Why Agentic-First Startups Won't Disrupt Enterprises as Fast as Everyone Thinks | Kris Lovejoy

0:00You know, Agentec AI is a hot new thing. Every large enterprise in the world is trying to figure out how to use it. It's really easy to build really cool things. It's really hard to run them at scale securely, compliantly, resiliently, reliably. My prediction is that by about 2031, about half of all of the kind of the traditional IT systems administration tasks, the line one, line two tasks, they will be provided by agentic AI and then humans will be in either in the loop or over the loop, if you will. so chris it's great to finally talk i know that this has been in the scheduling scheduling uh mode for for months now i usually start by having you introduce yourself and give some of your background if it's relevant and what you're doing at kindrel Sounds good.

1:04So my name is Chris Lovejoy. I am the global strategy leader for Kindrel. Kindrel is a company that is fairly new on the scene. We've been around for about four years now. We're actually a spinoff from IBM. So about four years ago, IBM took the organization that did a combination of strategic outsourcing and IT integration projects spun it out and we became a publicly traded independent organization. And so today we are really focused in three, I'd say, major areas. One is in IT modernization programs. So think about that as, you know, modernizing your legacy infrastructure data centers, whatever the case may be, moving those workloads to the cloud.

2:01Second area is in and around security and resiliency, which is kind of my domain or historic domain. And then the third is really specializing in data and AI and bringing, you know, generative, agentic, all forms of AI to, you know, enable customers new ways of working, if you will. yeah and on um exactly on that point um you know agentic ai is is the hot new thing uh i would i think it's safe to say every large enterprise in the world is trying to figure out how to use it Yes. There have been many pilots over the last year, but very few systems have gone into production, at least in a way that's of any scale that's significant.

3:01You know, we're coming out of the age of experimentation and, you know, where we've all been dabbling in using AI and, you know, agentic AI in particular now. And we're seeing the benefits of the capabilities, you know, but not in production. We are finding it very costly. We're finding it somewhat insecure. We're finding it to be unreliable. And most importantly, it's not scalable. Oh, and if you're in Europe, you know, there are sovereign concerns as well. So, you know, there are some things that are preventing us from going into what I'd call the age of industrialization around agentic AI. And so I think, you know, it's going to be a little bit that we're in this transition period and we'll get there eventually.

3:50Yeah. And so what are those? And actually, I'm just going to ask before we get too far into it. When you're talking in your role to enterprises about Agentex systems, what is Kindrel's role? Are you building stuff? Are you acting as a consultant? What's Kindrel doing in those conversations? So quick analogy, if we were to say that, you know, there's a bullet train, bullet train can go 150 miles an hour and it's sitting on tracks that limit it to about 30 to 60 miles per hour. And this is actually a true story about, you know, trains in the eastern seacoast of the U.S. Legacy infrastructure set up in 1930s, you know, keeping us back.

4:48In that context, what do we do as Kendrell? Our job is to build the train so we can work with the customer and build the agentic AI, that workflow. The other thing that we do, and this is where we really specialize, is we will modernize the infrastructure on which that train is going to actually operate. So our real focus is on preparing the IT infrastructure layer, getting the data layer ready, making sure that you've got all the monitoring, the registration, all the stuff that one would need to actually use agentic AI at scale. That's where we play. Okay. And are you seeing any enterprises adopting agentic AI at scale yet?

5:33It is very rare. to see one at scale. I'd say that what I see is vertical use cases. So for instance, there's some use cases like know your customer or within the healthcare space, invoicing, or if you have a sales cash order process that you can automate using Agentic AI, That seems to be successful. But those are vertical use cases, horizontal use cases where there is an integration of AI workflows across the business horizontally. I have yet to see that. Yeah, yeah. And that's the big promise, right, that you take a business process or workload, break it down into steps, decide what a genetic AI could handle and what a human has to handle and presumably speed up the whole process.

6:32Let me ask, are you talking to any enterprises about that more horizontal application? We are, as a matter of fact. It's beginning. And I think, you know, for most, it is interesting as to what they're looking for when you're talking about the horizontals. A lot of our customers we're talking to right now, obviously, there's a lot of economic insecurity. And so there's sort of this combination of nobody really wants to spend capital because they're worried. And they're not only worried about spending the capital, but they're being pressured by their boards to implement AI. And so there's this combination of, I don't have any money, but they want me to spend money on AI.

7:19So, you know, what do I do? And so where we're beginning to talk to customers and because we're beginning to see some success on our own is we've been, you know, four years ago when we came out of IBM, we created an AIOps platform that enabled us to manage, you know, our, manage our customer systems with a high level of automation. Over time, we've begun to agentify. And so where we're beginning to integrate or work with customers is in thinking about IT service management. So those are all of the processes that are required to manage an IT infrastructure. So we've got a kind of a maturity model, if you will, that allows customers to begin to integrate agentic AI to automate all of those IT service management processes.

8:13Now, the benefit is that it radically reduces the cost of IT service management, in some cases up to 90%. And then you can take those funds and you can use those funds to do the modernization. So it's like a modernization dividend, if you will. So that's where we're really beginning to see customers looking at that from a horizontal. But I'd say it's a bottoms up. it's going to take some time before once you get the foundation ready you know the tracks laid if you will before you can have that the uniform approach to integrating from a vertical perspective yeah yeah and that's interesting that's what i've heard too that a lot of the early applications are in the the it service area partly because those guys understand that uh and it's it's easier the change management's easier.

9:07When you talk about IT service management processes, just give me one of those processes that would be ripe for agentic adoption. Sure. So problem and incident management, configuration management, patch management, those are all good examples of things that can be agentified. Okay. And is Kindrel building that you said that you do build the infrastructure? Yeah. In that case, let's take that case of IT service management. How do you go about that? I mean, do you look at the process, break it down, decide what can be identified? And then do you build the agents? Does the customer build the agents?

10:06Do you recommend off-the-shelf agents? How does that work? Well, today what we're typically doing is so a customer, when they work with us, we have our processes, we have very, very, very well-defined runbooks for how we would implement those processes. So the ontology is defined. And then through AIOps, we also have a lot of information about how that process actually is executed in practice. So we can see the environment in which it executes, what the problems are, et cetera. So what we do is we have what we call an agentic framework and a policy kind of ingester. So what we do is we take that process and we take the data, the insights, we feed it into our kind of engine that crunches the workflow and crunches the insights and builds the agent.

11:01And so we build those agents that allow for the sort of the automation of the process. Now, what differs from customer to customer is how much true autonomy they want to give these agents. So there's this question of, is it an autonomous agent with a human on the top or is it an agent with a human in the middle? There's a big difference there. And so we're seeing some differences there in kind of the way customers want to see those checkpoints be instantiated. But we will build those agents and then we will actually run those agents. We'll register them and run them on our platform. We are beginning to look at running our agents on other people's platforms as well.

11:46And not to be announced, but this is, you know, a direction that we're thinking about right now. Yeah. And you said at the beginning that you've spent, you guys have spent time integrating, or maybe I misunderstood, but integrating agents into your workflows. Can you talk about how extensive that has been, particularly on the horizontal? You know, it's always a, it's a funny thing. We were just talking about this the other day. It's like, how do you actually describe this? Because the definition of an agent is so different. Is it, you know, is it like the orchestration agent or is it the incremental agency?

12:32It's so hard. is the number of tokens that are digested. What I would say is go back to the kind of the ITIL processes. So there are 34 ITIL processes. Of those 34 ITIL processes, there's a subset figure 20 that can be identified. Of those 20, we have... And just define ITIL for... Oh, I'm sorry. I'm sorry. So IT infrastructure library, think about that as the best practices, the implementation blueprints, and the maturity model for how you manage IT. So when you're in kind of the business, you're an IT administrator, you know ITIL. That's kind of like your Bible. You know, you're going to go and patch a system.

13:19ITIL tells you, okay, if you automate this, then you've gotten to ITIL for level three maturity because you're using this level of automation. Very well constructed and defined. So we think about everything within the context of ITIL working toward an ISO certification. So ISO becomes kind of the pinnacle of how you operate. So when we think about agentic AI, we are breaking down those processes into kind of discrete components that are mapped to the IT infrastructure library, and then we agentify those processes. So in answer to your question, we are working through the agentification of those processes right now, and we have agents that will help get to various levels of maturity within that context.

14:15Yeah. And are your agents all based on, on like the granite models or are you model agnostic? Can you go out and find the best model or do you, do you have models in house that you built that are underlying, that are the brain, so to speak of the agents? Yeah, our models, we tend toward open source. So anytime that we can use open source models, that's what we will use. However, you know, there is an IP question for customers and sometimes they want us to use their models. So it really just depends on the customer with whom you're actually interacting. Right, right. And these, you were talking about, you know, there's a big difference if you have a human in the middle of an agentic workflow or on top sort of managing the agent or agents.

15:16uh which which uh do you have and have you built an orchestration layer that gives visibility and management tools to a human uh to to keep track of what the agents are doing what they're allowed to do and that sort of thing yes absolutely so that that is we have something we've we call the Kindrel agentic AI framework that was released, I think, about six months ago. That's exactly what it does. So part of it is the policy and sort of digestion and agentic creation engine. And then there's another component part of it, which is the orchestration engine. Now, think about, you know, in the orchestration engine, think about that as a tool which allows for a single enterprise to manage kind of agentic workflows.

16:16We've also got underneath that, we've got another capability we call Kindrel Bridge. So think about that as multi-tenants because there's a concept of reusability. So what we want to be able to do is take an agent that worked over here for this customer. Now we want to be able to reuse that agent over here for other customers because that allows us to scale the use of the agents in a trustworthy way. So bridge is the mechanism by which we will register. So discover, register, define the policy that needs to be monitored, monitor, and then enforce policy across a spectrum of agents that are operating across multiple orchestration frameworks.

17:03Yeah, and security is a big concern. And, you know, I've got OpenClaw running on this laptop on my main computer, and I know it's dumb and I keep on meaning to move it. But these, from what I understand, you know, agentic systems are full of security holes. uh so how do you address that or how do you uh you know instill confidence in your customers that that what you're delivering to them is secure you know it's really it's interesting um there was a there was an article today um one of the media outlets about um claude having done an exceptionally good job of, you know, identifying vulnerabilities within software.

18:03And it was kind of well touted. And I thought to myself, you know, it's good and bad, you know, on one hand, it's nice to know that, you know, these QA checking tools are improving over time. That's great. The problem is it gives people a false sense of comfort, right? Because, you know, it is, the challenge is context is, is, and this is where the human comes to play nowadays, is, you know, let's say you've got an agent that goes out and it checks to see whether or not there are vulnerabilities within a particular system. It finds, and I'm making this up, finds a communication protocol that is unpatched.

18:49So, you know, goes ahead and patches that protocol. Well, guess what? That protocol was configured the way it was because it's tagged to or it's tied to a 15-year-old legacy, you know, system in the back. And now you've patched it. Now it no longer works. And now you bring down the system. So it's not just a question of vulnerability. It's a question of context and configuration and the complexity of the underlying infrastructure. And as we started out with, most customers are working with a hodgepodge of stuff, multiple clouds, legacy stuff, some SaaS stuff. And the intricacy of that integration is where the problems lie.

19:30And that lack of understanding of how these things interact and why the context for why they are configured the way they are becomes really problematic and is why you need humans in the loop to make sure that you're not doing something dumb. Yeah. And I guess would humans necessarily be aware of of why that patch is not there uh i mean that you know people come and go and documentation is not always what it should be is rarely what it should be so uh yeah how how do you how do you know that context how do you communicate that context to an agentic system Therein is the question, right? So then you bring in generative AI, right?

20:23Because the knowledge base becomes very important. I think this is the interesting thing, is going back to you can't have a 150-mile-per-hour train on a 30-mile-per-hour track. I mean, these are the issues that are holding organizations back. And so when we talk about the need for modernization and the need for uplifting these legacy systems, what you're pointing out is exactly the practical reality of why you can't just dump agentic AI into your infrastructure and think it's going to work. Because it's more than just knowing where your systems are. You have to know why they were configured the way they were.

21:06Yeah. And where is that knowledge generally in an enterprise? Is it in people's heads? Is it in documentation? It depends. I mean, if you have a really good, you know, ticketing problem and incident management system, you know, ticket resolution system, oftentimes that history may be, you know, well-defined. Like if you've got a CMDB, service nail, that sort of thing, you'll have those records. Sometimes you don't. Smaller organizations won't. It just really depends on how robust your compliance has been. And let me emphasize that because this is the, this is the dual edge of, you know, compliance.

21:45It's like we say, you know, compliance is hard. Compliance adds complexity and cost. It also forces hygiene. And there is a reason why there is hygiene. There's reason why auditors look at the record keeping in and around things like, you know, managing critical infrastructure systems. It's because situations like this happen and you have to be able to go to the manual or go to the database or whatever it is so that you can figure out what happened and can unroll it again. But, you know, that's why I have a fraught relationship with regulation or love-hate, whatever you want to call it. It's good and also, you know, it's really painful at the same time.

22:26Yeah, so in modernization, when you talk about modernization, particularly with the IBM history, I go immediately to COBOL. What are the main challenges of modernization in general enterprises? Is it upgrading the language in which systems were built or does it have to do with microservices and containerization and all of that? Um, that's, that's a good question. Every, every organization seems to be a little bit different. You know, there's some common patterns. Let's say if you wanted to go to, as an example, SA, like you're going to be doing an SAP modernization. you know what you find becomes really really hard is so you know the sap you have to go from sap legacy into rise right so we all know we're going to have to move to rise and so there's a you know they're usually most organizations want to do like a two-step so they want to first move to the cloud and then they're going to move to rise after that and then what becomes the problem well the problem isn't necessarily lifting and shifting SAP the application.

23:50It's the networking. It's the provisioning. It's all of the stuff that is integrated into SAP that needs to be kind of reoriented to support a cloud environment. That becomes really painful. And so I'd say it depends. The answer is it really depends on the application environment that you're trying to move. and each of them have their own specific foibles, if you will, that you have to watch out for. Now, companies like ours who have the experience and what it means to lift and shift will know, like you have to look here, you have to look here, you have to look here, but it is, each one, as I said, is painful.

24:33Can I just add one thing though, because you're bringing up COBOL and this is, let's talk about quantum for a second, because this is a part that, you know, so your folks are listening to Agentec AI, and I'm going to wax rhapsodic about this issue. So today we have like 800 billion lines of COBOL code that are supporting critical infrastructure services. Of that, 400 billion lines of that code are supporting our COBOL code that was written 20, 30, sometimes 40 years ago. Now, most organizations would say, oh, okay, well, that's fine because I got a mainframe and the mainframe is the most secure, you know, and that's right.

25:18It is absolutely secure. Here's the challenge is it is built to be backward compatible. So when a mainframe runs, that 40-year-old code basically runs like it was 40 years old, right? So that, again, to keep that in mind. Now, why is that important? That is important because most of this old COBOL code has crypto relics in it. So what happens is the cryptography for that application is actually baked into the code and it doesn't actually manifest itself until it's in memory. And so there's really no way to find it. So one of the things that we're finding now is actually generating a lot of questions about modernization is actually going back to some of these legacy environments, COBOL, Java environments, where the cryptography was actually built into the code.

26:17And now you have to do this archaeological dig to figure out what are the crypto relics to find out now what do I do with those crypto relics? Do I have to refactor the application? Do I, you know, like what are the choices? But this is another one on the horizon beyond the Gen. I that we're all going to be thinking about. Anyway, thank you for letting me take a little detour. Yeah, yeah. Well, even in that, in reviewing millions of lines of COBOL, who, I mean, are at this point, are you using AI to do that review? I mean, are there still, you know, cobalt fluent engineers that can spend weeks reading millions of lines of code to follow what was done and, you know, where any vulnerabilities are or problems and that sort of thing?

27:22I think that this is one of the areas, this is one of the bright spots for AI is actually in doing code refactoring. We've seen a lot of success. We use those tools today because one of the challenges I'd say for a modernization program is kind of that front end, just that kind of the discovery process and understanding like what is it that you need to do. I think this speeds up the process. It doesn't make the, you know, the modernization program, it makes it less ugly. It's still ugly, but it makes it easier to undergo because you have the, you know, you have the tools at the disposal at your disposal on that front end side.

28:04Yeah. Yeah. So, so you have this orchestration layer, you're building the agents. on the customer side what is there some metric of how many agents a single human can manage I mean there's a lot of talk about how middle management is going to shift from managing people to managing hybrid teams, which include dozens, if not hundreds of agents. And then, you know, presumably there'll be a layer of management below mental management that is specialized in managing agents. How many people, I mean, how big is that specialized workforce going to be? Does it need to be specialized? Do you need to have people that really understand the orchestration software, understand how, you know, the fundamentals of how agents work and are competent to manage multiple agents?

29:24You know, it seems to be shaking out in a fairly logical way, which means that, you know, so going back to, you know, sort of if you're talking about using agents for IT, you know, service management, and your team is kind of organized logically. So you have one team that does kind of security management, which is security health checking and compliance and audit. And then you've got another team that's doing problem and incident management and another team that's doing provisioning. It seems that kind of the ownership of the agents is falling within those contexts because there's a logic to it.

30:07They own the process. They own the tools, the underlying environment. They know how it works. They know what good looks like. They know what bad looks like. And the scaling of managers, you know, kind of the ratio of managers to agents is really dependent. It's not clear right now. It's really not clear. It's more about, you know, the sophistication of the agent and that kind of defines like what the structure needs to be. As well as, honestly, the amount of auditing, you know, compliance and checking that needs to take place. you know the more heavily regulated the more you know you're talking about a critical infrastructure industry where uptime is everything the more humans you're going to have in that mix and so therefore like the scale you know in one organization might be one to a hundred another could be one to a thousand it just depends yeah yeah uh i mean it services is a is as i said a good place to start because those guys are are uh they're that's the language they speak right they spend time with technology but what happens when you're applying agents to a business process outside of it i've heard people say well you need you know the structure the id the t department has to become decentralized where you have, you know, in sales, you have an IT guy that sits in sales and can manage agents for the sales team.

31:50Other people say that, you know, you just need to upskill the entire organization so that everybody is competent in working with agents. How do you advise enterprises? So we've got a, actually, it's one of the interesting practices that we built recently is a kind of a reimagining workforce and kind of operations, because this is exactly the problem that a lot of organizations have. Now, I can't say that there's a perfect answer to any of this, because we're all inventing it at the same time. Right.

32:58and using these technologies, not only so they understand art of the possible, but they understand what's going wrong or can go wrong. But at the same time, then they have the specialists, the ones that are responsible for owning, implementing and running the tool just as if they would run like Salesforce. You've got a Salesforce admin. Now you've got the Salesforce, you know, agentic admin. Now, whether they sit in sales or they sit within IT. I think it really depends on the organization. In some organizations, Salesforce sits with the business line. Some organizations, it sits with IT. It just depends.

33:41Yeah. And so do you see this adoption accelerating or are we still at, you know, enterprise leaders are still kind of feeling their way forward you know educating themselves and uh leaning on companies like kindrel to to help figure this out or or is do you see it uh being adopted all over the place and you think oh my god these people don't really understand what they're doing, a lot of them. And, yeah. I think it's, I'd say the experimentation is accelerating. I think it means there's so many projects that are underway, and I think that the level of success for those kind of siloed projects is improving.

34:38So, you know, six months ago, I would say, very rare to see a successful use case. Now we're seeing more successful use cases. That's good. am I seeing more like large scale implementations? I'm seeing a lot more talk about large scale implementations. I'm not seeing a lot more large scale implementations yet. Will we get there? Sure. But not, you know, I think it's, it's like you're pushing this gigantic, like bubble down the, down the road, you know, and it's all this experimentation, but it's just getting pushed and it's getting bigger and bigger and bigger and bigger, but the road still isn't, it's still gravel yeah yeah yeah well that's uh what what i wonder and and and the you know the the media and and the hype always runs far ahead of reality uh uh you know research sort of shows the promise and then the press picks it up as, and the impression among the public is that this is happening.

35:48But do you think it's sort of like self-driving cars that we've got, you know, the basic technology figured out, but there's just a lot of other stuff that has to fall in the line before it becomes widespread? So maybe we're looking at 10 years before enterprises are really grounded in agentic AI. We're assuming, so I just, you know, and I'm going to give you my prediction. You know, my prediction is that if you look at the space I'm in, which the IT infrastructure services market, My prediction is that by about 2031, about half of all of the kind of the traditional IT systems administration tasks, the line one, line two tasks, they will be provided by agentic AI.