Cisco’s Jeetu Patel: Everyone Is About to Manage AI Agents

0:00In February of this year, Open Router did a study that showed that the token consumption by agents has now exceeded that of humans.

0:11Jeetu Patel: And since February, from February until now, you have seen the increase of token consumption go up 14x for the agents. So, not only did we actually get into this phase of agents pretty fast, agents have now surpassed humans in its consumption of tokens and are doing most of the work related to AI. Every task you hand to an AI agent comes with a management decision. What do you want accomplished? How much is it worth? And how much freedom should that agent have to get it done? And as the consumption of agentic AI accelerates, those decisions carry real consequences for your budget, for your infrastructure, and your business.

0:53Because a growing token bill tells you something about how much AI you're using, It takes a different kind of judgment to know whether that work is making your organization better or more profitable. On today's episode, Cisco President and Chief Product Officer G2 Patel shares his view for how enterprise AI can move from AI adoption to agentic delegation and why leaders need to get comfortable managing that shift. In the conversation, we explore how to choose work worth delegating, give teams room to experiment within a budget, and secure agents whose actions can reach across the business. G2 also offers a challenge that stuck with me personally.

1:28Use AI in ways that stretch your thinking. That means asking more of it than faster output and asking more of ourselves than simply passing that output along. This is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

1:51Mr. G2 Patel, how are you today? Welcome to the AI Proving Ground podcast. It's good to see you, man. How are you doing? I am doing just well. Well, we had you on the show about a year ago. And at the time, you said we were entering the second phase of AI, the first being chatbots that intelligently answer questions for a user, and the second being agents that are conducting tasks and jobs fully autonomously on our behalf or near autonomously. And 12 months later, That prognostication looks mostly true. But where do you see the industry right now? Are we starting to creep into a third phase even?

2:28Or are we still in that second agentic phase?

2:31Jeetu Patel: I think you'll have the second and third be parallel. So the physical AI will be the third phase. But I think on the second phase, it's a really interesting time. Because when we talked a year ago, we were talking about just entering into the agentic era. In February of this year, OpenRouter did a study that showed that the token consumption by agents has now exceeded that of humans. And since February, from February until now, you have seen the increase of token consumption go up 14x for the agents. so not only did we actually get into this phase of agents pretty fast agents have now surpassed humans in its consumption of tokens and are doing most of the work related to ai and that's you know like so what you essentially have right now is we're in this infrastructure super cycle and the infrastructure super cycle basically is completely in full effect with less than like a few percentage of humans using AI and power user mode, right?

3:46Jeetu Patel: With agents. Imagine what happens when 10%, 20%, 30%, 50%, 70 % of the humans are using AI and power user mode. what kind of infrastructure is going to be needed to go out and fulfill the needs of AI. It is going to be an enormously supply-constrained world for a long, long time. Yeah, it's pretty crazy to think about. You talk about the AI super cycle and the one that we're entering. I know you're bullish and you've talked a lot about how this most recent AI super cycle is going to be more defined by or less defined by how many people are using AI and more about what those people will actually delegate to AI.

4:29It starts to get into agents, of course. Explain to me that shift of thinking and what the implications are for enterprise leaders.

4:37Jeetu Patel: So I think like every human, we are now comfortably in the phase where every person entering into the workforce today will likely manage humans before they manage agents. I'm sorry, will likely manage agents before they manage humans. And so all of us are going to be managers of agents. And as it pertains to like, what do we delegate to these agents? You'll essentially start to learn to delegate most things that we do to agents, either partially or fully. But there's going to be very little that you're going to do without the help of agents. and that'll take a while because it's like it's a it's almost like the the thing that are is the technology there today to help you with anything yes but it's your instinct and your muscle memory is not quite there where you actually think about using ai for everything and it takes a while it's kind of like when you you know when when the internet first came about there was the first few months where i would ask a buddy of mine questions and he still works with me, but I'd ask him questions and he'd be like, I don't know the answer, but let's look it up.

5:50Jeetu Patel: And that notion of let's look it up was a very instinctive thing that happened to us over time. But initially it wasn't that instinctive because like you didn't really know what look up was. That's not how you thought. I think it's going to be the same thing. It's going to be like, let's make sure that we can delegate to an agent. Let's make sure that we can have the agent do this for us as a first, as a first pass. And so that, that will just get more and more kind prominent in the way that we work. And so if it's going to be less about user adoption, people adoption, I mean, certainly more and more people will begin or will start using AI, but the amount of agents that they utilize is going to extrapolate by quite a lot.

6:28I'm curious from your standpoint right now, and certainly still early innings, so to speak, what qualifies as a good task or workflow or initiative that is ripe for that delegation? How should users be thinking about what they should versus what they should not be delegating?

6:48Jeetu Patel: I think you should think about virtually everything that you do, not starting from a blank sheet of paper and saying, how do I get a jumpstart on whatever I do? Like, for example, I tend to delegate a lot of the research for the learning that I need to do on a day-to-day basis to AI. right and i have i tend to delegate ai actually consolidating and pulling together news from different news sources for me i tend to delegate aspects of ai where i need to create content but i want to get a head start on the content so i can actually have my brain start working because otherwise you might experience writer's block at times.

7:37Jeetu Patel: All of those things tend to be superbly valuable. I tend to use AI to delegate editing. But what I don't do is I try not to delegate something to AI that makes me intellectually lazy. I try not to say, I'm not going to think, I'm just going to have AI think and then do it. That doesn't make me smarter. That makes me dumber. And so what I tend to do is say, what's going to make me smarter and what's going to stretch my thinking? Anything that stretches my thinking by delegating, I love to delegate. Anything that doesn't stretch my thinking by delegating, I tend to hold off on delegating. And that's a good rubric for anybody out there that's looking to jump in or continue their maturity with AI.

8:27But there's a whole kind of other side of the story here, which is enterprise adoption, if we're starting to delegate some of these tasks or workflows or making something more, you know, agentically forward, how should leaders think about what's creating actual value versus what's just, you know, consuming tokens or consuming your time on AI? Because you can get lost on that ladder.

8:50Jeetu Patel: You can get lost. And I think your time actually matters more even than the time of the tokens. Like most people don't realize that sometimes you can just, it can be this thing that sucks you in. And then before you know it, like hours have gone by and you haven't really accomplished much. You've just been going back and forth and back and forth. And so I think knowing how to get the most out of AI in the most efficient way possible is going to be a superpower that people are going to have to learn how to develop. Because just like you learned how to get a lot from your people that worked for you and got very efficient with how you manage them and got very thoughtful about the kind of work that you kind of delegated to them, I think AI is going to be no different.

9:36Jeetu Patel: What I found at times that started happening when Claude really was starting to take off with Claude Cowork is a lot of people would use either Claude or OpenAI for creating content and they never even read it. They would just start sending. So I started getting like volumes of documents over documents over documents. I'm like, there's no chance I'm going to be able to read all this. And there's no point in me now paying for tokens to summarize all this and not even read it. Like, what's the point? Like, but have you read it? And what I found was a lot of the people that were sending me these documents themselves hadn't read it closely enough.

10:16Jeetu Patel: Like, no, well, I didn't read it closely enough. I figured I'd send it to you. Like, no, actually, in a time like this, when you can generate unlimited amounts of content, what you choose not to generate and how precise you are in what you generate actually conveying a point might end up taking just as much time, if not longer. However, you might create output that's at least 10x better. And so the goal should be create output that's 10x better rather than just I am in this mad rush to just save time. Do something that wasn't possible before. Create original insights that weren't accessible to us before.

10:57Jeetu Patel: I think those things tend to be far more useful and additive to society than the other way around. How do you account for the fact that a good idea with AI, we encounter this all the time at Worldwide Technology. I'm sure you all have the same kind of situation, which is an idea or a dashboard or something that you whipped up in codex or cloud code. It's not working one day, but then two, three days later, a week later, whatever your timeframe might be, it can work because of model updates. That seems like it's a real gap there to not necessarily put something out to pasture, but put it on the shelf, don't lose sight of it, and make sure that you revisit it.

11:37How do you all at Cisco strategize about that so that you don't lose any good ideas that your employees are bringing to you?

11:45Jeetu Patel: Brian, this is such an important question that you just asked because one of the things that we tell our teams is just because it doesn't work today, you should not assume that it's not going to work. two weeks from now. In fact, what you should assume that it's going to work two to three weeks from now. And assuming that it works two to three weeks from now, four weeks from now, whatever, pick a time period, but it's not a year. It's a short enough amount of time. What you have to do is, assuming it works, you have to start planning for it right now because your planning cycle will take longer than AI's fixing cycle.

12:27Jeetu Patel: and so it changes the mental model of like speed is the superpower and you have to be in my mind speed is the moat and being early to market as a matter of habit and moving at a very fast speed is fundamental to long-term durable success so you're so we're out here delegating now to agents, how do you account for the sprawl that may or may not be happening with Cisco, but it's certainly happening at other organizations around the world? I'll give you an example. I was reading a Reuters article that found one enterprise in an audit found over 400 agents when their CIO thought that they were only going to have 40.

13:16How do you harness all of this creativity, innovation? It seems like it's something that's hard to wrangle together at a time when everybody's just trying to move very quickly.

13:27Jeetu Patel: I think what you have to do is you have to teach people how to do things by making sure that you're optimizing for the right metrics. So I would not focus on how many agents they can spin up or not spin up. I would focus on giving them a budget and saying you have this budget within which you have to get this job done. And And so have at it. If you want to use 400 agents and, you know, expend your budget that way, that's up to you. If you don't want to use agents, 400 agents, but use three agents or two agents or one agent, that's up to you. How many tokens you generate is up to you based on the kind of model you want to use.

14:09Jeetu Patel: You need to have some kind of intelligent routing layer that knows which query goes to what model. but don't just like we don't want to micromanage people like i'll give you an example there are times when companies will say hey you know what i am not gonna i'm gonna micromanage every single nuanced detail about budgets because i'm trying to do cost control the alternative plan could be just give the people a budget and say just stay within this budget and then you have autonomy to work within that budget the way that you want. But what that does is it creates leaders. The opposite, it actually creates a huge amount of micromanagement.

15:11We can't necessarily go a time of talking about agents and their capability without talking about some of the security risks that are involved. And anybody that's been paying attention to the AI landscape is certainly familiar with those headlines, whether it's Mythos or Fable or now Astra coming out. How does Cisco, how do you think about what agents need in terms of permissions, identity, governance? how would you advise organizations out there right now to get that right so that they can move safely and quickly and confidently when they are able to move with agents i think like the first

15:48Jeetu Patel: thing you have to do on the agent front is you have to be able to know and have an inventory of what agents do you have in your environment and how can you discover them two can you actually give it an identity so that the non-human entities within your organization also have an identity. Third, you have to make sure that you give it some kind of permissioning scheme, like zero trust access, which is basically least privileged access that we give to humans, right? Which is Brian needs the least amount of privilege given so that Brian can get the job done. You don't need to be given the most amount of privilege to get the job done.

16:33Jeetu Patel: Like the least amount of permission. That's what zero trust is. Zero trust with humans was all about access control. I, I allow or deny you something based on your user ID or credentials. And that's what I do. Zero trust with agents is going to be about action control because we in, in, we, We underestimate the amount of common sense that we rely on of humans in the way in which we operate as a species. Let me give you an example. How long have you been at WWT? Coming up on nine years now. Nine years. When you started nine years ago, they had email. They gave you an email account. You got in. You had an email account.

17:23Jeetu Patel: You could email anyone you wanted. Right? I'm sure that WWT has a very good board of directors. But nine years ago, my guess is you were not the hotshot you are today. You came in and were kind of slugging it out a little bit. You probably were expected not to go start shooting out emails to your board of directors. But they didn't have to tell you that. You kind of just knew. Okay, I'm not going to do that. Right? With agents, that's not the case. So with agents, you have to actually think about them with a far different degree of granularity than what you think about for humans. With agents, you have to say, huh, so here's what I'm going to do.

18:08I'm going to make sure that every action is monitored rather than every access is monitored.

18:17Jeetu Patel: Like right now for you, once Brian gets into the system, Brian has access to do anything he wants in the system. With agents, I might not want to give the ability for an agent to send a last email to all my board of directors. So I'm going to make sure that I'm monitoring the agent for when it starts showing drift. I'm not going to allow it to do anything. I'm going to give it action control. control. So basically I'm going to give my agent just enough permission at just the right time or just long enough before I revoke it. Right. And if I can do that, then magic starts to happen because the agent can do what you wanted it to do.

19:00Jeetu Patel: And then it doesn't do anything stupid that you didn't want it to do because inherently agents today, at least are like teenagers. They're supremely intelligent, have no fear of consequence, and have a lot of exercising of poor judgment that they do all the time. And so what you have to have is almost like a monitoring mechanism for an agent. Think about a babysitter when you have a kid at home, or think about a teacher at school. what you have to have is a set of guardrails that the agent is given that says operate the way that you want it to operate to get the task done around these guardrails and if you start to see an agent showing drift from the expectation that you have then you have to be able to intercept it at runtime you can't wait for it you can't do it at design time if that makes sense it does and And I love the teenager analogy there.

20:03I've certainly heard you make that a timer, too. It's a great one. How do we start to drill in some of that common sense to agents so that they do develop a fear of consequence? You know, I go back to when I was a teenager, to your point, lack of fear of consequence. But you only get that through experience, right? So is that kind of where we need to go with agents?

20:23Jeetu Patel: You get that through experience. You get that through a reward system. like what happens when a kid does something right? You actually tell, hey, great job. That was a job well done. And that encouragement makes the kid want to do it again. It's the same thing with agents. When you have reinforcement learning, what is reinforcement learning? Agent does something and then you reinforce whether it's you give it a reward or you tell it that it's not doing the right thing. and that that notion of reinforcement learning is pretty important and then there's this notion of recursive self-improvement and the person of self-improvement says i'm going to recursively continue to keep self-improving by giving it a reward system in place that says you did something i think it was the right thing to do i'm going to give you an award or reward i don't think that was the right thing to do so i'm not going to give you a reward i'm going to correct you so that you know for the next time not to do that.

21:22Jeetu Patel: And you keep training the model until the model makes it in muscle memory. And that's basically, it's not that different from the way that humans learn in some way. It's just right now, AI is far more inefficient from a token usage standpoint compared to humans, right? So an agent might be like close to in the test that we've done 450 % less efficient and more consumptive of network bandwidth compared to a human for getting the exact same task done. Why is that? Because what is an agent? An agent is basically a collection of a skills markdown file, a memory markdown file, a cron job, and tools access, right?

22:09Jeetu Patel: And so if you as an agent have to get something done, you have to keep uploading the context to the model. Here's my context. Here's my skill. Here's how I need to do it. I have to load that into memory so that I can put it into KVCache right away, right? That's a pretty hard thing to do. It's a very expensive thing to do from a bandwidth perspective and from a token usage perspective. Whereas for a human, you just kind of recall it. You've got like this amazing brain that you can recall so many pieces of data. And so now once it gets to For memorizing large swaths of quantitative data, a machine is much better than a human at doing that.

22:53Jeetu Patel: But for instinct and for understanding how viscerally you understand something, I think that the agents are still getting better and better at it. They're not quite there yet. Yeah. Well, you talk about AI agents being extremely inefficient, at least in their current state, compared to what we'll experience hopefully in the future here. And at the beginning of this episode, you also mentioned how just imagine when agents extrapolate even further than when they are now and the type of resource and compute constraint that we'll have. So what is this explosion of agentic architectures, agents, tasks being performed autonomously for people?

23:33What type of impact is that having on the tech stack? And how should organizations be positioning their tech estate for the future, knowing that it's only going to have to handle more and more?

23:45Jeetu Patel: I think we are going to be in a perpetually supply constrained world because we're not going to be demand constrained. Agents are going to have a tremendous amount of appetite to consume more tokens. And we are not going to have enough compute, enough power, enough network bandwidth, enough data center shell space, enough memory to go out and fulfill the needs of AI. And so that's what has to be constantly thought through. You know, what do we do to ensure that that happens in the most efficient way possible and, and know that you're not going to be able to get everything that you need. And so you're going to use and prioritize just like you would prioritize for a human getting a job done.

24:37Jeetu Patel: On a given day, there's about 25 things I could do on any given day. I choose to only do three or four. Why is that? Not because I'm lazy. It's because I can get a few things done well with the amount of tokens I have in a day. Agents are the same way. My tokens are time. How much time do I have and how do I get it done? An agent's token is compute processing power. I like that. Tokens is time. Well, if we're talking about perpetually being in this supply-constrained environment, how would you advise organizations to prioritize what they need to do from an IT perspective? pick the most

25:26Jeetu Patel: pick the hardest problems to solve that have a degree of high data ingestion requirements that have a degree of repetitive work that needs to be automated that have a degree of creative infusion that needs to be put in and pick a few workflows that you can look at end-to-end and actually get them automated because there is no better energizer for the organization than finding use cases that make ai feel like magic and you have to start ensuring as a leader that your your job part of your job is to make ai feel like magic and then convey that to everyone else so that the enthusiasm doesn't wane i think right now what's ending up happening is the narrative on AI has largely become very negative.

26:21Jeetu Patel: You know, I'm going to lose my job. This thing is going to be terrible for safety and security. I'm not really going to know what to do. I'm going to have to be on welfare. I'm going to be on universal basic income. All of those things are very unsettling. And that's all that they hear all the time. What they don't hear is, my goodness, think about the kind of things I could solve with AI that I couldn't solve before. Think about the original insights that AI will generate that I had never even thought about before. Think about the ability for me to be able to solve problems that would have taken me days and days to solve that I can now do within four minutes.

27:04Jeetu Patel: And think about what that does to free up my time so I can do things that are much higher order a it and I can actually concentrate on the most important things that I need to get done and have deep thinking time. Those things can be very liberating. But right now we have actually by, by amplification of certain kinds of media narratives, we have made it very restrictive rather than liberating. We've made it like, oh, so therefore I should not use AI because I don't want to lose my job. well one of the things we did with our team is people were worried about that and what I did was I'm like I'll make it easy for you I will guarantee you that you will lose your job if you don't use AI

27:50Jeetu Patel: and so you decide there might be a chance that you might lose your job if your job is automated on one end but I will guarantee that you'll lose your job if you don't use AI so use AI and then work with us so that we can be more creative on putting you and your brain towards the things that actually matter the most for the company and give you fulfillment, give you joy. Because when something gives you joy and fulfillment, you end up doing your best work and that's actually great for the company. And so that's why you should do it, not because of altruistic reasons. But those things actually tend to be very useful.

28:24Well, Gitu, you've been super kind with your time already. I don't want to let you leave, though, without asking about your new podcast with Cisco, Inside Product. As I understand it, offers a little bit of behind the scenes. Look at what you are working on, some of the innovations you have at hand. Tell me a little bit more about the show and promise me you're not looking to replace me here in this host chair.

28:44Jeetu Patel: There's no chance I could replace you, Brian. You're one of one. But here's the way that I've, this is what we struggled with. We struggled because of the pace at which we were innovating being so fast. We struggled with keeping our field and our partners and our customers up to date with all that was going on. And all of these podcasts, if you don't have it easily accessible, then most people just don't bother going and taking a look at it. You know, and so what I decided was I'm like, what if we get all of our product leaders and come in and talk about the areas that they're working on? And we'll just have conversations with them for 30 minutes.

29:26Jeetu Patel: it's you know something that you can listen to you on a driveway back and forth from work if you wanted and the goal over there is just to get our customers partners and employees smarter about the things that we're building and we've been doing that now we've had 10 episodes that came out we've had fantastic reception on them we will probably have more another 10 that we'll drop in the next few months and and then there's going to be a run rate where we just keep doing them we're building a podcast studio so that we'll always have those. And then we'll just start doing these on an ongoing basis.

30:00Jeetu Patel: And as a customer, if you want to know what we're building, you can go to the podcast as a salesperson. If you want to know, you can go there. And we're being very transparent. It's not this contrived corporate talk. We're just going to tell you what worked, what didn't work, and just keep it very real for the customers, partners, and our employees. Yeah, certainly required listening. If you need a peek about what's coming up into the future, certainly no better podcast than that to understand where the market is shifting, which is already moving very fast. Well, G2, like I said, you've been very gracious with your time, always are.

30:37Hope to talk to you again in another year, see how conservative your predictions were last time around. But thank you so much again. And thank you for the partnership with Cisco and WWT.

30:47Jeetu Patel: we love wwt i i thank you and jim and the entire team for all the work that you folks have done with us and we had a fantastic fiscal year 26 which just ended in august or in july i should say and we will have an even better fiscal year 27 that we're in the first quarter of and wouldn't have happened without wwt you folks have done an amazing job and my only suggestion to you is it's going to be hard for you to keep it just a year apart because there's so much change that's happening that you might have to actually get your clock speed faster. All right. Thanks to G2 for joining. My biggest takeaway, delegation is a skill we all need to develop, but choosing the work, setting the boundaries and judging the results still require our attention, even when an agent handles the execution.

31:35This episode of the AI Proving Ground podcast was co-produced by Nas Baker and Cara Kuhn. Our audio video engineer is John Knobloch. My name is Brian Felt. Thanks for listening. We'll see you next time.