The Rise of Agentic AI as the UX of the Future with Phil Tee, EVP at Zscaler

Today, we're talking to Phil Tee, Executive Vice President at Zscaler, about how agentic AI is becoming the UX of the future. Thank you to US Cloud for being our podcast takeover for this quarter. To learn more about how you can save at Microsoft support, listen to the end of the episode or go to uscloud.com today. You're listening to Joel Beasley, Modern CTO.

You picked the title of today's episode, right? Josh said that. Can you share it? Can you rattle it off, or do you want me to?

So this is agentic AI is the future of UX. And it is a big subject as people are trying to understand and make sense of really what the enduring impact of generative AI is going to be on enterprise technology. I mean, the technology looms bigger than enterprise technology, but it's what I know about.

I know it instinctually in my gut to be true. Like, I can feel it. You know, how do you put data behind it? How do you track the progress? How do you show it actually happening?

You know, this is—I mean, in some ways, and I suppose the pause for thought there is in some ways, you know, it's already happening. So if you think about the speed of adoption of generative AI tools, whether it's the big foundational models and their chat tools that face the internet, like ChatGPT and Perplexity and Anthropic and so on and so forth, Gemini, of course—and my employer, Zscaler, we've witnessed a 3,000% plus increase in the amount of traffic to those sites over the course of the last year. That's a 3,000% increase in a year of the adoption of that. And, you know, what are people doing there? You know, they are going there to obviously answer questions. But a lot of that question and answer response is around how to make use of technology. You're talking to technology about technology. And in some regards, if you like, that is the definition of a UX, right? That is the use of technology to access technology.

So, you know, I would argue that the models themselves are driving that. But there's a very specific, if you like, implementation of generative AI that you'll hear people refer to as agentic AI. And that in itself, I think, is going to be where this really bears fruit and becomes heavily adopted. How would you measure it? There are many, many, many ways. Obviously, we are measuring it in terms of understanding traffic through our zero trust network that we sell to people. But I think that as this market evolves, you may even measure it just with the commercial success of the companies that build applications that are agentic in their nature and operate as an agentic OS or as an agentic UX. I don't think the terminology's stabilized on that yet.

We are definitely waiting for the terminology to stabilize. We're drinking from a fire hose right now with all sorts of new terms and figuring it out. I was talking with Josh yesterday. I said, look, one week Grok is ahead. The next week, GPT is ahead. Then Claude's ahead for different—I have different reasons why I want to go to each one of them. One of the things I've been particularly excited about, I don't know if this is the correct word, but chain of thought, basically watching them reason in real time before—is that what it's called?

Yeah, chain of thought reasoning. And, you know, some of the excitement around DeepSeek when it released its models earlier on this year was the fact that they were training on chain of thought. And, you know, when you ask a large language model to introspect, when it gives an answer, when you ask it to question itself, it gives you better answers. And so, you know, there's a sort of an interesting empiricism about this latest wave of NLP, of natural language processing embodied by transformer architectures and large language models that people don't really know what it can do. And so, you know, there's a lot of measurements being done of the capabilities of generative models to understand where it can go.

At the heart, they're very simple, but in their implementation, they're incredibly complex. You know, so if you think of, you know, what is a transformer model? I mean, it's really just a sequence predictor. So, you know, we think of language as a series of words, a series of tokens in the terminology of NLP. And if you take a body of training data or training corpus, really what you have is a large number of sequences of these tokens. And a transformer model is trained and actually auto-learns—it's auto-inductive on that data. And what it is learning is what are the likely sequences, the likely next token. So when I say, "The cat sat on the," in your mind you already thought "mat." That's kind of what's going on in a transformer model, but in a very sophisticated way.

So that's the simple bit. The complicated thing is, you know, the training methodologies, the optimization methodologies, the—you know, taking the output model of some of the larger current foundation models, which are measured in trillions of parameters, and quantizing that down so that you get a 7 billion parameter model or smaller versions of stuff that you can run on a laptop or a mobile phone. All of that is a huge amount of engineering in it. But what we don't yet know, I guess what has been super surprising over the course of the last, you know, eight to ten years of this technology, is its continual capability to surprise us with things that it can do, which nobody anticipated.

One of the most frustrating things for me is getting into the conversation where people will say, oh, it's not intelligent. It's not really any of that. What it is is it's just predicting the next token. It's just predicting—and given all the information it has, it's predicting the next—that's exactly what we do as humans. Like, you can—where do you want to start? Because you can go all the way down to my cellular biology level. Be like, oh no, no, no, no. He's not talking. That's this cell sending a signal to that cell. And then vibration happens through and then air comes out, and then yeah, but then that's talking. So—

Now, Joel, you're in a deeply controversial area there.

I hope to be. I—yes. You know—

Is a human being, you know, is consciousness an emergent property of a very complicated machine, or is there something else going on? And, you know, it's a—you get into super deep water very quickly because when you say something else going on, you're talking about stuff that is not scientific. You're talking about spiritual things. And, you know, there is no good way to have a scientific discussion about what actually intelligence is.

I think we're getting there. I think we can. I think there's enough people who are interested in it. And, like, for example, consciousness could be an emergent property of a complex machine. We also might think consciousness could be a force. And as you apply it through different substrates, you get different results. You apply it through trees, their communication and intelligence is going to come out one way. You apply it through silicon, it's going to come out another way. You apply it through advanced monkeys, it's going to come out another way. And so, we don't know. We're all talking about it. But I'm excited because, you know, if I were to go to you—I think you can tell I'm excited too, Phil—because if I were to go to you a hundred years ago, right, and I would just say, hey, look, there's this invisible light that's all around us. Yep. You can't see it today. There's nothing to measure it today, but in the future you will be able to see it. You'd be like, what are you talking about? Now we have IR, right? And so I believe that that trend will continue. I believe there's things around us that we can't measure that we will eventually be able to, and it'll tell us more about our universe.

Well, I mean, it's Arthur C. Clarke, right? Any technology sufficiently advanced appears to be magic in this era. And, you know, Joel, my cards on the table. I started out life as a theoretical physicist, and I still publish. I still practice in the space. And, you know, there are many, many very smart people who have said some very clever things about this kind of stuff. But my favorite one, controversial character, Werner Heisenberg, who once said, it's surprising how much you need to know to know how little you know.

And really, the whole space of understanding around, you know, what the technology is going to be capable of in the future is mind-boggling. It definitely is exciting. I do think that you have to channel a little bit, you know, David Bowie with some of these things. So, I mean, David Bowie is sadly deceased. He did a video blog, a podcast, an interview, I guess, before it was called this kind of thing, many years ago, '99, around about the turn of the century. And he predicts an awful lot of what we've seen happen. But the reason why he came to mind for me is there is an album of his called The Man Who Sold the World, and typically sort of dystopian Bowie, very early 1970s stuff. There's a song on there called The Savior Machine, and there's a line that always sticks in my mind whenever I think about generative AI, what's going on in the market, which is, "You can't stake your life on a savior machine." And the biggest danger, in my view, with generative AI is that people come to believe it to be authoritative when it really isn't authoritative. And to follow your analogy, which, you know, consciousness can be expressed in many different ways, you know, the one thing about intelligence is it's fallible.

Yeah, 100%. And, you know, I know—congratulations on being a physicist. That's very smart. Cards on the table, I watch a lot of movies. So that's what I got. That's what I got. Oh my goodness. This is fun. But, yeah, I find that it takes a certain special someone, and you're definitely one of those people, to have conversations with me in the sense that, like, I'm a layman. I'm not trying to pitch or sell a specific thing. I'm more of a "what if." Like, what if this could happen? I'm not saying that, like, I fully 100% believe it and I'm willing to defend it. I, for lack of a better term, like to play with ideas. I like to play with potentials and possibilities. And then some people don't like to play with them. Some people, I'll be like, want to play? And they're like, no, it's this way. I was like, oh, okay. We can't play.

And you should play. You know, and kind of almost circling back to what is intelligence and what sets human intelligence apart, which is, again, a very difficult thing to sort of nail down philosophically. But maybe play is a big part of it. The sort of the objective-free exploration of ideas and testing of ideas. You know, there's a—I mean, back in my original sort of set of endeavors in physics, again, littered stories of where great innovations come from, whether it's Richard Feynman observing a plate spinner. And, you know, what you'll notice if you see a plate spinning is that it's obviously going around in circles, but it's also wobbling on its lateral axis. And there's a very well-defined relationship between how fast it's spinning and how fast it's wobbling as a plate.

And so the story goes that he was observing this at a point in time in his career where he felt washed up, kind of post the Los Alamos stuff, and, you know, just decided to start working out on the back of a napkin, you know, what the relationship was between those frequencies. And it immediately took him down a path as he got absorbed in playing with that idea that ultimately resulted in, you know, quantum electrodynamics and eventually a Nobel Prize that he shared with Schwinger and Tomonaga. And, you know, that's an example, but there are millions of other examples, um, whether it's in hard science or, you know, engineering, or for that matter, you know, J.K. Rowling and Harry Potter, you know, where you play with ideas. Creativity is maybe the thing that comes from that.

I read a little bit of his stuff, Feynman, when in my early twenties. And because, you know, I was big into systems engineering, well, in software. And while he wasn't—I don't think he was specifically talking about software—he was talking about science, but he was talking about systems in general. And there were a couple of quotes from him. I can't remember them, but I think there was one—I want to say it was him—but something along the lines of, rather than fixing the broken system, you just build a new one. And that, to me, was—whether it was him or someone else—that was very useful because sometimes with that new perspective, I could now look at code bases or applications, whatever I'm building, and ask myself, is it worth the effort to fix this broken system versus build a new one?

That—oh, sometimes, I guess you might say that's kind of a version of the sunk cost fallacy. You know, I think for engineers, one of the most powerful tools that you have at your disposal is "rm -rf *," and people very rarely do it. There is a tendency to tinker on with systems rather than going, well, I'm going to be brave enough to say, what I'm going to carry over from the past is my learning, and what I'm not going to carry over from the past is my code. I'm going to think again. And whereabouts are you based, Joel? What's your—

Nashville.

Nashville. Ah, beautiful. If you get out to the Bay Area—

Oh, yeah. I spent some time there. Yeah.

There is a—it's a touristy destination, kind of kitsch really to go to, called the Winchester Mystery House. And so the Winchester Mystery House is the ancestral home of Sarah Winchester. I think it was Sarah Winchester, the heiress to the Winchester rifle fortune. Okay?

I'm already sold. I already like it. Yeah.

So she went slightly batty in her later years. And she became convinced that if she ceased building on this house, that she would be haunted by the spirits of the people that had been killed by Winchester rifles. So as a consequence, this building is just—it's bizarre. It's labyrinthine. There are stairs that go into ceilings. There's windows in the floor. There are hundreds of rooms in this thing. You know, apocryphal—people have been lost for days in it. It spawned many horror movies. So if you're a fan of Stephen King, there's a book he wrote called Rose Red, which is loosely based on the Winchester Mystery House story. But I always use this—and there you go. I always use this as an example for, you know, young engineers or people in my team to say, this is what happens if you don't "rm -rf *." This is what your code looks like as a building. You've got to have the discipline to reinvent. And it's a spectacular place to go.

Those stairs, by the way, that you just saw there that are kind of these really weird, very low incline path stairs—she also suffered very badly from arthritis of the hips as she got older, so she couldn't walk up normal stairs. Kind of wacky. But if you're out there, it is—you know, I guess on my sort of Bay Area computer nerd tour, it's probably that followed rapidly by the Computer History Museum, which is not a million miles away from the Winchester Mystery House either. And you can go look at some fantastic bits from our technological past, like the Cray supercomputers.

Can you, like, walk in them, the giant computers that are the size of a room?

(Phil Tee at 00:17:48) So the Cray YMPs, and these were Seymour Cray's amazingly high-speed computers, I guess, probably the seventies and the eighties, early eighties. Very famously, they built it as a cylinder so that the wire lengths to the bus from the various different PCB cards was minimized so as to reduce latency on the I/O bus for the computer. So you have this cylinder, and it's an act of genius. Around the cylinder, they put benches. So there's like a circular bench.

(Phil Tee at 00:18:24) Now these things change hands for a lot of money now to have them in your front room is kind of a sort of a, you know, kind of—

(Joel Beasley at 00:18:31) Entry.

(Phil Tee at 00:18:31) Yeah. Yeah. It's like, look at my YMP that I have as my sofa.

(Joel Beasley at 00:18:38) My wife would say no to that.

(Phil Tee at 00:18:40) Yeah. Well, yeah. Along with the life-sized Dalek, if you're a fan of Doctor Who, right? From this planet Earth, you can go and buy an actual BBC Dalek. So that's maybe my sort of nerd fantasy den. So that would be a Cray YMP and a Dalek from Genesis.

(Joel Beasley at 00:19:00) What's the nerdiest thing that you own?

(Phil Tee at 00:19:03) The nerdiest thing that I own. So it's a physics nerdy thing.

(Joel Beasley at 00:19:10) Okay.

(Phil Tee at 00:19:11) So I have a Physical Review bound copy from 1948, which has the original papers that were actually spawned from Richard Feynman's PhD thesis, which establishes something called the path integral formalism for quantum field theory. And, you know, occasionally I'll get that out and sort of flick through it and marvel at the fact that sort of active in this period—it's 1948, of course. It's a long time ago. As well as that one particular paper, which is kind of special, you know, there are publications from Albert Einstein in there and Victor Weisskopf and, you know, all these kind of sort of giants of theoretical physics from the era.

(Phil Tee at 00:19:54) And it's just kind of fun to flick through and go, you know, what do we know now that, you know, would seem kind of crazy to them? So that's probably the nerdiest thing that I own.

(Joel Beasley at 00:20:08) I want it. That is so cool, by the way. It's—you're definitely my type of people. I want to know about agentic AI in the business setting today. So we talked about it at the beginning as a high level. Yes. It's coming. Yes. It's like a whole new civilization of people, but with artificial intelligence entering into our ecosystem, all that good stuff. And I'm very excited about the agents, and I'm very excited about all of it. But what is happening today?

(Joel Beasley at 00:20:38) Like, where's the "you are here" point?

(Phil Tee at 00:20:42) Let me just contextualize it with a super condensed run through to the where we are, because I think it helps explain it. So as we talked about earlier on, you know, the transformer models and the large language models, in essence, what happened from 2017 onwards was a massive scale up in the size of these models and the amount of training data that we use to train them. So we went from, I think, you know, GPT-1 was like a hundred million parameters or something like that. GPT-4o is 1.7 trillion. So it's just a huge scaler.

(Phil Tee at 00:21:22) And the training data that is used to train these things, you know, is all of the internet. So as much of the digitized content of the library of all human activity has been sort of fed into this beast and trained. Great stuff. The problem, of course, is the time and expense it takes to train a foundational model is astronomical. Hundreds of millions of dollars, months and months and months.

(Phil Tee at 00:21:52) You know, when Elon Musk got xAI off the ground, you know, and hunted around for a final location to do it, you know, they caught it very quickly because, you know, they were trading on effectively what other people had done. But, you know, a lot of it was being able to find the necessary square footage and power supplies, access to water cooling to be able to run these hugely intense computational tasks. The upshot of all of that, of course, means that if I give you a model of any sort, it is effectively frozen in time. So it is trained on everything that was, you know, up to maybe, let's call it January 31st of this year.

(Phil Tee at 00:22:37) So, for example, it wouldn't know about the outcome of the Canadian election that happened yesterday. So if you went to one of those foundation models and say, you know, how did Mark Carney beat Poilievre in the Canadian election? It'd be like, I don't know, and give you sort of weird answers. So how people started to engineer around that was they started putting context into prompts. So, you know, the prompt is just the text you type in ChatGPT, but there is a whole science around constructing prompts to get highly reliable answers from it.

(Phil Tee at 00:23:18) So you can provide hints and tips to the large language model about how to treat your question and the data it's put in there. And importantly, one of the things that you can do in a prompt is you can provide examples. So people started deploying an architecture sometimes referred to as RAG, which stands for retrieval-augmented generation, where you would take the question. So in this case, why did Mark Carney beat Poilievre? And it would initially take that question, and it would go to another index of documents that you maintain to find relevant documents, summarize that probably through another language model to generate a prompt which would provide more context for the large language model to be able to answer the question.

(Phil Tee at 00:24:09) Now that became almost a de facto architecture about a year, year and a half ago. Many people have deployed that, particularly in the enterprise, as a way of really indexing and searching, making it straightforward for people to query things like, you know, documentation, as a good example. But you still don't have really up-to-date information. So we still probably struggle with the Mark Carney, Poilievre question. So this is where agentic AI comes in.

(Phil Tee at 00:24:41) And this is where you build maybe an agent that knows how to go and query, whether it's the CNN websites or the BBC website or the Canadian Broadcasting Corporation websites to get up-to-date real-time information. And then these agents are fronted by a thing called a planner. So I send my query, why did Mark Carney beat Poilievre? And what the planner's job to do is to look at that question and go, what tools, what agents can I ask to go and retrieve relevant information to help answer that question and assemble it in such a way that I can present it back to the user? So this planner might say, go do a web search for recent news on Mark Carney and Poilievre.

(Phil Tee at 00:25:34) Another agent might be tasked with, you know, going to run some statistical analysis on, you know, the six months of polling prior to the election. Another agent, you know, might be tasked with going to read the last hundred days of Donald Trump's Truth Social posts because guess what? He's responsible for Mark Carney winning the election. Maybe not go there.

(Joel Beasley at 00:26:04) I don't know anything about Canadian politics, so—

(Phil Tee at 00:26:08) It's the whole tariff thing that's kind of got Mark Carney elected. But, anyway, you pull all this together, and now you're dealing with a language model and real-time information that the model wasn't trained on, but is good at summarizing back into an answer. So agentic AI is all about the idea of having multiple semi-autonomous software agents that may themselves use a language model in the operation of their task. But the idea is that they're able to retrieve data from a given particular source or perform a data processing task or an analysis task or wherever it might be. And they are coordinated by what is often referred to as a planner, which is kind of like an agent manager that forms the primary interface.

(Phil Tee at 00:27:06) And the reason why this is starting to gain such a lot of traction is all of a sudden, you kind of square the circle of you get all of that amazing goodness from the large language model in terms of, you know, its capability to understand language, to reason, to build tasks together with real-time, up-to-date data and specialized active interfacing into other tools. And it's suddenly becoming a huge market in its own right. There are early attempts to standardize, for example, the agent frameworks. Google's got one protocol with Gemini 2.0. There's also Model Context Protocol as we're talking to agents that's out there.

(Joel Beasley at 00:28:00) There's, yeah, it's getting—

(Phil Tee at 00:28:01) There's LangChain. There's all kinds of tools that, you know, that people use to build these agentic frameworks. And I think that's going to be the Wild West for quite a while. But where I'm excited about it is the prospect for this to yield the killer app for generative AI because all technology for it to become widely adopted requires a killer app. So, you know, famously, the PC was nowhere until Lotus 1-2-3 was invented.

(Phil Tee at 00:28:45) And then all of a sudden, this spreadsheet drove the adoption of the personal computer. So you need this killer application for it. And I think when I was talking to your colleague, Josh, yesterday, you know, I talk about the three-bubble Venn diagram for all technology. There's, you know, capability. So what is this technology capable of doing?

(Phil Tee at 00:29:09) There's utility, which is how useful is that thing that it could do? And there's impact, which is how much does it change? How much does it disrupt? How much does it alter the world by being in place? So if you think back to the spreadsheet example, capability, well, PC is very capable adding up lots and lots and lots of numbers. The spreadsheet is a tool for making that an easy way to unleash that power.

(Phil Tee at 00:29:39) So a spreadsheet is very capable at accounting functions. Utility, well, you know, if you've ever seen inside of a boardroom before Lotus 1-2-3—I'm not actually old enough to have seen inside of a boardroom—but people literally did do business planning on chalkboards with grids marked on them and manually calculating numbers in sales to, you know, what happens if I open a sales office in Kuala Lumpur? Right. Change the numbers and see what happens.

(Phil Tee at 00:30:08) So utility—it's very useful for business functions. Then impact. Personal opinion, this is unscientific, but I'm sure there's data out there that would substantiate it. I think it's probably the biggest boost to productivity in business in a hundred years. It's just utterly changed. I would never have been able to do what I did in my business career were it not for Excel as a—obviously, the successor to Lotus 1-2-3.

(Phil Tee at 00:30:41) So when you get the alignment of those three things, the magic happens in disruption and dislocation in the marketplace. So back to generative AI. Capability. A lot of what people are focused on at the moment are capability-demonstrating-only uses of gen AI, which is, you know, hey. Look.

(Phil Tee at 00:31:03) It's amazing. I can ask it to compose a Shakespearean sonnet. It did it. Or hey. Look.

(Phil Tee at 00:31:09) It's amazing. You know, I can get it to make this robot walk around, you know, my factory floor without bumping into things. That's amazing. But it's just capability. So, you know, utility—yeah, ChatGPT is really useful if you want to avoid having to do a homework assignment and get somebody else to do it for you.

(Phil Tee at 00:31:32) But, you know, maybe arguably the impact for that is negative.

(Joel Beasley at 00:31:42) It's positive.

(Phil Tee at 00:31:42) It's what you're spending your time on. Yeah. Yeah. That it liberates. So my view with the agentic AI—and I have this sort of vision of the agentic OS as the new UX—and it goes a little bit like this.

(Phil Tee at 00:31:57) I'll spin a story up. I'm an employee of Zscaler. Love the company I work there. It's fantastic. I've worked in a bunch of businesses in my life.

(Phil Tee at 00:32:09) Zscaler's awesome. Big plug there for Zscaler. In a prior life, I worked in a very large corporate very briefly after they acquired my company, and they had a whole bunch of systems, some of which were kind of old because the company itself was pretty old, but I would have to deal with to, you know, to be a good employee of the business. So, you know, you'd have expense management tools, you'd have accountancy management tools, you'd have HR employee management tools. And each one of these is a piece of enterprise software that does a thing.

(Phil Tee at 00:32:48) And I would have to learn how to use it as a user. And once I'd learned how to use it as a user, I would then go there to do that specific thing. So if I've got an expense to submit, I go to the expense system and I submit my expense. If I'm updating my budget, I go to the accounting system, I'd update my budget. If I was, you know, filling out a review on somebody, I'd go to the HR system and fill out the review.

(Phil Tee at 00:33:12) Now maybe the last fifteen years of enterprise software, driven by the mass consumerization of enterprise software, is trying to elevate those individual applications to like your iPhone. You know, your iPhone doesn't need a manual. It's beautiful. Maybe you don't have an iPhone, but you got something similar.

(Joel Beasley at 00:33:32) Oh, I got an iPhone.

(Phil Tee at 00:33:33) Yeah. It's an amazing bit of technology. So consumer electronics is fantastic. It's deeply intuitive, you know, so easy that every person over a certain age knows somebody under the age of 12 that can help them get it.

(Joel Beasley at 00:33:48) I've got three of them. Yeah. I like the—there was a book way back in the day called Don't Make Me Think. Yeah. And that was one of my favorite books.

(Joel Beasley at 00:33:57) I was like, yes.

(Phil Tee at 00:33:58) On UI design. Yeah. Absolutely. Brilliant book. So I would say that there was a wave of enterprise software, and I like to sort of typify it with a vendor like Datadog as an example, that just made monitoring like beautiful.

(Phil Tee at 00:34:17) You know, their UI was much acclaimed. They pioneered a way of selling called product-led growth. And the idea behind that is the product's so simple to use that, you know, they put a free trial out there. You go to the free trial, you know, because you want to, I don't know, manage your Kubernetes cluster. It tells you how to do that.

(Phil Tee at 00:34:37) It builds journeys through the tool till you get to a point where it's like, wow. I need to buy this thing. And then it spends its entire time encouraging you to buy more of it. And then you wake up on a Thursday, and you got a $20 million subscription to Datadog. So beautiful elevation of the art of enterprise software.

(Phil Tee at 00:34:58) My point with the agentic UX is that at the end of the day, that is still a fixed collection of UX assets, which I have to learn in order to conduct a fixed mission. Well, how about if the world was like this? How about you throw all of that away? I'm back to being an employee of this company with all these disparate systems. And I log on to my agentic system, and it says, you know, Phil, it's Tuesday morning.

(Phil Tee at 00:35:28) It's getting close to the end of the month. We should probably do your expense report. Would you like me to help you do that? And I go, sure. And he says, okay.

(Phil Tee at 00:35:37) I've been through your phone. I've collected some photos of some receipts. So I noticed that you took an Uber ride on Tuesday. Was that business? Yeah.

(Phil Tee at 00:35:46) It was. Yeah. Okay. I filled all that out and I submitted it. Oh, and by the way, it looks like you don't have your corporate credit card in your Uber account.

(Phil Tee at 00:35:54) Would you like me to update that? Yes. I would. Thank you very much. And off it goes and does that.

(Phil Tee at 00:35:58) Then he comes back to me and he says, you know, Phil, you spent quite a lot on travel last month. So maybe you need to go and update your budget projections for your group. Would you like me to walk you through that? And so I, you know, said, yeah. Of course.

(Phil Tee at 00:36:12)
And it starts putting in front of me some analysis of my budget and showing me how things are trending. And it does some deep math on all of that and says, "You know, you should probably just throttle back a bit on the hiring in September and maybe push a few of those heads out to January. Would you like me to do that and produce a report for you that you can take to show your boss to get them on?" You know, because you need to convince these three or four people. You see where we're going with this?

(Phil Tee at 00:36:39)
There's no mission that I've gone to that site with. It has understood what I need to be effective and has abstracted away all of the disparate systems into a single, all-encompassing, dynamically evolving UX for me. And I think that if you look at that capability-wise, agentic AI can do that.

(Joel Beasley at 00:37:05)
That's a natural progression. I mean, haven't we always been abstracting? I mean, look at ORMs when they came about, and then everyone's like, "Oh, you can't write SQL." It's like, you don't need to. We abstracted it.

(Phil Tee at 00:37:17)
Exactly.

(Phil Tee at 00:37:18)
Yeah. You know, capability tick, utility. Clearly it's very useful. It helps me do my job more efficiently and quickly. Impact, holy cow. Like, it utterly changes what I spend my time in. And, you know, here's the thing, and it's the lesson of enterprise technology since there's been IT—since Tom Watson Jr. sat next to the CEO of, I can't remember, American Airlines, and told him about this thing he invented called a computer in 1955 or wherever it was. If you own the interface, if you own the method of interaction with the tech of an individual, you own the individual. It really is eyes on screen and time spent. And, you know, I'm so excited about that as a direction. You know, it's obviously something which I'm clearly directly researching in terms of how we can make that happen in a cyber world inside of Zscaler. You know, but I think we're gonna witness over the course of the next year to five years the growth of these agentic UXs to various bits of enterprise technology.

(Joel Beasley at 00:38:43)
I think I am—well, I know I'm a 100% on board. The moment that the LLMs hit the mainstream with GPT about three years ago or whatnot, immediately, I was like, "Well, you know what comes next? The App Store." Because that's what always happens. It's like, then you'll have this app store. But then it doesn't always happen the way you think, because then, like, they started coming out with some apps, but then, like, agentic kind of—and it's everyone. It reminds me a lot of, for my entering into the internet world, all the CRMs—or not the CRMs, but the content management systems, the CMSs. There was a million of them. They all had their different ways. They all had their different styles and heuristics. And then all of a sudden, like, WordPress came out. And for a long time, there are many other brands, like, "We're gonna roll our own internally." And then eventually, we got to the point where just, like, everybody just agrees, like, "This is just the standard." Now it takes time.

(Phil Tee at 00:39:38)
It does. Yeah. And, you know, maybe there is a number of forcing factors that will cause an evolution quite rapidly this time. I mean, for a start off, there are a bunch of foundational models out there, and they are all blowing a humongous amount of money at the training and operation. So you think of, you know, whether it's Gemini or OpenAI, I mean, they're intrinsically unprofitable operations. I mean, they spend a lot of money building a thing which people are not demonstrating that they're willing to pay for yet, nor is there a well understood monetization methodology as well. You know, the interesting thing there is if you think about Google—I mean, some people, you know, what was the genius of Google? Who was the genius of Google? People go, "Oh, you know, it was Brin and Page because of the PageRank algorithm, which completely blew away AltaVista, if you remember that, or any of these old search tools." And I would argue that as cute as the algorithm is, and it is pretty cute, actually, the genius of Google was Eric Schmidt, AdSense, and AdWords. And, actually, even that was really bought in from a company called DoubleClick. So it was—and one of the founders of DoubleClick, he's doing well. But the bottom line about it is great technology is awesome. A way to sell it and grow it profitably, you know, makes it ubiquitous. So I think there's gonna be a lot of commercial pressure for these people that are driving the foundation model players to find that killer app and lean in. And also, what will happen is the market has got to consolidate. I mean, there can't be—in exactly the same way that, you know, there are fundamentally two operating systems today. You're either running Unix or you're running Windows or you're not running anything, more or less. You know, I mean, the variance of, you know—

(Joel Beasley at 00:41:52)
Yeah. Yeah. That's covered 99.9% of the computers.

(Phil Tee at 00:41:55)
I mean, even the thing I'm talking to you on, a Mac. That's a UNIX box.

(Joel Beasley at 00:41:58)
Yeah. Yeah. Of course. Yeah.

(Phil Tee at 00:42:00)
You know, under the skin. So there's—you know, the foundational models will consolidate down to a Pareto. There won't be three. There'll be one big one and another one that, you know, there's a bunch of diehards that are still using, and then nobody else. So there'll be that consolidation. And then I also think that there are—there's a sort of a—there are dark clouds on the horizon economically. Not—I'm not trying to search for those. It's not geopolitically. It's geoeconomically. So, you know, I grew up in a world where globalization was a given. Highly distributed supply chains were a given, and, you know, the only way is up into the right. And I don't think that's true anymore. I think, you know, the world economy, with everything that's going on, seems to be teetering on the edge of a nasty recession, and supply chains are disrupted, could get even more disrupted if there's instability in the southeast of Asia. And, you know, what all that means—and the reason why I bring it up is—you know, let's say there's a world in which there is a genuine shortage of NVIDIA chips because there aren't any being shipped to the US. That will force a kind of a make-do constraint on the market. And you might go, "Well, you know, what the hell is he going on about? I mean, this seems to be completely irrelevant." Until, you know, I tell you that all of the really big companies that you can think of—the market winners grew up in the teeth of very bad economics. Cisco, in a recession. Amazon survived the dotcom bust. You know, all of these—you know, there needs to be a kind of bit of grit in the oyster.

(Joel Beasley at 00:44:01)
Yeah. It's like a pruning. It always happens. Yeah. The market shrinks back up, and then people get really efficient. And it's almost—and when I look at it happen throughout my, you know, almost 40 years here, there's a couple points I can point to, and I'm like, "Okay. I remember that because I lived through that, and I can draw from those experiences." But it's almost like a beautiful thing that happens. It's a way that we're all communicating without, like, direct—it's not like a one-to-one. It's like the hive mind. Like, as a hive mind, we do these expansion and contraction things, and I love that.

(Phil Tee at 00:44:35)
There's a guy called Philip Kaplan who wrote a book after the dotcom bust called F'd Companies.com. He used to—in fact, he used to write a blog about this. And it's a very snarky, amusing canter through, and he basically lists out, I don't know, might be 30, 40, 50 companies that went bust in the dotcom bust. And so, you know, companies like Webvan, for example, and Boo and all the rest of it. And what's really interesting is if you read that now, for every one of those companies, "Oh, don't be so ridiculous. There's no way that anybody's gonna run an online grocery delivery service ever." And, you know, there are three that are worth billions now. You know, there was Drive.com. Online file storage. Well, guess what? Everybody pays to store data in the cloud. You know, Boohoo, even the brand's back. You know, it's—and one of the great lessons from this—and, you know, you ought to be an eternal young entrepreneur—if you wanna become a billionaire, just wait for the next big bust in tech and look at the companies that are failing miserably and steal their idea and reappear in five years' time.

(Joel Beasley at 00:46:03)
I'm writing that down. I want riches. We have about 10 minutes left. I know you got a hard stop, so I'm watching the time. A couple questions just so I can filter them down to the things I'm most interested in. Where are you spending the most time in your personal life as far as AI? I'm using Grok probably the most personally. What tool are you going to in your—outside of Zscaler and work—just your life?

(Phil Tee at 00:46:31)
So probably, I'm fairly conventional. So Perplexity, ChatGPT, Gemini is probably the three places I go. I don't use Grok, largely because I'm sort of off Twitter and all of that kind of stuff for no other reason that, you know, well, let's just say I preferred it when it was under its old ownership.

(Joel Beasley at 00:46:58)
Yeah. I, um—it's actually separate now, Grok. So, I mean, it's, like, within the X, but I didn't use it as much when it was in the Twitter app because I don't use Twitter a whole lot. It's just not my—I like Instagram. I'm a very visual person. Anyways, they broke it out into its own app. And, you know what? I would say for a couple weeks, I thought they were, like, way far ahead of GPT, and then GPT kinda caught up. And to be honest with you, they're all pretty great. But one of the things I do like about GPT is its ability to remember across conversations pretty well.

(Phil Tee at 00:47:37)
Yeah. The—I mean, what I do find is on any given day, like, I don't spend my time in one. I'll spend my time across multiple. On any given day, one is running slow or at least one is slower than the other. So, you know, I'll exhaust ChatGPT and it's like, "Oh, it's taking ages to answer this question." So I'll just go over to Perplexity or Anthropic or whatever and, you know, give it a go. But you're right. I mean, they are, I think, extremely comparable in terms of—I mean, I'm sure they would dispute that and say, "Well, you know, you're an idiot because this graph or that graph, and this is why we're best," but I don't notice much difference.

(Joel Beasley at 00:48:22)
Intelligence. Now we noticed, like, humans—here's just a big question for you. I'm just gonna ramble for a second, and you try to make it sound cool. Humans, the way that we interact, we have specialists. We're specialists. Right? We're getting really smart. So you bring the right specialists together and organize them. We're also starting to see that happen in the large language models space, where the larger you get with that context, under the—it almost gets stupid the longer the conversation goes. You need a new instance of it. And so is—do you think this is a byproduct of the way we created the LLMs or the way humans think? Or do you think this could be just, like, a property of intelligence?

(Phil Tee at 00:49:04)
Wow. So this is the—you know, like, so let's say there are, you know, little green aliens on that planet where they discovered, whatever it is, methyl disulfide last week. And, you know, this is the—you know, with their—would their ChatGPT look like our ChatGPT type question? Is there a sort of an inexorable teleology—there you go, got the dictionary out—of technology? Do we all end up converging on the same model because that is a reflection of what it means to be an intellectual reasoning being, or, you know, is it just not that constrained? You know, does it—you know, would our—you know, they'd look at our stuff and go, "Why are you doing it that way? You know, this is the way of doing it." And I have to say that, banti Richard Feynman, he was very, very famous for coining the phrase "shut up and calculate." Because at around the time of when he was influential in theoretical physics, people started talking a lot about the philosophy of quantum mechanics and, you know, what does the measurement problem mean? And, you know, what is—what is entanglement? This entanglement thing seems very sort of mystical Eastern philosophy that people like, I think it was Murray Gell-Mann, was very influenced by. And he used to just say, you know, "It's impossible to understand it. You know? Philosophy is not what we're paid to do here. We're paid to calculate nature, so shut up and calculate." And in some regards, you know, I like to think about, you know, what does it mean to be intelligent, and is a machine possible and capable? Is Geoffrey Hinton right? You know, in 10 years' time, my computer will experience love, or is he just, you know, making it up? But on the other hand, the deeply pragmatic and practical side of me comes out and says, "Well, you know, really, the ultimate litmus test is, is it useful? And, you know, can you build a tool out of this technology which enhances the human lived experience?" So whatever that might mean. Where it might mean, you know, make bankers more efficient at trading derivatives, and some people might say, "Well, that's not very good for you." It might be curing cancer, which some people might say, "You know, that's a far better use of your time." By the way, who am I to judge? Because unintended consequences.

(Joel Beasley at 00:51:51)
We don't know. And I think the humility of us not knowing and us just making the next right move—actually, so I'm 37 just to give you context. And where I'm going through in life, I've got three young kids and a wife. And my current philosophy, been operating it for about maybe six months, way of dealing with the world, is make the next right move. So I had to simplify it because there's so many different situations, and the variables are significant. I got business, family, life, faith. I got all of these different things, and it's like, I need some rule set for myself, some startup prompt sequence. And I was like, "All right. Do the next right thing." And I instinctively know what that is most of the time. Even when I feel uncertainty, it's more uncertainty about doing it, not knowing which is the right thing to do.

(Phil Tee at 00:52:39)
Yeah. And, you know, look, I always say—I teach digital entrepreneurship at the University of Sussex on and off. And, you know, one of the things I say to the young students there is, you know, it's about the art of letting life happen to you. You know, as Woody Allen once said, 80% of life is showing up. You have to be—rule nothing out, but try and respond to situations in a quality way.

(Joel Beasley at 00:53:09)
The most frustrating thing for me, Josh, is when I call support and I can't understand them.

(Intro Narrator at 00:53:14)
Yeah, man. I hate that.

(Joel Beasley at 00:53:15)
That's why I like US Cloud. Not only is it better, faster support, but all the engineers are US-based engineers, and it's also a lot cheaper. 94% of US Cloud's clients report saving a third or more when switching from Microsoft Unified Support to US Cloud. So now you'll just have to figure out what to do with all that extra money. If it were me, I'm responsible, so I'd reallocate the money to improve my team. Josh, what would you do?

(Intro Narrator at 00:53:40) I'd probably just buy more guitars.

(Joel Beasley at 00:53:43) More guitars. Visit USCloud.com to book a call and figure out how much your team can save. Thank you so much for listening, and if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected].

(Joel Beasley at 00:54:06) Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.