20. AI Helps Your Operations in Enterprise IT – Tech Field Day Podcast

On this episode of the Tech Field Day podcast, we take a look at ai, specifically how AI tools can help your operations inside of enterprise it. The answer is quite simple, but more complicated than you might think. Welcome to the Tech Field Day podcast, where we bring together a group of IT experts to discuss an idea about key topics in the industry.

This podcast features a variety of perspectives from members of the Tech Field Day delegate community, and is often recorded in association with one of our events. Tech Field Day is a part of the future and group, and this podcast is also published on our sister sites Techstrong tv. I'd like to take a moment for our guests to introduce themselves before we jump into the topic or the premise for this episode.

Uh, hi, I'm Kerry Culp. I'm a founding partner at Velas Span, uh, an enterprise mobility and cybersecurity consulting firm. Uh, my name is Keith Parsons.

com. Thank you Tom. Ron Westfall, research director here at the Futurum Group.

I lead our networking practice and always a pleasure to be joining a Tech Field Day podcast. Alright, thank you all for joining us. Let's jump into the premise for this episode.

You've probably seen a lot about AI in the news recently, and you may be confused because AI is everything, but it's also nothing. And when something is everything and nothing, is it really anything at all. The key is to figure out what AI can help you do.

And that's why we brought together some professionals today to talk about this, because it turns out AI has a lot to do with the way that we do our jobs, especially in the operational side of enterprise it. And we need to get a handle on what that is. The premise for this episode is that AI can help with operations.

Now, I'll be the first person to admit that last year, 2023, everyone was talking about putting AI in everything. It all had to be AI enabled. There had to be LLM support for my recipe application.

And it's dumb. Let's, let's be fair. I I don't really need that.

But as things move along, we have found ways to use specific aspects of what AI is capable of doing to make our operational lives easier. So I kind of wanna open up the floor to you gentlemen. What are some ways that you have used AI in enterprise IT operations to make your lives easier, Like probably everyone else?

Uh, as you said about a year ago, we kind of dove headfirst into ai, generative AI as a concept and understanding what it can do for us. Uh, obviously we've had exposure to AI for quite a long time, but generative AI tools really kind of changed the dynamic quite a lot. Um, we quickly realized that, like you said, it's everything and nothing all at the same time.

Uh, but fairly quickly we started to evolve how we engaged with it, and we really learned that it's good at a lot of things that maybe humans aren't as good at. Analytics is an example. So feeding it large data sets and letting it parse those data sets and pull things out of 'em that are meaningful or I, I search for those meaningful items.

I think that's how we pretty quickly figured out where AI can help us. So from that perspective, um, log file analysis, pulling logs, feeding it into a generative AI tool set, giving it some parameters, giving it maybe a, a, a persona that we'd like it to adopt and helping it help us identify trends or anomalies or errors or warnings or something like that in that log file, which could be millions and millions of lines of data that would take us, you know, mere humans hours and hours to parse it does it nearly instantly. It's not perfect.

It doesn't necessarily land on exactly the answer, but it is so much better than starting from zero. It just gets us down the road farther faster and then the brains have to get in involved in and really dig into it. That's gold, Carrie.

'cause uh, what I am seeing is that AI can do just what you're pointing to that is improve the workforce experience, and that includes, uh, making operations more efficient, automation of data gathering, uh, getting data insights, and also quite simply improving business outcomes. And this is the good news part of ai. Yes, there's challenges.

It's still, you know, the data and data out challenge. It's been with us for decades. AI is only good as the data management tools that you have in place, but it's still, we're seeing, I think, tangible improvements in outcomes.

Uh, for example, uh, by using AI powered predictive maintenance models, I've seen that there could be up to a 30% reduction in unplanned downtime and also like, uh, up to a three fourth improvement in service resolution times. And this is data coming, you know, that's, uh, readily available. It's like, uh, that source was field acts, uh, and but I think the bottom line here is that AI is going to deliver more good news and bad news is not going to be this technology that's gonna cause in itself widespread, uh, reductions in workforces is going to actually improve the experience for their workforce.

And it's really the outcome is that whoever can leverage AI the best is going to quite simply have a better workforce experience and also improve, you know, what the business can do. Uh, well I I, I like what you said, Ron. I just wanna add, add a little warning caveat.

One of the things I've seen that AI can do is very quickly solve the easy problems. The ones that that, that an experienced person would look at and go, yeah, I got that. So it starts doing those easy ones, meaning it professionals who are working on a system don't have experience on those little things that, that, that break and that you get really good at fixing all the time.

And it's kind of boring. So it takes that away. And what happens though is it's now AI is fixing so many of the issues that the issues it runs across are really difficult.

And I'm, I'm concerned that practitioners won't have that, that baseline understanding by the time it gets to a really what AI can't do, no one has that experience that that, that all us old folks have that we've, we've figured out over, over the years. So there is a little gap in between there. One of the things I've tried is using AI to do packet analysis in, in wireless, we have to point the finger at someone.

And if it's not the client, it's the infrastructure side. And neither of 'em will admit until you can say, look at this frame, did this thing and yours did the wrong answer. Finding those is really difficult.

It takes a lot of experience to, to get into the packet analysis. Ai, even in packet analysis, can take you like, like Carrie said, 80% of the way there, but that last 20% is even harder than it was finding the other. So I, I, I agree it can help the workflows a lot, but that last little bit is still even more difficult than it was before.

So a thought there, Keith, because this is something I hear a lot from people who are starting out in the industry that, you know, why am I solving this problem over and over again? Why, why can't I automate this? You know, uh, if there's an issue with a service not starting, is there a way for me to basically say, okay, if if I have to wait five minutes to start this service until the system reboots, then why not put a timer on it?

These kinds of simple things. Now obviously you have to understand the underlying infrastructure to know, well, why does this servicing to wait to start or something like that. But I think kind of coming back to it where you said being able to do like log analysis or packet capture analysis, like, like Carrie you were saying as well, um, is giving our people a chance to focus on those hard problems.

Maybe they do need to spend a little bit more time studying so that they understand seeing this log entry followed by this log entry that's correlated with this other log entry means that this is this problem. But at the same time, by being able to correlate those things together and feeding that feedback back into the system, we're effectively making the AI tool smarter by saying, the next time you see this pattern, this grouping of messages, it's indicative of this problem. So we're, we're basically training ourselves to make those problems easier to spot because we're using knowledge that can be gained by filtering out the noise effectively.

Yeah, and I think Keith, you, you touched on a problem that I don't know that I really ever put into words. It's a, it's kind of like a future problem, right? If we're not, if we're not letting the junior engineers fix those simple problems, that would be the, would start to form the foundation of knowledge that they need to dive into the deeper, more complex problems later.

That probably really is a bit of a, you know, down the road, let's say expert shortage, you know, foresight that we might, we might have. Now that said, I, I think, uh, so for us, we are using it a little less to solve problems and a little more to speed up the, an analysis. So it's, it's less to say, tell me what's wrong and more about giving it a, a set of parameters for what we're trying to identify and helping it parse the data or having it parse the data for us.

Another use case that we're, we haven't actually really nailed down just yet or, or gotten it to work quite right just yet is for change management. Because we do, we do an awful lot of large scale enterprise deployments that will often have us making large scale potentially very impactful changes to big fortune 25 kind of companies networks. Well, when we cause a sev one incident that's not great, that doesn't look good, right?

So we're working on kind of building out a, uh, uh, a not a model of its own, but effectively, let's call it just a GPT of its own, where, where we can use it to feed in information about the current state, the change state, the change, the, you know, the, the projected changes and have it help analyze whether or not we're going to introduce a problem. And I think that hits a key point about, you know, what is down the horizon. I think Carrie, uh, that's very, uh, valuable to bear in mind what is going on to make these AI models better, particularly on the training side, but certainly also on the inferencing side.

So we're seeing RAG capabilities join with the vector databases to do that specific, you know, customer aligned, uh, language model training so that the data sets are truly correlated with what the business actually needs and can then be automated and then, you know, make everything else run quite simply smoother. And yes, uh, to Keith's point, it's always going to require that balance between here is the AI enabled capabilities, but you still need a human safeguard or fail safe, okay, the AI engine comes up here is something that is an emergency, something an anomaly, uh, something that needs to be looked at. Do you want to authorize this recommended fix or is there something else we need to do?

The bottom line is AI is inevitable. I mean, we're seeing all the investments flowing in that direction and we just have to figure out how to optimize these AI capabilities, you know, throughout the organization. So I, I wanted, I just wanna interject there one quick thing.

Ron, you, you made a really good point, I think about that. You know, having the human involved, one thing that we did before we even allowed our team or, or anybody inside the organization besides a couple of us that were testing things before we allowed them to use it, we implemented an, uh, generative AI acceptable use policy. And that basically dictates the framework with within, you know, that, that, that our users have to work within to interact with generative AI tools.

And probably the key tenet of that policy is that you, the human are still accountable for the output. So it's not, it's not the ai, it's not the tool, it's not the GPT that you used. You can't point back at that and say, that was chat GPT or it was clawed or it was whatever that made the mistake.

Nope. You, the human are still accountable for the output. You need to review it.

You are the expert. You need to make sure that what is coming out of that and what you're then putting into whatever, whatever format is going outwardly is actually accurate and true. We also set some rules around how you can use it with regard to imagery and things like that.

You can't use it with real, real people's images. You can't use it with anything like that because we're trying to kind of put the guardrails up that keep us, you know, from, from doing silly things that, that we don't want to have happen. Those, those are good things, Carrie, for your business entity, yet I see vendors moving past that.

They're taking the, the human out of the loop, especially for the low tier things. I, I need to change some, uh, channels to fix a problem. So it's, it fixes itself.

It then does a pre and a post, did clients improve after I did this or not? And then it sticks and it goes, and there's no human in that loop. So I I, I don't think we want humans in all of the loops.

When the AI is smart enough to know, and when they train it on their own vendor code, their own vendor data sets, they, they get really accurate at the good stuff, at the easy stuff. So I see that a lot of vendors are moving away from that. The human's not in the loop on the easy pieces.

And that human only comes in when the AI doesn't have data to support. We did this automatically. So I, I don't have any fear of the AI doing automatic things.

Um, when it's the vendor's ai, 'cause they have, they have way better data because they trained it on their own internal thing. Some vendors are even using it to write their own code that because we know how we code and we have all these little things done, we can have it on the fly build something. Um, and an example, Hamina built a tool to do wifi design in AI in mere minutes.

Now, it was not even close to, I wouldn't even say 80%, maybe 50% of the way there, but AI self coded itself. And then they went, well, we'd also like to add heat maps and within a couple minutes it was adding heat maps. So there, there's things that can do that can give us better tools that the human can use, but it can also do solve the easy problem.

That's where I want to focus on yes, easy problems. Let the AI do it. I don't want to have it bug me every time.

You know, a door lock changes, just, just fix it. So who's the, so the question ultimately becomes, uh, so, so with a lot of our customers, you know, these are massive enterprises that have really comprehensive change management processes. They have to go through the, you know, the cab and get approval and they have to have rollback plans and everything else.

And, and I don't think, while those organizations are almost all using generative AI tools at some level, I don't think they've matured enough yet to turn over change management to something fully automated. I Guess it change it, it depends on what we're talking about in change management. I, maybe not, not switching VLANs around, but changing a channel on access point.

We've been having that automated for years now. Yeah, maybe changing power settings, changing whether or not we, uh, I I saw one where the MCS rates were at a certain average and then we would change automatically the data rates that were supported until that dropped and then you changed the data rate back down to get the maximum throughput of your airtime. Uh, I used to do that manually and it took hours on site.

And vendor, vendor AI can do it really, really quick and have that feedback loop that said, we tried it, things got worse, so we put it back. That's exactly what I would do personally over a long period of time. And yet AI has way more data set data to make that decision than I would, and I really don't wanna do those dumb things anymore.

So in, in, in the important things that could cause major failures, I, I see what you're talking about, but there's a lot of pieces in our networks that just need to be tuned. And I think that's an important point, Keith, because when you look at the way that things like LLMs and, and current generation AI work, it's very much focused on correlations, right? I we see this, which means this, or you know, more appropriately for a thing like an LLM, statistically speaking, this word follows this other word in a chain of words.

And so there's a strong correlation that this is a sentence you wanted to type, but one of the reasons why that works so well is that you're getting feedback, like you said, tune this and if it doesn't work, turn it back because obviously that didn't work and then the AI eventually learns, okay, that's a bad correlation. I shouldn't recommend that anymore. But how can we as practitioners avoid those kinds of traps?

Because we can see tons of examples of people just reading through the output of chat GPT for example, and saying, oh, well I'll just go ahead and do what chat GPT said, because obviously it's brilliant and as, as Ron alluded to, hallucinations are still a thing. Whether or not you are using rag, you have to know, wait a minute, configuring that command or doing that thing is kind of like putting Elmer's glue on my pizza, even though the system says I should do it. I know better than that.

How can, how can we train people to effectively avoid those traps? Well, I think part of it is chat, GTP or clot are the big LLM models that are more generic that they've been, they've been trained on some huge, you know, everything that's ever been written on the internet compared to a vendor's own model that's only looking at its own dataset. And the one things I, I'd like to begin up and get Kerry and Ron's feedback, what do you think about digital twins?

Where I, where as a vendor, I know the whole code. I wrote all the firmware, I know where it is. I'm gonna make a digital twin of that and let AI play all day long and then it's not gonna hurt the real world.

And then maybe the result of that has the human back in the loop to make the big changes. I, I like that it can do things really fast. I I just don't want it to do it and break real systems.

Yeah, that's music to my ears. Digital twins, uh, I think we have all firsthand experience with it. GPS maps, you're able to get that real time interactive information and then act on it accordingly.

Ultimately have to use your own experience to enhance that information that's being provided to you by, you know, an AI capable or AI enhanced, uh, engine, uh, or platform. And I think, uh, what's also important here is that I think it's about right sizing the model. Uh, that is absolutely right, Keith.

It has to be according to the specific needs of the customer, of the vendor and so forth. And well ultimately, yes, you know, the, the broad-based L LMS will become smarter reducing hallucinations to the point where, uh, say a year or two out from now, they could even be natively not requiring a rag or vector database capabilities because of all of this intense training that's being accumulated, just getting smarter and smarter. So I think all these will be, uh, factors into answering Tom's question.

Yeah, I think so Keith, the digital twin piece is really critical when with it, sticking with the concept of change management, if we have a digital twin that we can let the AI kind of turn it loose on that and let it do its thing and, and then we can actually start to work up the chain to more and more complex changes because it's not breaking anything. I mean, it might break something in the digital twin, but that's okay because it's just the digital twin. If we can use, if we can let it work its way up the chain there, prove out that it's working and it's not introducing problems, let it go on, let it run that same process on the production network at that point.

I think that's where it really gets to the point where it can be more autonomous. It's still kind of, it needs that human oversight I think, you know, to make sure that for the really big impactful things that nothing's going sideways. But then if we have that digital twin and it doesn't break that, turn it loose on the production network and let it go implement that same thing, Or even multiple digital twins that are iterative and it could, like, like chess playing AI from a generation ago, it goes down a path until it breaks, it goes down another path.

We, we could let it run across a hundred thousand, 10,000 digital twins and run it for a year or two years into the future. And it, it's all about how much compute you have to pull that off. Nvidia Liked that comment And, and I agree that a lot of companies that are kind of focused on the AI aspect of things are leaving off that digital twin capability, whereas companies that are developing digital twin technology are finally starting to embrace AI to kind of augment what they're doing.

In some ways they're using current generation AI for things like querying for network information and stuff like that. But obviously that next step is going to be offering capability to do stress testing or to do, uh, change management and things like that. Do you foresee a time though where we will get to a point where we trust the AI enough to turn it loose to say, okay, if the change checks out here on your digital twin, go ahead and implement it in production.

And this kind of goes hand in hand with what Carrie was saying about this generative AI use policy. Will we ever trust AI enough to say, I'm ready to let you kind of run the show at a certain level of confidence with non-complicated tasks? I I think it's, the definition is the non-complicated task.

The other thing I just wanted to throw out and see, I think this is an advantage for NAS vendors where they own the whole stack. They can run a digital twin from soup to nuts the entire thing. Whereas in the real world who people who are aren't with, with a Nile or a meter or a ramen or something, you have, you have a heterogeneous stack and there's a breaking point as it switches vendors.

So their digital twins are, are way more complex as, as it changes vendors in the stack. Those that hold that own the whole thing have a little advantage that their digital twins can be fairly accurate 'cause they own everything in there. So, so Tom, I I wanna answer your question specifically.

I think there's two parts to your question or two answers to it. Will we ever trust AI enough to let it go? Depends who we is in that statement, right?

Yeah, I think, I think to some degree, yes. That and, and that will progress further and further through the population who the we is is representing. On the other hand, I think there's going to, there is going to be a probably the biggest hurdle or biggest roadblock to just enabling us to do that and to trust AI to do things is going to be more of a legal question.

Who's accountable? Who do I hold responsible when this thing goes sideways? Now is it, so if it's inside the vendor, pick the, you pick the vendor that's running some AI tool set, well then the vendor's responsible, right?

It's, it's their, their AI is their tools, it's their network, they broke it. But what if it is a third party or maybe, uh, uh, not even one tool, maybe it's a, a, a mix of tools that are doing things, who is actually held accountable for it? And I think that's the kind of what I was getting to earlier on with regard to the cha, you know, getting a change approved, getting it through cab, making the changes and, and making sure it, you know, everything works.

We are responsible for that. And if something goes wrong, the fingers get pointed at us. We have to accept responsibility.

We have to, you know, we, we, we get called to the carpet and have to fix it and explain why it went wrong if we just turned it over to a trusted AI and it broke something who's ultimately responsible. And to follow on that, I think, uh, winners, uh, from digital twins and, you know, AI capabilities include the cybersecurity realm, basically, it's essential. Uh, this is the week of Black hat and we solved, like the announcements are cybersecurity is going to require AI capabilities just to be able to, you know, make the good guys help fight that battle.

And for example, we saw HPE Aruba networking coming out, uh, with the addition of NDR capabilities and ZTNA enhancements, and that puts a spotlight on network detection and response and zero trust basically are gonna require these digital twin capabilities and then these AI capabilities just to stay on top of the telemetry, particularly when it comes to proliferating iot devices, which are notorious for being, you know, targets, uh, real cybersecurity threats to an organization. And so this is coming together. This is demonstrating why ai, you know, why digital twins, well, certainly cybersecurity, I think brings that to the forefront.

Yeah, I'll, I'll second that. I mean, we, we look at it from our cybersecurity practice viewpoint. One of the biggest challenges that we see in the industry is that the r in all of the, you, you say whatever the, the EDR, the NDR, the, the, you know, MDR, the XDR, the r is lacking in all of them.

The response is truly lacking across the board. And if we can start to leverage AI first to help on the detection side, so on, on the, the detect and identify the threat or the breach or whatever it might be, but to the point of then trusting the a AI to participate in the response, theoretically, the response starts to, to get better and, and more, uh, let's say more responsive. Well, as you can hear from the discussion, there is a lot of potential with generative AI tools being used to support operations in an enterprise environment.

The key, of course, is that you use AI like any other tool that you would use. You need to have a firm handle on how it operates. You need to find the best possible role for that tool.

And you need to create guardrails policies in place to ensure that if something doesn't work, then you are not left with a giant blast radius. And as the tool proves itself more and more capable of doing things, you can widen its scope, you can add additional, uh, capabilities that you are adding to the policy to make sure that your employees are using it to its most effective capabilities. Maybe one day down the road, AI will take care of all the easy stuff and let us focus on the hard things.

Or maybe you find that AI isn't a good fit for your role, whatever it is, you're gonna have to do the groundwork to make sure that you're using it effectively. Don't believe the hype, believe what it can do for you. Thank you for listening to this episode of the Tech Field Day podcast.

If you enjoyed this discussion, please subscribe to our YouTube channel or subscribe to the podcast in your favorite podcast application so you don't miss an episode. Make sure to leave us a rating and a review. This podcast is brought to you by Tech Field Day, the home of IT experts from across the enterprise, which is a part of the Futureum Group.

com/podcast or check out some of our episodes on Techstrong tv. Thanks for listening and we'll see you next week.