WEBVTT
00:00:07.620 -->
00:00:19.289Hello and welcome to our podcast. Today we're discussing one of the most important questions in technology right now. How do you build AI that professionals can truly trust?
00:00:20.309 -->
00:00:48.179AI tools are becoming widely available, but in fields like law, science and healthcare, accuracy, security and privacy are critical, and there's no margin for error. At RELX, we believe our unique contribution lies in how we combine deep customer understanding, proprietary data, and advanced technology to create AI solutions that professionals can rely on for the high-stakes decisions they make every day.
00:00:49.439 -->
00:01:03.780So today, to explore this, my guest is Jamie Buckley, the chief product officer across both RELX's legal business called LexisNexis and our scientific, technical and medical business called Elsevier. Jamie, welcome.
00:01:04.469 -->
00:01:06.299Hey, YS. Thank you for having me on.
00:01:07.469 -->
00:01:16.379So, as we always do, before we dive in, can you tell us a little bit about yourself, both professionally and personally, and what you're working on today?
00:01:17.040 -->
00:02:00.989Yeah, sure. I've always been involved in technology, and I taught myself how to code when I was pretty young, and I got bachelor's and master's degrees in computer science. That's where I first learned about AI. Obviously, not generative AI that we see today, but more traditional AI. On my first job out of school, I was writing code at a B2B startup, and I worked with some customers there at the startup, and I really enjoyed it. And so I moved to product management early in my career, and I've been doing kind of product and product leadership really ever since then. So I've been at a few startups, also some larger companies, eBay and Microsoft, and I moved to Silicon Valley in 2010, and I really love it here.
00:02:00.989 -->
00:02:17.099I really feel like it's at the center of technology. I joined RELX as chief product officer for LexisNexis Legal in 2015, and then just last year in 2025, my role expanded to include CPO for Elsevier as well.
00:02:18.240 -->
00:02:20.789I can't believe it's been more than 10 years, Jamie.
00:02:20.849 -->
00:02:22.020I know it's crazy.
00:02:22.289 -->
00:02:25.979Oh my goodness, time sure flies when we're having fun, right?
00:02:26.520 -->
00:02:26.759Yeah.
00:02:26.759 -->
00:02:27.689Has it been fun?
00:02:28.169 -->
00:02:33.990It's been really fun. Yeah. As you know, it's an amazing company, amazing culture, amazing products. It's been a lot of fun.
00:02:35.099 -->
00:02:39.090And how have you been able to manage your personal life through this?
00:02:39.870 -->
00:02:48.810It's a lot. I mean, it's really two more than full-time job. I feel like I had more than a full-time job before, and so now we have two.
00:02:44.550 -->
00:03:00.509The nice thing is there are a lot of similarities between kind of how the companies run and how we build products and how we use AI across the companies, and that's actually made it a lot easier to share information between the two.
00:03:00.509 -->
00:03:09.585Yeah, let's get into that a little bit. Tell me, you know, doing two jobs in two different companies, similar job, obviously.
00:03:09.585 -->
00:03:09.645Yeah.
00:03:09.645 -->
00:03:16.365What is the rationale of doing this thing, and how is that helping you, or how is it challenging you?
00:03:16.365 -->
00:03:58.724Yeah. If you step back, my primary job is to work with our teams to build great products for our customers. I also spend a lot of time driving AI across the company to help employees. If you look at both Lexis and Elsevier, we both have leading high quality content, both have deep domain expertise, and both have a state of the art technology platform. Customers and products are different, but our approach to building products is actually really similar, and how we integrate AI across the company is also really similar. That's kind of the rationale for both roles, and that's what makes it a little bit easier to take on both of them at the same time.
00:03:58.724 -->
00:04:03.465When you look at how our customers are kind of using AI.
00:04:03.465 -->
00:05:04.889A lot of them use AI from frontier companies. This is kind of mostly from OpenAI and Anthropic, and lawyers from LexisNexis Legal, and then scientific researchers and health professionals from Elsevier. The most common complaint I hear is that they're not sure that they can trust the output, and we had a recent Elsevier study that showed that 86% of researchers use AI from frontier models, but only 22% of them trust it. he Frontier models are incredible, but they're trained on content from the open web where the quality is a little bit spotty. There's almost 1,600 official cases of lawyers filing documents in court with hallucinated citations. That has serious consequences for lawyers and their clients, so that's not good. The base LLM output is well constructed and confident, but it really lacks the grounding from high quality legal content and high-quality journals, just as an example.
00:05:00.839 -->
00:05:19.800For us, we integrate the latest frontier models into our products. We ground it with our content enrichment, so then automatically validate the output. So, all of this drives a high degree of trust for our customers. That's kind of a long answer, but it's key to what we do, and it's what I really focus on every day.
00:05:19.829 -->
00:05:34.004I think you put your fingers on really important things right there. Just going back to the data you shared about the trust percentage. The usage has gone up tremendously in one year time, but the trust hasn't has it.
00:05:34.004 -->
00:05:34.605That's right.
00:05:34.605 -->
00:05:37.185If anything, it's dropping.
00:05:37.185 -->
00:06:08.295That's exactly right. Yeah. The problem is if you're not able to trust kind of, the quality of the content that the answers are grounded on, it's really difficult. It's kind of like a garbage in, garbage out scenario. You really have to have a strong set of content base, and then ultimately you have to use state of the art technology on top of that, right? So you have to kind of fully integrate, and ultimately it's not just the content, it's all the enrichments that we apply on top of the content.
00:06:08.295 -->
00:06:08.475Yeah.
00:06:08.475 -->
00:06:36.209And those enrichments were really done over many many years. First, they started with humans. We have thousands of subject matter experts across both LexisNexis and Elsevier, and then especially over the last decade or so, it's been using AI for a lot of these enrichments with humans in the loop. And then, in the last few years, a lot more has been done with Gen AI to create even stronger enrichments, and then building more connections within our content.
00:06:37.139 -->
00:07:05.879Yeah, and I think about this from someone who loves food, cooking. You're saying that both have really amazing ingredients, both have the same process, and that they have same tools, right? So you're running two kitchens: one cooking maybe Western Italian food, and another cooking Eastern Chinese food or something like that, but you're supervising both kitchens.
00:07:06.930 -->
00:07:09.120Yeah, exactly.
00:07:06.930 -->
00:07:15.089Like the pots and pans might be similar, right? Like maybe there's a couple small differences here and there.
00:07:11.579 -->
00:07:27.660Maybe the appliances are really similar. But a lot of, there's so many similarities ultimately between the two companies, kind of how we build products. But just as you mentioned, our customers are different, and our products are ultimately.
00:07:27.660 -->
00:07:29.639Right. Customers come for different reason.
00:07:29.819 -->
00:07:30.600Yeah, that's right.
00:07:30.600 -->
00:07:38.310But they do have high standards, which is the next question that I have for you.
00:07:33.000 -->
00:08:06.000Which is that, it's okay if you want to know who won, who were the four semi-finalists at 1952 World Cup? It's okay to be approximately right, but when you're dealing with lies, when you're dealing with cases, when you're dealing with research material, you can't be approximately right. You've got to be precise. And how do we know that our customers do understand that difference?
00:08:06.360 -->
00:08:31.350Yeah, it ultimately comes down to kind of the three pillars that I talked a little bit about. There's kind of articulating the kind of fact, and we have a long history, like decades and decades of history of both companies on high quality proprietary content. That's critical. Also, the deep domain expertise that we built. I mean, that's not something that you can just pick up in, you know, in a couple years. Having that really deep...
00:08:31.829 -->
00:08:35.159Yeah, both companies are over 200 years in existence.
00:08:35.000 -->
00:08:38.899Yeah, exactly.
00:08:35.000 -->
00:08:50.269Yeah, so having that deep domain expertise that comes from all that customer understanding, and then building the state of the art technology platform on top of it. That's really critical.
00:08:44.960 -->
00:08:50.269We continue to communicate that.
00:08:50.269 -->
00:08:53.720I'll just give you an example of on the content side for legal.
00:08:53.720 -->
00:09:55.504We have over 200 billion documents, with a B, and that's updated millions of times every single day. Our deep customer understanding we talk about, that's built off of decades and centuries of interacting with our customers, and then we have all of the thousands of subject matter experts that help to build AI skills. But they're a critical part of our AI human in the loop testing and kind of how we build products. It's really important to make sure that humans are part of that. Then on our technology platform, we have the benefit of being multi-model. There's so much competition that's happening with the frontier AI companies, and we're able to pick the best model for the task. Depending on the specific workflow that our customers kind of working under, we can decide that we need, okay for this one we need the best logical reasoning model and we're going to use that. Or for this one maybe there's a simpler task, maybe it's like a simple summarisation task for example.
00:09:55.504 -->
00:10:12.950Where you don't have to take the best model. You know you can just take a model that's, let's say, really fast and maybe not quite as strong, but the quality itself is going to be almost just as good. Anyway, so we're able to mix and match different models from different sources.
00:10:09.710 -->
00:10:18.710That includes you know OpenAI, Anthropic, Google, and also even open source to ultimately produce the best output.
00:10:18.710 -->
00:10:57.259That's very interesting you talk about this choice that you're making behind the scene to maximise the customer experience. I'm going to jump back to something that you alluded to but did not mention specifically, and you were involved from the very beginning, along with several of your colleagues in different parts of the company. The RELX Responsible AI principles. Tell me, I mean, this was something concocted just as you started in this company, long time ago, before the question of trust was ever an issue, right?
00:10:57.259 -->
00:10:57.620That's right.
00:10:57.620 -->
00:11:11.269Where did you think of this having to be put in place so that our work thereafter for the past 10 - 11, years was indeed different from others.
00:11:07.399 -->
00:11:19.534That we do eliminate hallucinations much better than anyone does, and that there are these trust factors that are built into our services.
00:11:19.534 -->
00:12:30.200Yes there are a lot of core principles that are really critical core principles that across RELX, that we apply to all our AI products, especially. And we've been kind of working, and I said, you know what? We should write these down, and we should communicate these externally because it's something that we really strongly believe in as part of the culture of the company. I think about our responsible AI principles in kind of a few different areas. One is that we actually build our products in a really responsible way. This is without bias, with privacy protections. We're aware of the impacts that it has. That's a key thing. Second is that it's grounded by trusted content. We talked a lot about our proprietary leading content that we have across RELX, and ultimately the answer must reflect the underlying data that's represented by our content, and not just the most popular piece of evidence. It's really important in that sense to, we have to reduce hallucinations, which is a key part of that kind of grounded by trusted content, right? Then third is that it's empowering.
00:12:26.360 -->
00:12:55.835So we help humans to think critical. Let's say when we build on the Elsevier side, when we build a product for scientific researchers, we're not trying to spoon feed a manuscript back. We are trying to help the researcher to think more critically, to provide them with additional literature that they might want to look at. With maybe opposing points of view from something that they were considering, and ultimately this helps humans to think more critically. That's the whole point of what we're doing.
00:12:55.835 -->
00:13:11.254We're living in a world where we're kind of used to always ranking things. You've given me three key factors that have to be present. Are they equally important? Is there part, is there a piece that's a bit more vulnerable?
00:13:12.440 -->
00:13:14.570You're talking about content...
00:13:14.570 -->
00:13:14.899Content.
00:13:15.019 -->
00:14:12.649...and expertise, technology platform. Yeah, I think all of them together are really important. I think, if you don't have, let's say, if you don't have really strong underlying content. If you just have, let's say, some domain expertise in a technology platform. I mean, that describes a lot of the domain-specific startups. I think the problem there is, they're more of like a LLM wrapper, and it's really hard for their customers to kind of trust the output. If you just have the technology platform, and then you're kind of like a frontier model. The frontier models are incredible, very powerful. They do amazing work, but just kind of by themselves or trained on the open web, it's just not trustworthy enough, for knowledge professionals that really rely on these tools every day.
00:14:12.649 -->
00:14:19.370So, you now touched on something I wanted to ask you a little bit later, but let's jump onto it right now.
00:14:19.370 -->
00:14:20.000Yeah.
00:14:20.000 -->
00:14:30.965These large language models are, to me at least, beginning to encroach a little bit into some of our strength.
00:14:31.625 -->
00:14:51.559They're trying to gather the data themselves. They're trying to get more vertically expertised. So, is there tension between us and them, or is that a relationship that can continue to be partnership driven?
00:14:51.559 -->
00:15:22.294Yes, so we work really closely with the frontier model companies, and we have a strong relationship with all of them. That includes OpenAI, Anthropic and Google. And then we have long-standing relationships with Microsoft and Amazon. So we actively use models from OpenAI, Anthropic and Google in our products. We work really closely with Microsoft and Amazon on model hosting and cloud hosting. We have written partnerships with all five companies, strong partnerships. I'll give you an example, like a Gen AI example.
00:15:22.294 -->
00:16:23.090We signed a partnership with Anthropic in early 2023, and we were one of the first large companies actually to sign a partnership with Anthropic, and we've had a great relationship with them ever since then. So 2023 doesn't sound like that long ago, but it was actually kind of very early days in Gen AI since ChatGPT only launched broadly in late 2022. When we have similarly strong relationships with the other companies also, and they're all innovating. They're all kind of leapfrogging each other with different things that they're doing with their models. So, for example, they're competing on kind of quality, they're competing on price, they're competing on memory and other features of the model itself. A lot of competition. So the multi-model architecture that I talked about, lets us kind of use the best of what's available to us. Now, as you mentioned, OpenAI especially and Anthropic, are investing more in verticals.
00:16:19.190 -->
00:16:26.419Like Anthropic's Claude for Legal. That's kind of a vertical investment that they've made.
00:16:26.419 -->
00:16:47.674Just one. Both companies are doing more and more in verticals. So because of our really strong partnerships, we always talk with those companies before they launch. They reach out to us and they let us know what they're planning, and we always have a discussion with them to see if there's an opportunity that makes sense for us. You know, a partnership opportunity that makes sense.
00:16:44.884 -->
00:17:32.494Sometimes there is, and sometimes there's not. I'll give you an example of one partnership. In May, we integrated the plugins from Claude for Legal directly into Protégé, which is the flagship product that we have at LexisNexis Legal. We took the Claude legal skills that Anthropic launched, and we integrated those directly into Protégé. You can use those directly in Protégé, and that benefits. It's basically using those legal skills with our, the benefit of our content and enrichments, but it's not running on Claude, it's running within Protégé. So it's running, kind of on our product. That was an example of something, a partnership that we did with Anthropic.
00:17:32.494 -->
00:17:38.855Did that partnership protect our data from being reversed entered into their model?
00:17:38.855 -->
00:17:41.704Yes exactly.
00:17:38.855 -->
00:17:54.079Yeah. So we don't, they're not using our data to train the model. They're not using our content to provide answers. It keeps it all self contained within our ecosystem, within our products.
00:17:55.710 -->
00:18:14.309So while you're being a CPO for us and facing the customers, what about your internal customers? Are they as much of opportunity for you to improve our process, to be more efficient, effective?
00:18:15.119 -->
00:18:58.079Yeah, to me there's a huge opportunity for us to continue to use AI internally across, basically across all functions in RELX, every single employee. For the last few years I talked about the strong partnerships we've had with frontier companies. The last few years we've had leaders from LexisNexis and Elsevier visit San Francisco for executive briefings, with OpenAI, Anthropic and Google. So those are helpful to strengthen the partnerships and for us also to learn more about how they operate. When you step back and think about how do companies best use AI internally for their employees? There's really no better company to talk to than OpenAI. Like how are they?
00:18:54.599 -->
00:19:09.630They're really at the frontier, so how are they using it to help their employees? Anyway, they have a really interesting model called embedded innovation, where they have small technical teams that are basically embedded with different functions in the company.
00:19:10.140 -->
00:19:24.660There's an embedded innovation team for sales, one for marketing, product management, engineering, HR. You kind of go through the list. The sole focus of each of those embedded innovation teams are on improving productivity and quality just for that function.
00:19:20.700 -->
00:19:55.200They learn about pain points that folks in that function have, could be HR, marketing, whatever. It's a technical team, so they're building agents, they're helping to test them across the company, then ultimately to evangelise them across the group, and then measure the success of those agents and adoption. Anyway, so we have a similar embedded innovation setup at both Lexis and Elsevier. We have these small AI technical teams that are working closely with various functions, and it's been working really well so far, and we continue to invest in it.
00:19:56.670 -->
00:20:04.980How is this all being kind of accepted and embraced by our own people, both at LexisNexis and at Elsevier?
00:20:04.869 -->
00:21:13.599It's a huge advantage for employees that really want to improve their day to day. Just like any group of people, there are early adopters, and then there are folks that you have to spend a little bit more time on explaining things. But by and large, the employees that we have across both companies really embrace AI. It's not just something that's used to build products just for our engineering and product management and UX teams. It's really every function. So what gets me really excited is when I see something amazing that someone on our sales or marketing team did with AI. I get all these examples kind of coming in, and we just had a leadership conference across Elsevier where the top leaders across all of Elsevier. That you were part of also YS right. You joined that, we had that kind of who quite to trust. We have a lot of conference just recently, and as part of the conference we were sharing a lot of the great learnings across different functions, and it was really empowering just to hear how a lot of the different functions across RELX were using AI for their own productivity and output quality.
00:29:42.355 -->
00:30:09.849educational material that we have on our marketing sites around these. We get questions all the time around what's an agent and why is it important and what are these skills that are launched. We have information on all of that, and that might be helpful for someone even if they're not a customer of ours, right? Just kind of reviewing a little bit of like what's the state of the art in legal technology as an example. We have a lot of information on that's linked off of our website.
00:30:09.849 -->
00:30:21.549You said something very important right there. When technology is moving at this pace, we have customers that are quite varied in their readiness to adopt it, right?
00:30:21.549 -->
00:30:21.819That's right.
00:30:21.819 -->
00:30:41.275Some people who really need to start from bottom and move up gradually, and others who have already leaped and are way ahead, and are expecting things to move even faster. I think that our approach to our varied customer has to be that we don't assume that they're all the same.
00:30:41.275 -->
00:30:43.704Absolutely, yeah.
00:30:41.275 -->
00:32:08.904That's really important. That's really important for us. There are some customers that are early adopters that are kind of going to jump on the latest technology, and there's some that are a little bit more wait and see, and then there are some where we take a little bit more convincing that yes, this is the way to go. We have folks that are early adopters, those are kind of a maybe a little bit easier because they'll, if you have, that's something that even before we release, so that we have different customer panels that we take advantage of. Whereas like a bunch of early adopter customers where we have early releases of products, especially with AI based. We're doing AI based software development across both companies, and what's critical there is that we get a lot of feedback from customers. Those groups, the early adopters, are critical to give us feedback as we're iterating on software. Let's say for sets of customers that want to wait and see, or they really want to better understand before they adopt something. It's really important for us to understand what are some of the things that they're hesitant about, and talking through that.
00:32:08.904 -->
00:32:31.750Then there are some, sometimes we learn something there, and then we incorporate that into our products, or on our
00:32:28.599 -->
00:33:09.913But I know you're a very busy man, doing two jobs at the marketing sites, or in our discussions with customers to same time, sitting in the Silicon Valley, dealing with your colleagues from Asia to Europe to East Coast, so I'm gonna have to let you go at this point. But I think you've really made an important point about the three ingredients that are essential to differentiating LexisNexis and Elsevier from its peer group, and also to serve the customers, kind of intended needs. I really want to thank you for taking the time today.
00:33:08.440 -->
00:33:13.059make sure that we're addressing the things that are really For the audience, this is a real privilege to have someone who important to them.
00:33:13.059 -->
00:33:14.200Well, Jamie, when you and I sit down, we can talk forever.
00:33:14.200 -->
00:33:14.410That's right.
00:33:14.440 -->
00:33:36.609Absolutely, thanks so much, YS for having never sleeps here with us and share his experience. And I hope that when we talk again, that you will give us some, more kind of shockingly pleasing experiences that we've had with me. Always great to chat with you.
00:33:30.958 -->
00:33:32.589product management.
