A (3:18)

And you know, Joel, that's exactly what came out. And both of the key roundtables that we participated in, it was a little bit of confusion. And how can this some old world concepts like knowledge management and even older IT service management or itsm, how does that make sense in this modern avant garde, fast moving AI world? But it turns out that it does, that we perhaps have fallen into the trap of reinventing the wheel on how it is that we teach these non human workers agentic systems about the world that they live in. So I wanted you to pick up the story there, tell me a little bit about the state of understanding who's involved in knowledge management from your perspective and how do we begin to integrate this into this vibrant, fast paced, high stakes game of developing agentic enabled transformation?

B (4:12)

Well, and let's start with the stages I think, and I'll give a good example, you know, we walked into a customer with 20,000 plus knowledge documents. So this is, I call this stage one. Now how often are each one of those accessed? Or how many duplicates or how many different versions of that? And so stage one to me was is there a way to use large language model or some of our more generative AI to relook at that? And you could rewrite a bunch of knowledge based articles. You can create new articles today using large language models or generative AI. But how do you actually look at the relevance of the article? How are you seeing which ones are tagged correctly or which ones are created? So the first step to me was we threw all those into a graphical database and tried to understand the correlation, the relationship between tickets and those articles themselves. Then we created a generative knowledge management creation tool that rewrote those articles, looked at the duplicates or looked at the different versions and tried to write it in the most complete one. And that really was from a human consumption standpoint, it was for the agent to get something. So that's stage one. Stage two is as we, we learned about it. I had a hypothesis that we are writing all these articles for that human consumption or for that decision tree consumption at some point. The holy grail of all of it is for self healing machines. So you have to write that in a way for human consumption. Sorry, for compute consumption, meaning a script, a language or something to solve the problem. I kind of remembered back in my last 10 years, 10, 15 years. And we talked about, you and I have talked about this lane, but RPA or robotic process automation. And one of the limitations, either it costs more to develop the robotic process or worse, it solved 80% of the problems, but introduced another 20%. And so you didn't have the confidence to really roll that out. And so the next step was, well, if we can start tagging the points that this article solved the problem and now we know the importance, the effectiveness, all of those others, we can translate that article into machine consumption and scripts and really get to. If I got 98% confidence it solves the issue, maybe that's worth me running the script and I can do it much effective, much more effectively and cheaper than I could have five years ago with all the new AI tools. And then the third version of it is the next step. When we think about an agentic world, we're creating a whole bunch of little mini agents. You don't want to give an agent all jobs. You want to create lots of little agents with a very specific rule set and a very specific job. And so then the next translation of it has come. How do we create knowledge base articles for the rules or the binders of the agent? How do we translate kind of a workflow flow process and not just from an IT perspective, but how your business runs? How do we understand what things are done and create? Can we create a digital twin or knowledge from that to then give access to an agentic world and, and start to replace some of those repetitive tasks that people are doing.

A (7:17)

It was really interesting to see the realization among our guests as they really began to see a two way street sort of evolve. Right? You have had knowledge management traditionally seen as experts. Write something, someone runs into trouble in the world, they look up what these experts wrote and they fix their problem or, or they document an additional fix if the problem has evolved and you need to update the knowledge management article. The problem, of course, was that people just wanted to get up and go on with their lives, right? The process of documentation would often just not get done. Or as you've pointed out in the past, just a fraction of their knowledge base would get recorded. This twin idea which observes the actual human being is capturing all of the processes. Tell me a little bit about how that really enhances not only the speed, the fact that you don't have to document it yourself, that the AI is contributing to its own knowledge base, but also to the speed with which it can be then reapplied and how it can evolve to really capture the challenges that organizations need to address as they move forward.

B (8:28)

I think of that in two levels. Again, none of it's super easy. There's no magic, magic, magic thing for this whole part of it. So let's talk about whether in a B2C standpoint or you're an IT person in a service standpoint, me as an expert in whatever specific tower I might have, you know, I'll write that knowledge base article in a way that I understand it, and maybe my third level or fourth level engineer will understand that same facet, but my first level engineer won't. And more importantly, this is the really big nuance there. The user, when they describe their problem, it's a much different description than we as technologists would describe that problem. And so it's hard to reference the knowledge with the answer or the resolution when you have that. And so the beginning step to me was how do we take real time voice translation and describe in the article and how we organize and tag and create that article in a way that the user is describing their problem and via text or chat or via the voice itself. So that's stage one and that's a really important one. That's often the nuance that is missing in this whole element. What you were talking about when you talk about digital twin, that really comes from a big part of our background. You know, we've been in business for 153 years now, where we started life at, with, with typewriters and adding machines and creating that. But a big part of our Internet 80s and 90s was really about mainframes. And we still have a very large portion of our business where we do financial transactions and scheduling and reservation systems and energy control that are a mainframe operating system. We no longer make them. But how do we take those people which often that is developed in cobol, for example. So there's not a lot of people learning today on how to do cobol, but that is one of the most secure applications. It's not hacked. It stays up and running for six seven nines of uptime. But how do we take the knowledge that came out of those COBOL engineers, apply it in this knowledge base with the concept of a digital twin that's seeing what they do every day and then we can train new engineers or now we can use AI tools to rewrite that COBOL code in a different way. Or maybe you don't even have to rewrite it because you can just do an interface to it in a different way and use that really valuable data that sits in these mainframe applications. And I think that can be applied from an industrial standpoint. Right? How is it your person that's doing the maintenance in a mill or a shop or something, what are they doing every day? If you ask them, they're going to give you 30% of what they do. But if you're watching them, what you're going to learn all the nuances, the specialties, the little things that they do that make them really good at their job. We can start capturing that and find those redundant, repetitive tasks.

A (11:16)

Right. And I think that's the other, you know, we're talking about capturing legacy knowledge. And speaking of legacy, I was a rookie reporter when Sperry and Burroughs came together to form. I remember that very well. But speaking of legacy. Right. You know, it's this idea that because knowledge management has always been about capturing the problems that we're dealing with today, but the idea to go back and leverage the capabilities of AI and update that information, as you just described with Coordinates cobol, it really begins to, I think inform how it is that we can execute complex strategic initiatives. And I'm thinking about sort of digital transformation or business transformation. Notoriously low success rates right up there with mergers and acquisitions, very difficult to do. And it's typically that legacy stuff that can hang in the balance. How should we begin to sort of bring this new technology together with this knowledge management discipline? How are you seeing that strategically affect or does it have the potential to strategically advance the mission critical objectives of CEOs that are trying to get some value out of these emerging technology investments?

B (12:27)

I think let's use the example of if you've ever tried to implement a new ERP system in an enterprise, that's another one, right? I mean, one of the hardest things to possibly do because it's not a technology answer, you have to understand the business process and what I mean by that is not necessarily even financial terms, but what's happening for whatever it is that you're producing and how does that go through the cycles? What are the interaction points? And if you don't get that right, and you don't, you people either pivot and say, this is exactly how I want to do it. And so they make this very bloated ERP system that's never worked, has too many asks, or they revamped their entire process and how they go through them. Both of them are very difficult, both of them have potentially negative outcomes. But if you, if you take a process engineer and you start digitizing, creating the knowledge management of what they do in each step along every facet of how you build or produce, whatever you're selling to the market now, you can start capturing it. And frankly, you know, that's the big advantage of AI. Now you have this massive amount of data set that five years ago would still took a human to go through and try to figure this out. But now we can throw that into an AI algorithm or a machine learning algorithm to say, is there a better way? Is there an efficiency? You can start asking these questions. At the core, it is that knowledge management that starts that whole process.

A (13:51)

It just makes, and this was a point that was made by one of our participants, roundtable participants. It really does make you start understanding that the raw AI, that's when we think of these large language models or even these small language models that get deployed. The missing link that in many cases is preventing these AI initiatives from getting out of proof of concept hell is the internal processes. So this sort of enterprise specific knowledge that is the missing link in making AI deliver, that's the hole that knowledge management can plug if we begin to think of it in a modern way. And I guess my question is, who's got their eye on that ball? Do we need to sort of change who's running knowledge management? Do we need to rethink the mission of knowledge management and therefore the personnel and the processes? How do we begin to really execute from an operational standpoint, an effective knowledge management standpoint strategy with agentic, the agentic economy in mind?

B (14:56)

I would like to challenge this with a concept we talk about this week actually in the sense of there's a difference from a large language model that's sitting in a chat, GDP or something that you will go out to, and a difference from that knowledge sitting within a organization specifically. And many times you don't want to marry those two, right? Those are sometimes the ingredients to your Coca Cola or whatever it might be. I'll use a very specific example. So today I was creating in Microsoft Copilot Studio, creating a connector link through Salesforce, our CRM system to Copilot. Now you could go out to GROK or chat GDP or pick your flavor and say how to do it. But there's some uniqueness. Whether it was ours happened to be multi factor authentication and I needed API access and a whole bunch of other components. There's some uniqueness to my organization that many enterprises have in different flavors to make that enabled. And this is a perfect example of where you can create knowledge and your own knowledge management and keep it internalized that allows certain things to happen along these lines. I'm sure in my organization you probably wouldn't find people in my title doing these things, but in my organization I'm not the first one to come up with this problem. But if you have a knowledge management system that's creating that, this problem would have been identified. I would have been able to search using my own tool internally and say here are the steps for unisys to do that. And I could have kept that internal within the boundaries of the sovereign AI or within the boundaries of my own firewalls or whatever it might be. And it would have been quicker for me to finish that task there today. So we're still even learning this ourselves within unisys, which I think is pretty powerful. We have a lot to go.

A (16:45)

Yeah. One of the most interesting comments that made that were made were actually from someone on the security and the risk management part of the organization. I wanted to, you know, in the few minutes that we have left, really talk about the governance piece of this. Right. How do we make sure that we have the right people accessing the right information? And if knowledge management turns out to be the linchpin that drives success in enterprise AI, how do we make sure that we protect it from potential bad actors? So help me understand what UNISYS is thinking in terms of the governance wrap that we want to put around there to keep it focused and secure and safe.

B (17:21)

Yeah, I think about you have to treat this mixed bag of governance, this mixed bag of security. If you are so one sided on the governance side or so one sided security, you won't start with the AI implementation because you'll wait till your data is clean, you'll wait till you have all these components. From a government standpoint, the governance of the data is a very crucial step. And so you can set that parameters. You still have to think about the old Document control systems and where you point to these repositories of data and then even things like location, right? You want travel policy for the US versus maybe the uk. You want specific financial data or data that can't cross the ocean in the sense of where you're pulling the data from different countries based on those rules that come out of it. So it is a really important step. But the opposite side of it is, I think this is one of those items. Whereas people start to put a lot more of this AI where their data is, which typically is on prem of the type of data that we're talking about because it's revenue data or the secrets to your thing. We have to have two models, one that will search internally, one that will have the right permissions, one that might be all of Office365 and the copilot examples. And one might be, there are cases where you want business insights from a lot of other companies. And so you want to search the big databases, the chat GDPs of the world or others that come into it. And so at least having a decent framework of a maturity level, a decent framework of talking about, you know, what are the rules of the road, who can do what is really important. In my specific example where I was giving that I could have set up a security account that had admin permissions to all of CRM. That's dangerous. You don't want me to do that, right? You probably want to limit it to what only I have and make sure that connector is set up to only the person logging in. And so you really have to think about the usage and the data points that you're pointing the AI to to go capture those things.

A (19:25)

I mean, again, you're taking me back to old school, right? We're talking about authentication, authorization, role based access control. But in this case, I think the new wrinkle is that you're not just applying that to humans, you're applying that to the agents that are going to avail themselves of this knowledge management, this new application, this new way of using knowledge management. Am I on the right track with that?

B (19:48)

Absolutely are. And it's kind of like we've heard you can watch a movie about AI, you can watch Matrix or Terminator and you can say if, if, you know, if you just say go fix it, then AI will go figure out some workaround and some of it won't be the right answer. But I think that's the power of what we're talking about in agentic AI. If you set up many agents with A small rule, your only places you can go look at the local travel policy, your only thing that you can do and you have other agents checking the other components. That is the right answer. And I think that's the way the world's moving to. And that kind of puts a little bit more safety and control or ethical AI impact to it.

A (20:30)

So unbelievably we're at time, but I wanted to give you an opportunity to sort of share what is your message to folks that are looking to secure strategic advantage by deploying all of the promise of AI to get out of this practice, this experimentation phase of where we are, into this operational phase. Sort of help me connect those dots as we part, as we wrap this interview up and the role that knowledge management can play in that.

B (20:59)

We asked ourselves that very question, Lane. And one of the things that we've put together is kind of a rapid value assessment. Just assessing where your state of knowledge is is really important. And so we, we've got that down to now. We can do it in about a week's time to look at your. And you can point it just to your ITSM data or something along those. But what are, what's your data look like, how relative, how use it, how usable is it? Throw it into a graphical database, see the context, see the relationships and see what the next step is. I think that's a good start. Rather than waiting for you to clean up data and so how bad is it or how good is it? And then kind of look at how then you curate over time with that natural language lookup with that real time voice communication and how my users are describing or whoever the endpoint is describing the problem. And I think you can do that in an AI world much differently you could have six months ago. And so taking those first steps, finding the first examples, we all are searching for an ROI AI tool. We all are searching for real examples that leave the proof of concept stage right now. And my concept always been find the low hanging fruit, find the easier ones to prove this out because that'll generate the next evolution of ideas and thought processes about how we can accelerate in the AI world.

A (22:21)

Outstanding. Joel, thank you so much for taking the time to chat. I really appreciate the insight and the perspective as we begin to bring in really a modernized version of knowledge management as we move into this agentic economy. Thank you so very much for taking the time to chat.

B (22:37)

Thank you, Lane.