers essentially creating two variations of the content, one that gets surfaced through the UI and then one that gets surfaced through an agent-oriented way?
Joel Hron (22:10)
Yeah.
I think this is a really good question, Greg. And I would look at ourselves the same way I think Anthropic looks at themselves. They are a model provider first, and their job is to build models and tools around the models and make them available to builders. And then they’re building Claude Enterprise, the application. And Claude Enterprise, the application, is their best expression of the model. So this is a user interface that expresses the capabilities of the model in a way that allows the user to get the most out of what that model is capable of doing.
And I see our job very much the same. We take models from providers, but we build harnesses and tools around them and under them to be able to have legal capabilities that the base models themselves don’t have. And then our UI, and that’s sort of what is available via MCP, CoCounsel Legal is the agent, and it can do a variety of different things from a legal capability standpoint. Users can access that via MCP and plug it into different places, but CoCounsel, the UI, is our expression of that agent and how we believe a lawyer can get the most out of that agent for certain types of tasks.
And so that’s really how I see it. I don’t think they’re conflicting in any way. I think they’re both useful. One team is really optimized on how do I hill-climb the capabilities of this legal agent by giving it access to expert-level tools and systems. And the other team is focused on, how do I build a UI that expresses the capabilities of this agent in a way that is most useful for a human to interact with it?
Greg Lambert (24:05)
It seems like we have the two teams that are doing that. How well do they learn from one another? Because it would seem like there are certain ways that you’re surfacing information to a human that may also be relevant to the agent, and vice versa.
Joel Hron (24:22)
100%. They work very, very closely with each other. I would say most of the development we do today really starts with the agent. Most of how we think about solving legal problems starts with what is the agent capable of doing? And I think as we build UIs that express those capabilities, those UIs convey obvious gaps.
And some of those gaps can be filled or mitigated by the UI and how we construct the UI and how we construct the workflows within the UI and how we construct things like customization via skills and stuff like this. And some of those things need to be fed back into the agent team to say, okay, well, we need better tools to handle these sorts of edge cases, or we need better behavior for XYZ sorts of use cases. And so there’s a two-way conversation that happens between those teams.
I think the other thing that’s important as you think about agents is the context engineering for the agent is incredibly important, right? The agent is operating off of what it is discovering throughout the process of doing its work. And the human has a lot of context that the agent does not have. And in many ways, the UI itself is a way to help the human user convey their context to the agent in a way.
Just like if you were to hire a new intern at your company, you would probably set up some shared folder with them. And you would say, okay, here’s some recent documents we’ve put together, and here’s an onboarding document. That’s you conveying context to this intern to help them understand, well, this is what we’re doing, and this is why we’re doing it, and this is how we’ve done it in the past. And that’s the same thing that you want to elicit between a human user and an agent. And that’s what I think you guys are helpful at exposing as well.
Greg Lambert (26:33)
One final question on this topic. Just curious, if you were to put a percentage on it, on the coding that your developers do, how much of that are they relying on the AI now to do?
Joel Hron (26:47)
Yeah, this is a great question. Honestly, I have had other podcasts about this topic solo, and we could spend hours on it. We have an OKR in our organization that more than 50% of the pull requests that go into our codebase get written by AI. And I would say some teams are north of 80% at this point.
There’s a really interesting reason, though, we set this OKR. And sorry if I’m taking a tangent. You can pull me back into legal at some point if you want.
Greg Lambert (27:21)
We love OKRs on here.
Joel Hron (27:44)
Okay, so there’s a really interesting reason we chose 51%. And one of the engineers that is on my leadership team mentioned this to me back in December, and it really stuck with me. But he said something really changes about your mindset when you get to 51% of the code being written by AI, because now you, as the human user, are no longer in control of the code that gets written. You have ceded controlling interest of your codebase to something that is not you.
And it really is a good signal for, okay, well, how does your role as an engineer now change? Your role now as an engineer is less about writing lines of code and it’s more about building systems around how code gets written. And those things are governance systems and tests and guidance documents and architecture principles and things like this that help constrain and steer and guide the agent to do the right thing along the way.
And I think it’s a really good analog for how a lawyer should think about their role changing or how a tax professional may think about their role changing. As AI sort of picks up more and more of this grunt work, if you will, your job is more about how do you build systems around AI to do the work you want it to do in the way you want it done, rather than doing the work itself, right? And I think that’s really the mindset shift for an engineer that is taking shape right now. And I think it will likely take shape in other industries over the years to come.
Greg Lambert (29:06)
Yeah, that’s a good parallel. Thanks.
Marlene Gebauer (29:09)
Anthropic just shipped 12 practice area plugins covering everything from corporate law to litigation, and some deploying as managed agents. There’s a fine line between being the partner and being the competitor. So when a lawyer is using Anthropic’s native open-source corporate legal plugin versus routing that workflow through CoCounsel Legal, what’s the functional difference in output trust and defensibility? And I know you’ve talked a bit about this, so maybe you can do a compare and contrast.
Joel Hron (29:48)
Yeah, for sure. What I would say is that if you go look at these plugins, they are nothing but a rudimentary set of instructions for how to do a certain type of work. So I would say the plugins themselves have no concept of validation or verification or groundedness in factuality in any way.
They are helpful guides to an agent to help it meander through a task, but they do not have any concept of these principles, I would say, of fiduciary-grade AI. Now that’s not to say that one couldn’t take a plugin and say, hey, use these tools to validate your work along the way. And those tools could be CoCounsel MCP, living in a plugin.
I think you could do something really well in that context. And I think that’s how we think about MCP in the context of Claude, if you will. You can bring CoCounsel’s capabilities into a lot of these workflows in a native way and get the best of both. But the plugins as a standalone, again, are nothing more than a couple of instruction documents for the agent in terms of how to follow a path. And again, I don’t think they get to the level of depth and trust and transparency that we hear our users are looking for.
Greg Lambert (31:16)
I’ve heard people joking that these are the skills that their in-house legal department uses at Anthropic, but with all the really good stuff pulled out of it. It’s very basic.
Joel Hron (31:29)
Yeah. And look, I also don’t think Anthropic’s goal is to build a legal product per se that covers the spectrum. I think that’s why they are basic. I think they’re meant to be indicative and instructional around, here’s how you build guardrails for an agent, or here’s how you build workflows for an agent. You can take this and then make it much, much better. But here’s the seed of an idea, and you can then go use the platform to grow and expand and think about it in different ways.
And so I think that’s more of the message to take from the plugins than, here’s a legal product that stands on its own two legs. And again, I think our goal is to build tools that work in the context of that system and can be used to add validity and trustworthiness to whatever processes are happening there.
Greg Lambert (32:27)
Let me tag on to that, because one of the interesting things that they did put out was this thing that they call the cold start interview, that you can take a firm’s specific playbook and put it in, and Claude will write that to their Claude markdown files. And suddenly, when I say this out loud, I just envision my security ops person coming in and ripping my computer out of the wall and shutting everything down if I were to do this.
Marlene Gebauer (32:55)
It’s like, no, no, no.
Greg Lambert (32:57)
But the firm’s institutional memory can get encoded directly into these LLM instructions rather than vendor software. Putting back on your CTO hat, not that you’ve taken it off, how do you advise firms on things like this and what it is they should be doing with their proprietary information?
I’m sure you have a preference that they put it into what you’re calling the fiduciary-grade system rather than the LLM itself.
Joel Hron (33:35)
Well, certainly. I would say security is one of the things, I think probably security and trust are the two things that stand between…
Greg Lambert (33:47)
Yeah, that’s why we don’t talk about Grok as an enterprise tool, I think.
Joel Hron (33:50)
Right. So it stands between the real-world application of AI and the capability, perhaps, of AI today. And so I think certainly part of what the definition of fiduciary-grade is does speak to security and how we use that information. And we’ve been very clear that we don’t use that information in terms of training our models and products and things like that. And I think customers really appreciate the stance that we’ve taken on that. And I think they honestly trust us quite a lot because of the decades of work we’ve done with them on that front. And I would say we’ve earned that trust in many ways, and we try to re-earn it every day to keep it.
And so I think as a firm, this is your competitive advantage. In the future, if you can codify your knowledge, and we’ve talked about knowledge management as a domain of law for a while, but if you can codify your knowledge in a way, and you can do it better than the next person and make that knowledge available for AI in agent-native ways, then I think you have a tremendous competitive advantage.
And I would be very reluctant to, A, take that task lightly. And I would be very reluctant, B, to open with that approach. If I’m a law firm or a big corporation, this is something I want to be an expert at, because it is your lifeblood as a company at the end of the day. And I think if I’m making investments as a company anywhere, it’s going to be in this area. And then how do I serve that knowledge and intelligence then? I can choose a million different tools to serve it into. Owning that knowledge is really what differentiates you at the end of the day as a company. And certainly if I was leading a law firm or a big corporation in that sense, that’s what I would be focused on a lot.
Marlene Gebauer (36:00)
It’s funny, we just had this conversation yesterday with Ryan McClead about making sure that knowledge and content is AI-accessible in addition to being people-accessible. So, yeah, I agree. It’s going to be something important.
Joel Hron (36:16)
Well, and I’ll use this analogy, I don’t know if it will stick, but you could think of TR very much as a law firm. So a law firm has a lot of experience and matters that they’ve worked on over the course of time, that they want to index and organize and make sense of to inform what future work they do. That is the same job that we do with case law and statutes and regulations. We just happen to do it with most jurisdictions across the world.
And that is absolutely what we still believe differentiates us as a company and differentiates our products at the end of the day. And the better that we can organize that information and make it available, as you say, for AI, I think the better our products become. And in many ways, we have changed the foundational aspects of how Westlaw works for agents versus how it used to work for humans, and the APIs that the agent calls are different than the APIs that power the application today because agents work with the content in a different way, at a different pace, and at a different rate than humans do. And they need to look and feel and act differently for an agent user versus a human user.
Marlene Gebauer (37:33)
So Joel, you’re a bit of a unicorn. You sit at this intersection of legal content, AI infrastructure, product development, and figuring out what lawyers actually need from these tools. So what are a couple of resources, signals, or conversations that you rely on to separate the real movement in legal AI from a lot of the noise?
Joel Hron (37:56)
Yeah, it’s a really good question. I mean, I would say I read a lot. So I still enjoy following LinkedIn or Twitter. I do wish I practiced fishing more than I read about it, but I try to do both.
Greg Lambert (38:07)
And say you read a lot about fishing too, right? I see the books.
Joel Hron (38:21)
But I would say I do read a lot, and I think there’s a lot of really good content, particularly academic papers, that come out and people reference them on Twitter and things like that. It’s a good source of figuring out what to go read. But I think that’s a good way to stay up to speed on what’s happening in the market.
I think the most important thing, though, is that I use it. I think the most amazing thing for AI for me as an engineer has been, and I used to run a startup before we were acquired by TR, so at a startup with less than a hundred people or so, it’s quite easy to stay deep in the code and involved in how things work. Then you come into TR, and we’ve got 140 products and thousands of engineers. It’s impossible to maintain the level of depth in code. But with AI tools now, when I’m talking to a team and they have an issue, I can immediately go understand the code and what’s happening and what has happened in the last few weeks.
What commits have been made? What issues have come up? My level of depth in the code itself is far more than it could have ever been otherwise because of the ramp-up time and context switching. And that has been such a blessing for me as a technical person, to be able to do that. But I also learned so much by doing that about what is possible and what is capable. It gives me an intuition as well for where things are going and how I think I should be directing teams and this kind of stuff.
And so that would be the best piece of advice I would have for people: whatever it is, go use it, and use it a lot. The more you do, the better your intuition becomes for where this is going and what impacts you think it might have on your teams, your talent, on the products you’re building, etc.
Greg Lambert (40:16)
Yeah, I couldn’t agree more. Well, Joel, it’s time for our crystal ball question. I think we’ve thrown this at you before, but there’s so many things going on with the AI harnesses, the agents, the MCPs, the collaboration between foundational models and products. So what do you think is something on the horizon that legal professionals need to be looking out for and preparing for?
Joel Hron (40:49)
Yeah, it’s a good question. I think, obviously, agents is the trend. I think it’s easy to say that, but if you think about what does that mean, that means now that I am delegating actual work and decision-making to an AI, delegating actual work product to a non-human. And this is exactly like what I said with engineering as well. Your role now changes from doing the work to governing the work.
And so I think the change that you need to be anticipating is, okay, in a world where I’m no longer doing the work, how do I build the systems that give me the trust to stand behind the work that gets created? And for engineers, those are things like architecture design principles, high test automation coverage, etc. For lawyers, it’s other things. And I think for those people doing the work, it would really behoove them to spend a lot of time thinking about that.
And again, that’s what we believe we’re building in our tools: systems that can elicit and elucidate that transparency and trust that’s necessary to build systems around that kind of environment. But I think really thinking about what it means to have an agent do work, and what are the implications if you think three or four steps down the road, that’s what people need to be preparing for today.
Greg Lambert (42:23)
Tell everyone to keep track when the AI takes over 51% of your work. Your role changes, right?
Joel Hron (42:27)
And whether 51% happens or not, it’s kind of irrelevant. It’s more of a helpful mental exercise to say, what if I am no longer the controller of this system? What would I do? And that’s probably a good mental model for some actions that you might want to take.
Marlene Gebauer (42:47)
And I like the way you phrase it. You’re governing the work. I’ve heard, okay, you’re supervising or you’re checking. I think governing is a better word for it, given what you’re saying about you really have to think about these systems. What’s going to make you trust it? And think about that in terms of how to govern it.
Joel Hron (42:51)
Right.
Right.
Well, and again, I mean, the accountability does not shift away from the human in this process, right? And so that’s why I think governance is a good word. Because at the end of the day, if we ship a bug, it’s not like calling up my AI model and telling it how bad of a job it did. Engineers are accountable.
Greg Lambert (43:29)
Claude, you’re fired. Codex, you’re in.
Marlene Gebauer (43:29)
You messed up, Claude.
Joel Hron (43:32)
That’s not what happens.
I think humans maintain accountability. And so it’s incumbent on us to really lean into, how do we maintain that accountability in a way that we can stand behind and trust?
Greg Lambert (43:43)
Joel Hron, Chief Technology Officer at Thomson Reuters. Thank you very much for coming in and unpacking. Man, there’s a lot going on. It’s an exciting time to be in the industry, isn’t it? You bet.
Marlene Gebauer (43:54)
Hmm.
Joel Hron (43:55)
Absolutely. Thank you for having me, Marlene and Greg.
Marlene Gebauer (43:59)
And thanks to all of you, our listeners, for taking the time to listen to The Geek in Review podcast. If you enjoyed the show, please share it with a colleague and don’t forget to like and subscribe. Joel, where’s the best place for people to follow your work and learn more about what Thomson Reuters is doing with CoCounsel Legal?
Joel Hron (44:17)
Yeah, I try to put a lot of updates out on LinkedIn pretty regularly about new things that we’re doing and shipping and partnerships and things like that. So definitely I would say follow me or follow Thomson Reuters there. And I would also say you can gain access to the next versions of CoCounsel, which are in beta here pretty soon. And you can check that out on our website and try to get access. It’s something we’re quite excited about right now.
Marlene Gebauer (44:43)
All right, I’ll definitely encourage everybody to check out what you’re doing and talking about. And I should note, as always, the music you hear is from Jerry David DeCicca. Thank you so much, Jerry, and bye, everybody.