S7 #6 AI doesn't fix information problems. It exposes them.

[00:00:00] INTRO: Welcome to The Power of Digital Policy, a show that helps digital marketers, online communications directors, and others throughout the organization balance out risks and opportunities created by using digital channels. Here's your host, Kristina Podnar

[00:00:19] KRISTINA: Hi, everyone, and welcome back to the Power of Digital Policy podcast. Today, we're talking about why concepts like taxonomy, ontology, and information architecture suddenly matter again in the AI era. What many organizations are discovering is that AI does not fix fragmented information environments, and in many cases, it actually amplifies the weaknesses that were already there.

That's why I'm excited to have with us today Ian Davis. He is an amazing taxonomist, ontologist, and knowledge manager. He has been with the Dodds Group for about three years or so, but he's been in the industry much longer, which means that not only does he have the ability to help organizations structure knowledge in a way that makes information more usable, discoverable, governable, but he also has a lot of successes and a lot of failures that he's observed over time. So he's gonna be able to give us a really nice deep dive on the subject. So Ian, thanks for joining us. Welcome.

[00:01:15] IAN: Thank you very much for having me. Very keen to joining the conversation.

[00:01:20] KRISTINA: And you know what? This is such a buzzing area right now. It seems that for years, taxonomy, ontology work was often sort of in the background inside of organizations. What do you think has suddenly really become strategically important again? It feels a little bit like a full circle moment.

[00:01:36] IAN: Yeah. I've kind of been in the industry long enough, 30-odd years, to have been around when, as you say, it was very much in the background very little understood, not especially valued in a lot of businesses. And I think AI, I mean, essentially AI and the ability to create conversational AI and the expectation that business has that they should be using that and also the thinking that it was a bit of a magic bullet. Everybody's doing AI. We should be doing AI. Let's not really think too much about it. Let's just get in there and launch stuff and have conversational search and have generative AI and all the rest of it. And I think that prompted then some pretty high-level views of the output from that and of what that was generating and how employees were working with that. And I think that began to show that things were not all, all as they would have liked them to be within these businesses. AI essentially highlights the problems. It's like just turning on a very, very big blinding light, and all of a sudden you see what's really going on in the business. And I think that's been the big driver in people then revisiting taxonomies and ontologies and thinking more about metadata.

[00:03:02] KRISTINA: You mentioned that, it's like having a big spotlight on some of the problems within the enterprise. What are you seeing emerge as those problems as organizations attempt to scale on top of fragmented information environments?

[00:03:16] IAN: I think it's the classic problems that taxonomists and information professionals have been shouting about since the early '90s and probably before the early '90s. Metadata is messy, fill rates are low, data is duplicated. Often it's hard to understand where data is, where information is, who owns what, whether it's current, whether it's out of date, what it's being used for. I mean, it's been historically, I think a lot of it's been good enough. Businesses have been surviving on the good enough principle, and they can get by with knowledgeable staff, and they can get by with various mappings and various kind of IT workarounds. But now they can't all of a sudden, and they really have to look at it. And I think probably some C-level executives, some maybe CIOs are surprised at the fact that they have so much siloed data, they have so many duplications. They don't really have any guardrails, any context and they're struggling to identify entities, to create relationships between things, and to use their data and their metadata and their information to create knowledge for the business and to create useful outputs for the business and for the customers that they have.

[00:04:42] KRISTINA: You mentioned taxonomies. For listeners who may not live in this world every day, how would you explain the difference between taxonomies, we often hear ontologies, semantic models thrown around. How should they be thinking about those terms?

[00:04:56] IAN: Yeah, I think even within the world of taxonomy and information management, people argue about how to define one or the other. But where it differs from an ontology is within a taxonomy, you know something is a broader term or a broader concept. You know something is a narrower term. You know something is related to something else, but you don't know how it's related. How is it broader? How is it narrower? The relationships in a taxonomy are not explicit, and they're often not even defined. You can have a taxonomy which is augmented, which is enhanced by ontological relationships. So you still keep your hierarchies, you keep your structure, which is very foundational, and then you improve that by relationships which are ontological, so they're explicit, they're clear. You can see the, for example, you could have the name of a member of parliament In a people taxonomy, and you can have them in a hierarchy so that they are a member of a particular parliament, but you can also relate them to other things. And so you can say they are a member of this committee, they are a member of this political party. And the member of becomes the ontological predicate or property which defines the relationship between that person and that political party or that committee. Whereas in a taxonomy, you just have related to, now you can have member of.

[00:06:25] KRISTINA: And how do semantic models fit into taxonomies and ontologies?

[00:06:30] IAN: I think, again, it's a matter of kind of definitions and usage. To me, when you talk about a semantic model, I think about taxonomies and ontologies themselves, and I think about those things stitched into the content of a particular organization to create some kind of a knowledge graph. So if you have some kind of a sem- semantic structure, taxonomies, ontologies, connected to your content because you've essentially tagged your content with that structure, then you have a knowledge graph. And I would think of it as, that's how I would think of that kind of world. But, you know, in your world, in other people's world, maybe they have a different view of it, but that's kind of how I think of it.

[00:07:14] KRISTINA: So thinking about this from a large enterprise perspective, what does good look like, both in terms of, taxonomies and ontologies, but also that semantic modeling?

[00:07:25] IAN: I think good looks like what it needs to look like for your business. I very much think that what you should do is start with what you've got and start with why you've got it and what you want it to become, and not over-engineer things, but also not kind of under-engineer things, and not assume that the latest technology will solve all your problems for you without doing any work. So I think, generally think, generally speaking, in the world of AI, if you're using AI, most businesses will need a few basic things. They will need to have done as much as they can to clean the data that they have, the data that they know is valuable to them, and they need to remove the data that they know is not valuable to them. So that's kind of a whole metadata audit, metadata cleanup process. So then they have that foundational metadata supporting their content. Then on top of that, they have whatever kind of semantic model they need. It might be a taxonomy. It might be a taxonomy enhanced with ontological relationships. If they work within kinda medicine or pharmaceuticals, they might have, like, a huge ontology linked to all the different entities that they need, and that will basically form the knowledge structure, the brain of the company, the brain of the business. And then on top of that, they will be using some kind of AI tool, and that will be using the taxonomy and the ontology and the knowledge graph and inferring and identifying entities. And on, once it's got that knowledge, you're, you're basically putting your brain into the brain of the AI rather than allowing the AI just to make it up as it goes along. You're giving it structure. You're training it in the same as you train a child to grow up in a particular world. You're expecting an AI to exist within your world, and then you can create whatever products that you need to create. You'll be tagging your content. You'll have your conversational retrieval. And of course, the world in which a lot of the people listening to this probably live, you'll be worrying about governance. You'll be worrying about who controls what. Is there a board that reviews your big kind of st- semantic strategic decisions? How can you set up your kind of audit trails so you can go back from your products and trace things back and find out, "Well, how did we get that? Why did we come up with that response? How did that data get into the public domain?" So those to me are kind of the main building blocks that spring to mind.

[00:10:11] KRISTINA: You mentioned it's like having a child in our midst. One of the things about children is that they natively learn about their environment, and they adapt, and they grow over time. And I'm thinking about what we're talking about in terms of semantic modeling and taxonomies, ontologies. How do you balance consistency with reality that language evolves differently across departments and regions, especially in large enterprises which you've been a part of, 'cause it's not a set it and forget it. That child actually grows or needs to grow,

[00:10:40] IAN: yeah, and I think another big thing that people need to get away from is the idea of a project-based mindset. Let's just build a taxonomy, let's build an ontology, let's build compliance. Whatever it is, you don't just build it and walk away. You have to maintain it. It, it's an ever-growing thing, and you have to accept that it's not a project, it's an ongoing initiative. And I think that, yes, you have to kind of manage a few different things, and you certainly have to manage that the semantics remains aligned with the content and remains plugged into the outputs. So that's kind of a bit of the governance world, where you need people who know what they're doing to create the initial structures and to, to some degree make decisions, key decisions about the semantics and semantic modeling and the taxonomy and the ontology. But also you need to delegate some of that, so you need to bring in some subject matter experts, and you need to bring in local teams who are consuming the taxonomy, the ontology, and you need to allow them the flexibility to, for example, to add synonyms into a particular knowledge graph where basically there's one preferred term, but this team has got this synonym for it that goes into their system, and that team's got a different synonym for it that goes into a different system So I think you can have that kind of hub and spoke model where you have something centralized, but it's not overly centralized. In my experience, some of the problems large companies have is they swing over a period of years from being very highly centralized and essentially ignoring their local businesses to then being, going the other way and being hugely localized and ignoring the center. And they never often get the balance right. And it's the same with the semantics. You have to get that balance right. And also, AI can help with this because at scale, you can never really get enough people to maintain all this. One thing we were thinking... I was thinking about a few days ago was if you could get your AI, once you've got all this set up, if you could get your AI to review your content set, to review the data in your business, and to constantly sanity check for you how aligned are our semantic structures to our content, so I think you can do that kind of thing.

[00:13:13] KRISTINA: I think that's actually a really important part because as you were speaking, I could imagine several executives saying, "Oh, this is great. Let's just deploy AI, and it'll take care of everything for us, and Ian can retire to a lovely island somewhere- ... and enjoy retirement." But then you said, "Wait, somebody still has to review." So it's not just automatically, set it, tag it, forget it. There's still a role for humans in this process, it sounds like.

[00:13:37] IAN: Yeah, and I think, people say humans in the loop, and obviously that's important. , And I very much come from a perspective of AI, and I think that's where it should be. The people should not just be in the loop. The people should be at the heart. We should remember that AI is our servant, not the other way around, and I think top executives need to remind themselves of that, and they need to say that AI is just another tool that we use in our business in a sensible way, and it is not in order to replace the business. And if you point AI at something, it will always do its best to give you what you've asked for, although it sometimes won't understand what you ask for, and it will reinterpret that for you. And if it needs to, it'll lie to please you. It will twist the truth to give you a positive outcome, and you can't stop it doing that. The best way to stop it doing that is to control it, is to give it the knowledge graph, to give it the rules, to basically train it on your world and essentially say, "This is the world in which you are going to operate. These are the guardrails. These are the limits. These are the entities, and this is how we think." And then you'll get a better output.

[00:14:56] KRISTINA: So thinking about that at scale, because you've been inside of some really, really large enterprises, it takes a lot of different people getting involved to set those guardrails, to share the knowledge, to steer the technology in the right direction, and we always say technology is easy. What does that look like when we start to scale at a multinational? And how do you actually overcome some of the challenges maybe that folks have, either being entrenched in their own day jobs, they don't want to necessarily talk about knowledge graphs or what you wanna talk about, and also maybe being fearful for their jobs. They don't wanna be replaced by a tool.

[00:15:33] IAN: Yeah, I think it comes down to trust. I think people have to trust the executives who are making these strategic decisions, and the executives have to genuinely be doing it from a good place, and genuinely keeping people at the heart of the business. 'Cause they need to believe that people will always be at the heart of every business. I mean, there's not many businesses out there where you can take every person out and it will still be efficient, effective, and it will still be good. In terms of, like, the multinational nature of it, historically, one of the biggest issues with multinational scaling is cultural and translation. In the 1990s, I created an English thesaurus to catalog digital images. We then had to do that for our German office, for our French office, for our Spanish office. So we had to translate the thesaurus, so I had to get people translating every single term for me, and then we had to map the different language versions together in order to use them on our websites, and that was a very intensive process and very reliant on skilled staff. Luckily, with AI, it's almost kind of like language agnostic now. You know? It can speak pretty much every language. You don't need to worry so much about the language issues. It can even, to some degree, pick up on a lot of the cultural issues and the cultural whys and wherefores if it has the right training. So I think with most things, you wanna take it slowly. As I, I always think, like, proof of concept, start small, sell it to the business, and then if you've got a working proof of concept... And, and if you can get that proof of concept where you've got a bit of everything, you know, you've got the, a bit of governance, a bit of tagging, a bit of retrieval, you're pulling in representative data sets from the whole business so you can see how the AI and the tagging and the knowledge graph and everything work with a bit of everything, and then you can prove that it actually works. Not only does that build confidence at a high level within the business, but that also then enables you to spread the word. And if you can start to get positive stories about how, you know, "The German office did really well using this," "So-and-so in London was able to get an extra, like, ten percent of it, this, that, and the other out of that particular sale," then it will soon become apparent to the wider business that it's very much in their interest. So not only are the top level executives pushing it because they know it works and they know it scales, 'cause they can prove it and they can take it step by step, but also you wanna be involved. Because people love success, and they want to be involved in that success. Although, I think one thing I'm becoming increasingly aware of is the cost of AI. And the cost of embedding it into your organization. And one thing I'm worried about now, which I'm definitely making sure there are dots we don't do, is we just, we just go mad with AI. We put AI everywhere, whether we need it or not. We encourage the staff to use it whether they need it or not. And It racks up a huge number of tokens, which are very, very cheap at the moment, and we embed processes in the business which are highly token-centric, which are essentially consuming huge number of tokens. But then over time, we will get more monopolies coming. We will get increasing costs. People will be putting up the price of tokens. And can you scale that? I think you'll get businesses where all of a sudden they realize that they've embedded AI at a very expensive level deep into their business, and they can't afford to operate it anymore. And I think some businesses might go broke because of the shock of that if prices of tokens rise too quickly. And other businesses will immediately go for a trend of efficient AI, cutting back. How can we get the same output with fewer tokens?

[00:19:44] KRISTINA: That's a good reminder because I was talking to somebody last week in a workshop, and they were saying that their top two queries across the enterprise were, "What is the weather going to be like tomorrow?" And, "Should I have chicken or beef?" And as you're talking about this, what I'm wondering is, is there a role for taxonomies and ontologies to function as practical guardrails for the AI system, not just in terms of getting us to the right answers, but also from preventing the dumb questions? Or maybe not dumb questions- Yeah ... but just things we shouldn't be asking. You should be able to decide if you want chicken or not.

[00:20:20] IAN: Yeah, and I think you can get tokenizer websites now. So you can basically copy and paste a prompt into a tokenizer, and it will tell you how many tokens are in your question. And if you get the prompt response, it will tell you on how many tokens came back. And obviously, within your company, you can monitor that. And I think part of governance is potentially saying that if we go over so many tokens for a particular question, we need to consider whether that actually should be run or whether that should be allowed or whether we have a conversation with the department that is consistently using a high n- number of tokens. And again, as the prices rise, the pressure will increase to take these things more seriously. And while this is cheap, people don't really care. But not really caring now will kind of embed bad habits in the future which will then be harder to eradicate. And on the consumer side, people are gonna start paying for AI. There's gonna be ads. There's gonna be paywalls. But on the enterprise side, obviously the company will always pay, but I think it won't be quite so keen to allow its staff to ask the weather of ChatGPT and Gemini and Perplexity when the executives start seeing the costs of those tokens coming through.

[00:21:42] KRISTINA: Are you seeing executives starting to pay attention to that, or is that still a governance issue that needs to be tackled?

[00:21:48] IAN: I think I'm beginning to see some companies keeping an eye on that. But also I'm seeing a lot of companies just encouraging people to use AI and almost, you know, racking up... basically creating a score. You know, who's used AI the most this week? Which person, which department has got the most out of AI? But it's one of the world's... It's one of the par- parts where taxonomies and ontologies can help the business because if the knowledge graph, which is the ontology and the taxonomy connected to the content, if that has the relationships, if that has the entities, if that has that has the guardrails, that has your rules, so that helps AI when it's coming out with its responses. So it consumes fewer tokens because it knows what it's doing. It doesn't kind of make so many mistakes. It's not so wordy. It doesn't say, "Well, I think it might be this or it might be that," or, "You, you asked me this, I'm not sure what you mean. Can you clarify?" So that knowledge graph, that semantic relationships, the semantic modeling, helps save money and makes AI more efficient and more effective. And I think at the moment we focus more on making AI more effective, and , in, I say in a few years, because I'm, I started work in the '90s, but now I should probably say in a few months, it will be about making AI more efficient and consequently a bit cheaper.

[00:23:19] KRISTINA: Which also means that stewardship becomes central to the conversation. From your perspective, how should leaders think about ownership and stewardship of, certainly taxonomies and ontologies, but also AI as a tool set that relies on those taxonomies and ontologies?

[00:23:37] IAN: I think they need to see all of this as key intellectual property to their business. I think they, they have to stop seeing it as kind of like boring data that the IT department owns or a load of policy documents that's on a SharePoint site that nobody looks at. They can't get away with that anymore. This has to be front and center for them, and they have to value it. They have to value it, really value it, and they have to put money into the right staff and the right systems to maintain it and comply and audit and all the rest of it. They could get away with it in the past. It was kinda good enough. You could ignore these things. But I, I was reading again recently, you know, you can... and again, this is your world, Kristina, much more than it is mine. You can create a policy document which defines a particular compliance area or something like that, but that doesn't mean you're compliant. You know, that just means you've got a document on a SharePoint site.

[00:24:39] KRISTINA: Yeah, you were doing a great job, Ian, I think, of describing shelfware, right? Where we see organizations writing AI policies, putting them out on SharePoint as a PDF. And a lot of times what I see is that there's a lack of visibility into how information actually moves and what you described, which is how organizations behave or how you want them to behave across enterprises in different offices and different locales. I'm curious to hear your reaction on how people should go about addressing what seems to be an overwhelming problem. Because I know that for a lot of the executives that I talk to, they say, "You know what? It sounds really easy. You know, a lot of people just want a policy. The AI board is talking about the policy. I think we've got it covered." But when it comes to information management and really thinking about what it takes to get this right, it becomes overwhelming

[00:25:31] IAN: Yeah, because I suppose you think of the big picture, but again, it, it comes down to trying to create some basic things that you can scale and growing from a small size up to what you need. And again, not overtooling things. Data, information, it's there for a reason. It has to earn its place. If it's not being used, it should be removed or parked or whatever, but it shouldn't just be left to, to kind of sit on a shelf kind of thing. And I think there are people out there who've been banging on about taxonomies and ontologies and semantics and data and metadata for 30-odd years who have kind of got used to not being listened to and got used to sitting in a small room typing away, knowing that nobody knows or really cares about what they do, even though what they do is quietly keeping everything up and running, but it's not valued. And people are beginning to talk to these people now, and they're beginning to actually listen to what these people have got to say. And I think most organizations, you will have people like that. You will have people who know your data, know about taxonomy or ontology, know about kind of semantic modeling. And I think you just need to go and, and talk to them a bit more, give them a bit more notice, and then start to put in place the structures that they've been asking for in many respects for a long time and the investments that they've been asking for for a long time. You know, fix the old taxonomy software that they've wanted to replace in years, deal with the siloed data, see if that relational database can be fixed, do some metadata cleanup projects. You've got to start somewhere. So I think you, you either start small from the top, from the bottom up, looking at your data, or you kind of start in terms of what are our pain points or what new product do we want to produce and what will we need to do that. And kind of however you approach it, you just need to understand that semantics and ontologies and taxonomies and knowledge graphs have a huge amount to offer to a lot of businesses and make AI far more effective and far more efficient than it would ever be without them. But again, this can all grow with the business. You don't have to build a knowledge graph of a million entities. You might only need, like, a thousand people in a particular knowledge graph or a thousand products that are related together ontologically. You kind of build what you need, and then when you need a bit more, you build a bit more. One of the beauties of a knowledge graph and an ontology and a taxonomy is that it's easy to extend it. In, in the world of a relational database, creating tables, relating data together between those tables, you tended to build a table structure for a reason, and then, like, five years later, if the reasons changed, you had to rebuild all of the tables. It's a really clunky way to work, especially now. But with the semantics, it's just relationships. It's just a web of relationships. So I say, like, this MP is a member of this committee. If I suddenly decide, "Well, actually, I want to record what this MP likes to eat," I just add a new property, "likes to eat," and then that's there. All you're doing is stitching things together, and it's far more flexible. So it's a lot easier to adapt your business requirements using this kind of semantic web structure than it ever used to be. Your AI can help you with that because it can analyze what you need, but you just need some staff as well. So keep your employees at the center. Find out who in your organization knows about this stuff anyway and put some small things into practice with that kind of basic building blocks

[00:29:45] KRISTINA: The question that comes to my mind, because I'm hearing the excitement in your voice, I dare say it's optimism. If there is one thing that gives you optimism about where the space is heading, what is it?

[00:29:57] IAN: It's the fact that for 30-odd years, so many people around the world have been working in taxonomy and ontology and metadata and modeling all of this stuff, and they've been plugging away, usually, as I said, in very, very quiet, very, very little rooms with very little natural light, and yet they've still been doing it, and they've still been working for businesses who very often don't really value , their outputs. And now all of a sudden, light's being shone on them, they're being given investment, they're being listened to, And that gives me a lot of optimism for the future. And also, even when monopolies start to form, token prices go up, probably the governments take over the big data centers, businesses probably rely on more open source, smaller AI models, these people will still be there to help with all of those transitions and to maintain the business products and the information. So that's really for me the optimism, the fact that we have people who can do this stuff, as long as other people let them and other people value their staff over their AI models.

[00:31:14] KRISTINA: That probably also means, Ian, that you have to be a little bit more pessimistic about ending up on that island that I mentioned earlier, because you have to retire anytime soon.

[00:31:23] IAN: Well, that's the annoying part, 'cause I've spent 30 years in this. And, , I've been sitting in a dark room for 30 years, and now suddenly it's kind of a booming thing, and yet I am towards the end of my career. So I'm beginning to think, you know, how good would it have been, how good is it for people who are just at the start of their career in the information world, that there is no end to how valuable they are?

[00:31:46] KRISTINA: That's such a great note to leave things on, and hopefully also a nod to anybody who's wondering what career field they should choose next. Ian, thanks so much for joining us today. Definitely a great conversation, and such an important one too as organizations move from experimenting with AI to operationalizing it at scale. So thanks everybody for joining, for listening to The Power of Digital Policy. We'll see you next time. Take care. Thanks, Ian.

[00:32:11] INTRO: Thank you for joining The Power of Digital Policy. To get access to policy checklists, detailed information on policies and other helpful resources, head over to thepowerofdigitalpolicy.com.

If you get a moment, please leave a review on iTunes to help your digital colleagues find out about the podcast