David Pessis: Hey, everybody. Only a lot of early birds here, I'm impressed.

Pirasanna Sivalingam: Probably because we're just having lunch, Dave.

Marie Thurston: Yeah.

David Pessis: Yeah. To get that to him, Sri Ping, he'll be here in a moment, so you guys got a few more, you know, a few more minutes to eat peacefully. I'm having yogurt.

Pirasanna Sivalingam: I just had a very spicy tomato soup.

David Pessis: I'll wake you up, huh?

Pirasanna Sivalingam: Yeah. Okay.

Srinivasan, Sri: Um... I haven't seen a more chaotic place, workplace in a long time. And yet they've just breached north of 30 billion. About 14 months back, they were about a billion dollars in revenue. They've gone 30X in such a short period of time. You can kind of relate and understand what they've done. And on the flip side, if you look at AI adoption in US companies, it's about 47% of companies has at least a 10 user license of ChatGPT, that's the measure. up from 26% for a finance guy that is like, oh, amazing growth. But for a person like me in the enterprise, that's tepid. What's wrong?

Srinivasan, Sri: The models are doing really well. The adoption in the enterprise is low, and I chalk it up to organizational inertia. What is lagging now across enterprises is basically a plan to diffuse AI in the enterprise. And that's basically opportunity space for all of us. Okay. Now, the models have been around for 3 1/2 years. And for the first three years, it was like autocomplete and a bunch of other things. So let's take software engineering as an example. They were not useful. They were good at finishing up your line, finishing up your code. And then suddenly something happened. And that something happened at the end of December, November, December. And what ended up happening was...

Srinivasan, Sri: These guys created this concept of a harness. Claude code is a harness, codex is a harness, cursor is a harness, and what these harnesses do, the best thing, the best analogy I have is they are the thing that hides the works of the model and does all the dirty work on top of it, the software layers on top of it. that basically manages retries, gets you to the outcome you want, has the context and the memory and all those things. And it led to, for example, cloud code. You know, now the code is all out there. Some of that code is pretty crappy, you know, deep retry loops and all that stuff. But, despite all that... There are hands-free engineers at places like Anthropic.

Srinivasan, Sri: Zendesk, one of our portfolio companies, has 300 hands-free engineers. 300 are the 1800 are hands-free engineers. Now, it has to be done very carefully. But remember, these AI labs are leading indicators to behavior change that's going to be upon all of us. So a few months back, there was this discussion of skeptics, AI skeptics. If there are AI skeptics in the room, I will ask you to kind of think about it. This is the same skeptics who basically thought that calculators had no place in a math class.

Srinivasan, Sri: So keep that in perspective as we go. But these automation levels, if I were to look forward a year, you're going to have 20 times the code. A large portion of that code, my crystal ball says, is going to be AI slop. And if us as engineers show up well, not as authors, but as reviewers and agent bosses and commanders of these fleets of agents, if we show up that way, we'll have lower amount of slop and drive higher degree of precision in what we do. But there is no scenario in which I see. No agentic factory in successful organizations. Put another way, successful software shops will become agentic factories. That is a, that is a must.

Srinivasan, Sri: So. So the thing I wanted to put is a very, very simple graph. You know, you guys can guess which wave is which, but this is basically the race to 100 million in revenue. And the latest one is the fastest, 200 million happens to be cursor, not OpenAI or. Or anthropic, or any one of these, but these guys, these are the they are the latest and the fastest 200 million, uh, and that basically AI is now. faster than any other wave. You know, you had the Industrial Revolution back then, which kind of as a concept started really well. And it took like 150 years before kind of the manufacturing plant got automated or this, you know, 30 years for the Steam engine to go from demo to actually be real and viable between Lancaster and Manchester for those of you who are kind of history buffs on this space.

Srinivasan, Sri: And the challenge remains, like I said, diffusion lacks capability. And that diffusion is something that has to be top of mind for all of you. It has to be, you know, charity begins at home in what you do every single day. You need to make sure AI. is playing a relevant role. I've had this happen to me at home. I have two daughters. One of them is a software engineer at Microsoft. And, you know, she's a... She's an AI skeptic. Rather, she was an AI skeptic. And one of the challenges I saw with her was, I don't really want to do this. I love the art form of coding. I don't want to give that up.

Srinivasan, Sri: And what I found is, you know, in her groups, they started measuring how many people are using what AI, what percentage of AI contributions are people making. And they made it a gamification exercise. My daughter, everything I talked to her did not resonate. Let me call myself an abject failure on converting her to an AI user. But what changed was the gamification. She's a highly competitive person who basically said, let me see how this works. And now she's forgotten that she was a skeptic. And she tells me all the virtues of AI and all that stuff. So that diffusion is something each of you needs to think about every single day. Try it.

Srinivasan, Sri: The first tries are going to be a failure, persevere, and then you will get to see more outcomes. Because it takes six or seven tries before you actually can get it right. And once you get it right, and once you give it a lot more context, remember, it's like training a middle schooler, eventually growing them into college, and basically making them effective for you. Spend that time so that. You can now say you're using AI and it takes even more work to see the art of the possible. So diffusion is your responsibility in how you do things. And when you do that, the products you build for your end customers, the nursing facilities, the ambulator care facilities, will see the benefit of doing so. All you have to do is, is it effective or not? I ask you all, I implore you to look at the API volumes last year and this year. Compare the two. It's nearly 20x in your environments, and it's not from customers querying you more. It's from AI startups using your data.

Heather Addyson: Yep.

Srinivasan, Sri: What is rightfully yours? Kate.

Srinivasan, Sri: A few of them have shown how to gain revenue traction, and you'll see many more, but it's a massive hit and miss, closer to like lottery ticket odds in terms of successful startups. But for every incumbent category, there's at least 20 startups hovering over you. At least 20, if not more. One of the thought exercises for the product team and your engineering team should be writing down the map of startups that are hovering around you and what is it that they want from you? What part of your body is it that they want to eat? So, if you kind of use the Pirasanna knowledge analogy, it will be very useful for you to basically map this out very quickly, so that you know the existential threats can be handled.

Srinivasan, Sri: Incumbents, the third category I'm going to talk about, they have all the advantage. Successful incumbents today have understood 2 things. how to use their proprietary data, how to use their distribution, and a third thing they've learned is speed in the market. Those 3 are the only things that matter. Proprietary data. your distribution, that is your access to your customers and your network, and finally, speed. Speed is an ally for startups, not for incumbents. And for each one of you in this room, you've got to figure out how you increase velocity in what you do.

Srinivasan, Sri: So there's some learnings from consumer and enterprise AI. Up until now, most of us, I would say most, unless you're a cave dweller somewhere, you're now changed your Google search one word to like 8 to 10 words, and you're talking to an LLM whether you like it or not. Whether it's Google search and AI board or ChatGPT or Clark, all of us have got used to it. As a consumer, probabilistic works. The risk is like completely yours. It's personal. And you can take hallucinations because you kind of recognize it. And if you don't recognize it, you get kind of miffed. And it uses public data, who cares?

Srinivasan, Sri: Within the enterprise, there are two forms of engagement. One is your internal business processes, incremental productivity. Inaccuracies are okay there. Whether you're using it in engineering, I'm sure every engineer who's used it has seen that AI will make up crappy code. I've had this situation where I actually wanted it to implement a semaphore instead to manage like 2 threads that were going to conflict. It basically put a sleep function. Come on. So we all know that we have to kind of manage these inaccuracies. So it's a judicious mix of context, internal and external, that we have to manage in the middle bucket. The third one, for a mission critical business yourself, as yourself, deterministic outcomes are non-negotiable. You absolutely have to head there. Inaccuracies are unacceptable.

Srinivasan, Sri: And there are tools now available for you to move to deterministic, but it requires a lot of work in terms of harness development, context graphs, feedback loops, and so much more. And your canvas and your instincts need to completely change for you to be successful with enterprise AI. Completely different topic. I can spend hours on that topic with you guys if you're interested. And I actually ran a session earlier that maybe some of you attended. That goes deeper into the topic of deterministic outcomes, maximizing the potential of proprietary data. I'm going to address a different topic in the slides that ends you.

Srinivasan, Sri: So, let's talk about... you know, netting it out, netting out what is going on in terms of enterprise adoption. Firstly, nearly all of Fortune 500 are using AI tools. Everybody is. There is the model companies. The AI Frontier Labs, the infrastructure providers have tripled the amount of investment last year to this year, and that trend is only growing. This is not frothy or a bubble anymore. Their revenues are growing, and the data centers are running at capacity. This is not like the internet revolution where Corning

Heather Addyson: Aha.

Srinivasan, Sri: laid out a bunch of pipe that nobody was using. That is not happening. Um... My slides are dated and tropic is at 30 billion now. About middle of February, they hit 20 billion. So please think about what is going on. Somebody added 50% in like 45 days, not in millions, in billions. Kind of think about that scale. And they're only touching the absolute. tip of the iceberg of the enterprise. That's an enterprise-focused company. Just think about... where this is headed.

Srinivasan, Sri: I think GitHub publishes this data. You can look at it. 4% of all global code commits on GitHub. Game via Cloud Code. Four percent, that's millions, like you've got 100 million engineers worldwide. And... That number is headed to north of 20%. Agentic commits at the end of the year are 20 to 25%, if not higher. Not assistive, I'm talking agentic.

Srinivasan, Sri: Now, the three stages of agentic AI, I think we are all living it. The agent is there, but the autonomous orchestrated agent with obviously the right human in the loop is where we are headed. And we were starting to see it. Engineers, parts of engineering is autonomous. Cloud code co-work was 100% agent built. A human did not put a line of code. I'm surprised how crappy it is. Obviously, it will improve. But basically, there is... Working software today. That is 100% agentic.

Srinivasan, Sri: And that trend is only going to accelerate. I don't mean to be 1 to scare you all. I think this is reality. You have to basically morph it to what you want for your category. I'm not suggesting, please go to agents and shun all humans in your code. That's not what I'm suggesting. With the right guardrails, maximize the potential of these tools. and be ready for 10x outcomes, which these tools are going to provide to you. You're able to code up and create not only demos, but working software at a much faster clip than ever thought of before. But it requires a bunch of work to set context, to set these tools right, and to most importantly, change your culture.

Srinivasan, Sri: So, let's talk about engineering. I think I talked about the 4%, the other thing that basically is happening is... The engineer is now an orchestrator, and I'll kind of talk about it. The headcount implications the world over is much more flat than growing, and the role of an engineer is shifting. So before we go there, Each of us needs to basically say where we are on this curve and the stage of maturity along the way. And by the way, please stop me if you have questions. I'm happy to make it bidirectional. David, if anybody has questions I can't see, please ask. and stop me.

David Pessis: Yeah, just, you guys, I just want to raise your hands, um, you know, go for it and I can uh, MC it.

Srinivasan, Sri: Okay, so I think the key thing here is where are you in the maturity curve? My ask of all of you is be deeply self-reflective and skeptical. I'm sure you all had this professor or school teacher who was like the worst grader. You all remember that moment? Please be that on yourself on this curve. There's no point in kind of playing nice music and saying, hey, I'm transformative and you're like doing autocomplete in your code. Don't work. understand where the market is headed. If you spend time in the Silicon Valley bubble, that's kind of different, but they are a leading indicator to what's happening in the rest of the world.

Srinivasan, Sri: So, one of the big shifts that has that is happening and has happened in large part is engineers are no longer authors, engineers are orchestrators, your agent commanders, agent bosses, and the process is changing. I would... say Agile and you know waterfall and all these other methodologies and methods we had are giving way to agentic factories. Agent tech factories are the way to go. Technology is no longer the impediment. Ideas are the impediment.

David Pessis: Can you say, yeah, can you explain more what an agentic factory is?

Srinivasan, Sri: David, is there a question? Yeah, agentic factory is kind of the representation I've laid out. What these harnesses do is, for example, Cursive built a browser. They created a swarm of agents, agents that were that did that planned the work. Their only focus was. planning and breaking up the work into a plan. Then there were coding agents that basically took pieces of the plan and architected it or wrote down the code. There were design agents that basically looked at viability. There were judge agents that basically did code reviews. There were security agents that basically looked at... prompt injection and other methods and mechanisms. So there were testing agents along the way.

Srinivasan, Sri: So think of it as... different past human archetypes where these are agents that basically never tire, they work 24-7, they don't complain. and all they do is work. They act on your command. So think of it as a factory of agents. You know, one of these, you know, I'm sure you've seen these memes where, you know, robots take care of everything. It's very similar to that, only digital. And you are the one watching through the lens of six different cockpits as to what they're doing. And that's real now. If you sit with OpenClaw, which I do, in a sandbox, I don't suggest using it in your production environment. Not yet. But I use OpenClaw in a sandbox environment, and it is amazing. It is, I'm so happy it's a sandbox environment. It does things that I can't control. And it's basically running into every nook and cranny of my desktop and finding things that I did not know existed. It's because there are purpose-built agents doing different things, working in unison. That's what makes them a factory. David.

David Pessis: Thank you.

Srinivasan, Sri: And the shape and size of the team is changing. It's going, you know, basically your scrum team is basically becoming two to four people. Remember I said technology is no longer the impediment. It is the impediment if you have a lot of tech debt. You got to basically whittle that down. And ideas are a big blocker. You need more product managers to kind of think about ideas, viable ideas. The other thing I would recommend with product managers, I've seen a whole bunch of product managers use Claude and Gemini and ChatGPT to create specs and other things. I'd say there is more AI slot there. Product managers, be skeptical of your work. Engineers, be even more skeptical of that work. The best way to find out the percentage of use of AI in a particular spec is to put it through Quill Bot and you'll figure out whether it's like 100% generated or is there judgment in it.

Srinivasan, Sri: So there is many other techniques I would use. I would, you know, my big question to all of you and the big ask of all of you is, skepticism, healthy skepticism, contrarian thinking will make us better as we all figure out how to work in this new world. The gains in productivity within, oh, there is a question.

David Pessis: Oh yeah, go ahead, Lucas.

Lucas Westervelt: Hey, just wanted to kind of follow up on that idea of the agentic factory. How does that sit with the traditional silos we have a little bit around like software engineering versus DevOps and SREs? Like do you envision those roles also being rolled up together?

Srinivasan, Sri: Yeah. I think it's idiosyncratic, Lucas. The answer can be... purpose-built, I would expect. your CI portion of it to be a factory, your CD portion of it to be another factory. If you were asking me to design it, I would design it that way. But you can design it any which way. If it's a small team with a bunch of microservices, I would kind of design it as one factory. But if you're managing multiple data pipelines, multiple... code bases, it could be purpose-built factories within Lucas. But remember, you are designing the factory for your customer outcomes, not for yourself.

Srinivasan, Sri: OK. The productivity gains in engineering are uneven. So in this chart, I'm going to basically say existing code bases, new code on the x-axis, complexity on the y-axis. Existing code bases with high complexity require basically are the biggest bottleneck. Because tribal knowledge is in your heads. So I don't know how many times you all will say, only this engineer knows this. I don't want to touch this area of code because it's brittle. And the engineer who knew it left, and I don't want to touch it. You know, the same thing. You don't want AI to touch something. Please don't fall for this thing saying, hey, I can read the code and figure it out. It cannot. It cannot figure out permutation combinations with data. You have to give it the context. So the benefits are the lowest in that quadrant of high complexity existing code base.

Srinivasan, Sri: Incumbents have the same benefits that startups do when building new stuff. If you're building new stuff with existing code bases, context matters again. So what I have seen as best practice in many places is there are teams of engineers that basically their sole purpose in life is to build context into these tools. The one recommendation I have is please don't have every engineer figure out context by themselves. That is a loss by itself. Centralize the act of context building. Okay.

David Pessis: So, Tri, I actually have a question because we have such an emphasis on operational excellence. It really means like cleaning up the tech debt. Because like, as you've heard us talk about, like we need to become operational excellent before we become a high performing AI first organization. The low complexity stuff.

Srinivasan, Sri: Yeah. Mm. Yeah.

David Pessis: I guess like, would that be like the bottom right little complex? Like when I think about fixing the legacy code, would that be that bottom right 20 to 30, 20 to 40% effectiveness using AI?

Srinivasan, Sri: I would say both 20 to 40% as well as on the low complexity new code side because when you're rethinking stuff, for example, unit test coverage, I think of that as new code low complexity.

David Pessis: Yeah.

Srinivasan, Sri: Let's assume you're re-carving A monolith and creating a new service. I think of that as somewhere in the middle of those two. Um... And slowly but surely moving it further left than right.

David Pessis: Yeah, okay.

Srinivasan, Sri: So, that's the way I think about it.

David Pessis: Thank you.

Srinivasan, Sri: And by the way, I didn't make up these numbers. This is empirical data from a bunch of surveys that McKinsey did, a bunch of other guys did, but I kind of mined it into this type of a framework. So coding, engineering is the first domino. Okay. There's a lot to learn from coding, software engineering per se. I think the main things to learn are engineers built it for themselves, but every keystroke was used as feedback to make it better. So if you talk to the Frontier Labs, I often feel that I'm talking Klingon with these guys. Those of you who know, you know what I'm talking about. Or else you can Google Klingon and find out what I'm talking. Because they speak a different language. They talk evals. They talk primitives.

Srinivasan, Sri: Eval is critical in basically figuring out whether your agentic journey is working. Whether it's in your product or in engineering, you need to write the, get the right evals. Think of evals as your ultimate test to make sure you're making the right progress. And they're not only tests, they tell you qualitatively what you need to improve. So getting your evals right, and we're running a few sessions on this topic. Join if you're interested. But basically getting your evals right is like 70 to 80% of the problem. And that applies to not only your product journeys, it applies to your product management, use of AI, AI and sales, customer service, all those parts, cloud ops, all those areas.

Srinivasan, Sri: Second part, I would say, is... Data, really, really important. Domain experts matters so much more. Technology is not the impediment. Context is the impediment. Getting, you know, the things that are floating around in our heads and bringing that in so that rising tide raises all boats. You know, this whole thing of, I'm valuable to an organization because I'm an expert. is getting debunked at a very fast pace. Your value is different. It's not expertise, it's judgment. And there's a big difference between the two. High quality judgment. So before I go any further, I'm going to flip a little bit to a different part and talk about the cognitive revolution. Any questions, David? And hopefully I'm doing okay on time.

David Pessis: You're at 30, you got 24 minutes. I don't see any questions right now.

Srinivasan, Sri: Third, OK. Okay, so software engineering is just a start. Okay. You got huge white space in all the categories and you can see your category also somewhere actually it falls in the other category. The other and there's back office automation, RCM falls in that category per se. But basically all other categories, if you look at API counts as published by Anthropic. They're lagging behind. The right harnesses haven't been built yet. That's our opportunity. Not acting has an implication. There is a compounding effect in silence, and I'll be careful in describing that. What I mean by that is... You may think all of your competitors are not making progress. Silently, someone is making progress, and when they come out with outcomes, you're going to be caught flat-footed. So please act, be the one who breaks away from the pack. Um, in your area now.

Srinivasan, Sri: For a long period of time, software has kind of tried to automate. It's only flatter to deceive. It's required configuration, integration management, data quality, all the things that you all deal with on a daily basis. And then there has been services that basically Bain, BCG, McKinsey, a whole bunch of people have provided humans as thinkers charged by the hour. Those 2 twins are coming together. And the 10 trillion economy that's the focus of AI is services software. That is outcomes, high quality outcomes delivered by agents with high margins. That's basically the center of the universe. That is pretty much what's going on.

Srinivasan, Sri: happened with coding, happening with sales. If you go through Gong, for example, they're basically selling you inbound leads. That's what they're selling. By market category. Configure it, voila, magic happens. So you're going to see this happen. Harvey in legal, they're selling you paralegal as a service. So you're going to see services software pretty much happen all over the place. AI is nothing but a service. What's going to happen is these agents are basically purpose-built for an outcome, for a service that was otherwise done by humans in the market.

David Pessis: Right. Hey, Sari, real quickly, there's a lot of chatter about expertise versus judgment and like how to think about that.

Srinivasan, Sri: Yeah. Yeah, a huge amount of chat. I think I would say. Judgment is the exposition of expertise in context. That's the best way to put it. In other words... judgment is being applied to make your factory better with agents, directing well in the right places at the right times, or judgment that you basically allow your nursing staff to provide or your doctors to provide in your flow at the right time at the right place. If a doctor never uses their expertise, that's not judgment. It's exposition of expertise at the right time at the right place is judgment.

Srinivasan, Sri: What can and must you do today? I'd say, you know, deploy AI and engineering, be very skeptical, have the right measures in place. Two measures that really, really have been useful for me, it's lead time from idea to delivery. and regression rate. Those are the only two things that matter. Lead time also has a velocity element to it. If you want to add a sub-element, those two will basically tell you whether your product and engineering is basically doing well. The other thing I would say is invest in your people. as AI orchestrators, it's not like if 100% of your team uses AI and they all use it poorly, you're not getting great outcomes. But if you get 5% of your team using AI effectively, you will create outsized economic impact. So I don't want to leave anybody behind, but focus on the people who really care and use that as example to turn the rest around. There is no room for skepticism. David, I apologize if I'm saying it's loud. There is no room for that.

David Pessis: No, I think we're doing a lot of that. You know, there's a lot of experimenting going on, right? Everybody's using Claude, like the vast majority. And like, we're sharing lessons every week, Sareen. So I think we are, I can't point to like that 5%. I think it's a, we're on that journey right now. And so, you know, it'll be, you know, Maybe we can think about how we do that more efficiently, like identify these power users who are, you know, part of that 5% and how do we multiply them and share that knowledge across the org.

Srinivasan, Sri: Yeah, that's fantastic. And in the productivity piece, the thing I would leave you with is back to the same incremental to transformative productivity. Map out where you are and what business process and how far you're taking what. It's really important to understand. What is the impact of your investment and what is the outcome you're generating? So pretty much for each of these, create an eval, kind of describe the eval and measure the eval on a regular basis to see progress as you kind of bring AI into the fold. And the only way AI works is if you rethink a process. Remember, all business processes were laid out in the 1890s. The ones that are in use, they were laid out in the 80s and 90s by the KPMGs of the world, and they were human-centered workflows.

Srinivasan, Sri: When you have an agentic factory, I'm not sure a human-centered workflow works. It doesn't. So you've got pretty much rethink it. And unless you rethink it, please spend a lot of time in the design side of things. If you don't spend time in the design side of things, you won't get a lot of benefits. So don't jump in with these tools and say, oh, I'm going to apply it. Pixie Dust AI leads to what I call AI whitewashing. Okay. It's A multi-decade opportunity. I think you all know this. The one part I would say is progress will not be linear. Please don't be disappointed by lack of progress in a particular area. It's just the nature of AI. Takes 5, 6 tries to get it right. And suddenly, ohh my god, magic moments happen. And it can be very disappointing along the way. Show up as leaders and persevere through it. Because remember, there's a bunch of competitors in spaces and rooms who are plotting against you at this moment. And make sure their breakaway moment is not earlier than yours.

Srinivasan, Sri: I think as you reimagine your business, I would say think about your value proposition, revisit your product roadmap. Obviously, think about pricing models and such. And think about the opportunities. Don't just think tech debt and your current models and things like that. Think of all the differentiation that's going to come because you have the data, you have the understanding of distribution, and you have the ability to kind of... Innately automate every piece of cognitive labor and more in the organizations and business departments you serve. And please double down on your mods. Distribution, your customer trust. Only happy customers will buy from you. Only happy customers. Regulatory, you guys are in the sweet spot, man, like unbelievable mode. And data is only a mode if it's usable and if it's truly proprietary. It is proprietary if somebody can't rebuild it in years. If they can recreate it in like weeks, uh-uh, ain't proprietary.

Srinivasan, Sri: So, I think I'd say... You want to be a leader who embraces and drives change. I think leaner, meaner, nimble organizations are the need of the hour, and any amount of raising ambition. you're going to find that it's insufficient in the long haul. So I'm going to basically stop at this slide and ask question, answer any questions you guys have in that, you know, three things to keep in mind are. The 10 trillion plus services software market is where the attention and focus is going to be. We live in amazing times. Each of us should be excited to be in this journey where the next decade is an AI transformation decade. Believe me, we're at the start of that AI transformation decade. And let's make sure we leave an imprint in reimagining our businesses, in recreating in winning categories that could have otherwise not been done. PCC is such a prime example of that. You guys created this category and grew it from nearly nothing. And it's unbelievable what kind of a market share you have. But somebody else is going to reimagine if you don't.

Srinivasan, Sri: It's yours to lose. It's yours to not only win, but expand and expound. Remember, all the services dollar around your software will be yours if you become services software thinking. And most important, I would say, this is time for wartime mindset. This is not a time to basically relax and say, oh. We're a billion dollar business, you know, nobody can touch us. This is the business, this is the time to think $5 billion, whatever the LRP means, it is. Think big, think large, be ambitious, because the technology allows you to do things that were otherwise thought impossible. And now's the time to do it. You don't need to depend on, you know, some ambient transcription like in your space. You could build it yourself. And that's perfectly fine. I'm not saying build all the Lego blocks yourself. All I'm saying is the cost of software development, the cost of initiating a new feature or creating a new module. went down precipitously.

Srinivasan, Sri: And you can now take advantage of it. I'm not espousing cost cuts. I'm basically saying maximize the potential of what you have and expand productivity. The biggest measure is revenue per employee. And companies like Cursor are measuring it at about 6 million, 7 million per headcount. There are companies that are now measuring it in north of 10 million. That's where things are headed. PCC has an unbelievable opportunity. Backed by. decades of proprietary data. access to the best nursing facilities and the ambulatory care facilities, nearly all of them. These researchers, physicians, and clinicians, the thing that you need to grow is taking advantage of it with speed. I'm going to pause and take questions.

David Pessis: Once you get while we everybody, let's just use the hand queue and I'll. Maeve, go ahead.

Maeve Ruggieri: Sorry, my mute is fighting with me. Thanks so much, Sri. This is really helpful. I'm wondering if you could give us a little bit of your point of view and insight on the tension between moving, you know, as fast as possible. You called it kind of wartime mindset. as well as kind of the conflict between in a very highly regulated industry with regulation that does not match kind of the state of the world, right? HIPAA is a very old set of rules. but we are duty bound to follow them to protect, you know, patient information, the most sensitive information. So how do you kind of meet in the middle of that tension around HIPAA compliance, part 2 compliance, but and not having, I would say, you know, federal level or oftentimes state level regulations or in guardrails, how do you match, how do you balance that with, you know, the need to build velocity?

Srinivasan, Sri: Yeah. Yeah, Maeve, thank you for asking a fantastic question. Are you a product manager or an engineer?

Mehrshad Setayesh: Yeah.

Maeve Ruggieri: Product Manager.

Srinivasan, Sri: Product manager, okay. I think the... It's a vast topic unto itself. I do not espouse lowering your guard down on regulation any bit. I do not. I think well-documented regulated markets are a massive advantage with AI. because there is no room for doubt. You basically have to code your agents with all of that in mind. So it's happening in legal, it's happening in healthcare, major portions of healthcare. I actually spent time at the GP Morgan Healthcare Conference.

Maeve Ruggieri: Mm.

Srinivasan, Sri: And the amount of work happening here is unbelievable. And the only reason they believe this will be successful is because it's well documented, it's regulated, and you know what determinism looks like. So I think that's an advantage, Maeve, not a disadvantage. But please don't think of the moment you think of that as a as a we cannot do because because we have regulation, you are it's a it's a problem. It's how can we do with it? How do we build the right?

Maeve Ruggieri: Mmh.

Srinivasan, Sri: It's what happened while they were in the band. That is the model you actually want to train or create at some point in time.

Mehrshad Setayesh: Got it. Thank you.

Srinivasan, Sri: Todd.

David Pessis: Todd. Alright, let's go to yogurt.

Yogesh Rao: Hey, quick question, Shaik. In terms of how you think about estimation and time to completing software development and bringing a software product to market, a lot of our pointing and other mechanisms that we use for estimations are still very grounded and agile. What do you see like evolving in terms of like how we measure or how we look at estimations?

David Pessis: Whatever.

Srinivasan, Sri: Yogesh, thank you for asking. The first thing I would say is... David, this is the last time he's going to invite me here. So I would say if you're using Agile for... Estimating now in an agentic world, you're making a huge mistake, big mistake.

Yogesh Rao: Aha. Saah.

Srinivasan, Sri: spend as little time, basically make the estimation more scientific, and ask yourself why do you even need an estimate if you can actually code it in a couple of days. So when I talked to UKG, they've taken a different approach. They've gone back to like Kanban and a couple of other methods and they've ripped apart their Agile book and they said, don't need it anymore. So I'm not saying go do that now. Please figure out your method and mechanism. But in the time you do estimation, often you can actually code it up and be ready. And it's not just coding, obviously you have to do testing and other design elements and all that stuff. Take a different lens and approach to this that is ensanced on using the agent as an estimation method.

David Pessis: So, so this, so what we're saying is if it can be done in a week, you don't need to estimate, so everything should be done in a week. Yeah. Yeah.

Yogesh Rao: Ohh.

Carol Lewis: Yeah.

Srinivasan, Sri: David did not pay me for this.

Yogesh Rao: Budd.

David Pessis: Yeah.

Yogesh Rao: Sri, you said David might not invite you again. It's like we, you know, like what? No, it's me, like I wish I could and ask that question, no.

David Pessis: The.

Srinivasan, Sri: Okay. Okay, Bernardo.

Bernardo Sanchez: I loved your presentation. Thank you for it. I really got two questions. One's security and the other one is more of a personnel or staffing. I'll start with the security. From a security perspective, in your opinion, would it be better to regenerate many third-party libraries with their own code using AI? obviously not the ultra complex ones, but a lot of the smaller libraries so that we're not prone to CVEs. In a second, yeah.

Srinivasan, Sri: Yeah, it's such a fantastic question, Bernardo. It's a... With the amount of code that's being generated, I'm not sure it's a scalable answer to say let's just generate those because. The dependency matrices are expanding even more because most of the generated code is actually using some library or the other. So I think you've got to be careful about it. I don't think it's scalable, but you have to spend a lot more time. So when we say judge agents, I would say.

Bernardo Sanchez: Care.

Srinivasan, Sri: The act of coding goes down precipitously, the review time goes up immensely. And security reviews are going to be massive, and unless you staff for it and train for it. You're just a walking, talking breach waiting to happen.

Bernardo Sanchez: Yeah, that makes sense. Second question is, should we be hiring more physicians and nurses and experts in the fields that we want to expand into? To get. Great, thank you.

Srinivasan, Sri: Care.

Carol Lewis: Oh, hi. So I wanted to piggyback on Marshad's question and maybe double click a moment when were talking about regulatory and how we handle regulation and training your own model. With that, we have a very complex regulatory system when it comes to jurisdiction. We have multiple countries and that It segments there, and then it further segments into cohorts of state or province, et cetera, which gives it a bit more complexity than normal. And then the question becomes, if you're training or fine tuning a base model, the training corpus that was used to train that in the 1st place, can you address those?

Srinivasan, Sri: Aha. So I think the regulatory and the geo stuff that you're asking about. It's much like standing up clouds, Carol. Those implications are going to be there. But those implications are far more amplified because it's also what data you use for training, where the supply chain of data matters a lot more, so my suggestion would be design. up front for the regulatory landscape. and the geopolitical landscape. and make it very much part and parcel of it. If you think for every feature, then you have a bigger problem. One of the pieces of guidance I have is designed for agency. That is your architecture has to be thought of for agency. And that basically means like agentic management and data and models and things of that sort. So I'm trying to give you the most abbreviated answer because it's like a multi-hour conversation by itself.

Carol Lewis: Yeah, I realize that and we don't have multi-hour, so thank you.

Srinivasan, Sri: Yeah. Brinthan, your last question and then we'll hang up. Obviously you all have work to do beyond talking to me.

Brinthan Yoganathan: Right. Yeah, thanks. So, as models are becoming more much more capable, but also very expensive, has any other portfolio companies in your area been able to use one of those reasoning and capable models and convince customers of course to value?

Srinivasan, Sri: Of what to value?

Brinthan Yoganathan: a cost to value because an expensive model that means they had to pay a lot more.

Srinivasan, Sri: Yeah, absolutely, without a doubt. I think cost is where tokenomics is something that's an emerging art form. The better ones are doing it. I'd point you to Game Men for Coders is an example that came out. like couple of days back, take a look at it. It's basically, people have been able to reduce cloud token usage by like 60% by cutting out niceties and focusing on just the work. So it's a big facet. I think cost, like any other software project. you need to manage. So by default, we'll all use the opus models. But for 95% of the work we do, we only need a haiku or a simple open source router model. So we just need to kind of have the right mix in our agentic architecture to kind of manage that. And that has to be thought of from day one.

Brinthan Yoganathan: Thank you.

David Pessis: Sri, thank you so much. Really appreciate it, everybody. Thanks for all the engagement and questions. And if you guys have hard questions that you weren't able to answer, email them to me and we can, you know, we can flip them over to Sri if he has time to address them. But thank you so much, Sri. That was phenomenal.

Srinivasan, Sri: You're welcome. Thank you. I'm counting on you guys to win. Thank you.

Mehrshad Setayesh: Thanks, Tri. Thank you, bye bye.