Agentic AI: What is it AND Does it Matter? (Part 2)

Make sense of the hype around agentic AI. On CXOTalk episode 868, get expert insights on build vs. buy decisions, pricing models, and the strategic implications for your organization.

00:00

Jan 24, 2025
18,088 Views

In CXOTalk Episode 868, host Michael Krigsman interviews Praveen Akkiraju, Managing Director at Insight Partners, to explore the world of agentic AI. Praveen explains AI agents, their reliance on large language models (LLMs), and how they differ from traditional applications. 

Praveen discusses AI agents' current state and reliability, explains the layered architecture involving user interfaces, reasoning, and dynamic querying, and highlights the importance of evaluation loops and human-in-the-loop systems for consistent output. The episode also discusses practical use cases in various sectors like coding, customer support, and IT operations. It explores the complex economics, security considerations, and future trajectory of AI agents in both startups and large enterprises.

The conversation includes a discussion of the report "The State of the AI Agents Ecosystem" and the Agentic AI Market Map.

Agentic AI market map

Episode Chapters

00:00 Introduction to AI Agents
00:37 Understanding AI Agents
01:26 Role of Large Language Models (LLMs)
02:38 Layers of AI Agents
06:19 Current State and Challenges
07:09 Spectrum of AI Agents
09:42 Design Considerations for AI Agents
16:53 Human Element in AI
18:03 Training and Evaluation of AI Agents
23:50 Addressing Bias in AI
26:50 Navigating Constant Change in Business
28:08 AI's Impact on Fortune 500 Companies
29:21 The Evolution and Integration of Large Language Models
30:33 Introducing the AI Agent Market Map
34:20 Key Use Cases for AI Agents
40:19 Economics of AI Agents
43:52 Security Concerns with AI Agents
47:43 Future of AI Agents and Their Evaluation
50:41 Final Thoughts and Farewell

Episode Highlights

Leverage AI Agents as Strategic Tools

  • Consider AI agents as software applications that utilize LLMs in the areas where they provide the greatest value rather than viewing them as isolated LLMs. Incorporate these tools strategically into existing workflows and systems.
  • The technology is still early, with ongoing work needed on reliability, scalability, and consistent performance. Start experimenting while being realistic about current capabilities.

Design Multi-Layer Agent Architectures

  • Effective AI agents require multiple layers, including user interface, reasoning layer, dynamic querying, and evaluation loops. Each layer serves a specific purpose in creating reliable agent performance.
  • Strong evaluation mechanisms and reflection loops are critical for testing output quality and enabling continuous improvement of agent capabilities.

Balance Human-AI Collaboration

  • Current AI agents work best with humans in the loop, particularly for tasks requiring deterministic outcomes. The human role includes approving plans, providing oversight, and ensuring quality control.
  • Carefully design where and how humans interact with AI agents to maximize the benefits while maintaining reliability and accuracy.

Focus on Specific Use Cases

  • The most promising early applications include coding assistance, customer service, IT operations, and specialized analyst support. Identify focused use cases where value can be measured.
  • Success requires grounding agents in company policies, data, and specific workflow requirements rather than deploying generic solutions.

Address Economic and Security Considerations

  • Pricing models are evolving from traditional SaaS to outcome-based approaches that depend on the ability to measure value. Develop clear frameworks for evaluating ROI.
  • Security represents both an opportunity (better threat detection) and a challenge (new vulnerabilities). Carefully consider security architecture when deploying AI agents.

Key Takeaways

Focus on enterprise applications. There are significant opportunities for AI agents in enterprise environments. Prioritize use cases such as coding assistance, customer experience enhancement, and IT operations optimization to drive business value.

Acknowledge AI Agent Gaps. Recognize the evolving nature of AI agent technology and its challenges, such as non-determinism and security vulnerabilities. Adopt a measured approach focusing on security and clear metrics to assess economic impact.

Design for Reliability and Control

AI agents often behave in non-deterministic ways, creating risk for mission-critical tasks. Use reflection loops, human approval steps, and scaffolding techniques to ground the agent in company data and policies.

Target Clear, Measurable Use Cases

Organizations see immediate benefits by focusing AI agents on well-defined tasks such as coding support, customer service, and IT operations. Measure their impact through cost and time savings to justify further expansion. Aim for outcomes you can evaluate clearly, then scale as agent capabilities and reliability improve.

Episode Participants

Praveen Akkiraju is a Managing Director at Insight Partners. He brings a product and operational lens to investing in Automation, Data platforms, DevOps, and Infrastructure software. His investments include companies such as BigPanda, Bardeen, Reco, Rudderstack, and Workato. Praveen spent the early part of his career in the trenches, building products and scaling engineering teams to build world-class platforms in highly competitive market segments. He holds BS and MS degrees in Electrical Engineering and is an alumnus of the Harvard Business School.

Michael Krigsman is a globally recognized analyst, strategic advisor, and industry commentator known for his deep digital transformation, innovation, and leadership expertise. He has presented at industry events worldwide and written extensively on the reasons for IT failures. His work has been referenced in the media over 1,000 times and in more than 50 books and journal articles; his commentary on technology trends and business strategy reaches a global audience.

Transcript

Role of Large Language Models (LLMs)

Michael Krigsman: What is the role of LLMs, large language models, in this whole agentic AI world?

Praveen Akkiraju: It's essential to understand that AI agents are not just about large language models. You have to think about AI agents as essentially another software application that now incorporates large language models in areas where they add the most value.

Let's take a step back and think about how applications were previously defined. You used to have a database, a system of record, used to have a software workflow that was built on top of this, and then you had a user interface. That's typically what your classical application looks like today in the SaaS context.

The power of large language models is that they can be inserted into the stack at various points and can dramatically improve the productivity and capability of these application software models. You can think of large language models as playing different roles in this new paradigm we call AI agents.

Layers of AI Agents

Praveen Akkiraju: The first layer is the user interface. We are all now used to the ChatGPT interface where we go in, and you essentially type in something, or you can even speak to it. It's able to understand that context and comes back with an answer as opposed to a set of reference points or ten blue links.

Similarly, in the application context, you have a user interface now where the user can interact and basically say, "Hey, show me the data for a particular region for a particular product," and that LLM is then able to take that input and synthesize that and understand the context of it as opposed to the user having to figure that out.

That's one layer. The second layer, which is a really important layer and kind of in some ways sort of almost the core of an AI agentic architecture, is the reasoning layer - the ability to take the task given to the agent and break it down into a set of subtasks.

For example, if you're saying, "Let's go build a research report on a particular stock." When the user provides that prompt, the AI agent breaks it down and says, "How do I build a research report on this particular stock? I have to go to Yahoo Finance or one of these public websites, get all the information about that stock, be able to synthesize all of that information, create a report and then publish a report." So it breaks that down to specific tasks. That's the reasoning layer.

The next layer where the LLMs play a role is the ability to go out and dynamically query information. In the past application context where things were static, the application could only look at things where it had an API call built in. If it had access to a particular data store and had an API call, it would use that API call.

Here, we have an LLM that can reason and say, "Wait a minute. This information is available to me externally. This information is available to me through this API call internally, and I need to get both of these pieces of information to execute my task, which is to build this research report." So it launches a web search API and also launches an API to the internal research store to put those data points together.

The LLM has the ability to understand what sources of data it needs to get information from to execute the task. Once it does that, the application software synthesizes it, and then you come to the next step, which is the output.

A lot of really well-designed agents have strong evaluation loops, and this is really important. Once the agent actually comes up with the output, you test that output against what essentially is the gold standard that you set up and say, "This is how a research report looks like." The agent compares the output to the gold standard and says, "Does this match it?" And then it corrects itself if needed.

So as you can see, you're inserting the large language model at different layers of the agent in order for it to be able to make the entire process a lot more dynamic and interactive. That's basically what is different between a classic application flow and an agent application flow.

Michael Krigsman: So we have this reasoning, we have access to this broad body of human knowledge, and of course that is different from traditional applications.

Current State and Challenges

Michael Krigsman: How accurate are these agents today? What is the state of the art? How useful are these agents in practice?

Praveen Akkiraju: We are quite early on this journey. When you read the press, it feels like AI agents are everywhere. If you look at a website today for any software company, essentially AI agents are the way they are now expressing their products.

However, I think in terms of the maturity of the technology, both from the important question you talked about - how reliable they are - but also in terms of the customer end user side, the experience of the end users in terms of the deployments and the scalability and reliability of these, we're still in the early days now.

Spectrum of AI Agents

Praveen Akkiraju: There's a spectrum of agents. At one end of the spectrum, you could have these consumer agents, which are really focused on individuals. We were talking about this before the show started. OpenAI just launched their consumer agent called Operator. Very interesting. It's for us as individuals, part of the pro plan.

On the other end of the spectrum are essentially the next phase of evolution of RPA or robotic process automation in the enterprise - agents that take on complex enterprise workflows.

We have tremendous amount of activity all along the spectrum in terms of startups as well as incumbents advancing the state of the art, experimenting with agents. We are in the early days primarily because there are a few things that we're still trying to figure out.

Large language models by themselves are non-deterministic. What I mean by that is - and we've all experienced this - when you ask ChatGPT a question a certain way, and you ask the same question in a slightly different way, you may get a different answer. Now, that's getting better with some of the newer models, particularly some of the reasoning models that can correct themselves.

But the fact of the matter is that the core of the large language model is this sort of non-determinism. And so a lot of the AI agentic designers today, builders today, are working on what we call scaffolding in order to take the power of these large language models and harness them while making sure that we understand that the non-determinism exists, and we have the right architecture to handle that so you get a reliable, consistent and most importantly, scalable AI agent.

Michael Krigsman: Where are we there because of course that non-determinism, if you press enter, submit again on a prompt, it's going to give you a different answer. That's great if you're wanting help writing an outline or some type of summary of a document, because there's different perspectives you can take, but if you want it to run a task, like book me an airline ticket, you don't want a lot of stuff all over the map. You just want one thing done. You want that ticket to be booked in the right place with the most efficiency and so forth. So it becomes a big problem, I would think.

Design Considerations for AI Agents

Praveen Akkiraju: Yeah, and I think this is, again, a central design consideration when you're building AI agents. And I think we can maybe break this down into three parts.

The first part is the task itself. And you gave an example of a task. So how important is it for you to get the task absolutely right? Now, in ChatGPT, if it gives you a slightly different answer, it's like search - you're getting information. You're not essentially deciding. So you're okay with some level of non-determinism because the human mind can correct for that.

The other end of the spectrum: if you're essentially betting on this agent to execute a task consistently, it could be an enterprise workflow, a back-end workflow, code generation, or a customer service interaction agent. You cannot have that level of non-determinism. So that's one thing: How do you define the task? And I think that's a key question to ask.

The second aspect you want to look at is, okay, now that you say you have a task that you need to be accurate - and it's also essential to understand that AI agents are not all about just LLMs. There's a lot of existing software, there's a lot of reusing machine learning models, predictive models, which are deterministic, and the classic things that you as a software engineer do to build applications that go into making an agent.

So I'll say this again: LLMs are a tool. They're not the product. They're not the agent, they're a tool. So you have to understand, as with any tool, what the capabilities are.

The third piece is, okay, so now you've figured out that your large language model will help you build a plan, for example, in the reasoning layer that we talked about earlier. Most AI agents today essentially propose a plan, and you have a human in the loop that then approves the plan.

A good example of this is a lot of the coding copilots and coding agents that you have today. They're very good, particularly with GPT-3.5, which was like a step function jump in the ability of large language models to generate code accurately. The way they work is essentially the programmer's interacting with the large language model. It proposes a plan, which is approved and edited by the programmer before it actually goes out and is executed. So the user in the loop, human in the loop is a very critical design component today in AI agents, particularly in that planning layer.

Now, there are other things you can do, and we talked about part of this, which is what we call a reflection loop. So once the output of the agent comes out, you test it against another large language model, which is trained with the right output. So the agent essentially tests itself to say, "Did I get this right?" And it's able to think about whether the output is correct and then make those changes and go back and iterate again.

So these reflection loops, the way you build evaluations, which is how you determine the output is correct and using that data to continuously improve the agent, is again part of the design process.

So I gave you like three different things to consider. There's a lot more we can go into in terms of depth, but at a broad level, the takeaway is that you can design AI agents to be deterministic, assuming that you understand the task, you provided the right scaffolding, the evaluation loops, and you essentially involve the human in the loop. Today, AI agents work well when you have a human in the loop, and I think that's going to be the case, particularly on this end of the spectrum where output needs to be deterministic and decisions need to be a lot more accurate.

Michael Krigsman: Subscribe to the CXOTalk newsletter so you can join our community and we can tell you about our upcoming shows, which we have great ones.

We have questions that are stacking up on LinkedIn. So let's jump to a few questions right now. And the first question is from Ravi Karkara. And he says, where do you see American universities in creating a skills workforce for the AI-driven world economy? How will they learn to work with and deploy AI agents?

Praveen Akkiraju: I consider AI large language models a tool, right? Just as we had cloud, mobile, all these fundamental platform shifts. AI and large language models are tools to help us accomplish our tasks. So what I mean by that is it still is important for you to deeply understand the problem you're trying to solve with this particular tool.

Are you trying to book an airline ticket? Like Michael, you had mentioned earlier. Are you trying to execute a payroll function in an enterprise? Are you trying to respond to a customer support question? So understanding the actual task, which is what good product management is, is foundational to understanding how this tool - this new tool, much more powerful, dramatically more impactful than anything that we've seen in the past - is going to change how we as educators, as knowledge workers, as consumers, leverage AI.

In terms of how universities approach this, it depends on how you participate in this. On one end of the spectrum, education in sciences and math and good grounding in that helps you be part of the design process. On the other end of the spectrum, if you're more business-oriented, a deep understanding of your problem and essentially how the tool works is important.

We all are experimenting with ChatGPT today. I use it differently. My daughter uses it differently. You know, hopefully none of our teachers are listening, but she uses it sometimes to help her with her homework. And so she's learning as part of that process, just as I'm learning to use it for my use cases and so are all of us.

So I think it's a tool that we experiment with, but a deep understanding of the problem and figuring out how to understand how to use this tool is foundational. Product thinking is another key aspect, which I think is something we should emphasize - if you're not deep in the algorithms and math, you still have a tremendous role to play by understanding how to build products and solve customer problems.

Human Element in AI

Praveen Akkiraju: And I think the third aspect of this is there's a huge human element to this. We just talked about how AI agents function today. And one of the things I said is that humans are in the loop. Will AI agents get better at reasoning? Probably. The GPT-3 models are amazing, and you've seen huge advances just in the last few weeks, last part of last year in terms of the reasoning capability.

However, throughout the steps, you have to think about where the role of the human is. Is it an evaluation function? Is it a reasoning function? And so humans are always going to be involved and being able to understand and engage with the technology is effectively the most important thing. There is a massive human side of this as much as there is a technology side of this. So those are some of the areas - I'm not an educator, so that's the way I think about it.

Michael Krigsman: This is from Suresh Babu Madala, and he says, do we need to train agents or do agents interpret the question as a step-by-step task? That's a great question.

Training and Evaluation of AI Agents

Praveen Akkiraju: You do need to train agents. Again, there's a spectrum of what these agents can do in terms of simple tasks versus complex tasks. The level of training obviously will depend on what you're expecting the agents to do.

The first phase as you're inserting the agent into a particular task, there's a certain amount of grounding that needs to happen, and the grounding typically is connecting it to the right data sources, connecting it to the right application sources, connecting it and understanding, making sure that it understands and is grounded in the policies of your particular use case or your particular enterprise.

However, training is just like we are as human beings - we're constantly learning. If you onboard a new employee, a fresh college grad, they're in a constant training process. They learn, you have somebody who watches their output and you give them feedback, and hopefully you observe them continuing to improve.

It's the same exact thing for AI agents. That's why these evaluation loops are such an important aspect of designing an AI agent. I can't stress that enough. You have to be able to constantly understand the output of the agent, figure out where you can correct it and continue to improve.

And I think the last part of this, the observability of how these agents are functioning - where are the broader efficiencies and inefficiencies of what they are doing and not doing - is a constant part of your architecture.

Michael Krigsman: This would be an excellent time for everybody listening to subscribe to the CXOTalk newsletter. Go to CXOTalk.com. If you're watching on LinkedIn, look at the address on your screen and subscribe to our newsletter so we can notify you of upcoming shows and you can be part of this amazing CXOTalk newsletter.

Our next question is from Greg Walters, and you kind of addressed this a little bit, but he says, can't non-determinism be prompted into existence?

Praveen Akkiraju: The first aspect of this is to really understand the role of the LLM in your task and what is the level of reliability that you expect from an LLM's task. Now, there are certain things where the capability of the LLMs is getting constantly better. There's a reason why a lot of large language models are fundamentally trained on math and on coding tasks because there's a lot of transfer learning - as they get good at coding, as they get good at math, they're also able to get good at reasoning tasks, which are much more broadly applied.

So the first point I'll say is these models will progressively get better at managing the hallucinations, whether it is through reasoning loops, or whether it is through better post-training of the models in your deployments or whether it's inference time reasoning. This is sort of another scaling level that we now have. There are different techniques in the model itself which allow you to decrease the aperture of the non-determinism.

The other vector is, as I mentioned earlier, building the scaffolding around it, understanding that you're going to get an output from the large language model that needs to be synthesized into something that's a little bit more reliable and gets to the level of output that is acceptable for you. That's a reflection loop. That's your evaluations. And sometimes you may just have a deterministic runtime that you need to design to the AI agentic workflow.

Michael Krigsman: And Gus Bekdash says, it's interesting that agentic AI goes around the ridiculously frustrating prompt user interface that is not integrated with any workflow. This is huge. What are your thoughts about agentic AI simply around the user interface and integration with workflows?

Praveen Akkiraju: This is the quantum leap, in my opinion, right, in terms of the user interface and user engagement. In many ways, I think we've had some form of natural language processing, NLP, type interfaces. You can think of chatbots today. It's hard to escape them - if you're trying to book a ticket or anything, the first thing that pops up is a chatbot that's trying to get your information.

So that's natural language processing. It's able to understand voice, translates it into text, does a search, responds back to you. I think the opportunity with large language models is the ability to infer context. What we had in the previous generation of chatbots was a literal translation and a static sort of rules-based interpretation of that translation.

So what large language models have now, because they've trained on like the entire corpus of human language data, is they can infer context, they can infer tone, they can infer the particular intent and they're able to then appropriately translate that into their query and get you back a response.

So I think it's game-changing that you would have a large language model from a user interface perspective. Now remember, today we're still very text-based. Most ChatGPT interactions, though they have voice mode which is amazing, are still text-based, but think about the ability for us to be multi-modal - our ability to do voice, which we're there today, ability to input images and videos, which we're going to get to. Over a period of time, these models will effectively get to this point where we call they have a world understanding, understanding of the world model. And I think that could be game-changing in terms of how we interface with this technology.

Michael Krigsman: And this is from Arsalan Khan, who asks a very interesting question.

Addressing Bias in AI

Michael Krigsman: He says we want bias to be removed from data when it comes to AI; how do you remove human bias if humans are in the loop? And who decides what's a bias or not? It's a really interesting point.

Praveen Akkiraju: If our vision eventually is that, as some companies have articulated, every company has an AI agent and that's sort of the first interface point for a customer interacting with the company - let's say you're an insurance company or a government service, like maybe the DMV at some point - that interface is really important.

So, I think, look, obviously, the model companies are doing a lot of great work in making sure that we are conscious about bias. No, it is a fact that that is not a perfect solution yet. We're not quite there yet. In certain instances, you're not getting these perfect answers.

So part of this is the context I talked about regarding how you ground the agent is really important. And so what does context really mean? Context means examples. And so if it's a customer agent, for example, you could train it on the policies that you have, you probably have a lot of voice recordings of existing agent calls. There are great examples of how to handle situations in bias or confrontational situations.

So the training of this agent, grounding it in the policies and the rules of the particular use case is a particularly important task. Now, I think, look, the human in the loop is really about how you reason things. And obviously, look, the way a human interacts ideally is a positive in terms of our ability to eliminate bias. By inserting human in the loop, you're actually adding a step that improves the ability of the agent overall to correct its biases.

But I would say good training, grounding and policies, and obviously having responsible humans in the loop are the things that will help us get there. But it's an imperfect process. And that's why I say it's early days yet.

Michael Krigsman: Arsalan Khan comes back, and he says if subject matter experts train AI to create AI agents, would we really need the subject matter experts? What happens when AI agents come across a scenario that they haven't encountered yet?

Praveen Akkiraju: Training is a constant process for AI agents. What you get when you initially start this process with a subject matter expert helping you ground the model, ground the AI agent, is you're giving it a certain rubric. You're basically saying, "Hey, here's basically what is expected, here's a basic task, here's how you perform the task."

Navigating Constant Change in Business

Praveen Akkiraju: Maybe you launch a new product. Maybe you expand into a new region. Maybe you make an acquisition or you have new employees onboarded. It's a process of constant change.

So to some extent, I think you can design the agent to say, "Okay, I can expand into a new geography. This is how I understand it." But there may be different policies. As is often the case, if you're operating in a different country, they may have their own rules and regulations that you need to now ground the model in.

So I think the subject matter expert is really - as it is, and we all do this - we're all subject matter experts. We're not static. We're constantly learning, evolving, understanding new technologies, new pieces. And it's the same thing for the model.

So yes, the idea of the AI agent is to take away these mundane things - go get this document, put 10 documents together, create a research report. Yes, you don't need to retrain that thing, but being able to say, "Okay, how do I operate in the European Union, which may have a different set of rules, or in Asia, in Japan or in China or in India, which may have a different set of rules" - those are things that there's a certain amount of requirement of understanding those rules and regulations that need to, again, you need to train the agent on.

AI Impact on Fortune 500 Companies

Michael Krigsman: Let's start talking about business and Mario Garcia asks, he says, it's inspiring to see the impact of AI in fortune 500 companies. What insights about this stand out to you the most?

Praveen Akkiraju: A lot of these large companies today, I mean, it's been amazing to watch how rapidly both large companies as well as incumbents and startups have really embraced AI and large language models. And it's largely, I think, the ChatGPT moment which unleashed this because it was so accessible.

So you take a step back, you know, AI has been around for a long time. I did a course in neural networks back in my grad student days. What's changed in this generation is that today really powerful, complex AI models are available on the other end of an API call. That level of simplicity in terms of access to really powerful technology is essentially what enabled us to unleash large language AI as you see it in a lot of these use cases.

The Evolution and Integration of Large Language Models

Praveen Akkiraju: Now, I will again caution you to say that we are still very early. If you talk to a lot of these large corporations, they do have large language models integrated. Most of them have rolled out, for example, some form of a coding copilot. They've rolled out some form of customer support function. They've rolled out some form of an analytics sort of use case with large language models.

But we're still not deployed at scale primarily because we're kind of still tuning, tweaking, learning in terms of how do you manage AI agents? How do you manage biases? As one of your audience members just asked, how do you make sure that they are current? How do you make sure that they don't hallucinate?

The most important thing is fundamental - we all know this. It doesn't matter if there's a lot of really fun, exciting, interesting demos on X. Those are great, you can say, "Oh, wow, this agent can do this." What's really important is can it do it consistently and can it do it at scale? So those are the questions we'll answer this year, hopefully, and that's why we're so excited in 2025 about the trajectory of these AI agents.

Introducing the AI Agent Market Map

Michael Krigsman: Praveen, you and your team put together what you call a market map of companies involved with AI agents. Can you tell us about that? And I can bring it up on the screen so everybody can see. And there it is. Praveen, can you talk about the market map you've assembled?

Praveen Akkiraju: The market map is a dynamic living thing, in the sense that it will evolve constantly primarily because we are seeing so much activity and energy around the AI agent space. So what we try to do is to construct this in layers.

There's a foundational layer where there's a lot of these kind of data sources, integrations and such. There's this new sort of bucket we call the agent computer interface which is the ability for AI agents to use computer tools. And it could be integrations, it could be web tools. Some of the stuff is integrated in models. Some of these are interesting platforms that are created for this.

So we try to construct the model layer by layer - what's the bottom layer? Where are the data platforms? What is the middleware layer, if you will, which is the agentic computer interface, as well as a lot of these agent frameworks?

You know, we are investors in a company called Crew AI, for example. And I'm sure anybody who's talked about AI agents probably knows about Crew AI. It's one of the most popular open source frameworks out there. You can use Crew AI to build agents. Others like Langchain do something similar.

Above that, then what we try to do is to say, "Let's try and map out where is the energy in the AI agent space." And it's important to understand, I think it's on the left side of the market map, there's a lot of AI agentic products and offerings from incumbents.

Obviously Salesforce launched their own agentic workforce AgentForce, as they call it. Microsoft has copilots. We just saw OpenAI launch Operator, which is a more consumer-oriented agent. So effectively, all the incumbents have said, "Hey, we've got these great customers. We've got all these great use cases. Is there a way for us to improve our customer experience or productivity by creating an agentic workflow on top of our existing software?"

On the rest of the market map, you can see tremendous amount of energy in different verticals, particularly in coding, for example. There's been huge amount of great work. Cursor by far seems to be the most popular among developers today, but you also see very specific vertical agents - sales, marketing, legal, finance. So you can take advantage of almost each of these different functions and you can see companies building agents which are customized for that particular vertical use case.

This is really interesting company called Samaya, for example, that's doing some amazing work focusing on building agents for the finance workflow. So you see that the idea of a market map was not to be precise and capture the entire view. And I do apologize to a lot of the builders out there, some of whom we missed in the market map clearly. The idea really was to kind of give you a perspective of what this landscape looks like, where is the activity, how are builders approaching the agentic space?

Michael Krigsman: What are the opportunities and the use cases, the predominant or the most important use cases for AI agents?

Key Use Cases for AI Agents

Praveen Akkiraju: They're probably like three or four buckets. The first one, obviously, that everybody knows and understands very well are these coding agents, coding copilots, coding platforms, whatever you want to call them. And depending on the particular style, Cursor has a particular way of working - it's more like a copilot. If you take something like Devon, it has a different sort of way it engages, it fires off a bunch of agents that execute your plan and such like. But there's a lot of energy around developer facing AI agentic work.

The second area is in the customer experience. Customer experience, obviously everything from customer support agents, which obviously is the biggest use case. As we were talking about earlier, chatbots are already a fact of life. Can we make that experience much more realistic, much more engaging? So you're not basically saying "agent" as soon as you get a chatbot. So there's a ton of energy in that space. Lots of great companies building interesting products.

The other area which is kind of interesting is in the operations space. So if you think about operations, broadly speaking, it could be IT operations, it could be security operations. You have sort of this needle in the haystack problem. You have a lot of data, you have a lot of alerts that come in and you're trying to figure out, "Okay, which ones do I pay attention to?"

So this is actually a perfect use case for agents - the ability to synthesize all of that information. And if it's grounded in your policies and in a company's particular way of doing things, it's able to say, "Hey, here's maybe the top three to five alerts you need to pay attention to. This is the problem that they're articulating. And here's a few ways to remediate this."

So we're talking to some interesting startups that are actually focused in this particular space. So I think those are the three. And there's a lot more, but I'll just pause there in the interest of time.

Michael Krigsman: We have a question from Elizabeth Shaw, and this is on Twitter, who asks, how are organizations using agentic AI in their business and their ecosystem?

Praveen Akkiraju: Let's just take, for example, a customer service AI agent. That's a use case that we're seeing a lot of customers experimenting with. So what does this customer service agent do? You effectively, again, think of a chatbot today. The customer service agent is able to first of all be grounded in all of your data, your FAQs - how do I reset my password? How do I recover my password? Or how do I add my child to the insurance policy?

These exist as continuous tasks that you were able to then interact with an agent about. The agent understands what you're trying to do. Maybe you're looking for a form. Maybe you're looking for a website. Maybe you're looking for a particular quote or something like that. So I think that is a use case that is getting a lot of traction. We're seeing a lot of our customers looking at rolling that out into production.

The other use case, as I mentioned, is on the coding side. And we're not just talking about like an IDE, like Cursor, which obviously has a lot of broad adoption, but things like automated testing and provisioning. So you have software, you need to roll this out. Testing is a really critical part of that process of rolling things out. You're able to actually use AI agents very effectively in testing on red team kind of use cases where you can see if you can break it. That's a very critical function where you're seeing some level of deployments happening.

The other one, as I said, is IT operations. And this is very exciting because again, if you know, these are mission critical. You need to be constantly up. Most of these IT teams have 24/7 coverage because you cannot have critical systems going down. And so an agent is perfect because it has the ability to synthesize large amounts of data. It has the ability to solve the needle in the haystack problem.

Now again, work in progress. I wouldn't say all of these things are at perfection, but we are definitely seeing these. And then I mentioned this company, Samaya, which is very interesting. They are actively building like a finance analyst kind of agent, which is pretty accurate in terms of being able to extract context out of research and provide you really focused information.

Michael Krigsman: So you say that it's really accurate. I'm assuming that what you also mean is that it is consistently reliable and predictable.

Praveen Akkiraju: Exactly. And it's much harder to do that straight out of the box. I think a lot of times there's a lot of confusion in the market about like, "Hey, wait a minute, the large language model is just going to keep getting better and better and better. And ultimately they'll just be like one model that solves all."

And it is true in certain simple use cases. I mean, absolutely. The models are getting better. They're reasoning better. Maybe we will get to a point of artificial general intelligence where these models can just use computers like we do. And maybe that's the bar. But I think when you're looking at complex enterprise workflows, as I mentioned, the ability for the agent to be grounded and to be accurate and to present data in the way the end user expects will require some amount of post training, some amount of inference time reasoning as well as some amount of scaffolding in order for you to build the perfect agent.

Economics of AI Agents

Michael Krigsman: Can you talk about the economics of agents and then we'll jump back because we do have additional questions that have come in, but the economics are really important. So what are the aspects that drive economics and what do enterprise buyers need to think about when it comes to the economics of agents?

Praveen Akkiraju: You've seen sort of the whole spectrum of conversations, right? Everything from, like, with AI agents, we're just going to go entirely to outcome-based pricing to like, well, it's still software, so we have to kind of figure out how do you make sure that you're able to charge for it appropriately.

The fundamental question to ask when you think about pricing is can you measure the value of the AI agent output accurately? So what do I mean by that? Let's say you have a customer support agent. You can basically say, "Hey, the customer support agent handled 100 calls. It ended up taking me X amount of dollars to handle those 100 calls." The customer agent handled those calls. So I can attribute directly a value to the output of that agent. It had 100 calls. Each call is worth X. So there's basically 100X is basically the value of that particular agent's task.

The other end of the spectrum: Let's say you're doing a research workflow. So you've generated a research report or you're helping an analyst basically with their research and you improve their productivity. How do you measure the value of that? Do you measure it by individually asking the analyst how much more productive they were? It's much harder to quantify certain tasks versus certain other tasks.

So I think the first question in understanding economics of agents is, can you attribute value in a reasonably accurate way to the output of the agent? Based on that, if you can, then outcome-based pricing is essentially where we're eventually going to go to. If it's a lot more nebulous, then I think what we're going to see is some form of an evolution of the existing SaaS pricing model. So you might pay a platform fee for the agentic thing, and then maybe you hire an agent. So you pay on the number of times you run the agent. So it's some combination of that.

So we see this again as a spectrum. There's no absolute here. There's a lot of experimentation today. In some ways, companies are still trying to figure out how the customer is getting value. Customers are trying to figure out - if you're a CFO, you're used to paying subscription software. Okay, I've got X licenses for one year, and I can budget that. Now, if you go to sort of this outcome-based pricing, again, if you don't have an accurate sense of value, how would you, as a CFO, budget for these agents? So I think there's a lot of these things that need to be worked out. We're experimenting and understanding eventually where this direct attribution of value, I think we will end up in the outcome-based pricing bucket, but there's also going to be a lot of these intermediary models where you want to make sure that the developers are getting a fair value for the product that they're building and the customers are paying a fair price for it.

Michael Krigsman: So ultimately, when we reach the point where agents have discrete measurable output results, then we can move towards performance-based pricing. And until then, it's essentially usage.

Praveen Akkiraju: I think that's a good way to put it.

Security Concerns with AI Agents

Michael Krigsman: We have an important point now raised on LinkedIn by Naresh Kumar, who is VP and general manager of product management at Zscaler, and he raises the question, what about security and agentic AI? And we haven't talked about that, so I'm glad you brought this up.

Praveen Akkiraju: Large language models help with the sort of needle in the haystack problem, which is inherent to diagnosing security problems. I kind of grew up in the networking world and we used to build these large global scale internet scale networks. And a big part of the task was like, if there was an outage somewhere, to debug that would essentially mean we synthesize tons of data and figure out where we need to focus our efforts.

Security is the same way - you have a large aperture of exposure depending on the type of company you are, everything from your network to your applications, your devices, to individuals, to identity. There's multiple layers when you think about security.

And it's been a tough challenge. In the security industry, we've had these platforms called SIEM, which try to bring all this together and be able to give you like a unified view where you're able to manage this. But a security ops center is essentially the nerve center of how most companies run their security operations.

So, I would look at the role of LLMs in security in three ways. The first, I think, is from an operations perspective - I think it could be a very useful tool because it has the ability to synthesize large amounts of data and help in that needle in the haystack problem or prioritizing. I think it's a great use case.

The second one is elements integrated into the security products. Will essentially, you know, you talk about again, a security agent, existing security software being able to dynamically understand policy, dynamically able to respond to you adding more users, etc. I think you're able to sort of build those. We're seeing companies starting to build large language models into their software stack, just as we talked about earlier. It's a tool, right, where it's useful.

The third I'd say is, look, LLMs do represent a new threat aperture, particularly if the models essentially hallucinate or, for example, in that psyops use case, ignore or highlight or miss critical threats. So, while you're designing, while we talk about all these agents being deployed in a customer support use case or a finance analyst use case, if those large language models are not sufficiently grounded and they're not, you know, the data, the training set that they have is not protected appropriately, you risk not just hallucinations, but effectively a hijack of the entire agent.

So it's early days in that. I think we've had some interesting conversations with founders who are thinking about this problem in deep ways and building interesting things. But it's a problem space at this point. We will learn more and we will evolve the security architecture just as we're evolving with the maturity of the AI agents itself.

Michael Krigsman: Okay. And obviously Zscaler is thinking about this because he asked that question.

Praveen Akkiraju: Yeah. They're an important player and they have a huge role to play in our overall architecture.

Michael Krigsman: We have another question from Arsalan Khan. I'll ask you to answer this really fast because we're just going to run out of time now, who says, should we create a time and motion AI agent that assesses other agents if they have saved money in an organization, obviously referencing back to the pricing discussion we had earlier.

Future of AI Agents and Their Evaluation

Praveen Akkiraju: If you kind of listen to some of the industry luminaries talk about, like, we all have some of this army of agents or we will have human employees and agentic employees. There is a requirement to train all these agents, ground all these agents as well as to evaluate all these agents.

So we talked about reflection loops in terms of the specific sort of output governing these outputs. So similarly, at a higher level of abstraction, which is the value, yeah, it is an interesting idea to be able to say you have an agent that's constantly measuring the value of the output of other agents to ensure that they are meeting a particular mark.

For example, if your customer support agent, if it's not deflecting whatever, 100 calls a day or something like that, and the metric may vary, then maybe it's not performing appropriately. So it's more like an operational function. So there are ways to, I think in this sort of agentic future, there are the worker agents and there are the evaluation agents. And you potentially will have manager agents at some point.

You can think of a future where there is some level of different levels of hierarchy where you are actively evaluating and governing these agents. But again, we're early days yet. We're a little bit sort of hypothesizing how this looks like.

Michael Krigsman: Gus Bekdash comes back and he says, is it really better to have the agency and the knowledge in one model or have them as separate loosely integrated systems?

Praveen Akkiraju: It'll come back to the type of problem space that you're addressing. You know, in general, we know from just the things around us that there's no such thing as one unified body of knowledge. We as human beings, there's so much complexity in our world, in our workplaces, in our consumer-oriented lives that there's no such thing as like one mega intelligence that essentially does everything for you.

So I think we always solve problems by breaking them down, breaking them down into smaller problems, and then using different tools to solve those problems and putting these things back together. That's sort of the way human workflow happens, irrespective of whether you're building a chair as a project, or you're building a complex application in an enterprise.

So, I will again hypothesize here and say that I don't believe in this sort of single unified agent. I do think agents will continue to get better. Maybe we will get to this threshold of artificial general intelligence. How will you define it? That's another hour's conversation. But I think you will always have this notion of taking a problem, breaking it down, using different tools to solve it, and putting it together and you can think of that same architecture applying in an agentic world as well.

Final Thoughts and Farewell

Michael Krigsman: Okay. And with that, this has been an action-packed hour. Praveen Akkiraju, thank you so much for taking time to share your expertise and knowledge with us today. I really, really do appreciate you.

Praveen Akkiraju: Thank you. Thank you for having me.

Michael Krigsman: A huge thank you to everybody who watched. Before you go, subscribe to the CXOTalk newsletter so you can join our community, and we can tell you about our upcoming shows, which we have great ones. You can ask your questions during the live show just as today, and with that, a huge thank you to everybody and to Praveen.I wish everybody a great day, and we'll see you again next time. Take care.

Published Date: Jan 24, 2025

Author: Michael Krigsman

Episode ID: 868