WEBVTT
00:00:07.860 -->
00:00:25.995Hello and welcome to our podcast again. I'm YS Chi, and today we're talking about technology at RELX. How it has shaped the business over many years, how it is helping our customers make better decisions, and how is AI changing what is possible. Technology has been a competitive advantage for RELX.
00:00:25.995 -->
00:01:27.465We spend $2 billion on technology every year, and we employ more than 12,000 technologists across the group, over half of whom are software engineers. Today, I'm delighted to be joined by all four of our divisional chief technology officers. Vijay Raghavan from Risk, Jill Luber from Elsevier, Greg Dickason from Legal, and Brian Brittain from RX. Between them, they bring over 60 years of experience at RELX, which is especially interesting in a field that changes as quickly as technology. So, rather than ask each of you for a long introduction, I'm going to bring you in as we go. Vijay, Jill, Greg, Brian, welcome. Vijay, let me start with you as our captain. RELX has been leveraging machine learning, big data for decades, way before AI became a buzzword. So from your perspective, what has made technology a competitive advantage for us, and what makes RELX's approach different from our competitors?
00:01:27.750 -->
00:01:33.420Yeah, thank you YS. Like you said, at Risk especially certainly across RELX, and at Risk especially.
00:01:33.420 -->
00:01:39.900We've been using AI in some form or the other for well over a decade, probably 15 years or so.
00:01:36.180 -->
00:01:49.140I would say, what makes our approach not just different but sustainable for our long-term growth, is that we don't get hung up on the latest tool just because it's the latest, right?
00:01:45.510 -->
00:01:54.855Many of our most effective or most predictive models over the past decade, you could say, they were basically linear regression or logistic regression models.
00:01:54.855 -->
00:02:32.909There's nothing fancy about those, but it's how we use those models on top of substantial data assets, using our domain expertise, using our deep understanding of our customers, and knowing how to apply our expertise to meet our customers' problems. That combination of things, along with the technology and along with AI, I would say that's what makes us different. If you look at the current age of AI, our approach at Risk and at RELX in general, it hasn't substantially deviated from that core principle. I use five tenets, and I won't go into each of the tenets. But the gist of how we use our technology stack is, at the bottom of the stack, we rent things that we have no business to be building.
00:02:30.210 -->
00:02:42.495The cloud being the most obvious example. At the top of the stack, we build things that constitute our intellectual property. And AI is a tool that we use in the appropriate way, in the best fit for the right purpose.
00:02:42.525 -->
00:02:49.425So, it's interesting because you touched on some really important things that we're going to touch later on.
00:02:49.425 -->
00:03:14.460But I'm going to jump first to Jill and say, you have a particularly challenging task of navigating very large volumes of centuries of content, right? And it's in the scientific and medical area, which is also sensitive because you can't go wrong. How does that shape the way you think about AI and the trusted technology in this domain or these domains?
00:03:14.460 -->
00:03:28.064Yeah, it's a good point. When generative AI first came out, our area of scientific research popped to the top as being an area that would benefit the most from this technology.
00:03:28.064 -->
00:03:48.090We saw the community quickly jump to using LLMs, but also quickly saw the issues with just using an LLM to ask research questions. As we know, LLMs, they do hallucinate. It's predictive. It's not fact based.
00:03:42.194 -->
00:04:35.040It's not grounded in truth, and nor was it designed to be, to be fair. So that immediately showed a need from the community. We want to be able to take advantage of the power of this technology because it is undeniably powerful when it comes to synthesising large That has really been the focus of what Elsevier's been doing, and our use of generative AI is to help researchers, to help clinicians navigate the literature, navigate the research that they want. Bring the value to the top faster, helping the researchers, but at the same time grounded in truth.
00:04:35.040 -->
00:05:19.904Always being able to cite your source, being able to cite where the information came from, providing the ability to dive into that source and verify. So for us, really, generative AI has been accelerating what researchers already do by hand, but there's still the human judgment in the loop, still the
00:04:59.410 -->
00:05:13.660Yeah, the amount of human judgment that you use, and then using things like trust card, trust markers. That's pretty interesting. Now, the cousin of your business is LNLP, LexisNexis Legal & Professional.
00:05:10.029 -->
00:05:23.785And Greg, you have a similar issue with lawyers and people in the legal field using our services. How do you see? Is there any distinction, or is it researcher that's making the decisions on what they find in very similar to what Jill just described?
00:05:23.264 -->
00:05:23.504the research.
00:05:23.540 -->
00:05:25.865I think we're very similar to Elsevier.
00:05:25.925 -->
00:06:53.404Lawyers need trust. They need to be able to rely on the answers they get, and I think our reputation is something we have to protect, and it's something that we have a very high reputation in the market, which I think is great. I think for me, AI helps make lawyers more efficient, but they also need a trust layer underneath it. We have Shephard's Verify, it's also a trust marker, and we're working quite closely with Elsevier and how we make sure that we have very similar approaches to this trust conversation. Because lawyers, in particular, they want to move fast. They really want to embrace AI. I've been amazed at how quickly the legal market has moved into AI, and as you know, we've got some very big startups that have come and compete with us. Our big differentiator against them is the fact that we do have great AI, and not AI, we have great trust and foundational trust layers, and we have that reputation. For me, the AI innovation, it's happening fast but it's also quite a commodity. So the whole idea of AI harnesses, the competition between the different model providers means we have a lot of powerful tools out there. But so does everybody else. What they don't have is the great grounded content that we've got. They don't have the 200 billion documents. They don't have the 4 million documents a day that we're pulling in and linking to our knowledge graph. I think it's that combination that's going to mean we're going to win in this space.
00:06:53.449 -->
00:06:57.019Yeah, a lot of really important words have come out.
00:06:57.019 -->
00:07:16.985Human intervention, co-opetitors who are both competitors but also collaborators. Sitting in a different corner, Brian, RX is different. You guys are in the events and human face-to-face, interpersonal connection. How does this technology apply to your domain?
00:07:17.740 -->
00:08:12.819Yeah, well, it's a great question. As with everybody else on the call, RX has been using machine learning, big data, those techniques, for over a decade. But generative and extractive AI has actually created some rather interesting opportunities for us, to connect our exhibitors and our visitors in near real time. If you think about the show floor and all of the interactions that are going on between hundreds of exhibitors and tens of thousands of visitors. Our products are utilising this tech to create real-time actions. So you're interested in this at this exact moment. We'll give you other suggestions about other exhibitors or other content you may want to take care of right then, because that's one of the things YS that is interesting about our events. Everybody goes to our event with an idea. I've got to do these three things.
00:08:12.819 -->
00:08:49.105I've got to do this. I got to do this. I got to do this. But we see over and over and over again in both the data and in the research, and the actual actions that you change over time. And that's where this tech really comes in, because we can detect that your thought, your approach is now different. You have a different set of priorities, and we can tell you right then in the moment that these are the three next best actions you can take, and that makes a massive difference for leads, for our exhibitors, and for our visitors to make sure that they're able to get the business that they need. It's been absolutely brilliant for us.
00:08:49.000 -->
00:09:06.571Different application, but again, this goes back to what Vijay mentioned earlier about the deep knowledge of the customers. The needs itself, and then applying the data sets that are so well structured in your databases to match those things.
00:09:06.644 -->
00:09:15.429I'm going to go back to Jill and to Greg and ask this question around the cost of doing AI.
00:09:11.219 -->
00:09:29.370We've been so enamoured by AI for now nearly four years, and we're beginning to hear, even as lay people like us, that this is not unlimited spending spree that can be sustained. Tell us a little bit about the cost of AI.
00:09:29.443 -->
00:09:31.404Either one of you go first.
00:09:31.720 -->
00:11:00.325Yeah, I'll jump in and piggyback on one of the things that I heard Greg talk about, which is a knowledge graph. So because we have the content, we have access to our data. We have the ability to precompute knowledge graphs. So what that means is, once we can look across the data using this technology, we can combine, figure out what terms are related. I'm sure in the legal domain they are different. For us, it's things like taxonomies in medicine, in ontologies, and understanding two words even in a vector space may look different, but there is an actual domain connection between them. We have the ability to pre-build these knowledge graphs, build them into our ability to retrieve the information, and what that means is, when we are retrieving information and we're using AI to answer questions, we are spending minimal tokens. If you're using an LLM, even if you bring your own data, even if you bring your own data, the LLM is going to spin up a lot of tokens to first understand that data, building those knowledge graphs on the fly, as opposed to using a precomputed version, which is what we have in house, that's very centered toward research, or with those domains in mind, with a chemist in mind, with a clinician in mind. And so really having that precompute, brings those costs down, and we can pass those savings on to our customers.
00:11:00.325 -->
00:11:01.495Greg, do you want to add something?
00:11:02.169 -->
00:11:04.659Yeah, completely.
00:11:02.169 -->
00:11:07.720I think maybe it's worth taking a step back and understanding what a token is just briefly.
00:11:07.720 -->
00:12:27.294Token is a portion of a word that a model uses as a mechanism of compute. When a model predicts the next word, it's effectively streaming tokens in and streaming tokens out. Now, when you start to have models reasoning, what they do is they stream those tokens out to themselves and they loop multiple times while they reason. So as a result, the more you want a model to think, the more you want a model to do, the more tokens it uses, and the vendors are charging us in dollars per million tokens, which is starting to get very expensive. Particularly when you see some of the foundation, the frontier models, your Opus 4.8, Opus 5, for example, from Anthropic, and your OpenAI 5.6 Sol for example. You've got these great models that can think really well, but they're costing more and more. That means for the first time ever, software is no longer just build at very big capital costs and pretty much a zero incremental cost for every new user you've got. Now you've got a point for every new user. It actually costs you money because it's tokens to do the AI for them. So Jill's point is so valid because it means we're reducing the amount of tokens we use to give them the same quality of service. On our side, we're doing the similar things with knowledge graphs, but we're also doing similar things with how we choose the right model. Often our competitors will say they've got the best model, and the moment a new model comes out, they'll be using that model.
00:12:24.414 -->
00:12:30.174We've discovered that in a lot of cases you don't need that.
00:12:27.294 -->
00:12:40.659You can use older models, you can use smaller models, you can tune your own models, and therefore you can have a much better environment in which you can choose the right set of models to serve the customer.
00:12:40.659 -->
00:12:48.850Often it's a lot faster and cheaper at the same highest quality, and for us only 1% of our total costs are in token.
00:12:48.879 -->
00:12:56.080It's still not a big part of our business cost, which I think is great because it gives us a great competitive edge in the market.
00:12:56.450 -->
00:12:59.299Wow, I didn't realise that it was that low.
00:12:59.299 -->
00:12:59.750Yeah.
00:12:59.870 -->
00:13:06.409YS, from an internal perspective we need to worry about or think about token economics, not worry about it.
00:13:06.409 -->
00:14:14.195Just take it as a factor into consideration, because you know the term tokenmaxxing came into vogue last year, and I think it's quickly falling out of favour just as quickly as it came into play, because the hyperscalers hadn't accounted for how much cost they would be incurring if every one of their customers, like us, started tokenmaxxing. They've decided to pass on the cost to us, to their customers in the form of usage-based pricing. In a sense, it's not unlike what we saw in the cloud, and Greg kind of alluded to this. We went from a fixed capital expense in our data centers to a variable operating expense. The same thing is happening when it comes to software development, where what used to be kind of a fixed expense, if you take into account a body, could be a variable expense, and so the discipline of FinOps that came into being. Financial operations in terms of cloud costs, that same discipline now needs to be extended to look at token costs, because in the end AI is going to help reduce costs in technology, but the costs are going to shift. It's going to reduce labour costs in technology in some aspects, and the costs are going to shift to token costs. We want to make sure that the cost of the tokens that are consumed is not greater than the equivalent labour cost in that essence.
00:14:15.320 -->
00:14:40.414Well, you know what? You actually anticipated my question and answered it first. Great, thank you so much. But I'm going to stick with you for a second more. You've always been my tutor for technology in the 22 years I've been here. Anytime I have a question about technology, you know you got an email from me. One of the things that I wanted to bring up here for this session is, you know, it's not just the components.
00:14:40.414 -->
00:14:57.004You've been able to help, at least Risk division and as the Chair of the CTO Forum, the whole of the process. All of the connectors and multiple times.
00:14:50.524 -->
00:15:15.080Vijay what did you see four years ago when ChatGPT became the buzzword for everybody in the world that we had to do differently, and now you sit here, you look back. Are you happy with the journey we have taken in these nearly four years since then, to be where we are in each of the four divisions?
00:15:11.509 -->
00:15:17.570Truly, in my opinion, really leading in this transformation.
00:15:19.159 -->
00:16:45.485Yeah, first of all, I'd love to pretend that I was a genius and had the foresight to predict exactly where ChatGPT would take us, you know, four years later. I didn't. At that time, it was an interesting thing, right. It was a novelty. It was only probably about six months later that all of us on this call realised, oh my goodness, this is much more than a curiosity that consumers can play around with. At that point, we started looking at all the different possibilities because it became much more than just chat. As agentic AI came into play, we started looking at, to your point, the transformation of workflows. In other words, don't just take one thing and do it incrementally better. Completely re-engineer the workflow to reconceptualise how a given body of work gets done. If I just take one concrete example, let's look at cybersecurity, which is near and dear to all of our hearts. If you think about how things have worked in the last couple of decades, vendors publish a list of vulnerabilities in their software. They announce patches for those vulnerabilities, and then customers like us schedule those patches on 30 or 60 or 90-day cycles. Those days are gone. We've got to think about vulnerabilities being discovered even before a patch is available, and there's a term for it. What used to be called zero day is minus zero day, because patch. We got to think, we got to use these tools just like hackers are using them to completely rethink the workflow of the software development life cycle, and think about how we move things further ahead so that we create fewer opportunities or fewer vulnerabilities.
00:16:45.440 -->
00:17:20.400It's a very good point you made. Speaking of jumping, Greg, legal and professional, LexisNexis Legal & Professional, was the first among the four siblings to jump on it, and very quickly, very boldly. Tell us what made you feel like you had to do it, and are you happy with the speed at which you're moving? Even yesterday you announced another innovation through the customer lab. Just in short periodic, tell us how you feel about speed.
00:17:20.630 -->
00:18:13.190I think for us it was very clear. First, we were using this before ChatGPT came out. It was inside our search engine, so we're using GPT-2 and earlier models. So we had an idea of what it could do. But when ChatGPT came out, the response from the legal community was so intense. The fact that they could summarise and do initial drafts became very quick, very very clear, very quickly that this is going to be something that's going to change the way lawyers adopt technology. Lawyers traditionally have been very sceptical, blockchain and all the rest before that. That would be yeah, everybody tells us this is going to change our industry, but it never has. But the change that we saw with ChatGPT was instant. What was great is that Mike, our CEO, picked it up immediately, and he pivoted our whole business and said, "We've got to do this." And therefore, we got ahead. As you know, that increased our revenue and everything. Everything flowed from that because we got ahead.
00:18:10.069 -->
00:18:19.490But I think the conviction was we were building, and customers were buying straight away. They were loving what we were doing.
00:18:15.500 -->
00:18:22.069But YS, we've got to move fast because we've got such strong competitors in this market.
00:18:22.069 -->
00:18:48.605We've done it, but others have done it too, and they're using agents to build their own software. For me, it's a case of we've got to get faster and faster, and that's why what you mentioned was our customer innovation lab, something we've launched in New York, and that's about collapsing the difference, the distance between the person who has the problem, the customer and the person who has a solution, our engineers and our data scientists to zero. So they're in the same room, they can learn from each other in real time, and that's going to help us produce product even faster.
00:18:48.750 -->
00:19:00.779I can't wait to see examples of problems being attacked at breakfast, and pilot is up by lunchtime, and then it's ready to go live by dinner time. I think that's amazing.
00:19:00.779 -->
00:19:02.220That's great. That's great.
00:19:02.819 -->
00:19:09.180I think the pilot by lunchtime is a yes. By dinner time, we need to make sure the security rails are in place. It might be slightly longer.
00:19:09.210 -->
00:19:37.694Our CTOs always think about something a lot more sophisticated than I can. The opposite end, Brian. We have to say RX was a little bit late in needing this thing or feeling like we needed it, but boy, you guys are now deep into it. What is it like to watch LNLP, Elsevier and Risk do this, and then say this is what we're going to do because we are different?
00:19:36.000 -->
00:20:29.019We were along on the journey with everyone at the same time. We started moving them into the product space in relatively short order. Coming out of Covid, which obviously was a challenge for our entire industry. When we started seeing it in 23, we started moving into it, into our product specifically. But YS, if you move away from cyber and you move away from product for just a second. If you think about the way any company does business, that's been the interesting part when you start thinking about how tokens work and where your real costs and where real benefits and how you can speed up your business process, whether it be sales or marketing or pick your favourite. We tend to think about things as far as AI assisted or agentic. Most business processes right now and inside the RX are AI assisted.
00:20:25.404 -->
00:21:38.277We're going to make a step better, so we're going to do it faster. We're going to get better call sheets. We're going to get better marketing campaigns, but it's making one step better. When you get to agentic, now you're switching this to much more of an outcome-based. You're starting off with a specific goal that you want to get to. You put the guardrails in to be safe, and then from there you let the systems run with your proper oversight to make sure that what you're getting out is what you want. One classic example of this is if you think about our world of 350 events, literally all over the world. Every event has their own website. All of us as technologists, we all have the same underlying technology that we have in place, which is great. But the look and feel of every website is different because you're in different audiences. To audit that, to check that it has the latest SEO capabilities, AEO capabilities, the right content, the right value props, make sure that's all on there, is incredibly time consuming to do. AI assisted, we could speed it up by about 10%.
00:21:34.281 -->
00:21:49.884When we switched it to agentic, so really change the business process. We saw upwards of 90% speed increase and accuracy go up, again with the proper checks, balances and guardrails.
00:21:45.824 -->
00:22:05.358So that's, although our product side to the beginning of your question is yes, we've embedded it in, in the places that it matters for RX. We're also seeing on the business process side, and that's something that we're sharing across all of RELX and can continue to leverage.
00:22:05.390 -->
00:22:38.704That's a really important point. I'm going to put it on hold over here about business process, and I'm going to just jump to Jill about something you mentioned in the middle of your discourse, which is about unintended consequence and the security that Vijay talked about. Jill, tell us. We are hearing more recently about problems with unintended consequences. These AI being embedded into our processes, right? To me, as a layman, this is scary. Please put me at ease.
00:22:39.920 -->
00:22:48.698Yeah, I mean, specifically within Elsevier, where this starts to raise its ugly head is in our submissions.
00:22:48.779 -->
00:23:03.465We get author submissions for new research, to publish science, and the ability for someone to create what I'll call fake science, fake research, has just exponentially increased.
00:23:03.548 -->
00:23:22.744It's so easy. It's so easy to create fake research. While we have these tools that can create, really harm to the scientific community if we're not careful and if we're not responsible. We can use these same tools though to detect it.
00:23:22.826 -->
00:23:32.425We can use the AI to find the fake science, to find when something's a little fishy.
00:23:27.503 -->
00:24:29.200Images is a great example, where now AI-generated images become tags. I think it's great that the LLM providers have basically all agreed that we need to tag any images that we helped create, just so that the world sees that there was some generative AI input here. But it's not always the case, and we do need to have the ability to find said images is a great example, but also the work from one author and it's in a domain that they've never really researched before. It can raise flags, and so we're using the AI to help really increase our research integrity because it's so important. I think the integrity of the research that we produce is, it's being attacked now more than ever, and I think it's such an important.
00:24:26.900 -->
00:24:33.500You can't let sneak in fake stuff into what is an unspoiled system of trust.
00:24:33.089 -->
00:24:49.019That's right, that's right because we talk about the trust that we create with our answers and out of the content that we have, but that's built on the assumption that the content we have is trustworthy.
00:24:45.059 -->
00:24:52.740And so if that ever starts to decay, we have real problems.
00:24:49.019 -->
00:25:02.295It's such a high priority for us, is to make sure that research integrity is higher than ever now that there are more abilities to create fake science.
00:25:02.329 -->
00:25:03.589Yeah, thank you for mentioning...
00:25:03.000 -->
00:25:04.619Just emphasise one point.
00:25:05.329 -->
00:25:09.920Yeah, Vijay I was going to ask you. There's probably more to say about security too.
00:25:10.710 -->
00:25:32.309Well, actually, I was going to emphasise a point that Jill made about using AI to find or reduce the severity of the issue that you talked about, right. Because we often talk about human in the loop, and there's a temptation to say that the human is going to somehow magically find the issues that the AI fabricated or hallucinated on. But let's take an analogy. I don't know how many expense reports you get YS.
00:25:32.309 -->
00:25:58.140I get quite a few, and I don't go through them line by line for each my direct reports and say yes, I'm the human in the loop verifying them. I sanity check them. I kind of skim over them and I hit approve. There's a term for that. It's called approval fatigue, right. And so if AI gets things 99% of the time, that 1% of the time that it doesn't get right, my approval fatigue is going to set in. I'm going to hit yes, I approve. And that's where Jill's point is really valid, which is you've got to help AI to actually figure out where that 1% is.
00:25:59.319 -->
00:26:24.289Good point. Very good point. I have my own filtering system for things like that. I look at anomalies, anyhow. But can you also talk to us a little bit about whether we should be more concerned about these models doing things on its own that is harming us, our customers, our society? Any one of you, or just one of you, can answer this, and I'm going to move on to the next topic.
00:26:25.039 -->
00:27:48.960I can jump in for a second. Yeah, I was just going to just highlight quickly what Jill was saying because I think in Lexis we actually, in legal, we actually have an opportunity because we can test and check other AI output from some of the model providers and not just the model providers, but the competitors who don't have a legal background, that don't have that trust layer underneath them. That's one of the things that Shephards Verify can do, is actually check that. Because on a lot of cases AI will hallucinate. But back to your earlier question, YS. I do think we're at that point where models are at the point where they're actually good enough to be able to independently hack into systems and find, as Vijay was saying zero day or negative day exploits. Exploits that haven't been found, that are novel. You probably heard in the last couple of weeks there's been a number of those both from OpenAI and Anthropic. Now I'm slightly sceptical. I think both of them, it helps their own brand to show that their models can break out and do this. I think there's an element of that, but I do think these models are very powerful now, and it means we're in a very different world from a cybersecurity perspective. That we have to be far better at how we patch, keep our systems up to date, and also how we dynamically respond to threats, not just statically respond to threats. That's why we've been working very closely with Vijay's team on newer tools, which I think are much more effective at this. Much more rapid response to emerging threats.
00:27:48.920 -->
00:28:20.255Thank you for all those input. I'm going to switch gear again. There's a couple other things I want to touch on before we run out of time, and that's to going back to Brian's point about process change. Internal changes because this has as much implication to everything that we do in our processes. A couple of you just jump in and tell us examples of how you are going about applying the technology power to do faster, cheaper, better every single day.
00:28:21.410 -->
00:28:24.736I'm happy to start.
00:28:21.410 -->
00:28:33.442In Elsevier, two areas that we're really leaning into internally. One is obviously the software development life cycle.
00:28:28.983 -->
00:30:57.704Using the AI to accelerate our ability to create products and features that our customers love, and some of that first part that Greg mentioned about having the ability to test with customers is because of this technology. What's happened in the SDLC, the software development life cycle, was traditionally design, build, test, deploy. We have swapped design and build because the build is so cheap, you can build multiple times. You can test multiple versions, and then you can basically land on a design that works best with the customers at that point. It's fundamentally changed the way we produce software. The other way that we're leaning in is with non-technical functions, so these tools allow. We talked about workflows. We talked about agents. They're really helpful for all areas: finance, legal, marketing was mentioned. These are areas where, what I'm excited about is these are non-technical people who now think code is very useful and cool, and just because they didn't know how to harness it before, and now they can. Now they feel more relevant in the conversation because they can contribute to really building something, and building code that fits their needs. What we've done, and I believe Greg would agree, he's done this in legal as well, is we've created teams. Whole squads, product and software engineers, dedicated to some of our non-technical functions. For the first time, I have technical headcount dedicated to the finance function. It's not really something I've had before, for the legal function, and we have lots of editors in STM, so that we have lots of folks who could benefit from having this kind of technology. Now, non technical are, more non technical teams don't really understand all of the nuance between we talked about security. Like these tools are dangerously acceptable as well. You can create workflows and not realise you've exposed your infrastructure. You created a new endpoint. So we have to take those same process of guardrail security, firming up, making things production ready.
00:30:53.599 -->
00:31:06.554It doesn't matter if we're talking about a product or if we're talking about a workflow that moves a contract from one area of the business to another.
00:31:06.624 -->
00:31:16.535You have to take the same kind of care and keep the humans in the loop there to make sure that we've tested and we've secured the workflows.
00:31:16.480 -->
00:31:40.576Yeah, I couldn't agree more. We call them AI enablement teams. I think you probably use a similar term or some variation thereof. But good developers don't worry about just writing code. They worry about writing maintainable, secure, compliant code. And we operate in a highly regulated industry, so all those things are important. The other aspect is, at the last RELX board meeting YS, one of the board members asked me an interesting question: "Are we at the point?
00:31:40.626 -->
00:31:59.163She asked, are we at the point where AI is so advanced that technical debt will go down to zero? And my response was, I thought about five seconds, and my response was, it might actually cause it to explode if you're not careful. Because unless you tell it to use these libraries and these parameters, and these are your guardrails and this is your paved road.
00:31:59.213 -->
00:32:18.058It's got to do whatever it wants in the in the interest of expedience. That's where the point that Jill made about making sure that we have engineers embedded with those business teams is really important. To say there's a structure to this. It's not just vibe coding where you say something and it magically appears. You actually got to put the structure into this.
00:32:18.000 -->
00:32:59.595This sounds a lot like me having to handle my five year old grandchild. It could actually...whatever I give her to play with. Great point. I'm going to jump to try to go forward from here because we've done well as a group, individually, all four of you in your businesses. I mean, what do you see, at least visibly, not too far in advance, but you know, ahead of us, the next transition that you feel like you need to help your CEO and your leadership team to make through the technology angle that you have the benefit of?
00:33:00.650 -->
00:33:11.450I think the earlier comment and conversation around tokens and sort of the ubiquity of intelligence is going to be the next transition.
00:33:07.069 -->
00:33:45.005We've gone from AI to agentic, and now it's going to come down to how efficient can we make these models? How can we make intelligence really work throughout our business? I think we're going to see open weight models and how we tune those and fine tune those. I think we're going to see having far more financial management of how we direct workflows. I think it's going to be the same as Vijay was talking about earlier, that cloud maturity model, we're going to see a very similar thing in AI. We're going to keep getting better. We're going to keep getting faster. But now it's going to be a lot more about rather than just this whiz bang cool thing on the side, how do we really embed this in our business, and particularly how do we manage this financially.
00:33:47.884 -->
00:33:48.934On the financial...
00:33:49.744 -->
00:33:50.285Go ahead, Jill.
00:33:50.930 -->
00:34:51.139On Greg's point of financial management, one area that we're finding is we would take small bodies of work and outsource it, especially in publishing. Right? We outsource a lot of the ability to format data. There's just things that just weren't worth our investment in time, we would outsource. Because of the technology and because it allows us to do things. I have got developers writing code in languages they've never learned before. They've never coded in that language before. Doesn't matter, right? They, you understand it enough to review the code that the AI has created for you with the direction you've given. There are lots of bodies of work that we can now in house and create. Move on it faster and for a lot more cost efficient versions because of this technology. It's about going faster, but it's also just things we couldn't do before.
00:34:48.139 -->
00:34:57.769Genuinely, things that we just was not in our wheelhouse and areas we weren't going to invest. We can now do it for ourselves.
00:34:57.769 -->
00:34:59.030Brian.
00:34:59.405 -->
00:35:13.175Yeah, I agree with everybody's saying, and one of the beautiful parts about being inside RELX is you get to share the expertise as you go around. Jill's figured this out, so we can figure out how to leverage it. Vijay's figured this out. RX figured this out.
00:35:13.445 -->
00:35:19.534Why do you do this to me all the time, Brian? That was going to be my next question.
00:35:20.110 -->
00:35:46.735Well, in this case, in this case YS, if you look at the future. RX, as with everybody, will always be customer problem based. But we are starting to see something that, if you talk about the business of technology, that we've talked about for literally decades, which is how do you spot problems on equipment? How do you spot problems on data flows? How do you spot problems before they become problems?
00:35:46.735 -->
00:35:57.175Stuff sends out signals before you get it. Interestingly enough, we're starting to see signals and patterns that we can actually call ahead. We know that your laptop has a problem.
00:35:57.175 -->
00:37:05.980We know that a network device is about to go down. We know that we have a cloud issue, and where this gets really interesting YS, is this is a situation where those signals are going to look very similar to those zero days that Vijay talked about early on. If you've got a bad situation occurring, those are the same type of signals you're going to pick up, and now you're talking about taking signals all the way into patterns, and patterns you can immediately execute against. It's an interesting if you look forward, not that far. The hope is that we'll actually be in a place where we can pick up those signals into immediate patterns that are shared across
00:37:05.980 -->
00:37:25.735a time when a culture of our company was total siloed. We didn't even know each other's names. Okay. I mean, 99% of the company did not know the name of the CEO of another division. To today, look at you. You guys actually like each other. You come from each other's business.
00:37:22.735 -->
00:37:37.719You've transferred from the division, and you actually like each other. Your CEOs like each other. How does this work and help in this moment in time.
00:37:37.719 -->
00:37:44.800When I came here 22 years ago, you know that Vijay. We were there. It was not imaginable.
00:37:40.840 -->
00:37:44.800So...
00:37:44.800 -->
00:38:20.335Yeah. First of all, we should give Kumsal the credit of starting the CTO Forum, to be fair. She was the one who initiated it to her credit. But one of the things we did early on was acknowledge that we can't get into you know each other's shots and solve each other's problems at a very detailed, granular level. We set out, some ground rules that said, this is the class of problems we want to collaborate on. It's because otherwise we'll be boiling the ocean and trying to do, trying to bite off too much, if you will. From that, we developed a very healthy understanding of what problems to collaborate on and where we can get maximum leverage.
00:38:20.335 -->
00:38:28.360Something very basic, you know, security. We talked about security a bit. That's clearly something we all need to collaborate on and learn from.
00:38:25.539 -->
00:39:08.605Procurement is another obvious thing. The collective power that we can negotiate with, whether it's an Amazon or Microsoft or OpenAI or Anthropic. Those are the kinds of things that we should be solving together. But then we get into more tangible things on specific projects. As we speak, there's an opportunity where Greg's team is collaborating with parts of the Risk team to see where we can use Protégé within the Data Services business. Those are very specific examples where you can say, "Hey, this looks interesting. You've already built something. Maybe I don't need to build the same thing twice. How about we leverage your expertise and some of the code that you've written to move faster on all working very well for a very
00:39:08.269 -->
00:39:09.204But YS, I'd call out, as being the newest member long time.
00:39:11.226 -->
00:39:14.967of the group, just how easy this group is to become a part of.
00:39:15.027 -->
00:39:36.269You know, how encouraging everybody is. How I'd use the term low ego, which really means, we're all there to help each other. This doesn't ever feel like this is anything more than a forum in which you're supported. Which I would say, especially for technology that's moving so fast, it's a really good place to be, for our business and for I think the
00:39:35.599 -->
00:39:36.994The last question is going to be something where I four of us.
00:39:39.577 -->
00:40:51.119ask you for a word or expression that comes into your head that you want to convey, and it's not going to be a gotcha kind of thing. I'm going to give it to you now, and while you're thinking, I'm going to ask Jill a question, a different question. When you're addressing people who are not in the technology team in your divisions or within RELX, what is the one thing that you would like to convey to them in one word or one expression, to think about in their daily lives as a non-technologist? Okay. While you're thinking, Jill, we have talked so often about the combination of data, content, unique content, unique data, structured, certified, plus domain expertise. Deep knowledge of what the customer wants, plus technology. That's what makes four of you be able to do what you do so well toward our customers. How can you add spice to that formula? Is there anything else to add to that formula going forward?
00:40:51.000 -->
00:41:01.338It's a good question. I think with where we're headed in Elsevier and what we see with these tools and the communities, really demand for use of them in a safe way.
00:41:01.402 -->
00:41:13.065It's around combining. You said it the domain expertise. So, in the future, you look down the road. What does research look like? It has autonomous agents doing research on your behalf.
00:41:13.128 -->
00:41:50.137Just an example of when the researchers, is doing the research in preparation for publication. That's a really long cycle. It can be depending on your area. It can be a very long cycle. What you really want to stay on top of is while you're doing your research, what other research has been published? What else is out there? That can take time away from what you're doing. So why not have agents go do that work, do that research for you on your behalf, bringing back the information. I think that's where we're headed, is around keeping the domain specific needs and having that users.
00:41:50.199 -->
00:42:21.344What does their whole day look like? What is the spectrum of work they do, and tuning our solutions to meet those needs as opposed to being a black box, just chat bot that starts with,"hello how are you and what's your name?" We need to have the profile of the user, what kind of work they're doing and really understand what their outcomes they're looking for. And then you can make it autonomous and then you can really elevate what it is they're trying to do and help their decision making.
00:42:21.329 -->
00:42:23.340It's very interesting.
00:42:21.329 -->
00:42:38.369One of the things I learned in today's session from all of you is that, because we've narrowed what it is we're going to tackle with our customer base, you can go really deep and detailed.
00:42:32.250 -->
00:42:50.894Whereas the broader models or broader businesses that are trying to tackle too many things at the same time, can't have that benefit of precision, trust, expertise and speed.
00:42:46.005 -->
00:42:57.585Okay, back to my question, now you have had enough time to think about it since Jill did a good job extending her answer.
00:42:53.474 -->
00:43:05.474So, starting with Greg, what's the one thing that you would like non-technology people in the company to think about as we go forward?
00:43:05.449 -->
00:43:08.420How do you keep yourself up to date? How do you learn?
00:43:08.750 -->
00:43:09.050Learn.
00:43:09.050 -->
00:43:13.670How do you stay curious? How do you learn? Use a model. Use AI to teach you.
00:43:14.269 -->
00:43:17.150Okay, Brian.
00:43:17.150 -->
00:43:22.699For us, it's magic. We're a face-to-face business. You have to keep the customer solutions up front.
00:43:22.699 -->
00:43:28.849Applying the technology is a secondary step. It's always about the customer solutions and bringing the magic.
00:43:28.849 -->
00:43:32.210Okay, Vijay.
00:43:32.210 -->
00:43:41.764I would say don't fall in love with a particular tool. Fall in love with the problem, and think about from the problem backwards, as opposed to getting wedded to a tool of choice.
00:43:42.574 -->
00:43:43.385Finally, Jill.
00:43:44.885 -->
00:43:48.545The AI is not going to replace people, in any way.
00:43:48.695 -->
00:43:56.614It's not here to replace people, and we're not using it to do that. We're using it to enhance our ability and what we can do as people.
00:43:57.530 -->
00:44:01.010Good. Well, thank you.
00:43:57.530 -->
00:44:28.474You know, if they didn't give me a limitation on this thing, I would have made it a three-hour podcast in four-part series. But it's always great to learn, to remind myself of what I should not forget, and thank you for doing that. I'm sure that the audience will say the same thing. Thank you so much for your time and your contribution in today's podcast. Thank you to the audience for joining us, and we will see you at the next episode.
