From Firefighting to Fire Prevention in IT Operations
Episode Summary
Airlines have been grounded and banks have gone dark for days, in a world that already has AI everywhere. Karthik SJ, General Manager of AI at LogicMonitor, argues the answer is not to slow innovation down but to build systems that survive the pace of change. He explains why root cause analysis has become process of elimination across ten tools, what it means to run investigations with agents in parallel instead of hiring more people, and why he would rather show a confidence score than a confident answer. His framing for the whole shift: firefighting to fire prevention.
Key takeaways
The surface area of the stack grew faster than the ability to diagnose it. Finding root cause is now process of elimination across network, application, third parties, and recent changes, and the answer is agents investigating in parallel rather than more people looking at ten tools
Pick a narrow scope and do not overpromise. The destination is self-healing infrastructure that fixes itself before anyone wakes up, but getting there means guardrails, high precision, and high trust, because being wrong can cost millions in outage
The reframe is firefighting to fire prevention. Fire is already too late, so what you want is smoke signals, and the reason teams ignore smoke is that they are busy with visible fires
Hard ROI is alert reduction: 100,000 alerts a day cut to 1,000 actionable ones. Soft ROI is real but harder to take to a CFO
Generative AI is not always the right tool. Clustering alerts with unsupervised models delivers 80 to 90 percent noise reduction out of the gate, and an LLM would cost ten times the tokens for the same result
Trust comes from showing your work. LogicMonitor exposes an AI workbench so customers can see the models and attributes, and attaches confidence scores to root cause findings, hiding anything below medium
About Karthik SJ
Karthik SJ is General Manager of AI at LogicMonitor, where he leads the company’s AI-powered observability work including its Edwin AI product. He has been deploying machine learning in enterprise software since well before the current wave, at SAP and then at a startup pre-ChatGPT, and describes himself as a builder at heart. He started as an engineer, moved into product management, and now runs the AI business, and he writes for the Forbes Technology Council.
In this episode
| 00:42 | Welcome and guest introduction |
| 02:21 | Airlines grounded, banks down, and why resilience is the answer |
| 03:57 | Speed and complexity as two separate problems |
| 04:15 | How the complexity accumulated |
| 05:01 | Access to AI is not the advantage. What you do with it is |
| 05:51 | Root cause as process of elimination |
| 06:52 | Agents investigating in parallel |
| 08:02 | Narrow scope, guardrails, and the road to self-healing |
| 10:03 | Observing beyond your own infrastructure |
| 11:49 | Firefighting to fire prevention |
| 12:44 | Why teams ignore smoke |
| 13:56 | Proving it against an outage that already happened |
| 15:21 | How customers get started |
| 17:32 | Who buys this, and the managed service provider fit |
| 19:00 | The invisible layer |
| 19:33 | Hard ROI and soft ROI |
| 21:15 | Time to resolve, time to recover, and productivity |
| 22:00 | Asking questions of the data in plain language |
| 23:19 | Where predictive AI beats generative, and by 10x on cost |
| 23:59 | Knowledge graphs |
| 25:07 | The AI workbench, and confidence scores |
| 26:33 | Where downtime is most expensive |
| 27:36 | Twenty twenty-six is the time to place your bets |
| 28:25 | Resources |
| 29:26 | Building a prototype on a flight |
| 30:51 | Leadership: bold vision and relentless execution |
| 31:43 | Wrap-up |
In Karthik’s words
“It’s not access to AI anymore. Everybody will have access to AI, but it’s what you do with AI that’s gonna count.”
— Karthik SJ (05:01)
“I call this from firefighting to fire prevention. What you’re looking for is not fire. The fire is already too late.”
— Karthik SJ (11:49)
“Do you need this Ferrari for the use case? Probably not.”
— Karthik SJ (23:19)
“If I give you a root cause, I need to tell you how confident I am. Anything below medium we don’t even show you.”
— Karthik SJ (25:07)
“The only thing stopping you today is your own limit and lack of imagination.”
— Karthik SJ (30:51)
Resources
Karthik SJ and LogicMonitor
• Karthik SJ on LinkedIn: linkedin.com/in/karthiksj
• LogicMonitor: logicmonitor.com. He points listeners here for the company’s observability resources
• Edwin AI: logicmonitor.com. LogicMonitor’s AI product, named on air as the tool customers test against a past outage
• Karthik SJ at the Forbes Technology Council: forbes.com. Where he publishes
Tools he uses personally
• Replit: replit.com. He built a working prototype on a flight with it after a customer advisory board gave him feedback
• ChatGPT: chatgpt.com. Used alongside Replit for that same prototype
Also referenced
• AWS re:Invent and Gartner: Two conferences he attended back to back before recording, and the context for his place your bets advice
• SAP: sap.com. Where he worked on what was then called machine learning
Related AI Realized episodes and events
• AI Governance as Code: From PDF Policies to Pipelines: Ken Johnston and Bob Rapp on making governance executable inside the deployment pipeline.
• Connecting AI Agents to Live Enterprise Data: Deepti Srivastava of Snow Leopard on the gap between AI agents and dependable structured data.
• Analytics as Code: Why AI Stops Guessing With Data: Chris Parmer of Plotly on verifiability, and why code changes the hallucination question.
Frequently Asked Questions
-
Because the surface area grew faster than the tooling. Karthik SJ describes the stack as having so many layers that when something breaks, whether from a manual change, a system, a partner, or a third party, finding the cause becomes process of elimination: is it the network, is it the application, was it a change someone pushed. His answer is not to hire more people to watch ten different tools, but to let agents investigate in parallel across all of them.
-
It is Karthik SJ’s framing for the shift observability has to make. By the time there is fire, it is already too late, so what you actually want is to detect smoke in the IT environment. His observation about why this fails today is practical rather than technical: teams ignore smoke because they have visible fires burning, and the fire is more obvious. Agents make investigating smoke cheap enough to be worth doing.
-
Narrow scope, and no overpromising. Karthik SJ is candid that the vision is aggressive, self-healing infrastructure that detects, investigates, and fixes before anyone wakes up, but says getting there is a journey. The near-term focus is guardrails, high precision, and high confidence, because getting it wrong can mean millions of dollars of outage, and nobody can afford an uncertain answer at that stakes level.
-
Test it against an outage that already happened. Karthik SJ describes customers arriving with a specific incident from two weeks earlier and running LogicMonitor’s Edwin AI against it to see whether it would have caught the problem. That produces a concrete counterfactual, how many hours would have been saved, rather than a projection, and it builds trust before any commitment.
-
Alert reduction, expressed in numbers a CFO will accept. Karthik SJ’s example is a CIO’s team facing 100,000 alerts a day, reduced to about 1,000 actionable ones with the rest classed as noise. That is less time correlating and triaging, and it converts into time to resolve and time to recover. He distinguishes this from soft ROI, which he says is real but harder to defend in a budget conversation.
-
When the job is clustering rather than reasoning. Karthik SJ says the industry over-indexes on generative AI and warns about cost of goods sold: he asks whether the use case needs a Ferrari, and usually it does not. LogicMonitor clusters alerts with unsupervised models and gets 80 to 90 percent noise reduction out of the gate. An LLM could do it, but at roughly ten times the tokens for no additional return.
-
By showing the working and being honest about uncertainty. On the predictive side LogicMonitor exposes an AI workbench where customers can see which models are used, which attributes matter, and how correlation is done, and many build their own models using domain knowledge rather than data science degrees. On the generative side, every root cause finding carries a confidence score, and anything below medium is not shown at all.
-
Place your bets, and learn through deployment. Karthik SJ, fresh from two industry conferences, argues that reading material and talking to peers has diminishing returns, and that scoping a deployment correctly teaches more than further evaluation. His own example is being stuck on a flight with customer feedback and no access to a designer or product manager, and building a working prototype in a few hours with Replit and ChatGPT.
-
[00:42] Christina Ellwood: Welcome to AI Realized, the podcast for enterprise executives leading AI deployments. From tackling security, data, and operational challenges to navigating organizational transformation, AI deployment offers a unique opportunity to redesign our organizations from the inside out. I'm Christina Elwood, your host for today's episode, and today we are talking with Karthik SJ, the GM of LogicMonitor. In this episode, we're gonna talk about observability. IT leaders are being asked to do so much more today in running these AI operations in parallel with their traditional IT operations. They're increasingly complex hybrid systems that are running at the speed and efficiency that the business expects but are challenging with these new infrastructures. Observability is really critical. In fact, it's no longer just about dashboards, it's about turning telemetry into decisions and decisions into actions, and that's why I'm excited to talk with Karthik today. And welcome, Karthik, to the show.
[01:44] Karthik SJ: Thank you, Christina. I'm glad to be here.
[01:46] Christina Ellwood: You bring so much experience to this conversation, your experience at public companies like SAP and your experience in private companies as well. As we dig into the AI-powered observability and what that means in practice, and how enterprise, how enterprises that use AI can move from a reactive monitoring to a proactive operation, why don't you set the table for us about why this is such a challenge, and what is the first step for any enterprise executive to manage their AI deployment responsibly?
[02:21] Karthik SJ: Yeah, thank you for that. And I think we live in s- such a fascinating time where, you know, with all this AI and someone tells me my Zoom connection is running slow or I can't hear you when our outages have happened. We saw that couple of years back where airlines got crippled and had a-- lost like millions of dollars, or banks were, are down for five days in these times and people can't do payments. That's happening with all the AI in the world, and we are still vulnerable to some of these outages, and you notice it and it could bring things to a grinding halt. And so the answer is not to slow down and stop innovation, but the answer is, how do you build systems that are resilient with the pace of change that are there today and what to expect, right? And so I call it like really going from firefighting to fire prevention. And IT operation teams have been designed like twenty years back, thirty years back with this. Everything was meant and for these teams to look at alerts and systems and there's a lot of manual work in the mix. And I'm calling for a time to relook at this because we have normalized what I call three AM pages or two hundred people war rooms and just doing busy work. And we gotta just flip this and say, "You know what? Hold on. Is this the best efficient use of people's time?"
[03:37] Christina Ellwood: These are system issues. Something's happening
[03:39] Karthik SJ: in the network. I can't even find this network. I can't go to the data center. What better way to use agents than, than looking at system data? So that's what I'm really excited about when you look at observability, is it starts at not just about collecting the data, but what insights is it telling you that I call it the category of unknown unknowns.
[03:57] Christina Ellwood: It-- You bring up really two issues. You bring up the speed, how fast things are changing, but also how fast they're operating. And you also bring up the complexity. Is one of those driving the challenge of observability more than the other?
[04:15] Karthik SJ: I would say the complexity has happened gradually over time. Like people kept buying a lot of software and they had specialized departments to do the network and the security and the application, and we just created complexity gradually. And then suddenly with AI, everything became like yesterday. So speed, I would say, is both gradually and suddenly. But yeah, the AI thing has completely flipped the script on speed and expectations because it's no longer okay to ship your product once in six months and tell my customer, "You know what? Wait for six months for the roadmap to come," because we live in a dynamic time where if you don't do it, there's probably some kid out of a Stanford dorm building the next, uh, unicorn startup, right? And so you gotta be-- The cost of innovation or the cost of AI
[05:00] Christina Ellwood: is coming down and
[05:01] Karthik SJ: it's-- and so you gotta, really gotta figure out that it's not access to AI anymore. Everybody will have access to AI, but it's what you do with AI is gonna count. And so that's why speed is paramount in today's times. And we personally went from when I started from a quarterly to a week- weekly to almost now a daily shipment.
[05:19] Christina Ellwood: I see. So th- that reminds me of just CI/CD type infrastructures that with SaaS brought this model to the forefront that we could deploy frequently and we could deploy-- we could back up too. We could undeploy, if you will, if we deployed something that wasn't really meeting expectation or what have you. So how does that look? How is it different with AI than it was, say, with SaaS and, and SaaS models? You've pointed to the network a couple times. Is it related to the deployment architecture, or is it related to the technology?
[05:51] Karthik SJ: No, it's a very good question. And I would say fun-fundamentally, the layers of s-- the stack someone has, the surface area has increased so much, if issues happen today, which they will, because it could be a manual one or could be a system, could be a partner, could be third party. The time people are spending to get to the root cause, yeah, is a lot, right? Because they gotta-- it's like process of elimination when you start. Is it the network? Nope. Is it the application? Nope. Is it-- I mean, it could be a change that somebody pushed and, and that caused the issue. So my point is, if you have the way to scale this investigation, I call it AI investigations, is, is not to hire more people to look at 10 different tools. The way to do it is let agents go investigate in parallel with all your tools and all your infra and all your data, come back and tell you, "Hey, Christina, here's a hypothesis of what I think is happening." Is it gonna be right all the time? Probably not. But if 8 out of 10 times you got it right, I'll take that because that saved me millions of dollars in a potential outage.
[06:52] Christina Ellwood: Also, it saved a bunch of hours of time that could be put in on the 20% that it didn't find, right? The thing about the agents is they're the cause of the problem and the source of the solution, isn't, aren't they? The problem with the agents is they are not a static thing. They aren't like a SaaS platform where for a period of time it is a thing, okay? Until you deploy the next release, you know what you've deployed. All right? With an agent, it's changing all the time. The agent is never the same from the time you deploy it to the, to when something may be going awry. So it's a rapidly changing element. There are many of them. There are gonna be many more of them. Some of them inside our firewalls, some of them from outside our firewall or going outside our firewall. So the troubleshooting problem, I think, is really for agentic systems, is rooted in some ways in the beauty of agents, which is their ability to be context-specific. So y- that helps you when you're troubleshooting, but it also causes the problem. So how do you balance, or not even balance, what's your strategy for going about, I don't know, not just observability of agents, but actually troubleshooting problems related to agentic systems?
[08:02] Karthik SJ: Yeah. I think this is, uh, early days, and I would say in some ways, pick a narrow scope. Try not to overpromise in terms of the expectations. And so of course, our vision is always aggressive. Like, we wanna do self-healing ultimately, where agents can detect and tell, investigate and fix your infra before you even wake up and you gotta know there's an outage. So that-- imagine that kind of scenario, and it's, you're not too far away from that. But- To get there is a journey, and so what we've been really doubling down is on guardrails and making sure that whatever use case we pick, uh, it has high precision and high confidence and high trust because getting this wrong, it could mean millions of dollars in terms of an outage. So you cannot afford to be like, "You know what? I'm not sure if this AI is something that's accurate or not." So the num- our number one thing is, like, guardrails, high accuracy, build the trust. Once you build the trust of the operator, then you get the next use case. That gives you the right to the next use case. And I call this agency. There's a spectrum of agency, and we started with pretty much just predictive AI. It's not even generative because predictive AI, you hundred percent guarantee reliability in what it's gonna predict. Uh, it's using AI nevertheless, but it's still much more predictive. And then you can slowly start to relax that as you gain trust. And so we've been using agents for things like investigations. What is the alternative to that? The alternative to investigations is you having to pour yourself into hundreds of alerts from five different tools, which I don't think anyone wants to do these days. No one wants to look at alerts and start the day at 3:00 a.m., and it's all pretty technical and jargon. And so I think we've picked use cases where it's almost a no-brainer. But the holy grail, like I said, is really automation. Is-- Are people ready to trust AI to fully automate? I mean, I mean, not everybody, and so the question is, it's a journey, so you've got to get them through the journey. Otherwise, you can't just jump to automation today because they don't want to trust it yet.
[10:00] Christina Ellwood: It's not trustworthy yet either, is it?
[10:02] Karthik SJ: Correct.
[10:03] Christina Ellwood: There is that problem too. I'm struck by, by your description of the way in which you think about troubleshooting because it reminds me of some writing that Mahyar Hanafi has done and shared with our community about building AI applications. And he talks about, in the, in his AI framework, he talks about the three O's, orchestration, observability and optimization, and that sounds parallel to what you just described. Do you see those as two sides of the same coin?
[10:35] Karthik SJ: It's absolutely the whole loop, right? If you think about it, first, can I see what's happening in my environment? And now with the surface area in- increasing, we had an outage, I think, earlier this year with a bank, and it had nothing to do with their own observability. It had everything to do with the partner's observability. A main- mainframe system that they used, the backup went down, the power... They had a power failure. Not only the power go out, their backup also went out. So talk about everything going wrong. But the result was five-day outage almost. I think it was almost five-day outage, and you could not send money, payment services were disrupted. But the reason I give you that example is observability is such a key thing is first you gotta know what's your surface area, what are you, and also your external signals. And so we call that internet ISP providers, what's happening there. And so you really gotta increase the surface area of your observability, right? And
[11:30] Christina Ellwood: then- So in this case, the bank being able to observe beyond its own infrastructure and being able to observe the other banks or being able to detect when there is... Okay, that's a detection problem. And so how does that-- H- how would you orchestrate for the bank that was the victim here?
[11:49] Karthik SJ: Great question. I call this from firefighting to fire prevention. So what you're looking for is not fire. The fire is already too late. You know, you need, uh, CPO says, yeah, Smokey the Bear, right? And so you need smoke signals in your IT environment. And there's smoke. What happens today most of the time when you find smoke, people ignore it because I just have, I have fires burning. Should I look at the smoke or should I look at the fire? I'm gonna pay attention to the fire because that's more obvious. But guess what? This smoke right here could end up being a firestorm in two weeks if you don't pay attention. So how do you know? That's where AI comes in, is I can help you with smoke signals, and I can tell you, "Christina, go investigate the smoke. Is this something real or is this just a false alarm?" And it doesn't cost you to go investigate with AI. Like you could pour in-- It's unlimited elasticity. Like you could run as many investigations as you can. It's only a factor of compute and tokens,
[12:44] Christina Ellwood: but you don't have to hire two hundred people to do investigations all the time. So you're trying to keep your systems resilient
[12:50] Karthik SJ: and run the smoke, investigations on smoke. Because you know what? That bang, I bet you if it saw the smoke, but it could not react to it because it was expensive or ignored it, uh, and that ended up being an outage. So we want to get system-- companies to go from like reactive to proactive, look at the smoke, and then even optimize. Optimization is a continuous thing. How do I get my cost in order and use the right infrastructure, right-size it? So I see this as a continuous loop of observability, o-orchestration, and optimization. Yeah, that's, that's a great, that's a great take.
[13:21] Christina Ellwood: Yeah. I think it makes really good sense in both contexts. So going back to your smoke analogy, I agree with you that agents provide a highly elastic and cost-effective way of detecting the smoke and even potentially using predictive analytics to determine what is going to rise to the level of potential fire. That's a great analogy. However, it's not free, so how do you, how do you allocate the resources between firefighting and smoke detection?
[13:56] Karthik SJ: No, and I try to give this example of an insurance company, right? And as we are looking at agent pricing and consumption and usage, so we do offer customers a way to get started and offer them, "Hey, here's how many investigations you can run." Build their trust and proof points that you're looking. So a customer would come to us and say, "You know what? I had this outage two weeks back. Let's give LogicMonitor's Edwin AI tool a run and see if you're able to catch it, and so we can go proactively back and see if you caught it." And so that builds trust, and then they want to run a few of this, and they compare and see, had this outage happened two weeks back, and had I had Edwin, I would've saved five hours of a,
[14:38] Christina Ellwood: a downtime. Gotcha. Yeah, that's a smart way for anyone to, to sell a solution like yours, because of course, we all wait for the horse to be out of the barn before we shut the door. Exactly. Right? That's just human nature that we do, we do it that way. So if-- And I wanna take a moment to just give people a grounding in what is LogicMonitor, 'cause we haven't really talked about that. Um, but from a cost perspective, when you go in with that situation where you're doing that retrospective analysis to build the confidence of the customer that LogicMonitor could have detected this, this issue before it became a fire Is that a fixed price? Is it free? Is it somehow bounded in a way that makes it a no-brainer? What does that look like?
[15:21] Karthik SJ: There are various flavors. We are seeing how customers get to experience the software. We see a lot of companies today in the industry with agent host companies. I've never seen anything like it in my career, this amazing PLG, product-led growth. You could go in and I've seen Cursor do that pretty well. You could sign up and experience the software and get to users. Developers and engineers are bringing that back to the enterprise. And so product-led growth is-- with AI is very supercharged because I've never seen anything like it, where we did have premium and all in the days. But this is something else because you need to get to value realization really quick and cut through the noise. And so if you could experience value in a software, and so our goals are how do we get customers and users to experience value in a highly context situation? And that could be getting a-- and you can get a slice of it that depend- depending on what meets your budget, but it could be a PLG motion where it's for no cost. It could be a paid POV where you want something more bespoke, or it could be a full-on paid pilot. Or sometimes people just, "You know what? I have enough proof points, I'm willing to take the full plunge." But ultimately, what's gonna happen, I think, going forward, is there's gonna be more scrutiny on ROI and proof points because we have reached the point where I think the promise is there. But unless it gets-- it's married with a use case that can show what I call a painkiller category or is it a vitamin, right? So I say at least IT operations is a painkiller category, meaning it has all the elements of like high stakes. These are outages. It has a high volume problem, and it's highly repetitive. Like, there's absolutely no way anybody's gonna say, "I wanna hire 100 people to look at alerts for my tools." No one's gonna do that in twenty twenty-six. And so that's why I say, I tell people is when you're trying to apply AI, focus more on the problem, less on the technology. Because I get caught up with the tech and, and you miss the big picture. But follow the pain, solve the problem, and look at workflows that people are willing to replace and not wai- wanting to do. And you could also take them on the journey. Sometimes it's a change management, Christina. Our champions are so used to doing it a certain way that I have to tell them, "Guys, this is not best use of your time." And it's a change management process, so taking them on the human element is also important.
[17:32] Christina Ellwood: Who's the ideal buyer for a product like yours? Is it an enterprise, a very large, complex enterprise that would have a huge, even global footprint? Or is it someone who is more middle tier company or even small? What is the size and ICP of your target?
[17:48] Karthik SJ: Yeah, we sell to all sorts of customers and both, I would say enterprise customers and also managed service providers, and across different verticals like healthcare and financial services and manufacturing. And if you see customers like, yeah, IHG Hotels, if you have a bad customer experience in terms of things aren't working in your portal, you would definitely have a logic monitor alert there that can help tell you what's going on. Or if you play Topgolf, Topgolf is one of our, also our reference customers, and if a bay is down and you have a bad experience, then that affects their own how they're delivering their value to customers and end users. So we get notified and we can help bring those bays back online. So yeah, these are several industries and it's because, like I said, the problem is pretty horizontal. It appeals to anybody that needs to deliver a good end-user experience, and our mission is to tie the business and end-user experience to the IT infrastructure.
[18:46] Christina Ellwood: I see. So I should think of it as a customer-facing u- use case that if I'm doing customer experience, user experience, I would want y- your tool for monitoring that type of a system, as opposed to, say, an internal kind of-
[19:00] Karthik SJ: It's an invisible layer. You're right. It's an invisible layer. You don't see it, but it's what keeps things up and running.
[19:05] Christina Ellwood: Gotcha. Gotcha. And when, uh, so I always think it's interesting to, when you're trying to do value-based either pricing or just value-based retention of customers, and you're in the kind of business that you're in where you're trying to keep the fire from becoming smoke, how do you represent or how do you show through the product to your customers where you've prevented something so that they can measure the ROI in avoided problems?
[19:33] Karthik SJ: Yeah, great question. I would say that there's two categories of ROI and what we're looking at. There's the hard ROI, where people can measure those cost savings and say, "Hey, you know what? I can go back to the business." And there's a soft ROI, which is also important, but it's really something more maybe not CFO friendly when you talk about AI tools and ROI and productivity. So I would say from a hard ROI, look, the problem we're solving in terms of, like I said, if you have 100,000 alerts, Christina, that a CIO's team has to look at every day, and we reduce those 100,000 alerts to 1,000 alerts that are actionable and everything else is noise, that is less time for your team to, to correlate alerts and triage alerts. Because everybody's, "Oh, is this... Do I pay attention to it? Do I-- Is it a network issue?" So that's number one. The second one is really, we talk about tickets. Everything people start their work today in IT as it's a ticket. And so if those tickets go to an ITSM vendors, pick your favorite, there are lots of ITSM vendors out there. But the work typically starts with a ticket, and we help cut down the volume of tickets teams need to look now. So some customers are realizing savings of 70% fewer tickets to look at day to day, giving them time back. So I call it more avoided IT costs, so you can scale more with the same team, or you can repurpose them to other functions. And then there is these peak ones, peak events, which outages, where if it goes down and you're losing millions of dollars, we can help you get back online really quick. Instead of a five, five-hour outage, it could have been five minutes.
[21:13] Christina Ellwood: So you can measure that also in the time to resolve. You can
[21:15] Karthik SJ: measure that also with time. We call it time to resolve or time to recover. And the last one is productivity. I don't need to look at and stare at dashboards. They all light up like a Christmas tree, and I'm, like, scrambling. I don't know what do I pay attention to it. But instead of looking at 47 dashboards, Christina, I could use agents to investigate real time. I call it vibe debugging or vibe troubleshooting. You have vibe coding, you also need vibe troubleshooting. Get the agents to go investigate real time, and build you the dashboards you want in context. Why would you build-- why would you have your IT team backlog with 50 dashboards, and you don't even look at half of those? I never look at many dashboards these days. I just want real-time data and real-time answers to the problem I'm solving at this point of time.
[22:00] Christina Ellwood: Do you have an interrogation feature? This is one of the things that I think are, is really novel about these analytical systems, is you can ask questions of the data. Do you have that too, and, uh, do you find that to be popular with your users?
[22:12] Karthik SJ: Absolutely. So we have a query language, and in asking it in this query and learning the query language, you could just ask in, you know, plain language, and you know what to ask for. And, "Hey, which of my sites are affected, and which ones are really down?" And you could-- it would go query and come back with an answer. And so y- absolutely. And, and things-- where things are going is- The, what I'm realizing is the experience is less about speeds and feeds and more about give me the whole dashboard. Like, I don't even know what is a good dashboard. If you're, you're running a-- you're a CRO, and you, like, you need these five, six tiles and these trend charts, give me the whole thing. So what I'm s- telling you is, like, it's going from, like, an individual chart and an individual tile to the whole dashboard, and I'll take it. If it's good enough-- Like, I also have this concept of what I call good enough AI. Like, people over-engineer on trying to make the perfect AI. Humans are not perfect. Why should AI be perfect? I'd say if it's good enough AI, take it and run with it because that's saving you a ton of time.
[23:08] Christina Ellwood: I'm very interested in how you're combining predictive analytics with the more good enough AI side of the product. How is that happening in, in your organization, and where do you think that's going?
[23:19] Karthik SJ: Yeah. And we over-index so much on generative AI, and it's good in certain cases. I would also say you've got to keep a eye on the cost and the cogs, and is it really-- do you need this Ferrari for use case? Probably not. And so we've been pragmatic
[23:32] Christina Ellwood: in that sense, and that's really given us an opportunity to
[23:34] Karthik SJ: flex where we need to. But the predictive AI we're using is for clustering, so it looks at a bunch of alerts, and it clusters it. We just use unsupervised algorithms where you don't need training, so it, it works straight off the oven. But the point is,
[23:47] Christina Ellwood: like, yeah, we've been getting eighty, ninety percent noise reduction out the gate. Of course, I could
[23:51] Karthik SJ: use LLM for that, but that will just take 10X more tokens. I don't need to do it, uh, if I'm getting the ROI from this. We use a combination of both where it makes sense.
[23:59] Christina Ellwood: Yeah. That, that, that makes sense to me too. And also, I'm hearing more and more about people using what I think of as logic layers, where they logically connect the things that are related so that the, they're not looking at the LLM or the predictive analytics to do that because that wiring is fixed anyway. And so the ability to map those into some kind of, I don't know what kind of layer you would call it 'cause I'm not an architect, but
[24:24] Karthik SJ: it's a- Call it a knowledge graph. Yeah.
[24:26] Christina Ellwood: Oh, the knowledge graph. Okay. Yeah. So, like, in a knowledge graph, I suppose that's a great example of it. I think there are different strategies that I'm hearing people using. But yes, that makes sense to me. And how transparent do you, do your customers expect you to be in how that decision-making is going on? 'Cause in responsible AI, we also care about trans-trans-parency, and we care about accountability as well. And it-- the more you are being-- you're this type of strategy of looking for the source of the problem and bubbling up only the things that require human attention means we're relying a lot on the system to be making good decisions, and how do we know?
[25:07] Karthik SJ: Yeah, this is so critical to build trust again with your users, and so not to be overconfident. So we've been really focused on, I think for the predictive AI, Christina, we give what's called a AI workbench. So you could see what models we use, what's the attributes, how is it doing correlation. Guess what? Many of our customers build their own models, and that's fine. And like we give them the tools, and they don't need data science degree. They know the domain knowledge, and they can apply that and get the models going, and they get the output they want. On the generative AI side, with AI investigations, if I give you a root cause, and I need to tell you how confident I am. So we give confidence scores of, "Hey, it's medium or high confidence," and anything below medium we don't even show you. And, and then there's this feedback loops where with evals you can say, "I'm a domain expert. I would have written this root cause slightly differently." And so we capture that feedback, and that's something we are obviously investing in, is like, how do we get human feedback, incorporate that into the model and make it better, and make it more as part of the pipeline as opposed to a one-off thing which we are doing outside the product. You wanna make it a part of the product experience. So that's something we are doing, is we're capturing it more in, in part of the pipeline.
[26:16] Christina Ellwood: Gotcha. It sounds like this product is gonna be very valuable and important, especially to organizations for whom customer experience is critical. So I-- what I think of those as companies that are consumer-oriented
[26:33] Karthik SJ: Yeah. It could be consumer enterprises. We have customers that are like healthcare providers where they're manufacturing vials, and if they get that wrong and they're out is they have to literally throw away the vials, and that's super expensive for them. And so I would say yes, consumer enterprises. We also have managed service providers that are being also using this AI as a way to transform their own operations. What kind of work do they do? How do they give a better experience to their customers? How do they be more proactive to their clients? So it's all across the board. That's what makes me really excited. It's not unique to any industry or any category. It's pretty, pretty much horizontal.
[27:12] Christina Ellwood: Yeah. I think the managed service provider is such a natural fit, right? They get it coming and gone. So as we reach the end of our time together today, I'm curious about what you would most like our enterprise executives to take away from the conversation today as they think about their own AI deployments.
[27:36] Karthik SJ: I've been doing this for a while, I would say at SAP. Back then we called it machine learning, and then I was at a startup. This was pre-ChatGPT, so a lot of battle scars deploying AI. And then out here, and I think what's fascinating for me is I'm a builder at heart. I love building, creating value, and obviously transforming customers and users' experience along the way. That's great. And my thing is, I would say, and I was at Vegas for two weeks back to back, re:Invent and Gartner, and I would be like, "This is the time to place your bets." I think you must have read our material, must have spoken to a lot of your peers. You may have a few things on the sidelines. Twenty twenty-six is the time to go all in. Place your bets, and you'll learn more through the deployment with scoping it the right way than, than sitting on the sidelines. And I think there is more benefit than to sit out.
[28:25] Christina Ellwood: So do you have resources that you can recommend to listeners to act on that advice and to learn more about you and your work at LogicMonitor?
[28:35] Karthik SJ: Absolutely. And listen to your podcast and your newsletter, so that's a great source to start with. Yeah, also check out logicmonitor.com. We have some great resources there. And, yeah, no, I think in general, I keep a keen eye on, uh, Alessandro Lenzi's podcast. That's another one I like to understand how practitioners are really using it. And, and the great thing about being in the Valley is, yeah, I could go to a meetup, and people are more than generous to share, which is pretty fascinating. I haven't seen that, like founders and entrepreneurs just sharing what's going on, what's working, not working. And that's really-- that's what makes this so great of a place is like people are really happy to help out and share what's worked for them.
[29:12] Christina Ellwood: Yeah, I really love that too. And specific to your call to executives to get off the sideline, are there any resources that you think are particularly useful to taking the plunge?
[29:26] Karthik SJ: Yeah, I would say in terms of the-- I've been using tools like Replit personally. I mean, I mean, I started as an engineer, and then I became a product manager, and now a GM. Uh, there was a story I'll tell you when I was going to an EAB customer advisory board, and a customer gave me feedback. And normally, I would talk to my designer and my PM and get a prototype going and a spec. Uh, I had none of those options available because I was on a flight, and I couldn't get hold of them, and the only way was for me to use this tool, and I was like, I was forced myself to use Replit and ChatGPT, and I got a working prototype ready in a few hours through internet on the flight, right? And when I landed, that was a much better experience because he gave me feedback, and I was able to go back and prototype it. So I think the point is going from idea to code is really key. The more you are spending time in deliberating it, we're losing cycles. So people want to react to software, and I think idea to code is gonna be the big resource, I would say is the-- And as for leaders and organizations, the sooner you can remove roadblocks from your teams and processes to go from idea to code, that's gonna be game-changing.
[30:35] Christina Ellwood: Gotcha. Okay, good advice. So in the AI era here, the what it takes to lead a team or to be in a leadership role is changing. What is your personal defining edge as a leader in the AI era?
[30:51] Karthik SJ: Yeah, that's a great question. I think it's the only, the only thing stopping you today is your own limit and lack of imagination, right? And so I would say inspiring your t- I like to inspire teams with a bold vision and relentless execution, right? That's really what it is. I think y-you gotta dream big. You gotta think about where things are going because these are not incremental workflow changes. These are transformational workflows. So you gotta have-- Yeah, just because we did something thirty years doesn't mean it has to be the same. It's gonna happen really fast. Otherwise, somebody else will tell you this is a new workflow. So I would say dream big and then focus on execution where you're shipping code because ultimately that's where rubber hits the road. Otherwise, we have this great material, but if it doesn't translate to software or anything the customers can touch and feel where you can learn, it's the combination of the two that's gonna be critical.
[31:43] Christina Ellwood: Thank you, Karthik SJ, the GM of LogicMonitor, for joining me today on AI
[31:51] Karthik SJ: Realized. Pleasure. I really enjoyed those questions. They're super amazing. And yeah, thank you for continuing what you're doing.