Challenges of Operationalizing AI in 2025 with Mike Thomson

Hey guys, thanks for the throw. We're here with Michael Thompson, who is currently president and COO of Unisys, and we'll soon be the CEO starting in April. And we're gonna have a little chat about, well, what technologies are really the most disruptive for enterprises specifically in 2025.

Michael, welcome to the show. Thank you, Mike, for, uh, having me. Really a pleasure to speak with you today and look forward to the conversation.

I think we've got some great topics to, to run through. Alright, well, we always seem to be caught up in one technology hype cycle versus another, and AI appears to be no different, but I feel like people are, or organizations are having a little trouble trying to figure out how to operationalize this latest wave of gen ai. I wonder if we have not found ourselves stuck in some permanent experimentation loop.

I mean, what do you think is gonna happen in 2025? Yeah, look, that, that's the billion dollar question as, as they say, right? Um, I, I think your, your pulse on that is spot on.

Uh, I think there is still, uh, a very heavy discovery stage, especially when you talk about gen ai, right? But, um, in, in, in my opinion, and I think the way we treat it, there are, there are several elements of AI that are not hype, that are real and we've been working with for a decade, and they continue to expand and and extend. And so when I think about it, or we think about it from a strategic perspective, and I think it's indicative of the market as well.

There, there's kind of two vectors, right? The first is the AI that you're using embedded in your organization and in, for us as an IT solution slash services company that extends to how we deliver to clients and the, the things that we're able to do for clients. And it also is embedded in the back office functions, marketing, finance, advertising, et cetera.

I think that piece of the utilization of ai, whether it is the, uh, enhancement of AI operations and delivery of provisioning a cloud environment, uh, and, and, and kind of automating that component, whether you're focused on the data abstraction layer and how you actually can, uh, align these, these, uh, variable data points to get a better outcome, a better experience, predictive analytics, all of those types of things. That's pretty well defined known. And there's some really interesting and and utilization cases for gen AI in that I think about thought leadership, uh, and, and managing, um, frontline tickets and service and all of that.

So that, that's pretty well defined. And, and on its way. The, the thing that you mentioned and I think is really real and, and the billions of dollars that are being spent in this space is about how you commercialize ai.

How do you make it revenue oriented? How do you bring new products to a client that is definitely still in this kind of search and destroy mission, right? That everyone's looking for the ROI on the next big thing.

There are clearly some things out there where, uh, they're already known and they're moving forward. I would say financial services and and healthcare are probably the two industries that are on the forefront of how to use that technology to actually bring, uh, I'll say new and innovative products to the market. And I think the rest of the market is really in the other prong, which is around how you do what you do better and how you engage with the technology to be i'll, I'll say another tool in the toolkit, uh, to really either lower the cost of delivery, hire the, or, or raise the, the element of, um, utilization and or experience to, to the end user.

I think those are the, the two ways we look at that. But at you, you're spot on with this hype cycle in the sense of it's the issue du jour, but, but I think it's starting to come down to it's the next issue du jour and how are we using that and how will it ultimately, you know, reflect itself in, in the market prospectively, To your point, I feel like last year was the year of fomo, right? We had the fear of missing out.

And when I look at it this year though, oddly enough, I think things are starting to split into two categories. One is kind of like, what are the new AI table stakes, right? I gotta have this capability 'cause everybody's gonna have this capability.

And then how do I actually start identifying things that will provide, say, unique differentiated value from my company that's a competitive advantage. So how do you have that conversation? Yeah, again, like spot on, um, uh, question wise that everyone's wrestling with.

Uh, I love the FOMO analogy and, and it's true, right? There was a little bit of this, Hey, everyone's applying ai and if you don't, you're left behind, right? And so you're, you're running forward with it even though you don't know where you're going, but you have to do something right to stay competitive.

I, I think it actually ties into the second part of what you've described, and I think it ties into our lead in here on the technology in general, the application of it in how you deliver, whether you're, how you deliver your own framework, your platform or the differentiation of your platform to your clients is where it's actually being applied. So if I go back to my first, um, you know, commentary on that first vector on utilization of it, you know, if you think about the primary use cases that came out of the chute, um, one would be co-development from a software perspective. One would be, you know, how you're using it in advertising and marketing to create content.

One would be how you're managing your service desk and a and in an omni-channel way that includes the ability for digital agents to, uh, you know, support that those are elements that are real and differentiating in how you bring your solutions to market. No question. But now we're starting to get into, and, and this will continue, uh, by the way, I, I believe it will continue.

The first stage is how do you adopt the technology? Where do you adopt the technology? The second stage is really how it becomes part of your workforce, if you will, right?

So just another element of your workforce, whether that's the toolkit or actually as a digital agent and those types of things. The third is how it ultimately interacts with the unit, right? Instead of you prompting it to do something, how can it interpret what you want or need and ultimately deliver a better outcome without being asked and removing that kind of prompt engineering component of it.

We're not at that stage with this, but we are in our lives and typically how we operate with technology in our lives, we want in our work. So you, you know, simple analogy. You go into your kitchen and you say, Alexa, make me coffee, or turn on the coffee and it works, right?

If you think about a smart, um, conference room and all the data telemetry to understand the physical environment, the tools that you use and the user sentiment, in a perfect world, I'm booking a meeting and it knows it to me, it knows the type of projector I want, it knows the room temperature. I'd like an added nose where I want the shades. I go in, it's tested all the equipment, the zoom call is up and operational.

That is a great user experience, and I didn't have to ask it to do that from a prompting perspective because it knows who I am from a persona, persona point of view. And it, and it's delivering that type of quality. And the preventative measure there is, let's say it does, its pre-check and the projector's not working well, it ships me to a new room and does that same setup and I don't even know what happened and I just go to my new room, right?

So that to me is the experience you want in a business environment without having to pay, you know, additional, right? You, you still are managing your budget, but using technology to provide that level of client experience by using the data telemetry of all the different components that we just mentioned is how it comes to life, right? And, and I think that's where that's really going.

I think at the risk anyway of being overly simplistic, when I talk to some folks, they're trying to figure out where to insert these AI models and LLMs because they're probabilistic and a lot of the business processes are deterministic. They generally need to be done the same way every time. And the last thing an LLM does is the same thing every time the same way.

So, so how do we figure out where these things actually fit in a workflow That that's, that's a, again, a, a a great dynamic that we're actually all encountering right now, right? And we, when I think about, and I'll just, I'll, I'll point it to Unisys just for a second here, and then we could take it to the industry view, but so having a definitive AI practice that can help companies do exactly what you've described, right? And, and an LLM, uh, if you think about data in general, and you think about where data resides, the, the current statistics will tell you 70% of data still resides on prem.

So do you take the data to the AI or the AI to the data? Well, it's pretty costly to take the data to the AI and you know, and, and you've seen this, um, dialogue, especially recently with deep seek and you know, the, the elements of what's going on there from an AI perspective, that's that the heart of that is the question you asked. It's a data construct issue, and it's very similar to what we used to think about as master data management.

And, and when you think about an LLM and the creation of an LLM, that's really about master data management, the dynamic nature of what those data sources are, how you secure those data sources, how you power the ability to do that compute, and then you really think about, well, if you see how, how it's happened in the past and how it may evolve in the future, those LLMs become, you know, s SLMs or small language models and they get pushed to the edge and you do GPU compute at the edge at the server, or you do it at the device level with a small language model that's fit for purpose to be more deterministic as, as you've described, right? As opposed to, so it, so the art is really about establishing the data model, making sure it's dynamic, how it connects to the certain processes that you're really trying to own in on automation, and then cultivating that data model to a small language model via, you know, dynamic tokens where you can interact very specifically and get, you know, great speed, no latency. The answer you want predictive in a smaller subset that can help you be deterministic, right?

I, I, I think that's what everyone's trying to coalesce around. And, and if I just tie it into deep seek, which is, you know, obviously the, the, the new thing that came out in the last couple, you know, week or so in regards to what, what China's doing on their AI modeling, essentially, it's that, right? They're, they're, they're taking this large language model and they're building these, um, you know, partitions if you will, to, you know, minimize the throughput that's needed, and then they need less compute power to get to the same result.

And it's no different than those same premises that we've used in, uh, ma master data management and data oceans and data lakes and, you know, just environments that are smaller, more fit for purpose quicker and at the edge, right? And, and I think we're at the very early stages of how we do that. And, and, and the cautious thing here is obviously prevention of hallucinations, using it in a matter that's practical, using it from a point of view that's, you know, not, not causing the company harm, making sure that data is secured and that you really have rights to all of the component pieces in there.

So it, it's still a security issue. You've got a lot of issues around data sovereignty and where people are that are accessing that data. So there's all these different levers, and it's not just the data, but it needs to start with the data, if that makes sense.

You mentioned agent AI earlier, and I've been scratching my head about this. So I'm gonna have all these AI agents running around, they're gonna be optimized for performing a specific task, but those tasks are part of a larger workflow. So how do I orchestrate all that into some sort of end-to-end cohesive process?

Yeah. That again, like that's a, you know, obviously Mike, you're plugged into all of this, right? 'cause you really are hitting the heart of the challenges that all of our clients are facing and we as IT solutions providers are facing.

And when I mentioned earlier about having, uh, AI be a member of the workforce, that's exactly what I meant by that, right? It's very few processes where AI is going to do the process end to end. You, you know, there, there's this fear that you, you know, at at, at a manual level, AI is gonna take over all the manual jobs.

I don't think that's true at all. I think it's how we develop our manual workforce to work alongside of an AI component. You know, what you've described is your typical RPA or you know, component where you punch out of a process, you do something in an automated way, you come back into the process workflow and you pick up with that process.

And so it, it's no different than any, it goes back to TQM, right? You define your process flow, you find the nodes that are either repetitive in nature or pieces that you can save certain elements of time because you're doing the same things. You punch out and do that and you come back into the process and then you finish that process cycle.

I, I, I think part of the, the fear of adoption in some cases here is there's this big, this big ticket price, right? For putting AI in. And so everyone's chasing the big ticket, ROI on it.

And I think if we take a viewpoint that a whole host of small innovations in a process at the end of its lifecycle begets a ROI output, and we, and we think about it chunking up that way, it's actually a more practical way to adopt the technology. And as the technology evolves and our processes evolve, we'll change our processes. But to do all of that in one fail swoop to say, I need to replace this entire process with this AI to make the ROII, I don't think you can make a case for that.

And the other thing that I was wondering about is, have we thought through the security implications of these AI agents? 'cause I got a feeling that there's a bunch of cyber criminals out there licking their chops going, Hey, you built a what that I can hack into and make to almost anything. It, it again like that is, that is the quandary, right?

We, you hear a lot about ethical ai and that's about how people are using it. But what you don't hear about is the unethical hacking exposure to that, right? So that is the biggest issue with these data models.

And, and you know, the, the, if, if you think about hacking component into your LLM and what it can do to all of the outputs, right? Uh, it it's incredibly, um, uh, painful and, and, and something that we really need to think about, uh, in, in a holistic way. And it's not going to get less.

It's gonna get more, I mean, we, we talked about a little about the utilization and the familiarity with the technology. Let me just give you a brief example then see how these si if you don't protect it, what the output could be. So let, let's say we have facial recognition at a bank, and when I walk in the door, it knows who I am.

So before I even walk up to the teller, they know who I am, what accounts I'm there. And, and it's an automatic experience of, you know, good morning Mr. Thompson, what would you like to do?

And you know, I don't have my card. I'm not like they, they know who I am and they know everything or elements about me. Now, now think about if you were able to get access illegally to that same data and modified the recognition.

So your face became my face when I walked in. And there's no other form of verification. If you don't think that's a playground for nation state actors to spend billions of dollars to get into that on any scale, uh, you're kidding yourself.

They're very sophisticated. They're spending as much to use this technology in a nefarious way than we're spending to use it in an ethical way. And so, so I see that as being more and more problematic.

And I also see it, if and when it happens, it is more detrimental to the company because the, the proliferation of what they have access to in an LLM or in the environments that you and I have been talking about is exponential to just taking data, which is kind of the viewpoint to that. So there's one other topic that people are talking about, and it's this whole quantum computing thing, and I cannot figure out if the hype around that is just a case of AI envy or are there actual use cases for this stuff, and when might that manifest itself? Well, it's not AI envy, uh, look, I, I look at, I look at Quantum, uh, as a Y 2K event where we don't know when the actual trigger dates going to be.

I, I think it is, it is known and understood that at some point, uh, in the near future, and the near could be over the course of the next five years, Liz, if we just bound that, that quantum computers will be able to break the current encryption we have. And the second that that happens, and, and that's why I say it's a Y 2K event, if you're not prepared for that transaction, you're done. Everything we talked about from a security perspective is, is on the table.

It has two elements to it that I think are pretty interesting. There's a software element to that, and there's a hardware element to that. And, you know, so, so part of it is building the, uh, quantum defense mechanism in the proprietary software that many companies have, and that keeps the bad actors out from using the hardware to break that encryption.

The other piece of that is the infrastructure, you know, the VPNs of the world and the encryption embedded in that. And, and you really need to look at your entire estate and pinpoint the areas where encryption is critical. And it's critical to everyone when you talk about, you know, your, your access directory or, or you know, anything like that.

Uh, and the, and the gateway to get to that. So you've gotta pinpoint all of those areas and determine the fixes that have to happen now to prevent that level of forced entry. And you've gotta do that same thing on the software side.

It's coming. The same billions that are being spent in a nefarious way are being spent on that. And there's this whole concept of, you know, steal now decrypt later.

Right? Uh, when you, when you think about that, when that's also real. So, but, so you have to have a protection against that standing where your vulnerabilities are and, and having a, um, a plan to how you get to, uh, prevention and, and you, and you, and you damn well better have that before 2029 or 2028 in some people's cases, um, because once it happens, it's too late and it's too long a lead time to actually fix it.

There you go. As they say, Q day is coming. Hey, the, the Chinese have a proverb.

That's something that says along the lines of, may you live in interesting times. Well, we're here And we'll be here for quite a while, I'm sure. Go.

Hey Mike, thanks for being on the show. Thank you for having me. Pleasure talking to you.

And, uh, love to catch up as as frequent as we can. All right, back to you guys in the studio.