Latest / CXOTalk / An Advanced Look at AI Agents, Part 2 | #868
Transcript
- 0:00What are AI agents and will they really change the world as some
- 0:05folks claim? Today on episode 868, we're
- 0:11taking an advanced look at agentic AI with a prominent VC
- 0:17investor. Praveen Akiraju is managing
- 0:20director at the venture capital firm Insight Partners, where
- 0:25he's focused on AI agents. This is his second appearance on
- 0:29CX, so talk to discuss Agentic AI Praveen.
- 0:33Welcome back. Thank you, Michael.
- 0:35Excited to be here. What are AI agents?
- 0:38You can think of AI agents as application software that is
- 0:44able to understand a user's intent and be able to reason on
- 0:49that user's intent, come up with a plan and execute the plan.
- 0:53At a broad level, that's really what an AI agent is.
- 0:56So today, if you kind of look at how that task is done, you
- 1:00typically have an application, you have a human who comes up
- 1:03with the plan and executes the plan, right?
- 1:05So the, the transition that we are in the process of making
- 1:08with AI is that we're now offloading a lot of that
- 1:12planning process and the execution process to these
- 1:15agents and allowing them to use applications as humans would,
- 1:20right? So that's really the paradigm
- 1:21shift that we are in the process of making.
- 1:24That's why AI agents are so exciting.
- 1:26What is the role of LL miss large language models on this
- 1:32whole agentic AI world? It's important to understand
- 1:37that AI agents are not just all about large language models,
- 1:42right? You have to think about AI
- 1:43agents as essentially another software application which now
- 1:50incorporates large language models in in areas where they
- 1:54add the most value. So let's take a step back and
- 1:57think about sort of how applications were previously
- 2:00defined. You used to have a database, a
- 2:03system of record. You used to have a software
- 2:07workflow that was built on top of this.
- 2:09And then you had a user interface.
- 2:10That's typically what your classical application looks like
- 2:14today in the SAS context. So the power of large language
- 2:19models is that they are able to insert into the stack at various
- 2:24different points and be able to kind of dramatically improve the
- 2:28productivity and the capability of these application software
- 2:33models. So you can think of large
- 2:35language models as playing different roles in in in the
- 2:38sort of new paradigm we call AI agents.
- 2:40The first layer would be the user interface itself.
- 2:44You know, we are all now used to the ChatGPT interface where we
- 2:47go in and you essentially type in something or you can even
- 2:50speak to it and you know, it's able to kind of understand that
- 2:53context and comes back with a an answer as opposed to, you know,
- 2:57set a reference points or the 10 blue links, right.
- 3:01So similarly, in the application context, you have a user
- 3:04interface now where the user can interact and basically say, Hey,
- 3:08show me the data for a particular region for a
- 3:12particular product and it and that LLM is then is able to take
- 3:16that input and synthesize that and understand the context of it
- 3:20as opposed to the user having to go and figure that out, right.
- 3:23So that's one layer. The second layer, which is
- 3:25really important layer and it's kind of in some ways sort of
- 3:28almost the core of an AI agentic architecture is a reasoning
- 3:32layer, right, which is the ability to take the task given
- 3:37to the agent and break it down into a set of sub tasks.
- 3:41So for example, if you're saying let's go build a research report
- 3:45on a particular stock, right? So if that is the the, the
- 3:50prompt that the user provides, the AI agent then takes it down
- 3:54and says, OK, how do I do build a research report on this
- 3:56particular stock? I have to go to, you know, Yahoo
- 4:00Finance or one of these public websites, get all the
- 4:02information about that stock, be able to kind of synthesize all
- 4:06of that information, create a report and then publish a
- 4:10report. So it breaks that down to
- 4:11specific tasks. That's the reasoning layer,
- 4:14right? The next layer where the LMS
- 4:16play a role is the ability to go out and dynamically query
- 4:21information. So in the past sort of
- 4:24application context where things were static, the application
- 4:27could only look at things where it had an API called built in.
- 4:31So if it had access to a particular data store and had an
- 4:34API call, it would go pick that API call.
- 4:37Here we have an LLM that can reason and say, wait a minute,
- 4:40this information is available to me externally.
- 4:43This information is available to me through this API call
- 4:46internally. And I need to get both of these
- 4:48pieces of information in order for me to execute my task which
- 4:51is to build this research report.
- 4:53So let me launch a a web search API.
- 4:58Let me also launch an API to my internal research store, put
- 5:04those data points together, right?
- 5:06So it's, it's the LLM has the ability to understand what
- 5:10sources of data needs to go get the information from for it to
- 5:14execute on the task. Now once it does that right,
- 5:18they, the application software synthesizes it and then you come
- 5:21to the next step. It is.
- 5:22You have the output. A lot of really well designed
- 5:25agents have strong evaluation loops and this is really, really
- 5:29important, right? Once the agent actually comes up
- 5:32with the output, you test that output against, you know what
- 5:36essentially is the gold standard that you set up and say this is
- 5:39how a research report looks like.
- 5:40So the agent compares the output it got with the gold standard
- 5:45and says, does this match it? And then it corrects itself.
- 5:48Did it matches that standard, right.
- 5:50So as you can see, you inserting the large language model at
- 5:54different layers of the AI agent in order for it to be able to
- 6:00make the entire process a lot more dynamic and interactive.
- 6:04And that's basically what is different between a classic
- 6:07application flow and AI agent application flow.
- 6:11So we have this reasoning, we have access to this broad body
- 6:19of human knowledge, and of course that is different from
- 6:23traditional applications. How accurate are these agents
- 6:30today? What is the state-of-the-art?
- 6:32How useful are these agents in practice?
- 6:36We are quite early in this journey.
- 6:38You know, it feels like when you read the press you have, you
- 6:41know, AI agents are everywhere, right?
- 6:44Or if you look go look at a website today for any software
- 6:47company, essentially AI agents are the way they are now
- 6:52expressing their products. However, I think in in terms of
- 6:56the maturity of the technology, both from the the important
- 7:01question you talked about how reliable they are, but also in
- 7:04terms of on the customer end user side, the experience of the
- 7:09end users in terms of the deployments and the scalability
- 7:13and reliability of these. We're still in, in the early
- 7:16days now. I will say that there's a
- 7:19spectrum of your agents, right? So in, in, in the spectrum
- 7:23essentially, you know, at one end of the spectrum, you could
- 7:25have these consumer agents which are really focused on
- 7:28individuals. You know, we were talking about
- 7:31this before the show started opening.
- 7:34I just launched their consumer agent called Operator.
- 7:38Very interesting. You know, it's for us as
- 7:40individuals part of the pro plan and the other end of the
- 7:43spectrum are essentially the next phase of evolution of RPA
- 7:49or robotic process automation right in the enterprise.
- 7:52So agents that take on complex enterprise work flows, right.
- 7:57So we have tremendous amount of activity all along the spectrum
- 8:02in terms of, you know, start-ups as well as in comments advancing
- 8:06the state-of-the-art right experimenting the I agents.
- 8:09So we are in the early days primarily because there are a
- 8:14few things that we are still trying to figure out right now.
- 8:17Large language models by themselves are non
- 8:20deterministic. And, and I think what I mean by
- 8:23that is that and we all, we again, we've all experienced
- 8:27this. When you ask ChatGPT a question
- 8:29a certain way and you ask the same question in a slightly
- 8:32different way, you may get a different answer.
- 8:34Right now that's getting better with some of the newer models,
- 8:38particularly some of the the reasoning models that are able
- 8:41to correct themselves. But the fact of the matter is
- 8:44that the core of the large language model is this sort of
- 8:47non determinism. And so a lot of the AI agentic
- 8:52designers today, builders today are working on in what we call
- 8:55scaffolding in order for us to essentially take the power of
- 8:59these large language models and harness them.
- 9:01At the same time, making sure that we're able to understand
- 9:04that the non determinism exists and we have the right
- 9:08architecture to be able to handle that.
- 9:11So you get a reliable, consistent and most importantly,
- 9:15scalable AI agent. Where are we there?
- 9:19Because of course that non determinism, if you press enter,
- 9:26submit again on a prompt, it's going to give you a different
- 9:30answer. Is great if you're wanting help
- 9:33writing an outline or some type of summary of a document because
- 9:37there's different perspectives you can take.
- 9:39But if you want it to run a task like book me an airline ticket,
- 9:44you don't want a lot of stuff all over the map.
- 9:48You just want one thing done. You want that ticket to be
- 9:50booked in the right place with the most efficiency and so
- 9:53forth. So it becomes a big problem I
- 9:54would think. Yeah, and I think this is again,
- 9:57a central design consideration, right?
- 10:01When you're building AI agents. And, and I think we can maybe
- 10:05break this down into like 3 parts.
- 10:07So the first part is the task itself.
- 10:10And you gave an example of a task, right?
- 10:13So how, how important is it for you to get the task?
- 10:17Absolutely. Right Now in ChatGPT, you know,
- 10:20if it gives you a slightly different answer, it's like
- 10:22search, right is you're getting information, you're not
- 10:25essentially making a decision. So you're OK with some level of
- 10:29non determinism because the human mind is able to correct
- 10:33for that, right? The other end of the spectrum,
- 10:35if you're essentially betting on this AI agent to execute a task
- 10:40consistently, it could be, you know, a enterprise workflow and
- 10:46a back end workflow, right? It could be code generation or,
- 10:51you know, you're essentially a customer service interaction
- 10:54agent. You cannot have that level of
- 10:57non determinism. So, so that's one thing like how
- 11:00do you define the task? And I think that's a, a key
- 11:03question to ask. The second aspect that you want
- 11:06to look at is, OK, now that you say, let's say you, you have a
- 11:08task that you need to be accurate.
- 11:10And it's also again, important to understand that AI agents are
- 11:13not all about just LLMS. There's a lot of existing
- 11:17software, there's a lot of reusing, you know, machine
- 11:21learning models, predictive models, which are deterministic,
- 11:24right, as well as the classic things that you as a software
- 11:28engineer do to build applications that go into making
- 11:31an AI agent, right? So I'll say this again, AILLMS
- 11:37are a tool. They're not the product, right?
- 11:39They're not the agent, they're a tool.
- 11:41So you have to understand, as with any tool, what the
- 11:43capabilities are. So how do you so that's, that's
- 11:46the second piece. It's like when you think about
- 11:48the AI agent it you have to think about leveraging the
- 11:51right, the LM in the right places.
- 11:53Then the third piece is OK, so now you've figured out, OK, my
- 11:57large language model is going to help me with building a plant,
- 12:01for example, the reasoning layer that we talked about earlier.
- 12:03So most AI agents today essentially propose a plan and
- 12:09you have a human in the loop that then approves the plan.
- 12:12So a good example of this is a lot of the coding, you know, Co
- 12:15pilots and coding agents that you have today.
- 12:18They're very again, and particularly with the the the
- 12:20cloud Sonnet 3.5, that was like, I think a step function jump in
- 12:24the ability of large language models to accurately generate
- 12:27code, right? The way they work is
- 12:29essentially, you know, the programmers interacting with the
- 12:32large language model. It proposes a plan which is
- 12:35approved and edited by the programmer before it actually
- 12:39goes out and executes it, right. So the so-called sort of user in
- 12:42the loop, human in the loop is a very critical design component
- 12:46today in AI agents, particularly in that planning layer right
- 12:50now, there are other things you can do.
- 12:52And we talked about part of this, which is, you know, what
- 12:54we call like a, a reflection loop.
- 12:57So once the output of the agent comes out, you tested against
- 13:01another large language model which is trained with the right
- 13:05output. So the model, the agent
- 13:07essentially tests itself to say, did I get this right?
- 13:11And it's able to kind of think on whether the output is correct
- 13:15and then make those changes and go back and iterate again,
- 13:18right? So these kind of reflection
- 13:20loops, the way you build evaluations, which is how do you
- 13:24determine the output is correct? And using that data to
- 13:28continuously improve the AI agent is again part of the
- 13:31design process. So I gave you like 3 different
- 13:33things where you have to consider.
- 13:35There's a lot more we can go into in terms of depth, but at a
- 13:38broad level, you know, the take away is that you can and design
- 13:43AI agents to be deterministic. Assuming that you understand the
- 13:47task right, you provided the right scaffolding, the
- 13:51evaluation loops, and you essentially involve the human in
- 13:56the loop. Today, AI agents work well when
- 13:58you have a human in the loop, and I think that's going to be
- 14:01the case particularly on this end of the spectrum where output
- 14:04needs to be, decisions need to be a lot more accurate.
- 14:09Subscribe to the CXO Talk newsletter so you can join our
- 14:13community and we can tell you about our upcoming shows, which
- 14:18we have great ones. We have questions that are
- 14:22stacking up on LinkedIn, So let's jump to a few questions
- 14:27right now. And the first question is from
- 14:31Ravi Karkara. And he says, where do you see
- 14:35American universities on creating a skills workforce for
- 14:40the AI driven world economy? How will they learn to work with
- 14:46and deploy AI agents? I consider AI large language
- 14:53models to be a tool, right? Just as you know, we had cloud,
- 14:59mobile, all these fundamental platform shifts, the AI and
- 15:03large language models are a tool to help us accomplish our task.
- 15:07So what I mean by that is it still is important for you to
- 15:12deeply understand what is the problem you're trying to solve
- 15:16with this particular tool. Are you trying to book an
- 15:19airline ticket like Michael you had mentioned earlier?
- 15:22Are you trying to execute a payroll function in in an
- 15:26enterprise? Are you trying to respond to a
- 15:30customer support question? So understanding the actual
- 15:34task, which is what, you know, essentially good product
- 15:39management is, is foundational to understanding how this tool,
- 15:43this new tool, much more powerful, obviously dramatically
- 15:48more impactful than anything that we've seen in the past, is
- 15:51going to change how we as a we as basically educators, as
- 15:58knowledge workers or as consumers leverage AI.
- 16:01So in terms of, you know, how universities approach this, it
- 16:05depends on sort of how you participate in this.
- 16:07I mean, to one end, the spectrum, you know, the, the
- 16:11education in, in sciences and math and, and good grounding in
- 16:15that helps you be part of the design process of this.
- 16:18On the other end of the spectrum, you know, if you're
- 16:20more business oriented, you know, a deep understanding of
- 16:25your problem, how and essentially how the tool works.
- 16:29I mean, we all are experimenting with GBT today.
- 16:33I use it differently, My daughter uses it differently.
- 16:36You know, I hope hopefully when none of her teachers are
- 16:38listening, but you know, she uses it sometimes to do help her
- 16:41with her homework, right? And so she's learning as part of
- 16:44that process just as I'm learning to use it for my use
- 16:48cases. And so are all of us, right?
- 16:50So I think it's a tool that we experiment with, but a deep
- 16:53understanding the problem and figuring out how to understand
- 16:57how to use this tool is, is foundational.
- 16:59So, you know, product thinking is I think is another key
- 17:03aspect, which I think is something we should emphasize.
- 17:07You know, if you're not deep in the algorithms in math, you
- 17:10still have a tremendous role to play by understanding sort of
- 17:13how do you build products, how do you solve customer problems,
- 17:16right? And I think the third aspect of
- 17:18this is there's a huge human element to this.
- 17:21We just talked about sort of how do AI agents function today?
- 17:24And one of the things I said is human in the loop, right?
- 17:28And you know, real AI agents get better at reasoning probably.
- 17:32I mean, the O3 model is amazing, right?
- 17:34And you've seen huge advances just in the last, you know, few
- 17:37weeks last, you know, later part, later part of last year in
- 17:40terms of the reasoning capability.
- 17:42However, throughout the steps that you have to think about
- 17:47where the role of the human is and at is it an evaluation
- 17:51function? Is it a reasoning function?
- 17:53And so humans are always going to be involved and being able to
- 17:56sort of understand and engage with the technology is
- 18:00effectively the most important thing, right?
- 18:02There is a massive human side of this as much as there's a
- 18:06technology side of this, right? So those are some of the areas.
- 18:10And, you know, I'm not an educator.
- 18:11So as I approach this like, that's the way I think about it.
- 18:15This is from Suresh Babu Madala and he says do we need to train
- 18:20agents or do agents interpret the question as a step by step
- 18:26task? That's a great question.
- 18:28You do need to train agents. Again, there's a spectrum of
- 18:32what these agents can do in terms of simple tasks, which is
- 18:34really complex tasks. And the level of training
- 18:37obviously will depend on what you're expecting the agents to
- 18:41do. You know, the first phase, as
- 18:45you're inserting the agent into a particular task, there's a
- 18:47certain amount of grounding that needs to happen.
- 18:50And the grounding typically is, you know, connecting it to the
- 18:53right data sources, connecting it to the right application
- 18:55sources, connecting it and, and understanding, making sure that
- 18:59understand and grounded in the policies of your particular use
- 19:02case or your particular enterprise.
- 19:05So that's the initial part, however training, you know, just
- 19:09like we as human beings, right, we're constantly learning,
- 19:12right? And if you, let's say you
- 19:13onboard a new employee, you know, fresh college grad, they
- 19:17are in a constant training process, right?
- 19:20They learn you have somebody who watches their output and you
- 19:23give them feedback, right? And you hopefully you observe
- 19:26them continuing to improve. It's the same exact thing for AI
- 19:30agents. That's why these evaluation
- 19:32loops are so important aspect of designing an AI agent.
- 19:36I can't say that can't stress that enough, right?
- 19:38You have to be able to kind of constantly understand the output
- 19:42of the agent, figure out where you can correct it and continue
- 19:45to improve. And I think the the last part of
- 19:48this is the observability of how these agents are functioning,
- 19:51right, where are sort of the broader efficiencies and
- 19:54inefficiencies of what they are doing and not doing is a
- 19:58constant part of sort of your architecture.
- 20:00This would be an excellent time for everybody listening to
- 20:04subscribe to the CXO Talk newsletter.
- 20:08Go to cxotalk.com. If you're watching on LinkedIn,
- 20:11take a look at the address on your screen and subscribe to our
- 20:15newsletter so we can notify you of upcoming shows and you can be
- 20:18part of this amazing, amazing CXO Talk community.
- 20:23Our next question is from Greg Walters.
- 20:27And you kind of address this a little bit.
- 20:29But, he says, can't non determinism be prompted into
- 20:34existence? The first aspect of this is to
- 20:37really understand, you know, the role of the LLM in your task,
- 20:43right? And what is the level of
- 20:46reliability that you expect from an LLMS task?
- 20:49Now there's certain things where the capability of the LLMS is
- 20:55getting constantly better and that you know that there's a
- 20:58reason why a lot of large language models are
- 21:00fundamentally trained on math and on coding tasks, because
- 21:03there's a lot of transfer learning as they get good at
- 21:06coding, as they get good at math, they're also able to get
- 21:08good at reasoning tasks, right, which are much more broadly
- 21:12applied. That's why there's a lot of
- 21:13focus on those tasks. So, so the first point I'll I'll
- 21:16say is these models will progressively get better at
- 21:20managing the hallucinations, whether it is through reasoning
- 21:23loops, right, or whether it is through better post training of
- 21:27the models in your deployments or whether it's inference time
- 21:31reasoning, right, which is sort of another, you know, scaling
- 21:34level that we now have. There are different techniques
- 21:37right in the model itself, which allow you to decrease the
- 21:41aperture of the non determinism, right?
- 21:43So that is one vector. The other vector is, as I
- 21:47mentioned earlier, building a scaffolding around it,
- 21:49understanding that you're going to get an output from the large
- 21:52language model that needs to be synthesized into something
- 21:55that's a little bit more reliable, right, and gets to the
- 21:58level of output that is acceptable for you.
- 22:01So that's a reflection loop. That's your evaluations and
- 22:03that's, you know, and sometimes you may just have a
- 22:05deterministic runtime that you need, right to design to the AI
- 22:09agentic workflow. And Gus Speckdash says it's
- 22:13interesting that Agentic AI goes around the ridiculously
- 22:19frustrating prompt user interface that is not integrated
- 22:23with any workflow. This is huge.
- 22:26What are your thoughts about Agentic AI?
- 22:29Simply around the user interface and integration with workflows?
- 22:35This is the quantum leap in my opinion, right, In terms of the
- 22:39user interface and user engagement.
- 22:42In a lot of ways, I think we've had some form of natural
- 22:45language processing, NLP, right type interfaces.
- 22:48You know, you can think of chat bots today.
- 22:51It's hard to escape them. You know, if you're trying to
- 22:53book a ticket or anything, like the first thing that pops up is
- 22:55a chat bot that's trying to get your information right.
- 22:57So that's natural language processing it.
- 22:59You know, it's able to understand voice, translates it
- 23:01into text, you know, does a search and responds back to you.
- 23:05I think the opportunity with large language models is the
- 23:09ability to infer context, right? What we had in the previous
- 23:13generation of chat bots was a literal translation, right?
- 23:16And a static sort of rules based interpretation of that
- 23:20translation. So what large language models
- 23:22have now, because they've trained on like the entire
- 23:25corpus of human language data is they can infer context, they can
- 23:29infer tone, they can infer, you know, the particular sort of
- 23:33intent and they're able to then appropriately translate that
- 23:38into into their query, right, and get you back a response.
- 23:41So I think it's game changing that you would have a large
- 23:45language model in a user interface perspective.
- 23:47Now remember, like we're still today, still very text based,
- 23:51right? Most of ChatGPT interactions,
- 23:53though, they have voice mode, which is amazing, right, are
- 23:55still text based. But think about the ability for
- 23:59us to be multimodal, right? Our ability to do voice which
- 24:02were there today, ability to input images and right, which
- 24:06we're going to get to right. And over a period of time these
- 24:08models will effectively get to this point where we call they
- 24:12have a world understanding, understanding of the world
- 24:14model. And I think that could be game
- 24:16changing in terms of how we interface with this technology.
- 24:19And this is from Arsalan Khan, who asks a very interesting
- 24:23question. He says we want bias to be
- 24:26removed from data when it comes to AI.
- 24:30How do you remove human bias if humans are in the loop?
- 24:34And who decides what's a bias or not?
- 24:38It's a really interesting point. If our vision eventually is that
- 24:42as some companies have articulated that every company
- 24:45has an AI agent and that's sort of the first point of interface
- 24:48for a customer interacting with the company, right?
- 24:52Let's say you're an insurance company or you are a, you know,
- 24:57you're even like a government service, like maybe the DMV,
- 25:01right? At some point that interface is
- 25:05really important. So I, I think look, they
- 25:08obviously the model companies are doing a lot of great work in
- 25:11making sure that we are conscious about bias.
- 25:16Now it is a fact that that is not a perfect solution yet.
- 25:20We're not quite there yet in in certain instances, you know,
- 25:24you're not getting these perfect answers.
- 25:27So part of this is the context that I talked about in terms of
- 25:32how you ground the agent is really important.
- 25:34And So what does context really mean?
- 25:36Right? Context means examples.
- 25:38And so if it's a customer agent, for example, you could train it
- 25:42on the policies that you have. You know, you probably have a
- 25:46lot of voice recordings of existing agent calls, right,
- 25:50that are great examples of how to handle situations in bias or
- 25:53confrontational situations. So the training of this agent,
- 25:57grounding it in the policies, right, and the rules of the
- 26:03particular use case is a particularly important task.
- 26:08Now, I think look, the human in the loop is really about sort of
- 26:11how you reason things. And obviously look, the way a
- 26:15human interacts ideally is a positive in terms of our ability
- 26:19to eliminate bias, right? By inserting human in the loop,
- 26:22you're actually adding a step that improves the ability of the
- 26:26agent overall to be, you know, to correct it's biases.
- 26:30But I would say, you know, good training, grounding and
- 26:32policies, right? And obviously, you know, having
- 26:35responsible humans in the loop are the things that will help us
- 26:39get there. But it's an imperfect process,
- 26:41and that's why I say it's early days yet.
- 26:43Arsalan Khan comes back and he says if subject matter experts
- 26:49train AI to create AI agents, would we really need the subject
- 26:54matter experts? What happens when AI agents come
- 26:58across a scenario that they haven't encountered yet?
- 27:02Training is a constant process for AI agents, right?
- 27:07What you get when you initially start this process with a
- 27:10subject matter expert helping you ground the model, ground the
- 27:14AI agent is you're giving it a certain rubric.
- 27:17You're basically saying like, hey, here's basically what is
- 27:19expected, right? Here's a basic task.
- 27:21Here's how you perform the task right now.
- 27:24The task evolves. We are in a dynamic world,
- 27:27right? Let's say you're in a company,
- 27:29maybe you launch a new product, maybe you expand into a new
- 27:32region. You know, maybe you have you,
- 27:35you make an acquisition or you have new employees on boarded,
- 27:37right? It's a, it's, it's a process of
- 27:39constant change. So to some extent, I think you
- 27:43can design the AI agent to say, OK, I can expand into a new
- 27:46geography. This is how I understand it,
- 27:48right? But they may be different
- 27:50policies. I mean, as it's often the case,
- 27:52if you're operating in a different country, they may have
- 27:55their own, you know, regular rules and regulations that you
- 27:58need to, you need to now kind of ground the model in and such
- 28:01like. So I think the subject matter
- 28:03expert is really, you know, as, and we all do this, right?
- 28:08We're all subject matter experts.
- 28:10We're not static, right? We're constantly sort of
- 28:12learning, evolving, understanding, right, new
- 28:14technologies, new pieces. And it's the same thing for the
- 28:16model. So yes, the idea of the AI agent
- 28:20is to take away these sort of mundane things, OK, go get this
- 28:23document, put 10 documents together, create a research
- 28:26report, right? So that yes, you don't need to
- 28:29retrain that thing. But being able to say like, OK,
- 28:31how do I operate in the European Union, which may have a
- 28:34different set of rules, or in in Asia, in, in Japan or in China
- 28:39or in India, right, which may have a different set of rules.
- 28:42Those are things that there's a certain amount of requirement of
- 28:46understanding those rules and regulations that need to, again,
- 28:48you need to kind of train the agent on, right.
- 28:51Let's start talking about business.
- 28:54And Mario Garcia asks. He says it's inspiring to see
- 29:00the impact of AI in Fortune 500 companies.
- 29:05What insights about this stand out to you the most?
- 29:09A lot of these large companies today, I mean, it's been amazing
- 29:13to watch how rapidly, you know, both sort of large companies as
- 29:19well as incumbents and start-ups have really embraced AI and
- 29:24large language models. And it's largely, I think the
- 29:27ChatGPT moment which unleashed this because it, you know, it
- 29:32was so accessible. So you take a step back.
- 29:34You know, AI has been around for a long time, right?
- 29:37You know, I, I did a course in neural networks back in my, you
- 29:40know, grad student days. What changed I think in this
- 29:43generation is that today really powerful complex AI models are
- 29:50available on the other end of an API call, right?
- 29:54So that level of simplicity in terms of access to really
- 29:58powerful technology is essentially what enabled us to
- 30:01unleash large AI as you see it in in a lot of these use cases.
- 30:07Now I will again caution you to say that we are still very
- 30:11early. If you talk to a lot, a lot of
- 30:13these large corporations, they do have large language models
- 30:17integrated. Most of them have rolled out,
- 30:19for example, some form of a coding copilot.
- 30:21They've rolled out some form of, you know, customer support
- 30:24function. They've rolled out some form of
- 30:25an analytics sort of use case with large language models, but
- 30:31we're still not deployed at scale, primarily because we're
- 30:35kind of still tuning, tweaking, right, learning in terms of how
- 30:40do you manage these AI agents? How do you manage biases?
- 30:44As one of your audience members just asked, how do you make sure
- 30:47that they are current, right? How do you make sure they don't
- 30:50hallucinate? The most important thing of the,
- 30:52and this is fundamental, right? We all know this.
- 30:54It doesn't matter. There's a lot of really, you
- 30:57know, fun, exciting, interesting demos on, on, on X, right?
- 31:02And that's, those are like, great.
- 31:04You can say, oh, wow, this agent can do this.
- 31:06What's really important is can it do it consistently and can it
- 31:10do it at scale? And so those are the questions
- 31:13we'll answer this year hopefully, right?
- 31:15And that's why we're so excited in 2025 about the trajectory of
- 31:18these AI agents. Praveen, you and your team put
- 31:22together what you call a market map of companies involved with
- 31:27AI agents. Can you tell us about that?
- 31:30And I can bring it up on the screen so everybody can see.
- 31:34And there it is. Praveen, can you talk about this
- 31:39market map that you've put together?
- 31:41The market map is a dynamic living thing in the sense that
- 31:47it will evolve constantly, primarily because we're seeing
- 31:50so much activity, right, and energy around, you know, the AI
- 31:57agentic space. So what we tried to do is to
- 32:00basically construct this in, in layers, right?
- 32:03So there's, there's like a foundational layer where there's
- 32:07a lot of these kind of data sources, integrations and such.
- 32:10Like there's this new sort of bucket we, you know, we call
- 32:14sort of the agent computer interface, which is the ability
- 32:19for AI agents to use computer, you know, tools, right?
- 32:24And you know, it could be integrations, it could be, you
- 32:28know, web tools, some of the stuff is integrated in models,
- 32:30some of these are, you know, you have interesting kind of
- 32:33platforms that are created for this.
- 32:35So we try to kind of construct the model as, you know, layer by
- 32:38layer. What's the bottom layer, OK,
- 32:41where all the data platforms, right?
- 32:42What is what is this sort of middleware layer, if you will,
- 32:45which is the agentic computer interface as well as a lot of
- 32:47these agent frameworks, right? You know, we are investors in a
- 32:51company called Crew AI, for example, and I'm sure anybody
- 32:55who's talked about AI agents probably knows about crew AI.
- 32:57It's one of the most popular open source frameworks out
- 33:00there. You can use crew AI to build
- 33:02agents, say others like LAN chain, which are also do
- 33:05something similar. And about that, then what we
- 33:08tried to do is to say like, wow, let's try and kind of map out
- 33:11sort of where is the energy in the AI agent space?
- 33:14And it's important to understand, I think it's on the
- 33:17left side of the of the of the market map.
- 33:21There's a lot of AI agentic products and offerings from
- 33:26incumbents. So obviously Salesforce, you
- 33:29know, launch their own sort of agentic agentic workforce agent
- 33:34force as they call it, right. Microsoft has Co pilots.
- 33:37We just saw open AI launch operator, which is more consumer
- 33:40oriented agent. And so, you know, effectively
- 33:45all the incumbents have said like, hey, we've got these great
- 33:47customers, we've got all these great use cases.
- 33:49Is there a way for us to improve our customer experience or
- 33:53productivity by creating an agentic workflow on top of our
- 33:58existing software? On the rest of the market map,
- 34:02you can see sort of tremendous amount of energy in in different
- 34:05verticals. You know, particularly in
- 34:07coding, for example, there's been huge amount of a great
- 34:10Workman cursor by far seems to be the most popular among
- 34:14developers today. But you also see very specific
- 34:16vertical agents, right, sales, marketing, legal, right?
- 34:23You, you see finance, right? So you can take almost each of
- 34:26these different functions and you can see companies building
- 34:30agents which are customized to that particular vertical use
- 34:35case, right? This is really interesting
- 34:37company called Samaya, for example, that's building doing
- 34:40some amazing work focusing on building agents for the finance
- 34:43workflow, right? So you, you, you see that the
- 34:48idea was a market map was not to be precise, right?
- 34:53And capture the entire view. And you know, I do apologize to
- 34:56a lot of the the builders out there, some of whom we missed in
- 35:00the market map clearly, right. The idea really was to kind of
- 35:04give you a perspective of, you know, what this is landscape
- 35:08look like, right? Where?
- 35:09Where is the activity? Like where?
- 35:11How are builders approaching the AI agentic space?
- 35:15What are the opportunities and the use cases, the predominant
- 35:21or the most important use cases for AI agents right now?
- 35:24They're probably like 3 or 4 buckets and you know, the first
- 35:29one obviously that everybody knows and understands very well.
- 35:33Are these coding agents, coding Co pilots, coding platforms,
- 35:36What do you want to call them? Right.
- 35:39And depending on, you know, the particular style, you know,
- 35:43cursor has a particular way of, of working.
- 35:47It's more like a copilot. If you take something like Devon
- 35:50has a different sort of way it, it engages, you know, it fires
- 35:53off a bunch of agents that you know, execute your plan and such
- 35:56like. But there's a lot of energy
- 35:58around developer facing AI agentic work, right?
- 36:03And so you can see that in the market map as well.
- 36:06The second area is in the customer experience section.
- 36:10So customer experience, obviously everything from like
- 36:13customer support agents, which obviously is the biggest use
- 36:16case. We, we, as you, as we were
- 36:18talking about earlier, chat bots are already a fact of life.
- 36:22Can we make that experience much more realistic, much more sort
- 36:25of, you know, engaging? So, you know, like me, you're
- 36:28not basically saying agent as as soon as you as soon as you get a
- 36:33a chat bot, right? So there's a ton of energy in
- 36:35that space, lots of lots of great companies building
- 36:38interesting products. The other area of it's just kind
- 36:41of interesting is in the operation space, right?
- 36:43So if you think about operations, broadly speaking, it
- 36:46could be IT operations, it could be security operations.
- 36:50You have sort of this needle in the haystack problem.
- 36:52You have a lot of data, you have a lot of like alerts that come
- 36:56in and you're trying to figure out like, OK, which ones do I
- 36:58pay attention to, right? So this is actually a perfect
- 37:01use case for AI agents, the ability to synthesize all of
- 37:04that information. And if it's grounded in your
- 37:07policies and in sort of in in a company sort of particular way
- 37:13of doing things, it's able to like, say, like, hey, here's
- 37:15maybe the top three to five alerts you need to pay attention
- 37:18to. This is the problem that they're
- 37:20articulating. And here's a few ways to
- 37:22remediate this. So, you know, we were talking
- 37:24some interesting startups that are actually focused in this
- 37:26particular space. So I think those are the three.
- 37:29And, you know, there's a lot more, but I'll just, you know,
- 37:31in the interest of time, maybe I'll just pause there.
- 37:34We have a question from Elizabeth Shaw, and this is on
- 37:39Twitter. Who asks how are organizations
- 37:44using agentic AI in their business and their ecosystem?
- 37:48Let's just take for example, a customer service AI agent,
- 37:52right? So that's that's a use case that
- 37:55we're seeing a lot of customers experimenting with.
- 37:59So what is what is this customer service agent do?
- 38:01So you effectively again, you think of a chatbot today, the
- 38:06customer service agent is able to first of all, sort of be
- 38:10grounded in all of your data, your FAQs.
- 38:12You know, how do you, I mistype my password, how do I recover my
- 38:18password? Or, you know, how do I, you
- 38:20know, whatever, add, add my child to the, to the insurance
- 38:25policy or whatever. You have these things continuous
- 38:28task, you know that, that you were able to then interact with
- 38:31an agent. The agent understands what
- 38:33you're trying to do. Maybe you're looking for a form,
- 38:35maybe you're looking for a website, maybe you're looking
- 38:37for a particular sort of quote or something like that.
- 38:40So I think that is a use case that is getting a lot of
- 38:43traction. We're seeing a sort of a lot of
- 38:46our customers looking at sort of, you know, rolling that out
- 38:50into production. The other use case that I
- 38:53mentioned is on the coding side, right?
- 38:56And no, we're not just talking about like an IDE like like
- 39:00Cursor, which obviously has a lot of broad adoption, but
- 39:02things like automated testing and provisioning, right?
- 39:05So you have software you need to roll this out.
- 39:09Testing is a really critical part of that process of rolling
- 39:12things out. You, you're able to actually use
- 39:16AI agents, you know, very effectively in sort of that
- 39:19testing and sort of, you know, red, red team kind of use cases
- 39:25where you can like see if you can break it, right?
- 39:28That's a very critical function where you're seeing some level
- 39:31of, of deployments happening. The other one, as I said, is
- 39:34like in IT operations, right? And this, this is very exciting
- 39:40because again, if you, you know, these are mission critical,
- 39:44right? You need to be constantly up,
- 39:47you know, most of these IT teams, you know, you always have
- 39:50a 24 by 7 coverage because you cannot have, you know, critical
- 39:53systems going down, right? And so an AI agent is perfect
- 39:56because it's it, it has the ability to synthesize large
- 39:59amounts of data. It has the ability to basically,
- 40:04you know. Needle in the haystack, right
- 40:06problem now again, work in progress.
- 40:09I wouldn't say all of these things are at perfection, but we
- 40:13are definitely seeing these. And then I mentioned this
- 40:15company Samaya, which is very interesting.
- 40:18They are actively building like a finance analyst kind of agent,
- 40:22which is, you know, which is pretty accurate in terms of
- 40:25being able to extract context out of research and and provide
- 40:30you really focused information. So you say that it's really
- 40:35accurate. I'm assuming that what you also
- 40:38mean is that it is consistently reliable and predictable.
- 40:44Exactly. And it's much harder to do that
- 40:48if just straight out-of-the-box, right.
- 40:50I think a lot of times there's a lot of confusion in the market
- 40:53about like, hey, wait a minute, the large language models are
- 40:55just going to keep getting better and better and better.
- 40:57And, you know, ultimately they'll just they'll be like one
- 41:00model that solves all, right. And, and I think in it is true
- 41:05in certain simple use cases. I mean, absolutely the models
- 41:08are getting better, their reasoning better.
- 41:09Maybe we will get to a point of artificial general intelligence,
- 41:13right where, you know, these models can just use, you know,
- 41:18computers like we do. And, and, and maybe that's the
- 41:20bar, right? But I think when you're looking
- 41:22at complex enterprise workflows, as I mentioned, the ability for
- 41:28the agent to be grounded right and to be accurate and to
- 41:34present data in the way the end user expects will require some
- 41:40amount of post training, some amount of inference time, right
- 41:44reasoning, as well as some amount of scaffolding in order
- 41:47for you to build the perfect agent.
- 41:49Can you talk about the economics of agents?
- 41:53And then we'll jump back because we do have additional questions
- 41:57that have come in, but the economics are really important.
- 41:59So what are the aspects that Dr. economics and what do enterprise
- 42:05buyers need to think about when it comes to the economics of
- 42:08agents? You've seen sort of the whole
- 42:11spectrum of conversations, right?
- 42:14Everything from like with AI agents, we're just going to go
- 42:17and tally to outcome based pricing to like, well, you know,
- 42:20it's it's still software. So we have to kind of figure out
- 42:24how do you make sure that you're able to charge for it
- 42:28appropriately. The fundamental question to ask
- 42:30right when you think about pricing is can you measure the
- 42:35value of the AI agent output accurately, right?
- 42:40So what do I mean by that? Let's say you have a customer
- 42:43support agent. You can basically say, hey, the
- 42:46customer support agent handled 100 calls, right?
- 42:50And you'd have taken me X amount of dollars to handle those 100
- 42:53calls. The customer agent handled those
- 42:55calls. So I can attribute directly a
- 42:58value right to the output of that agent.
- 43:01It handed 100 calls. Each call is worth X.
- 43:04So there's basically a hundred X is basically the value of that
- 43:08particular agents task. In the other end of the
- 43:10spectrum, let's say you're doing, you know, a research
- 43:15workflow, right? So you've generated research
- 43:18report or you're helping an analyst basically with the
- 43:21research and you improve their productivity.
- 43:23How do you measure the value of that?
- 43:25Right? You measure it by, you know,
- 43:27individually, like how, you know, asking the the analyst,
- 43:31like how, how much more productive were you right?
- 43:33And you know, it's it's much harder to quantify certain tasks
- 43:38versus certain other tasks. So I think the first question in
- 43:41the understanding economics of agencies, can you attribute
- 43:45value in a reasonably accurate way to the output of the agent.
- 43:50So based on that, if you can, then outcome based pricing is
- 43:53essentially where we're eventually going to go to,
- 43:56right. If it's a lot more nebulous,
- 43:58then I think what we're going to see is some form of an evolution
- 44:02of the existing SAS pricing model.
- 44:04So you might pay a platform fee like for the agent thing and
- 44:08then maybe you hire an agent. So you pay on the number of
- 44:11times you run the agent, right? So it's some combination of
- 44:14that. So we see this again as a
- 44:16spectrum. There's no like absolute here,
- 44:18right? There's a lot of experimentation
- 44:20today in some ways, like companies are still trying to
- 44:23figure out like, you know, how's the customer getting value
- 44:25customers trying to figure out. And if you're ACFO, right,
- 44:28you're used to paying subscription software, you know,
- 44:30like, OK, I'm I'm I've got X amount of licenses for one year
- 44:34and I can budget that right now. If you go to sort of this
- 44:37outcome based pricing, again, if you don't have an accurate sense
- 44:40of value, how would you as ACFO budget right for these AI
- 44:44agents? So I think there's a lot of
- 44:46these things that need to be, you know, we're kind of
- 44:50experimenting and understanding eventually where this direct
- 44:55attribution of value, I think we will end up in the outcome based
- 44:59pricing bucket. But there's also going to be a
- 45:01lot of these intermediary models where, you know, you want to
- 45:05make sure that the developers are getting a fair value for the
- 45:08product that they're building and the customers are paying a
- 45:10fair price for it. So ultimately when we reach the
- 45:14point where agents have discrete measurable output results, then
- 45:21we can move towards performance based pricing and until then
- 45:26it's essentially usage. Right.
- 45:28Yeah, I think that's a good way to put it.
- 45:31We have an important point now raised on LinkedIn by Naresh
- 45:37Kumar, who is VP and General Manager of Product Management at
- 45:42Z Scaler. And he raises the question, what
- 45:46about security and agentic AI? And we haven't talked about
- 45:50that. So I'm glad you brought this up.
- 45:52Large language models help with this sort of needle in the
- 45:55haystack problem, which is inherent to sort of diagnosing
- 45:58security problems. I kind of grew up in the
- 46:01networking world and be used to, you know, build these large
- 46:06global scale Internet scale networks.
- 46:09And a big part of the task was like, you know, if if there was
- 46:12an outage somewhere to, to debug, that would essentially
- 46:15mean like we synthesize tons of data, right, and figure out
- 46:19where we need to focus our efforts, right.
- 46:22And security is the same way. You have a large aperture of
- 46:27exposure depending on, you know, the type of company you are,
- 46:30everything from your network to your applications, your devices
- 46:33to your individuals to identity, right?
- 46:35There's like multiple layers, right when you think about
- 46:37security and so, and it's been a tough challenge, right in the
- 46:42security industry, we've had, you know, these platforms called
- 46:45CMS which try to bring all this together and be able to give you
- 46:49like a unified view where you're able to manage this.
- 46:52But you know, a security OPS centre, essentially the nerve
- 46:56centre of how most companies run their security operations.
- 46:59So I would look at the role of LLMS in security in three ways.
- 47:03The first I think is from a operations perspective, I think
- 47:07it could be a very useful tool because he has the ability to
- 47:11synthesize large amounts of data and help in that needle in the
- 47:14haystack problem or prioritize it.
- 47:16I think it's a great use case. The second one is, you know, LMS
- 47:20integrated into the security products, right?
- 47:24Will essentially, you know, you talk about again, a security
- 47:28agent, existing security software, right, being able to
- 47:32dynamically understand policy, right, dynamically able to
- 47:35respond to, you know, you're adding more users etcetera.
- 47:40I think you're able to sort of build those.
- 47:42We're seeing companies starting to build large language models
- 47:46into their software stack. Just as we talked about earlier,
- 47:49it's a tool right, where where it's useful.
- 47:52The third I'd say is look, you know, LLMS do represent a new,
- 47:56you know, threat aperture, right?
- 47:58Particularly if the models essentially hallucinate or for
- 48:02example, in that psych OPS use case, you know, ignore or
- 48:06highlight or or miss critical, critical kind of threat.
- 48:09So while you're designing, while we talk about all these agents
- 48:12being deployed in a customer support use case or, you know, a
- 48:16finance analyst use case, if those large language models are
- 48:19not sufficiently grounded, right, and they're not, you
- 48:24know, the data, the training set that they have is not protected
- 48:28appropriately, you risk not just hallucinations, but effectively,
- 48:33you know, a hijack of the entire AI agent.
- 48:35So it's early days in that I think we've had some, you know,
- 48:41some interesting conversations with, with founders who are
- 48:44thinking about this problem deep ways and building interesting
- 48:46things. But yeah, you know, it's, it's,
- 48:50it's an problem space at this point.
- 48:52It's we will learn more and we will evolve the security
- 48:55architecture just as revolving with the maturity of the AI
- 48:58agents itself. OK.
- 49:00And obviously Z Scaler is thinking about this because he
- 49:03asked that question. Yeah, they're an important
- 49:05player and they have a huge role to play right in in in our
- 49:08overall architecture. We have another question from
- 49:12Arsalan Khan. I'll ask you to answer this
- 49:14really fast because we're just going to run out of time now who
- 49:16says should we create a time and motion AI agent that assesses
- 49:23other agents if they have saved money in an organization,
- 49:27obviously referencing back to the pricing discussion we had
- 49:31earlier? If you can listen to some of the
- 49:33industry luminaries talk about like we all have some of this
- 49:36army of agents, or we will have, you know, human employees and
- 49:40agentic employees. There is a requirement to one
- 49:44like train all these agents, ground all these agents, right,
- 49:48as well as to evaluate all these agents.
- 49:51So we talked about reflection loops in, in, in terms of like
- 49:54the specific sort of output governing these output.
- 49:57So similarly at a higher level of abstraction, which is the
- 50:01value, right? Yeah, it is.
- 50:03You know, you could, it's a, it's an interesting idea to be
- 50:06able to say you have an agent that's constantly measuring the
- 50:09value of the output of other agents to ensure that they are
- 50:13meeting a particular mark. For example, if you're a
- 50:15customer support agent, right? If it's not deflecting whatever,
- 50:19you know, 100 calls a day or something like that, right?
- 50:22And the metric may vary, then maybe it's not performing
- 50:25appropriately. So and you could have an agent
- 50:29that's it's, it's more like an operational function.
- 50:31So there are ways to, I think in this sort of agentic future,
- 50:36they are the worker AI agents and they're these sort of, you
- 50:40know, evaluation agents and you potentially will have manager
- 50:45agents at some point, right? You can, you can think of a
- 50:47future where there is some level of sort of different levels of
- 50:50hierarchy where you are actively evaluating and governing these
- 50:55agents. But again, we're we're early
- 50:56days yet. We're a little bit sort of
- 50:58hypothesizing how this looks like.
- 51:00Gus Beck Dash comes back and he says is it really better to have
- 51:05the agency and the knowledge in one model or have them as
- 51:10separate, loosely integrated systems?
- 51:13It'll come back to the type of problems space that you're
- 51:17addressing, right? You know, in general, we know
- 51:21from just, you know, the things around us that there's no such
- 51:25thing as one unified body of knowledge, right?
- 51:29We as human beings, there's so much complexity in sort of our
- 51:33world, in our workplaces, in our sort of consumer oriented lives,
- 51:38that there's no such thing as like 1 mega intelligence that
- 51:43essentially does everything for you.
- 51:46So, you know, I think we always solve problems by breaking down,
- 51:51breaking them down into smaller problems, right?
- 51:53And then using different tools to solve those problems and put
- 51:56these things back together. That's sort of the, the way, you
- 51:59know, human workflow happens, you know, irrespective of
- 52:02whether it's a, you're building a chair right, as a project or
- 52:06you're building a complex set of application in an enterprise.
- 52:10So I will again hypothesize here and say that I don't believe in
- 52:14this sort of single unified agent.
- 52:16I do think agents will continue to get better.
- 52:19Maybe we will get to this threshold of artificial general
- 52:22intelligence. How will you define it?
- 52:24That's another at worst conversation.
- 52:26But I think you will always have this notion of taking a problem,
- 52:31breaking it down, using different tools to solve that
- 52:33problem, to put it together. And you can think of that same
- 52:36architecture applying in an agentic world as well.
- 52:39OK, And with that, this has been an action-packed hour.
- 52:45Praveen Akiraju, thank you so much for taking time to share
- 52:49your expertise and knowledge with us today.
- 52:51I really, really do appreciate you.
- 52:53Thank you. Thank you for having me.
- 52:55And a huge thank you to everybody who watch.
- 52:58Before you go, subscribe to the CXO Talk newsletter so you can
- 53:03join our community and we can tell you about our upcoming
- 53:08shows, which we have great ones, and you can ask your questions
- 53:13during the live show just as today.
- 53:15And with that, a huge thank you to everybody and to Praveen.
- 53:21And I wish everybody a great day and we'll see you again next
- 53:25time. Take care.