Agentic AI in the Enterprise 2026 | CXOTalk #916

There was a Goldman Sachs study that came out enterprises are blowing through their token budgets a a year's worth of AI budget in three months. This is notion of token maxing, which is as many tokens as you provide will get consumed as quickly as possible.

Praveen Akkiraju of Insight Partners invests across the agentic AI stack. Praveen, what is the state of agent technology today?

So, if you kinda take a step back and and look at the evolution of just the models, the end of 2024 was, like, one of the most seminal moments in the e- in the evolution of models, right? Uh, it was when, uh, the four O model came out, which essentially introduced the concept of reasoning. Uh, and I think that was the first step towards truly making AI, uh, less of a chatbot and more of an agent, right?

And, um, so, from there, we saw the, uh, introduction of DeepSeek and sort of early or late, late twenty four, early twenty five, uh, and that fundamentally changed the economics of of, uh, large language models. Uh, and later part of last year was when we started to see the true sort of agentic capabilities, uh, in the model start to start to emerge. Coding became a huge use case, and, you know, all of a sudden, we, we start, now realized, like, wow, um, you know, you could actually use models to fundamentally, um, you know, write code.

And beginning of this year, I think, was sort of the the real inflection point, both in terms of, um, you know, the the capability that came forward from a coding perspective, but also the ability for, uh, open source frameworks like Open Claw to be able to use models to actually show us that work can be done. Not just, you know, uh, technical work like coding, but also just basic work, right? So, you see that arc of models and the evolution of these things, the pace has been dizzying, uh, towards unlocking the true capability of these, uh, large language models, to actually, you know, do work for us.

So, what is the state of AI agents today?

Think about, um, what agents is an agency is, right? Which is essentially the notion of being able to take a task, um, understand the context of the task, break it down into a plan, and then be able to use a set of tools to execute the task. That's essentially what, uh, the definition of of agents and agency is. Now, the core of these agents is the large language models that we we've talked about, you know, the evolution of these things.

But fundamentally, you know, you think of large language models as these kind of stateless, um, stateless machines, right? In the sense that they absorb all the information you give them, you know, it's a compression algorithm. They essentially take all the world's information, compress them into weights, right? And, um, are able to sort of reproduce that, uh, or respond to prompts based on those compressed weights. Now, the issue here is that, um, because they're stateless, they don't remember, uh, or like humans do when we do tasks, the context of what the work that has been done before.

So, the real sort of, um, evolution of agents has been the combination of these really powerful large language models with, um, these kind of things we call harnesses, right? And so, harness essentially is, um, you know, is a set of tools, its context, its memory that is given to the model in order for it to perform a particular task. So, you know, we can think of these, um, harnesses in the context of applications, right? So, an application like, um, you know, cloud code, right, has a very well-defined harness that gives us the gives it the appropriate tools, uh, the memory, um, the the kind of guardrails, if you will, in terms of coding practices and such like that allow it

to then deliver code, right? So, combine the model with the harness, give it a set of tools, and that's how you kind of do work, right? And more, um, more recently, I think you started to see, uh, you know, applications beyond coding. A chat GPT in and of itself is an application, and open claw essentially is a personal agent, which created its own model, right? Uh, its own, sorry, its own harness that attached to the model.

And so, what you can think of these as if, uh, large language models are stateless, um, you know, powerful, um, AI engines, you combine that with harnesses to give it sort of things like in-context learning, so you can actually do work, remember, um, you know, the context, the work you're doing in a long horizon, and actually execute tasks.

We have all of this opportunity, and there's been evolution, uh, among the the the tools, and we all see the the promise, uh, or at least we're told there's all of this promise out there. So, what's real versus what is the hype in agentic AI right now? Can you help sort that out?

When you think about just the models itself, and going back to because that's the core of a lot of what agents are, um, you know, they have what's called jagged intelligence. In the sense that they are state-of-the-art, like, amazing, you know, surprise you with, uh, the capability that they have in terms of coding, or in terms of being able to, um, synthesize information. But at the same time, they make very common, silly mistakes, right? Uh, or they're easy to be tripped up when you, uh, when you change prompts.

And so, we're still on that frontier of jagged intelligence. So, I think, when you think about agents, you know, 2025 was sort of the the the year of agents, and rightly so, because, uh, the models advanced. I think we had, um, you know, huge advancement in terms of computer use agents, you know, browser use, things like that, that, um, that are, uh, that enable these agents to be real. Uh, and so you s you saw this significant inflection in terms of deployment of agents.

So, you know, sophisticated enterprises today, uh, have deployed over a thousand agents in production for various different tasks, right? Um, you know, at the same time, this jagged intelligence essentially means that, um, you really need to do your homework in terms of defining this harness, right, for your particular use case, uh, which essentially means that you have to, uh, make sure that the agent has the appropriate context. And that context essentially mean in in an enterprise's use case, access to the appropriate data, uh, the appropriate sort of guardrails, compliance, right, uh, and and security, uh, framework that enables it to then function within that that sort of, um, that, uh, particular use case.

And today, most agents still, uh, do incorporate a human in the loop, uh, for most tasks as a way for from, uh, from a final sign-off perspective. So, I think, you know, the the hype versus reality is that, uh, are we seeing agents in production? Yes, right? Are they perfect? Are they completely autonomous, you know, fully agentic? Um, no, right? You need, uh, to do a ton of work in terms of defining these harnesses, defining these sort of g guardrails, if you will, in order for you to get the appropriate, um, you know, output with a human, uh, in the loop.

And they're also, you know, significant areas of of concern, right? I mean, uh, you know, security being one of them, um, but, uh, but also, how do you actually, un- uh, you know, regulate how token use happens, right? Are are these agents doing useful work? Um, you know, do you, uh, do you actually have the right sort of governance, uh, of how you manage all these agents? If you have, like, a thousand agents in an in an organization, um, you know, how do you kind of how do you ensure that they are?

They have access to the right tools, um, they being governed properly. So, there's there's a lot of these open questions, which I think a lot of, um, uh, enterprises are still working through, but there's no denying the fact that, uh, just from a pure capability perspective, models, harnesses and such like, as well as from a, you know, deployment in the field, we've seen a huge step forward, uh, over the past year.

You have just laid out the agent world. Let's start unpacking this. Let's begin with applications. So, where are you seeing real applications for autonomous agents that are effective, that are actually delivering real value? Let's start there.

There's a spectrum here, right? So, you can think the spectrum as one end is what we as individuals experience, uh, in terms of agentic capabilities, um, and you can think of that as, uh, your, um, you know, uh, you might be using a tool, um, like, uh, you know, maybe you built your own open claw or something like manus. On the other end of the spectrum, you have enterprise, uh, class agents, right? Um, and, and there's a whole set of things in the middle, um, where you have, uh, vendors building these very specific agents for specific kind of use cases.

So, I think if you kinda look at look across the spectrum, um, I think one thing that obviously was like a huge seminal moment was, uh, open claw, uh, you know, earlier this year, and the reason why open claw kind of burst out was. An and if you kinda go back in history, uh, of of AI, there's been a lot of companies trying to build these personal agents, you know, agents that, uh, you and I could basically deploy, give access to our our contacts, you know, our email, calendaring, etcetera, and have them do work for us.

Uh, and open claw was the first kind of open source framework that actually started to work because, uh, it was able to bring, um, the the appropriate sort of ability to understand the work, ability to remember work, right? And ability to take the right context and plug it back into the model in order for it to execute a task. And I think that truly blew people's mind because it was so, um, it was so effective, right?

Uh, and simple to to to execute. So, you saw on the personal side, and I think all of us have been playing with, you know, cloud code, um, either on the CLI or in the application itself, and and doing, you know, interesting workflows in a very seamless way. So, I think the, uh, so we've seen huge amount of like personal productivity improvements. A lot, um, most of these tend to be like, um, single use case, uh, or more single user oriented, uh, uh, a agents, right?

The other end of the complexity is where you actually have an enterprise that's trying to build or or or buy an agent that needs to be deployed deployed across thousands of, uh, employees. So, here you're kinda seeing, um, there's there's two aspects to it. So, the first aspect is there certain, um, functional sort of agents, you know, you can think of customer service, for example, where you you clearly have, um, you know, well-defined agents, uh, that are effective, that are able to deflect a lot of, uh, first line kind of support calls, and and be able to kind of, um, demonstrate productivity, right, in that context.

And so you're seeing the the use cases where, um, you're seeing clear fit is customer support, um, obviously, coding coding, I think, has been the ultimate sort of AI use case. I think it's almost like it's estimated about 3 billion in revenue right now from just coding agents, uh, you know, customer support agents, I think, is close to about 500 million, uh, you know, in terms of revenue generated just for that for that sort of use case.

Uh, I think search has been another one that's, um, it's, uh, it's less of an agent, I think it's more of a tool, but, you know, you've seen good traction, uh, enterprise search, right, uh, as a as a fundamental, um, use case, right? And if you think about sort of the specific sort of verticals where you're seeing good, uh, good adoption and tools, you know, clearly there's been a lot of lot of momentum in the legal space, um, in in healthcare, uh, in in the tech forward companies are essentially deploying AI agents in a significant way, right?

So, um, I think when you look at the spectrum, you you see significant advancements in the personal sort of single user, single use case kind of agents, where you can build your own because the interfaces have become much more simpler. At the same time, enterprises are focusing on these specific sort of use cases where there's clear, uh, product market fit, clear value and measurable ROI, uh, for these, uh, deployments.

We have a very good question, interesting question from Anthony Scrifiniano on Twitter, who is the former chief data scientist of Dun and Bradstreet. He's been a guest on CXOTalk a number of times. And he says this one of the big challenges with agency in the real world is negligence. Can you talk about the danger of AI agents acting in concert in a negligent way and the impact on the enabler?

When you think about models fundamentally, I think they are prompt driven. And, you know, when you think about sort of the construct of an agent, um, you know, you give it a task, it creates a plan, and it goes out and uses a bunch of tools to execute that task. So, um, you're in that there are many, uh, many areas where you could potentially have this notion of in in the plan could be wrong, it could be using wrong tools.

Uh, it could take the input from the tools and, uh, and and go down a different path, right? Um, where it might actually do something malicious, or there might be a prompt injection attack that makes it do something malicious. So, I think this is, uh, the reason why this notion of a harness, uh, and guardrails is so, so critical, right? Um, and and almost say the the agent is actually the harness. Um, so the ability for you to cre- define clear guardrails, right, clear constraints, uh, and clear, uh, you know, evaluation criteria for the agent is extremely important.

And that's a fundamental, uh, aspect, uh, where you need to spend a lot of time when you build these agents, um, because, you know, e- essentially, these large language models, uh, are, uh, steerable. I mean, they're very sensitive to prompts, and they react very, uh, very, very, uh, immediately to the prompt that they're that they're given, right? Um, so the actual sort of constraints and the, uh, the steering of this agent has to happen in the harness.

So, um, you know, observability is is extremely critical to be able to incorporate into these. So, one one design aspect here that we's we're seeing is this notion of sandboxes, right? Uh, you know, if your agent is essentially, let's say, writing code to execute specific tasks, um, what we're seeing from a design pattern perspective is, uh, you know, developers are spinning up these sort of ephemeral sandboxes to first kind of make sure that the the code is that the AI agent has written is sandboxed, executed, is is essentially, um, measured against the guardrails to ensure that it conforms before you actually execute and push it out into production.

So, there there are ways to address this, but, you know, it's absolutely, uh, a critical concern.

We have a great question from Chris Peterson, and he says, giving agents access to our quote-unquote context is interesting, but how do we maintain boundaries between the context we own as individuals versus what belongs to others, friends, family, employer, and so on?

This is sort of one of those things where, um, from a data access perspective, uh, you know, being able to design this thoughtfully is critically important. Again, um, when you look at sort of the way an agent, um, accesses data, you either, uh, giving it an API, uh, or you're giving it access to an MCP server, uh, or maybe you're using e CLI if you essentially, um, you know, getting it giving it access to something that's running on on your computer.

So, uh, an API app uh, gives you the most granular ability to create, like, you know, uh, an OAuth, uh, architecture that allows you to be very granular about the kind of information, uh, the API can access. MCP tends to be a little bit more abstracted, uh, in terms of what, uh, what it returns because it's more of a generalized sort of access paradigm. So, when you give an agent access to data, I think you want to pick the right sort of interface that it uses in order for it to be able to get access to this this particular data.

Um, and again, you know, I'll I'll go back to, so this notion of, uh, creating the right guardrails, uh, you know, you have these tools, as I said, either using a an API or or, uh, the appropriate choice there, but also just in terms of policies, you know, you can define these dot MD files that, uh, effective effectively tell the agent, hey, check to see if this data is PII, right? Uh, or check to see if this data is sensitive before you actually use it.

You can define what the agent does in in these kind of dot MD files that are are given as a specification for the agent when it gets started.

Yeah, sure, except when the agent ignores the instruction and gives you back something altogether different. And then you look at it and it's not right, and you say, well, you didn't do what I said. And the agent comes back and says, Yes, you're right, I didn't do that. Here's what I should have done. And this is totally and this is seamless. So, you know, there's no way to know unless you know something about the answer that you've been given.

This is very dangerous.

You give it the appropriate guardrails, you give it the appropriate data, uh, it still produces, uh, a result that's, uh, that's not that's incorrect, right? Uh, and so, um, you know, what we are seeing again from a design, uh, perspective is, uh, you you need to create the right observability framework. So you're not just waiting for the output at the end of the agent's execution, but rather you're looking at the traces in at every single step, right?

And being able to sort of understand that. And again, combining this with sandboxes, you know, we have a company in this space called E two B that, uh, that that's been, you know, tremendously successful, kind of helping developers with this, right? To create observability at different stages of the agent's execution, so you can catch it. Because, you know, look, there's an important point, right? Like, error is compound. So, if you have, uh, uh, a large language model, you know, you have a multi-agent architecture, and an agent makes an error and feeds it into another agent, you essentially have a compounding effect of errors.

So, it's really critical for you to be able to have this kind of granular observability a and design the evals. Now, you know, I I it's easy for me to say to design the evals, but that is actually turns out to be one of the more challenging aspects of agentic design, right? And I think, um, you you see the the the sophistication of something like cloud code, right? Or a a a customer support agent, it's actually the brilliance of the harness, because they've actually done exactly this.

They've created the observability so it checks its answers, they create the evals so it actually knows what is correct and what is wrong, and and that's kind of how you get to, um, and and it's a fine tuning process. And I think a lot of enterprises are actually going through this, which is to essentially start this process of building the agent, and, you know, continually tweak it till you get to a point where you feel like, OK, you you've got this, uh, you know, you've got the the, um, the agent executing the way you want it.

Preeti Narayan, who asks a very interesting one, and she says, Most enterprises are built for AI assistance, that is tools that help humans decide, but agents actually make decisions and act on their own. That's a fundamentally different problem. So, when the shakeout comes, and parenthetically, it's interesting that she's making the assumption that a shakeout will come, but when she but she says, when the shakeout comes, do the survivors rebuild? Or do the ones who bolted agents onto old infrastructure simply not survive?

The rate at which model capability is advancing, right, is making a lot of the assumptions or sort of these notion of you can wrap things around a model and and deliver value obsolete, because that's been absorbed directly into the model. So, I think the the so they there's if you kinda like, take apart this question, I have a couple of things here. So, the first thing is how, you know, all the legacy software or the existing software that we have, you know, how does that transition into, uh, in in this agentic era, right?

And the second question is, you know, agents have autonomy, um, and you know, what does the future look like when when, you know, e- effectively, we do get all of these things figured out and and get fully autonomous agents, right? Um, I, you know, I kinda see the world as a bit of a hybrid, um, and and that's primarily because the challenge of actually, uh, allowing an agent to be fully autonomous, um, is i is a fairly challenging task.

I mean, in a bounded problem space, like, let's say coding, where you know what the output needs to look like, right? You can measure it, you can test it, right? Uh, and you can certify it. Uh, you the convergence towards autonomy is likely to happen faster. And we s we see this today. You know, uh, I think if you talk to some of the large labs, they'll tell you that they feel like coding is at AGI already, in the sense that it is at a level above human, uh, capability and is able to actually, uh, work as as well as humans do, right?

Um, but in, you know, more sort of unbounded tasks, or I think a task where there's a lot of lot of dynamicity, like, you know, let's say, you're managing the supply chain, uh, in an enterprise, and, you know, you have different SKUs, maybe different geo, maybe different sort of supplier onboarding with various different criteria. The ability to kind of, um, generalize, uh, what the agent does across these new capabilities pretty much again comes down to how well your harness behaves.

And so designing that is a lot more challenging than, you know, a bounded problem, like solving math, right? Or, uh, being able to, uh, uh, being to write good code. So, I think that's the way I would kinda think about this in terms of this, you know, how do you spread these across bounded and unbounded? So, for existing software, you know, if you're just to bolt on agent on top of, uh, software without fundamentally changing the the way, um, you know, software interacts with data, with with users, and being able to respond dynamically, then I think you you clearly are are gonna lose that battle to, uh, potentially a newer version of, uh, of software that's a lot lot more

sort of, um, agentic and dynamic, right? Ability to respond, take inputs, you know, be able to, uh, do work a and work, uh, and and be able to, uh, um, offload the human from a lot of the the work that an existing platform would do. However, I think I do see a future where, um, you we and we are seeing this, and we have, uh, a number of companies, obviously, that are doing, uh, you know, great work being able to sort of understand how an application works and figure out where can they insert an agentic, um, step that accelerates the time to value for the customer.

I'll give you one example, you know, um, Stampli, which is one of our companies, uh, they do, uh, accounts payable that are procured to pay company, right? So, one of the challenging aspects of, uh, of accounts payable workflow is, uh, reconciling invoices, uh, and, you know, it essentially is a manual task because unless you reconcile the invoices, you know, you you can't approve the payment. And so, uh, what the team's done is basically said, like, hey, let's let's understand the entire workflow of, uh, accounts payable and figure out where are the steps, like, you know, potentially there's form filling or there's invoice reconciliation, where we can insert an agentic step to accelerate, uh, the time to value for the customer.

I think that's a thoughtful way where you can actually take an existing software platform, right? And of course, it does require some surgery, right, to be able to insert these steps in the middle, but then you can accelerate dramatically the value for the customer because all of a sudden now you're you've eliminated, you know, days worth of work of an invoice reconciliation. The the AI does it for you, and then the human approves it. So, you know, that is real value in an existing software paradigm, right?

At the same time, there are, uh, other examples where it makes sense to kind of, like, bottoms up, uh, reconstruct sort of the the the workflow. So, a lot of these sort of, uh, operations oriented software, right? You know, whether it's security operations, IT operations, etcetera, uh, customer service operations, where, um, you have a a strong data platform that you're then able to now build an agent on top to essentially, you know, use one of the LLM's superpowers, which is discerning signal from noise.

So, I think that's the way, y- you know, that's why I I believe this is a hybrid world where they will be, uh, you know, companies that are doing things in a thoughtful manner to transform the existing software. At the same time, you have bottoms up, AI native approaches, which, you know, clearly will will either create new surface areas like coding. Nobody thought coding engines would be a a a brand new kind of market space. It's $3 billion worth of revenue in that space right now.

We have a question from LinkedIn on LinkedIn from Swami Vijayathan, and Swami says, How do you envision the target operating model? Number one, enterprise AI to be dominated by dominated by human in the loop by design or autonomous agentic enterprise with guardrails and governance, especially with the questions around AI rel- reliability. And obviously, the the human in the loop is what everybody is doing now because it's a more conservative approach. But how do you see this shaking out?

Human in the loop is the defined paradigm, uh, and in a lot of ways, I think that, um, I go back to sort of my reference of bounded, unbounded problems, um, the the more deterministic a particular workflow is, right, the more bounded, more well-defined the outcome is, I think the faster the opportunity for us to get to autonomy will be, right? Um, I think, um, I I don't believe that we will, e- at least in the near term, be able to get to complete autonomy across the entire enterprise architecture.

The enterprise is just way too complicated. Um, you know, you, you have, uh, a lot of very specific workflows that are unique to an enterprise. So, you kinda d you know, divide the spectrum of use cases, you know, there's some standardized workflows, you know, maybe front end facing workflows, which are fairly standardized across multiple enterprises, um, where you could potentially get to a higher level of of autonomy, um, but the it's it's a combination, I think, of three things.

So, the first thing is the nature of the workflow itself, right? Um, and the the you know, where you lie in the spectrum of, uh, you know, bounded deterministic workflows versus dynamic, sort of more unbounded workflows. The the second aspect, I think, is there is a regulatory compliance aspect to, um, a lot of these industries. So, if you're in healthcare, for example, where there's been a lot of AI adoption in a space that typically has been challenging in terms of, uh, technology ramps, you you see a lot more rapid adoption.

Uh, but there is significant amount of regulatory and compliance based, uh, requirements that essentially requires a human in the work, uh, human in the loop, uh, because you do need that that sign off and that accountability that is, that is mandated, right? And and the third part of this is again, uh, it's going back to the previous conversation we're having is how, uh, you know, how well-designed I is your guardrail, right? How well designed is your governance framework, right?

How well designed is your security architecture, i in order for you to have that confidence to let the model function? Um, you know, i if you kinda go back to the human way, listen coding as an example, um, you know, you always had this notion of you, you write code, uh, there's a there's a PR, and then there's a code review in order for you to be able to before you push code to production, right? And, um, you know, so there's always sort of, even in your traditional, you can take any workflow, I mean, it could be a legal workflow, it could be a finance workflow.

There's sort of there's amount of work done, and then there's like one step of approval, uh, could be a senior, uh, you know, legal professional or finance professional who signs off on the work before it goes out. So, that step of verification has always been built into a lot of these workflows. And so, um, the question is, um, you know, is that step, you know, is it worth eliminating that step, right? Or, uh, is, uh, you know, can I if you're able to get to increasing amount of accuracy on those, um, you know, that is essentially what determines whether you can get there.

So, the quality of the harness design and the quality of the governance, uh, and that'll be, I think, an experiential process. We're gonna keep experimenting till we get to a point where we get more confidence and, you know, it's very much gonna be use case dependent.

We have a really interesting question from Arcelon Khan on Twitter. Are you aware of, uh, any libraries of AI agents that are accessible to all organizations? So, organizations, uh, all organizations have to do is provide the data, and could industries work together to create this?

Personally, I'm not aware of, uh, a public library of AI, uh, agents per se, other than, you know, those that might be, you know, available. Um, I guess you you could, kind of i it there's probably a library of different open claw agents that, uh, you know, folks have built and published that you might be able to access to. But again, that's one of those things where, uh, you you really want to be very careful about, um, you know, deploying a third-party agent into your into your environment unless you you truly have confidence.

So, but what we are seeing obviously is, uh, within an enterprise, uh, the the notion of, you know, building agents and publishing those agents and making sure that, uh, the teams across an enterprise have access to the agents is is clearly the way, uh, uh, you know, uh, enterprises are encouraging the transition to an agentic architecture, right? Um, you know, I think the broader question of, you, you know, are there ways you can collaborate? And and I think you're starting to see this in particularly areas such as security, for example, um, you you know, you can think of, uh, the MITOS release whenever it's going to be publicly released, uh, could be a a very powerful tool for the community

to be able to sort of figure out how to protect themselves. Of course, it's a double-edged sword, and that's, I think, the conundrum that we're we're dealing with, uh, with this really, uh, capable model. You know, I think I just saw an article from, uh, Mozilla, the the the latest, uh, version of Firefox. They found 271 bugs, uh, using MITOS, which is kind of amazing. I think only three of them were probably, uh, CVEs, like truly high severity, but still, you know, the surface area of the power of that tool is obvious.

So, that is an example, uh, maybe an imperfect example because it's a double-edged sword, but, uh, I do think that, um, you you will see some of these general purpose agents that that will be out there that help the broader community. But I think, uh, given the non-determinism and and so many, you know, kind of open questions around agentic architectures that we just talked about, is governance and guardrails and such like it's, um, eh, you know, you should be very careful about just taking third-party agents and and and just using them, right?

So, there's we're still not at that level of maturity.

We have an interesting question from, uh, Sayeda Zinoth, and she asks, Do we need to ensure that organizations or do organizations need to ensure that they fix their processes that are not working properly before they integrate with agents, or can to what extent can they work with existing processes? And I'm gonna ask you to keep your answer pretty short, Praveen.

I'll kinda take this as an example of the data architecture, right? So, this is a typical conundrum because most data today in an enterprise is sort of pretty fragmented. Not everything has a nice clean little API, uh, that you can access it through. So, the conundrum's always been, should I first do a big data transformation project before I get started, uh, with with my, uh, agentic transition? And I think what we're seeing enterprises do is they can't afford to to wait, right?

So, uh, and, you know, you have interesting tools now, like computer use and browser use that actually enable agents to use your processes and your data and access your data like a human would, right? So, it's it's sort of a, um, you know, I think we've found these tools so you don't have to wait, and I don't think you can afford to to wait to kind of drive this huge transformation project before you go to AI.

I you know, I think you'll need to kinda do these things in parallel.

So, from that standpoint, it's somewhat different than historically implementing something like ERP or CRM, where you really did have to clean up the data and or clean up, clean up your processes, it it's what made sense.

When you say processes, you know, obviously the way, uh, you know, you have to rethink, uh, human centric versus agent centric policies, right? These are completely different sort of workflows. So, there is a level of sort of, uh, architectural design, uh, of processes that does need to happen, right? And so, you know, you start off with a few workflows, um, and you you assume that there's an agent in the loop, and, you know, design your process for that, and you learn, and you go along.

Um, so, you know, i it it they, uh, the answer here is really sort of it will be experimental, experimental, right? You're gonna learn as you go along, right? But my my broader point is that you I don't think you can wait to kind of do a process transformation or a data transformation. I think you need to pick a surface area and start to work on that, right? And ideally use the tools that that these, uh, AI models provide you in order for you to accelerate that process.

It's a really, really good point because the technology is changing so much more rapidly than anything in the past. I mean, ERP systems, uh, didn't change from one week to the next. They were stable, and so you had it was not a moving target in the way that the current AI technology is.

It almost feels like we're building on quicksand right now, right? Because, um, just given the rate at which model capability is advancing, or, uh, you know, new capabilities are being unlocked, uh, you you need to be able to, uh, be more dynamic. And and that's part of, I think, uh, the challenge. I think this is called the Red Queen effect, right? Which is, uh, effectively, every enterprise, every knowledge worker, every professional today, we all feel the same sense of urgency that if we don't change, if we don't evolve, uh, you know, we're not gonna keep up.

We're afraid about getting ahead, right? Just to kind of keep up, you have to, you have to keep, uh, keep changing, keep evolving. That's that's the state of things, uh, where we are.

Another question again from LinkedIn, and this is from Yaw N, who says, You've watched dozens of agentic AI companies hire. What's a skill that mattered a lot in new grad hiring two years ago that you now see hiring managers have stopped caring about?

If you go back two years, they were clearly defined roles, right? You had a software engineer, essentially, proficiency on code. He had all these like elaborate coding exams that you had to get through, uh, you know, designing walk a tree, you know, the th- these these kind of things that that were, uh, kinda bars for you to get through. That they were product manager roles, they were for deplo- you know, they were sort of field roles, uh, in a in well-defined sort of, um, you know, architecture of what an organization looks like today.

If you actually look at a lot of startups, um, all these roles are mashed together, right? Um, there is no clear product manager. A software engineer today is on site, uh, with the customer, uh, actually working with them, deploying code, learning from them, uh, developing almost, you know, g- be the customer in the morning, develop code at night, you know, go back to the customer, right? The iteration loop has gotten much faster, and it's broken down the silos.

So, you know, if you are a new grad coming out, essentially, that is what you're expected to do. You're not, you know, you're not going to be given a project and say, OK, here's a project, OK, learn the code base and then, you know, produce something or fix a few bugs before you write code. You're essentially gonna be dropped into a customer site and say, OK, let's let's go build a product and iterate fast. And I think that means that you have to be sort of a, uh, a a software engineer, a product manager, a forward deployed engineer, all of this stuff rolled into one.

Uh, I'm talking about a startup context, of course. You know, if you're a large organization, some of these roles might be further more well-defined, given sort of the complexity of what, uh, that that particular company might might deliver. But I think that's what we're seeing, you know, these these roles in in startups are, uh, are hybrid, right? And, uh, close to the customer and close to the code is kind of what what differentiates, uh, you know, uh, engineers now.

I had conversations during the past week with CIOs from two very, very large insurance companies, you know, with thousands or tens of thousands of employees, separate conversations. And one of the things that became clear is this melding of roles, the, uh, the dropping of boundaries between, uh, product design, software development, DevOps. These are all merging together. Now, these particular folks are highly, highly advanced and highly innovative, but it's a glimpse into the future where Praveen, what you just described will take place inside very large companies as well.

It's not just the startups, at least going into the future.

If you think about what we saw with Claude co-work, right? Uh, you know, Claude co-work essentially was a realization in Anthropic, uh, as well as other labs as well, that, um, hey, this coding harness that we built to write code can actually do work, right? Can actually do, um, non-coding work as well. And so what that means is that if you're, uh, a business analyst, uh, or a a financial analyst, or, you know, support engineer, whatever, uh, you now have, uh, a very powerful model that you can use to define your workflow and create your own sort of, you know, application agent, if you will, to to get your work done.

And and that's sort of the democratization of, uh, that that that AI has enabled, right? In the past, you probably needed to go to an IT, uh, you know, uh, team and say, like, hey, here's what I need to get built, and there's a certain amount of time. So, now if you're a business analyst, you say, like, OK, I just need to, like, whatever, um, you know, get all this information, synthesize this, put this into a particular spreadsheet, run this particular analysis, and I could create that entire workflow myself, and the AI would be writing code in order for me to do that.

So, that is actually the true power of, uh, of agents. It's democratized coding, it's democratized the ability for, uh, knowledge workers, whether you are, you know, an engineer or an analyst, to be able to get your work done.

Let's talk about the SaaS apocalypse, the SaaS apocalypse. Do you believe that AI agents are going to replace software like Salesforce, like Workday? Where do you fall down on the SaaS apocalypse argument?

We talked about this a little bit in in the previous question as well, where we were sort of talking about how you take existing tr- uh, software and be able to transform it, right, in a thoughtful way, right? Uh, essentially looking at where can AI accelerate work that's that that is being done, uh, and accelerate value to the customer, right? So, I gave the example of invoice reconciliation in an in a accounts payable or procure-to-pay platform like Stampli.

Um, you know, I think, look, system of records are are not going anywhere, in my opinion, right? And because fundamentally, you know, the one question we haven't talked about is there's a real cost to agentic software, right? And this is basically there's a variable cost to software. So, every time you are writing code or using a tool, uh, or doing reasoning, uh, you're burning tokens, right? And sure, I mean, the token costs have gone down dramatically, you know, uh, over over the years, as these model capabilities have advanced.

Um, but, you know, it it is still fundamental cost that you have to know, you have to know, uh, you know, think about when you start initiating a project. So, does it make sense for an enterprise to say, you know what, I'm just gonna bottoms up build a brand new CRM or bottoms up build a new ERP, right? Um, you know, is that worth the amount of tokens that you would spend, uh, to go do that?

Versus, you know, where is the real value for you as an enterprise, right? Uh, you know, which is basically accelerating your business, right, your use case, you know, if you're whatever insurance provider, you know, bank, uh, you know, whatever manufacturer, right? That's kind of where you want to spend your tokens as opposed to going down to the infrastructure level and saying, like, well, let me just get ro- go reinvent the wheel just because I can, right?

Uh, can you build it? Oh, of course, you can, right? And, you know, it's interesting because in a previous cycle, you know, companies like Tesla essentially rebuilt their ERPs because they felt like, um, the existing ERPs did not fit the kind of model that, uh, you know, Tesla wanted. I mean, they're fully vertically integrated car production systems, uh, and so that makes sense. I mean, there might be certain use cases where i it logically makes sense, but, uh, I I just I don't buy the fact that you're you're gonna go and reinvent the wheel here and try to take some of these really complex, must be super reliable kind of infrastructure pieces and software and and redo them.

Oh, I totally agree. I mean, you know, I personally, I think it would be insane. Who's gonna try to rebuild their Oracle or SAP or Workday systems? Now, there's going to be nibbling at the edges for sure, but we have a question relating to specifically this from Elizabeth Shaw on Twitter, who says, How is agentic AI pressuring existing enterprise software pricing and business models, which is which is really the the question.

We clearly are seeing budget shifting in in enterprises, right, from traditional SaaS, um, you know, per seat pricing to, uh, to more sort of agentic tools. Uh, and and a big part of that is in some way is just sort of a cleanup, if you will, uh, uh, there's a lot of sort of, uh, excess, you know, seats or excess sort of software buys that have happened, which are being kind of reconciled now based on usage, uh, and that budget's being redeployed for AI experiments, right?

Um, the one thing I'll say, and it's kind of important, right? Uh, a a lot of this is we're observing this in real time. Uh, none of this is, uh, none of what I'm sharing or, you know, uh, the conversation we're having is absolute truth, right? It's kind of a point in time. This is kind of where things are and and it's a big part of where we are as an industry right now, which is given the rate of evolution, uh, you know, we're trying to connect the dots and and extrapolate, but, uh, you know, we are still early in this journey towards sort of agentic architectures.

So, I I believe that, you know, today where enterprises are, um, they are, uh, very aggressively leaning into, uh, this notion of an agentic transformation for the whole architecture. At the same time, I think what we are also seeing is that most enterprises are running experiments, right, around, you know, where does the an agentic, um, solution make sense? How much is it gonna cost us? What is the true ROI, right? How do we govern these things?

So, these are all questions that are being answered in real time. Um, in certain use cases, obviously, you know, uh, the the, um, the answer is obvious, you know, customer support, obviously, we we're seeing huge amount of traction there. Writing, you know, marketing collateral, those kind of use cases are clearly demonstrating a lot of value, a lot of operations use cases where you can, you know, you have a ton of logs or a ton of data and you're trying to, you know, discern signal from noise.

Those use cases kind of make sense. Um, but I I still kind of w- we're still sort of in this early stage of adoption of agentic architectures, and, you know, still have a number of significant questions to be answered, right? So, we're not quite, you know, across the line yet.

How should CIOs make the build versus buy decision regarding agents?

You know, in talking to a lot of enterprises, what we see is sort of this the build versus buy is kind of breaking down around the front end versus back end kind of, uh, line. So, a lot of, like, front end customer facing tools where the workflow is perhaps a little bit more standardized, right? Um, you, you are, y- you know, um, enterprises are leaning towards a buy kind of decision. So, it could be something like customer support.

It could be something that's, you know, a standardized workflow, like, you know, if you have a finance workflow, you know, you're, you're, you're running your, your finance statements, it's kind of done the same way in most organizations. In the back end, there's a lot more variability depending on your industry, depending on sort of the compliance and regulations that you have to, uh, you have to abide by, uh, and also, you know, there's a lot of data that you have.

So, bringing those things together is kind of where we're seeing the buy, um, you know, leaning, you know, enterprise is leaning towards towards the buy, uh, side of decision. So, using frameworks, yeah, right, using these model capabilities to kind of build their own sort of customized agents. So, at a high level, that's kind of where the the line is. But again, you know, we're early in this journey, so it's something that we need to keep, uh, you know, following.

Where can CIO safely remove humans from the loop and where not? In other words, where does autonomous operation actually make sense?

It really comes down to sort of the harness that you're able to, you, you're able to build around this agent, right? And what I mean by that again is the ability to create, um, you know, a a well-defined reasoning, uh, and a well-defined sort of decisioning process for the agent, uh, and so the the context that you provide to the agent in terms of the data it has access to, the guardrails, the tools it can use, the kind of the, I guess, the parameters of how those tools can be used.

Um, I think defining that, uh, be being able to have very granular observability and evaluation frameworks in order for you to make sure that as the agent's progressing, you're able to understand, you know, and catch any, you know, drift in in its performance. And and being able to sort of have, um, a well-defined set of criteria of what good looks like, right? Um, this sort of again, the bounded, unbounded sort of spectrum, if you will, of workflows that give you confidence to assent, uh, allow the agent to be able to, uh, execute on its own, right?

So, the answer to that question again is there's no, you know, there's this, there's the ability for the organization, for, you know, for the team to be able to define these criteria, I think will is what, uh, is what enables, um, you know, that dial of autonomy, if you will. Um, I I think, um, we are, we're still, at least what we see today is still sort of significant amount of human in the loop, and that's probably going to be the case till we get to an inflection point in, you know, the way in, uh, the accuracy of these harnesses.

What practical advice do you have for CIOs on managing cost, risk, unpredictability, the variability of inference costs, your token costs going out of control? What it what can CIOs do? What should they do?

Let me focus on kind of the cost aspect of things we we've talked about sort of governance and, uh, you know, the, uh, the ability to deal with the unpredictability of large language models, right? I think, uh, as we said earlier, you have to, you know, software now has a variable cost, and and I think it's important to understand what that means. So, you know, you, you've seen a lot of things in the press, uh, or, you know, on on on X, where you say, like, oh, I'm measuring my team by how many tokens they've used, right?

And and that's sort of, you know, engineers are measured by how what their token consumption is. Uh, the one thing we've clearly seen is in organizations, and there's some data point, I think there was a Goldman Sachs study that came out, um, that basically said that, uh, enterprises are blowing through their token budgets a a year's worth of AI budget in three months. Um, you know, it it's, it, you know, this this notion of token maxing, which is, uh, as many tokens as you provide will get consumed as quickly as possible.

So, we're still in the stage where, uh, I think CIOs need to take a step back and say, OK, I have X number of projects, right? Because we've unleashed the beast, you know, everybody now has access to build their own agents. How do you prioritize these, right? And, you know, just old school sort of making sure that you're investing your tokens in the most highest priority projects. Um, I think that's, you know, what we've always done, uh, historically, in terms of defining software projects.

Just because you're able to build your own agent, uh, doesn't mean that you should. Again, going back to the example of do you want to go build your own ERP or a CRM? Does is it worth it? So, I think prioritization, um, clearly understanding sort of token costs, because now you have multiple agents, it's not just token costs, by the way. You might be using APIs to go access different tools. So, you know, those tools may have their own particular costs.

So, for an agent to actually execute, you have to now break break down the cost of the the platform, the the tokens, the tool uses, all of these things that compose it to figure out and then weigh that against your ROI, uh, criteria. You know, are you actually getting better productivity? How do you measure productivity in granular terms? Is it, you know, calls deflected if you're in a customer support, uh, area, or is it the speed to reconciling, uh, i accounts payable, right, i in the previous example?

So, creating those kind of, um, business level metrics, matching them to your the cost of what it takes for an agent to execute, I think is important. Now, I do think that, um, models are gonna get better, more efficient from a token use perspective, and you already see that happening, you know, all all the way from, like, Opus 4.5 to, like, 4.7 or even the GPT 5.5. They are getting much, much more efficient in terms of token use, and I think that's important.

But also, you know, being creative in the way you use memory and being able to sort of use smart ways to kind of ensure that you're putting the right context in. These are the kind of things that are design principles that help you optimize token cost. But you kind of weighing the token cost versus the ROI, I think is a critical kind of lens in addition to the other things we talked about.

Bridget, uh, Jaggie, uh, on LinkedIn says, How can selling companies charge for their AI solutions? Just in a sentence.

There's a spectrum of do you go to consumption based pricing? Um, you know, I think we've actually done some work at Insight, uh, our pricing, uh, team. And I think one way we're thinking about this is perhaps you think of this as a fractional FTE right um or a head fractional headcount because I think we are going to th- this notion of a hybrid human agent workforce where you know each of us as a uh as employees will have our own agent.

And maybe the way to think about this even from a pricing perspective is to price it as a partial FTE particularly if it's doing you know specific uh work right um but eh this is a point of experimentation again uh there there this is a longer answer but you said ten seconds so I'll let me just leave it at that.

All right Praveen Akkiraju managing director of Insight Partners thank you so much for being a guest today Praveen on CXOTalk. I'm very grateful to you once again. And thank you to everybody who watched before you go. Subscribe to the CXOTalk newsletter go to CXOTalk dot com and we have great shows coming up. Our next show is with the chief technology officer of Mozilla. So check it out. Thanks everybody have a great day.