Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board

[MUSIC] Brett Taylor is the ultimate Silicon Valley veteran. He was one of the greatest Google Maps, in fact, the like button was Co-CEO of Salesforce. He pushed through Elon's acquisition of Twitter. He was on the Twitter board. He's now the chairman of the OpenAI board. And his day job is founder and CEO of Sierra, which is bringing AI to customer service. He's one of the smartest people I know on the topic of how AI is changing established companies. Cheers. Cheers. [MUSIC] So most important question, how do you install the OpenClaw on your work laptop? I have not. Have you played with OpenClaw? I have played with OpenClaw. I haven't like bought them that many. I, you know, you can put these things in virtual sandboxes for less money. It's really interesting. I mean, it's very compelling. It's probably the first-- I wouldn't have predicted the first kind of broad-- I don't know. Consumers exactly accurate, but maybe a hobbyist, use of AI would have been this kind of semi-rogue, open-source project. It was through three name changes in three days. Yes. And I love it. I love everything about the chaos of it, just because all people in our circles have been talking about AI agents for consumer use and all these fancy computer-using agents and instead you're chatting over WhatsApp with the thing on a Mac mini that is mildly unhinged and insecure. It's just fascinating. The whole thing is fascinating. But isn't that-- OK, the thing that seems to me is funny is if you look at the landscape, still in 2026, if you open a new Gemini chat or if you open a new chat to read chat, it's basically a blank slate. There is no memory. And then-- I mean, club people talk about the WhatsApp and telegram integration and things like that. But it feels to me a big part of the value is not only can it do stuff proactively, but it has memory. But the way it has memory is this super janky. It's like the movie memento. It writes things to a markdown file. And it's just writing the things to remember. And the compaction is buggy. You don't always write down the exact right things to remember and stuff like that. But is it funny that you can get super polished mainstream consumer apps that have no memory at all? Or this wildly insecure three-name changes project that's kind of almost remembered things by scribbling notes in the margin. And that is the state of consumer AI. I have probably not very thoughtful because of technical theory on this. So coding agents have gone through transformation over the past four months. Like the difference between-- if we're here in October, versus now our conversation about the future of software engineering would be materially different. And how often can you say that about a technology? And people always-- in my circles, anyway, you look at a coding engine, you extrapolate to other domains. You could all digital tasks be like this. And the answer is obviously yes over some period of time. But it's really interesting because I think sometimes-- I think the hard part of engineering is in the details. And code repos have very specific qualities. One is all the context is in one place-- in files that are largely textual, not binary. And for most broad information tasks, that's not true. You're making-- like when you're writing your annual letter, my guess is the sources of information were in so many different systems, data warehouses. And so it's not impossible for an agent to use those things. But the idea that you can like straight line from coding agents to writing the striped annual letter, I don't totally buy. And then similarly, when the agent's actually performing work on a code base, there's feedback. There's compiler errors. There's often unit tests. There's integration tests. There's the history of every change every made in a really formal format along with code reviews. And so you can actually-- it's almost designed for a robot. And you can self reflect maybe we as engineers or sort of have always modeled ourselves after robots. And now we can actually fully realize that vision. So it's interesting about it is like the idea that it wrote a markdown file for memory I think is maybe more significant than a hack. I actually think to some degree-- Turning your life into code kind of. Yeah, it's like you almost like that at all. Everything in a file system that sort of looks like source code, not because that's the only way these agents can work. But actually, it's quite an efficient way to get a mix of context and random access memory. If you think about like a vector database, it's more random access. You have to know what to look for. But actually, that's not how real memory works. There's a mix of it. So you're loading a markdown file. And as you said, compaction only sings matter. But the messiness of it actually probably produces a more useful agent than a lot of the fancier things. And I use memory in chat to BT and I love it. But I actually think this idea that there's a directory of just everything you've ever done is actually maybe more useful to an AI than people think. And actually, if you follow over the past couple months, just this emergence of harness engineering where you're doling the harness around an agent to do work, I wonder in the short term, it might just be one of those idiosyncrasies in history. Mimicking a code base is actually the best way to make a general purpose agent work. And maybe over time, we'll get fancy than that. But it's actually like a relatively efficient harness for an agent. So anyway, maybe that's why-- maybe that's what it's called. Yeah, and it's very terminal centric. And yeah, it's kind of backwards compatible. You can use GREP. You don't need to make some vector database. And the AI is really know how to use all the unit tools. And so you've got a lot of lift from that. It's exactly right. I mean, software engineers were notorious for making tools for ourselves first. So there we just always bend every other domain in digital towards that domain. But the reason I brought up things like unit tests, integration tests, opening up to this blog post, I can't remember the engineer did the post, but on harness engineering. And one of the more interesting parts of it was documentation. So rather than just having a single agent's markdown file, they had a directory of essentially the entire product, the architecture, and their sort of fillingness out over time. And agent's markdown became sort of pointers to it. My hypothesis, having used codex a lot, I wonder of the output of session where you make a change to Stripes products should be a documentation artifact in addition to code, where the documentation artifact is actually what the product manager, version of John, and the code was the engineer version of John, where there's a lot in the code that is more transient. You might be fine to delete that. What was the intention? What was the PRD? What was the customer problem? It's actually the more durable asset. And I wrote this on X and one of the funniest comments would be like, it would be the greatest irony if software engineering agents made a whole of us just write documentation the whole time, just because notoriously, every good is your hate's-ready documentation. That's our job. But I know it is resonated with me. How much are you AI code? Like you're a very prolific engineer in the old-fashioned hand spun way of writing code. And so how is that situation? We spoke our text code. Pour over. Yeah. Pour over code. That's a really-- I would use that. I am trying to get to world where I'm not writing code. It's hard emotionally. That makes any sense. I have a hard time not caring. I don't care about the assembly language produced by the compiler. But I know why-- Why should you care about the code? Why should I care about the code? I care about correctness. I care about robustness. And I know intellectually, I don't need to look at how the compiler enrolled this loop to verify its elegance and correctness. Yet somehow I feel that way about code. And I'm not saying the code doesn't matter, but I've been trying to force myself to not care because I feel like I won't be a self-actualized software engineer in the future. If I'm too precious about that artifact, which used to be so central to me. Right now, writing Markdown files, maybe that's fine. It feels somewhat like a local maximum. And maybe we'll just be like, oh, of course, it's Markdown is hell. We work with machines. If you think about what a compiler does, there's this interesting mix of formality and informality. And if you've used Python versus Rasts or different into the spectrum. Now that you're not writing the code, I really wonder what that programming system should feel like and look like. And I don't mind chatting with codecs. That's fine. But I also think, as you imagine, all the tests that you care about, all the-- it's showing you demos and mockups. And I wonder what the future integrated development environment for lack of a better term will be in that world. So what I'm trying to do is force myself to not be emotionally attached to the code, which is very hard for me because that was my entire life before I was like proud of the elegance of the code that I wrote. But if I still care about the craftsmanship, what do I want? And I haven't quite visualized it yet. It feels to me like a very interesting time in agentic engineering, because you were talking about this in the Department of Harness Engineering and people having skills and MCP and everything like that. It's always interesting when not only is the leading product in the category changing, we're just figuring out what the categories are. We need things for MCP or skills or stuff like that. And it's all very fast-moving. And that just feels to me like a very interesting time where clearly a new way of engineering is shaking out. And 2026 is clearly not the final word. Absolutely, and in fact, I'm growing worse. skeptical, a MCP is like a meaningful part of the future. Not, it's fine as a protocol, but it's interesting. Going back to your joke around, open-clog is writing a big mark down file. I think it works better than a bunch of MCP servers, but going back to the point of every AI agent knows how to use Crap, but knows how to use all these things. I feel like this view of a multi-agent world was you have all these agents that do tasks, there's fraud detection, another one over here for personalization, and then you make a super agent that does all these things, and it looks really good on the whiteboard, like most elegant looking, but completely nonsensical architectures. Then you realize if you just imagine you anthropomorphize the stripe experience and you're like the checkout concierge, what information do you need to have know a priority to actually make that a humane experience? What ends up happening in the multi-agent systems is you stuff all the context in the subagents, and the one on top has no ability to actually not sound robotic. Then in contrast, you look at something like open-clog, it's just a bunch of mark down files, and the memory feels right, even though it's a little bit clueless. Similarly, if you go back to my arguments about a source control, it has so much context, so it's not like you just have the myopic view of the fire editing, like it really has some expansiveness. My sense is we're making true agents over time, the way we think about context, and how that context is sort of like shared, so that the agent that's orchestrated actually understands what's behind all these APIs and why in the history, well, maybe look a little bit more like open-clog and less like MCP over time, and I think these agents need a lot more context than what MCP affords. >> One thing we've noticed is, there's a bit of a what's the old is new again, phenomenon and the where, with this agent of commerce stuff that's happening. We actually built the APIs for this like 10 years ago, as part of do you remember that move of social shopping that was like, for a while, like buying on Instagram, buying Twitter? Yeah, and they're just kind of, it didn't quite happen for a few different reasons at the time, but the concept is very similar that you want some action of the distance you want to be able to go kind of manipulate stuff off-site. And similarly, I think Patrick has wanted for the longest time in Stripe is, do you believe you just s'sh into your Stripe account? What do you mean? It's like a very ergonomic way for developers to work, is you just be able to like log into your Stripe account, and you have a command line there, and then you can like list out all your, or you can like, tail the payments log, or you can use it. >> He wants to tail and pipe the crap. >> Yeah, exactly. >> And of course now we're building that, because it's like much more relevant in the Agentic world. But I find somehow, yeah, all the Agentic stuff is also bringing back, I don't know if you have this experience as well, it's bringing back a lot of ideas that you might have had before. >> Well, it is because to some degree, the elegance of Unix, which has sort of been the basis of why everyone wants s'sh, and like the curl command that was sort of famously on the Stripe homepage. For people who got it, it was remarkable because you could have all these tools that did something well that was small and useful, and you could chain them together to make something great. I actually, I wonder in the future, we've talked a lot about this, you know. If you look at the canonical software of the service application, like Stripe's console, and obviously you have your, what's the consumer see? But like the configuration that Stripe customer will log into, you would have a web app, and that's like the forms and fields and buttons and graphs. >> Yeah. >> And then you have the API, and it was typically like a REST API, or GraphQL API, and you could do stuff with it. And this is how computers talk to those human's used it. I wonder if the web application of the future will actually be, certainly you want a web app, you know, for the rare human, you know, who wants to sign in, but will you have an agent harness? And what I mean by that is something more than the API's, but just like if you think about the harness that you provide in a code base, the skills, the documentation, the rules, imagine the person who's the greatest Stripe expert. He knows how to extract the most value from the Stripe account. That's the harness. >> Yes. >> Not the API, that's just the button you click. And, you know, will that be an endpoint on Stripe.com, and so that your agent knows how to like just get the most value from Stripe? >> Yes. >> And I imagine you don't care of your merchants, you know, log in. What you want them to do is drive value for themselves, drive GMV, drive payments, and so I just like, I'm really excited about that because, you know, an API is great, APIs are awesome, but a harness is basically like, here's the instruction manual for all the Unix commands that power, you know, Stripe. That's very interesting. >> Yes, and I think if you look at the shape of a lot of APIs that services have, and like, there's an Stripe API cover, it's probably more complete than most, but ultimately it is a way to manipulate some of the highest value business object in the thing, whereas actually what you want is one, all of the data to be browsable in some kind of agent-grate-accessible or textual way, and then all of the actions to be, you know, able to be taken by agents. And it turns out there's a lot of switches in the dashboard. >> Yeah. >> There's no API for, and we are all as an industry collectively discovered. >> And it might be easier, we like imagine being a product manager in the future. You just need to add the switch to the dashboard. You're like, yeah, it looks like a Russian submarine to switch this, but who cares, right? Like, agents can handle it. And, you know, as long as the harness describes when to use that switch, you know, it's easier than your eye design in some ways. And that's fascinating to me. >> But one funny point, DarioMade, it's not clear, well, there's a race between people getting their stuff accessible via agents, and just can desktop computer use getting better. And so it's actually not clear where the approach be. Stripe builds way more APIs, and that's how your agents manipulate the Stripe account. Or you just give your agents access to Chrome and it's a login. >> Well, so actually I'll give a funny story here. So, Sierra, like, company-- >> Sorry, we'll get to Sierra. >> No, no, it's fine, but there's a real funny story here. So, Sierra powers a lot of healthcare companies. So, like, on the healthcare parasite health insurance-- >> Well, no, not the API quality. >> Well, so, first of all, actually, pretty sophisticated engineers of these companies. I really enjoy working with them. So, you're done with Sigma, Blue Cross Blue Shield, on the healthcare parasite insurance. Then you have healthcare providers, like, center health to be worked with. Then you have revenue cycle management. So, like, R1, a revenue cycle management, basically, help providers get paid by the insurance companies. And then you have a lot of other people in the middle, pharmacies, pbms, and they all call each other. So, like, healthcare provider has to call a pair because a procedure happened and they have to get paid. So, we have pairs with AI agents that pick up the phone. >> Oh, sure. >> And we have providers that have AI agents that pick up the phone and make phone calls. We have revenue cycle management companies that work to make outbound calls to do it. We've already had-- >> Do they switch to the agent language? >> They don't. We've done English over the publicly switched telephone numbers. You know, TCP/IP. >> Yeah. >> And English over PSTN. And it goes, I mean, it sort of reinforces, I guess, Darius Point, which is, you can engineer all these fancy protocols, but the rails that are already there already exist. >> Yes. >> And English is spoken by all AI agents. And the publicly switched telephone number has been around for a hundred years. And it all works. Which is fascinating. So, you have all these fancy MCP things. And we're doing, like, English over PSTN. So, on one hand, I think I actually agree the principle that one of the powerful parts about AI, with its ability to do text, do audio, and do, you know, I'll say, I'm not sure how you qualify computer use, but you can call it a form of image recognition and manipulation. Certainly, that's useful because you get to the point where you don't need to, like, fully finish the last mile to get value. The thing I'd say, they're going back to your-- like, talk about all the actions, and just, they're not all in the product. All the APIs don't exist. These visual interfaces were designed for us. So, think of, I mean, your stripe is sort of, I think, famously was one of the few enterprise software companies with good design for a long time. And I like, the stripe dashboards really elegant, right? And most enterprise software, you can't say that about their dashboards. I don't think the ideal agent harness will be that elegant, because it's optimized for something else. It's optimized for the context that you need to perform complex multi-step procedures on behalf of a person. And my guess is it's just very different. And I think seeing the way you write a harness for a coding agent is just so different than the way you do eye design. So, I'm certain that it's great that you can click around a green screen, or whatever, click around a green screen, because it's an oxymoron. But type around green screen or click around the legacy on premises, enterprise software system. I think these harnesses will be really good. And I wonder if, is there a world to yours from now where stripes ability to work with the agent that manages commerce for a direct consumer e-commerce company, that will be one of the ways you're evaluated. And in fact, if you're, for lack of a better word, harness is not compatible with the way their agents work, that's actually like that you're not compatible with them. Yeah. And I'm not sure that's right, but I so I think it's great that these things are backwards compatible. It's great that our agents have spoken over the telephone. phone already in English. >> That's pretty funny. >> But I don't think it's like the long-term future because there's so much value that you can provide. Put it in another way, the agents using a sophisticated application harness can just do a lot more and do a lot more like with higher fidelity. >> Yes. Well, we should get to see our, because you see a lot of real world AI adoption. So maybe start by grinding us. What is the, you know, the business skills scale very quickly. So one of the latest metrics as of that you can share because they keep changing for a month to a month. >> Yeah. >> So, Sarah, we help companies make AI agents for their customer experience. So if you have a big phone line, these AI agents can replace your IVR system and just pick up the phone. If you have a digital chat system, an AI agent can pick it up. You don't need to wait on hold. These agents can not only answer questions, but take action on your behalf. We work with, you know, healthcare companies like SIGNA. We just did a great case study with SOFI. And I'm really proud that we raised their net promoter score by 33 points just because it's just so delightful. >> Wow. >> It's really fun to see, you know, sort of all these different brands across a wide range of industries get so much value from their agent. With leader in the space, like just you talked about the metrics, we reached $100 million in ARR and 7/4, 150 and 8/4 were, I think, around 165 now, one month in February, you know, next quarter. So growing really rapidly and really proud of the momentum that we have. >> That's super cool. And what is the typical adoption? Are people using it for, you know, email chat support because that's the easiest modality. Do they adopt us for everything including phone and stuff like that? >> It's changed a lot over the past years, but I'll say the median customer in the little describes some, like, interesting outliers, which I hope are sort of gums as of the future. So most will start with one channel in a few use cases. So, you know, a lot of healthcare companies, phone remains. The dominant channel says, "Hey, for a few types of phone calls, let's have the AI agent take them and see how it does. Do people like it or are comfortable with it? Does it lower our cost? Does it raise whatever metrics, usually it's customer satisfaction?" And does it work more effectively? So for example, like for a car insurance company, it'll be a first notice of loss. You know, I got an offender vendor, you know, and that'll be the typical way you start. For a lot of more digitally native companies, they'll start with chat and kind of similar. But almost all of our clients will do both. So a serious XM, if you call them on the phone, the reagent harmony, which I love that name for a serious XM, will pick up the phone. And if you go to their homepage and you see the chat, that's also the same agent. So the neat part is, I think it's pretty neat because you have, like, literally, all of your, I'll say, customer experience team or, you know, whatever you might call it, your company, they can spend all their time on one thing. And it actually works over WhatsApp. It can work online. It can work in your website, work in mobile app. It can pick up the phone. That's a pretty big change. A lot of our clients, when we start working with them, they'll have like a digital team, a call center team, and all these different teams. And we've kind of gotten to the point because we've digitized the last remaining analog channel, which is the telephone. Those are all unified. When I said my sort of gums to the future, we have a few, like, really ambitious customers, like Rocket Mortgage, great Detroit company, they own RedFen. They bought a mortgage services company called Mr. Cooper. If you go to RedFen, you can search for a home using an AI agent and go to Rocket.com. You can originate a mortgage with an AI agent and you can service that mortgage. >> It becomes a product you say, rather than just customer service. >> And really end in sales service. And I think that's really exciting. I mean, our whole view is that if we're in 1994, you were doing cheeky pie in about this internet phenomenon. >> I'm going to find out. I was a bit young, but yeah. >> Yeah. >> I was just a bit of a bit of a, I want to have my nirvana sure. >> Exactly. >> You know, I, we would be talking about like, look, this is like your digital front door. Or maybe we wouldn't have the pressure to say that. >> Yeah, yeah. >> But I hope. >> On the information super high way. And I think the same is true of most companies, AI agents, singular. There are lots of agents, but the one with your brand at the top that your customers interact with is special. And that's the one we're trying to power for companies. >> So using this like what's the customer is built on Sierra? Your aspiration is that it just becomes sometimes the primary way people deal with the company. >> I think a company's AI agent will be the vast majority of their digital interactions. Yeah. And I think digital has come to include the telephone. And that's sort of a big shift because we think of that differently. And that's a huge change just because the bigger shift is so, customer service, which is one big part of what we do, but not the only thing we do, is traditionally been thought of as a cost center because it's really expensive. So I'm sure you have people answering the phone for your clients. And depending on where they're located and how well trained they have to be, like how simple the case, it can cost $10, $20. It can be much less of its more simple case. And you know, you have some customers who pay you millions of dollars. And you'll answer the phone any time they call. And you might have one that has not even started monetizing yet. And you might want to call them, but there's a limit to like literally how much you can afford to talk to that person and still have a profitable business. I always joke, it's probably easier for you and me to call soon to our than to get Google customer service on the phone. >> It's very hard to get Google customers on the phone. >> But it's not because they don't like you. It's just if you think about the average revenue per user of Google, they literally, I mean, they just can't afford to do it. So now if you take that $10 or $20 phone call and you make it $10 or $20 and over time, one to $2. All of a sudden, not only can you afford to provide a great customer experience to more people, even less profitable customers or in lower margin businesses, which I think is very exciting. So it's not just doing what you did before, but new. >> I'm probably a better customer service. >> You really can. And then just think about running like a subscription business, where you care as much about customer acquisition, you care a lot about churn because that's how your lifetime value equation works. >> Yes. >> And you think about, okay, if I had a budget of how much I spend on service and now I can do 100 conversations more than I could before, can I actually reduce my churn rate? Can I improve lifetime value? And then the interesting thing is then you realize that, wow, all of my competitors have access to the same technology. >> Yes, yes. >> And they're saying, okay, what are my competitors going to do to actually take my customers away from me? And then that's where you start to get things like, the ATM machine didn't actually reduce bank branches, because some bank had the great idea of, I'm going to put different people in this branch, they'll generate revenue and all of a sudden it wasn't job displacement, but something completely different. So I think the exciting part in our world is, you're taking something that's just so, so, so expensive that people literally hid their phone numbers so people couldn't call them, and you're making it inexpensive and delightful. And the thing I'm excited about is like the second and third order of the effects are going to be really interesting and very hard to predict, and that's pretty exciting. >> I want to come back to this idea of the agent as the UI, because I find it really interesting. Like we talked about this ambition on your letter in the context of agent to commerce. Again, I think people are trying to pitch too much of the end stage of, like, you know, fully autonomous, you know, the robots just choosing few. And the point we always make is like, let's just start with not having to fill out the web form. Like no one likes filling out forms in the internet. >> I don't fear so. >> I wonder just, we'll, we'll using websites have been actually a bit of a, like the fax machine, you know, we use the, from emails over the telephone lines as we have transmission information, or like, I wasn't working for this, but like there was an era of like voicemail memos for you, and the working world for that. >> So we'll still do this, where they, you know, do the voice. >> But like companies will like blast of voicemail memo to like employees at the company, and like that's a way of distributing information. And all these things are like very moment in time, and maybe navigating websites and filling out forms was like a bit of a moment in time. Is that how you see things playing out? >> I don't know. I mean, it's interesting because if you look at the past few iterations of technology, the, you had the PC revolution, then you're the internet and the browser, then the smartphone came out and the tablet. And I was more optimistic about tablets than sort of the way the world turned out. You know, I see more tablets on airplanes, but like, I don't, you know, I'm guessing if I walked around and strived, I would see very few tablets out. And similarly, there's more smartphones than people, but they're still about two billion PCs in the world. And I think it peaked some number of years ago, but it hasn't gone down as far as I know, and I haven't tracked this. That's interesting, right? We sort of added to our digital world, but I think the perhaps the more interesting metric is for like you and me, what percentage of emails were sent through each device? And certainly from 2010 to 2020, most of the world might have transitioned from like percentage of email on desktop to smartphone, you know, significantly. And so it's almost like market share of digital interactions, which I think is a really interesting way to think about it. And certainly as you think of like Stripes business, like where does commerce originate? And you saw that move to mobile, but it doesn't mean that people, it's actually a very big, you just, you wouldn't want to eliminate the PC commerce business, like that would actually be catastrophically bad. >> Yes. >> And so then you look at AI agents and I believe most businesses will be their primary digital interface. And it's because it works over WhatsApp and it works over the phone. If smart speakers make a comeback, they'll work over smart speakers. >> What's their name now? >> Like, big speakers were just too early. >> Yeah, I was like asked for the weather, just turns out to be like not the biggest market in the world. >> Exactly. >> But now- >> Seven-time-er. >> Seven-time-er. >> Yeah. >> I mean, it's amazing to be that much money off a timersetting speaker. [laughter] And so like it is very future proof because it's fundamentally conversational, but maybe it's like going from you know Punch bar punch cars, Mising Keyboards, touch screens now, voice and chat and and probably 3D immersive at some point Does it just sort of add and make the other ones less important? Is probably what I think about it I do wonder if we'll see the end of the smartphone at some point. It doesn't seem anywhere close to right now But it is interesting. I mean, I think most people don't love how much we're sort of addictive to staring at this glowing screen Yes, on the other hand, you can't talk to TikTok, you know, it's fundamentally visual But I wonder if there's a world where you could actually be really productive without such an invasive device on your body. Yes, and if that's the case, can it offer an opportunity to sort of like Unwedge some of the addictive properties of these technologies and get a lot of the benefits from it You know because like at least for me like I think all of us are so connected You sort of end up like I'm a check my email and then like you're like, where if I put it in the past I already know The fact that we actually have technology that affords that kind of innovation now. I think that's quite interesting So I don't know what the future is, but I'm very excited for it. I know that's actually cheesy But I sort of like we've now like changed the ingredients available and we have a lot more recipes we can cook and I think that's very exciting Yeah, I really am excited for not having to look at the screen for as many things for a variety of reasons and when a customer installs Sierra I know it's there's a significant customer satisfaction component as well as cost but I'm curious what kind of Costs difference to the make and maybe relatedly what's When a customer is fully deployed what kind of mix they see between Chris fully resolved agentically things that end up having a human who is presumably somewhat AI assisted But just what does a normal equilibrium look like there? Yeah It turns out most of our clients have pretty different priorities So some are very focused on cost savings and you can automate Very very high percentages of your cases a lot of There's a company called ramp that's a really impressive tech firm there. We had our access here. Oh, that's great Well, they're automating 90% of their cases They're really sophisticated though because they're like, you know They're basically getting in front of cases before they they escalate But I think it's kind of an example of just a really fantastic company. Yes, you know Improving really well and you can see anywhere between you know 70 90% which is really incredible The interesting thing though is the there's counter intuitive effects to it The cases that do make its way to your customer service team get it more complex sort of by definition So what's called average handle time will actually go up? Yes And we heard from one of our clients that actually their their satisfaction of their calls and agents went way up to because it turns out It's way more fulfilling totally to solve a hard problem. How do you try plugging it out? The other is you when we had one retailer whose volume Total volume went up almost as much as they saved from the AI I was paradox it was a form of that so You know if you've used a chatbot from three years ago They were really like like no one like years ago if you said like do like chatbots They're like zero people would say yes It's so funny that there was a Silicon Valley wave of hype around chatbots It was in earlier than that it was like 28 it was like pre-element pre-transformers Yeah, and they were just like multiple choice machines or something It was just the worst products of all time and so Replacing it was something that was like a delight way of people are like I'm gonna talk to this thing a lot more So they ended up keeping their cost and really go down but the volume of customer conversations went out You know two or three X yes and the CEO is incredibly happy about their like I've just we're now actually listening to our clients So it sounds funny. It's a little bit of a choice how much you want to drive cost savings You know with AI versus other metrics most of our clients are interested in the top line metrics and so If given if you could save $10 or save one dollar and improve your net promoter score and competitive positioning By a meaningful amount everyone in the world would choose the latter So that's the interesting going on right now because again going back to about it was 1994 and we're hawking websites on this this podcast I think if we were to go to a major bank and Say if you launch a website you're in a competitive energy gets every other bank With the benefit of hindsight that would have been over promising The correct thing to say was if you don't want a lot of Website and so This technology is broadly available. Yes, and so as a consequence You know you can't just sort of like yeah launch in all parts of AI not just our business You can't just launch AI observed the cost savings pass it on to shareholders Unless you have a monopoly. Yes, yes Most business is just a consumer surplus in exactly so you either going to lower prices But I think that's why it's an overused analogy, but the ATM bank branch thing is really interesting because If every single company in an industry has access to technology I would say it's an imperative not a competitive advantage. Yes And the more interesting I would say board discussion is when everyone adopts the obvious things customer experience customer service Software engineering legal just pick the ones where their solutions off the shelves solutions available now What will the industry look like and my guess is you could ask to have to be to think And my guess is there's some really interesting second order effects and when you have competitive markets You're gonna end up investing lowering prices whatever it may be and that's the thing I don't think it's talked about enough and I actually think that We just it happens with every technology change you projected through the lens of what you're doing today And you don't take an effect. It's like a multiplayer game that we're all in right now And that's fascinating to me and so the change is disruptive But I think it's gonna be like I'm very excited for the next few years as like the world absorbs the technology We start getting to some of the second and maybe the third order effects What is the most impressive AI adoption or kind of AI native behavior you've seen from a client? Well, that's a really good question. I'll probably say Rocket where we have a really great relationship. I think Barun is our CEO Shahn Mahatras or CTO two people who like really are I would say not only just like curious about AI but like very interested in like transforming the like home ownership experience with AI And I don't know like when you like got your first mortgage, but it's like super It's very intimidating and it's not a modern process They literally call it mortgage folders for a reason like you used to be a folder um And for me, it's an example of a company like trying to transform an industry And and the reason I brought it up in the context where our last question is it's not just saying How can we take AI to do this? But it's like if you were to think about the homeowner experience from searching for a home on reds and all the way through servicing it And you had AI available. What would the ideal experience be like and it's not really interesting to see rocket With their acquisition strategy to kind of like integrate that experience And that's why I'm excited. I think there's an opportunity for CEOs and like industries like that To have a bold vision of like what the future could be And and you know going back to my point imperative not competitive advantage. It is a competitive advantage right now So If you imagine like I haven't tracked like the market share of all the big US telcos But you know If you look at T-Mobile Verizon AT&T and you tracked it over the past 10 years You end up with like Surges and market share growth the iPhone came out You ended up with five G and you ended up these things where But it's my impression of the industry is you end up with these sort of like moments the drive market share and that ends up at an equilibrium I think that so it's interesting about it as like the iPhone moment for you know telecommunications companies like soft bank in Japan This is the moment where perhaps if you have a competitive equilibrium you can absorb this technology use it And you'll have this window where you can like actually like shuffle the data Is it taking a technology that shakes the competitive equilibrium? Yes exactly right Definitely not this time Yeah And so you talked about how coding is so Is such a domain that is suitable to AI Because all of the context you're working with Exists in the repo It is in text It's kind of neatly organized to be executed and read by humans and so this kind of good bounce there um The problems that Customer service agents are not uh Of that character and so how do you actually smush Everything Into a format where uriagent cancers Yeah, we uh we spend a lot of time thinking about that to some degree One of our engineers called all of us like we're creating like a domain specific language for specify and customer experience You know like what is the mechanism of specifying it We use this metaphor we call journeys um, which is you know, what is a customer journey end to end and what is the agent need to be successful in that journey What tools doesn't need to access what information doesn't need to access and you can if you think about the capabilities of an agent like skills and the coding agent you'll add different capabilities over time as the customer is talking to you. The key thing that's been a breakthrough, there's probably not surprising to like the technologists listening to this, but has been a huge difference between those like crappy chatbots of four years ago is the reasoning capabilities. You know, I think that, you know, well, we had one client who had acquired three companies and they had three identity systems, three CRM systems, three of everything. And so they had this big IT project where they were gonna unify all those systems. But I was like, why don't you just have the agent like go and all three of them and just think. And they're like, well, what if there's duplicate it? What if the data conflicts? They're like, you know, that's gonna fool you. - That's diffusion. And I was like, well, what is your person, what does the person do? Like, well, they kind of think about it. And I was like, let's just do that. And that's the interesting thing about these AI agencies, they actually, the basic human, basic reasoning, not superhuman ASI. - Yeah, yeah. - Turns out to be a huge breakthrough in customer experience. The other interesting thing is the innate knowledge of the all. You don't want an AI agent to hallucinate, obviously. But Sonos is one of our clients. And do you have a Sonos speaker? You probably have someone. - I have had the idea. If a Sonos speaker breaks, it's never the speaker. It's always Wi-Fi. That's what I've learned. And it's always true of me too, right? There's always some Wi-Fi. - Yeah. - You know, if you wanted to make an AI agent to help you with your Sonos speaker, like you obviously can give it all the manuals for the speakers or the technical spas, you can give it the device telemetry, all the stuff you need. Do you really need to give it the history of Wi-Fi? Well, now it turns out like large language models have encountered every possible Wi-Fi problem. So like, why does a Sonos AI work so effectively? Well, it does a lot about Wi-Fi in addition to all the Sonos things. And if you look for any given AI agent, all of the like, it turns out being trained on all of human knowledge is actually useful as a starting point for a lot of tasks. And I think that's been the big breakthrough. So how do you give it all of its knowledge? Well, first we've built, I think the best platform in the market to do so, where you can really narrow the guardrails for regular conversations, widen them for less regular conversations. But the fact that it starts with like knowledge of obscure Wi-Fi, idea of secrecy, turns out to be the greatest breakthrough of all of them. - Have you had the opposite problem where there's a customer who's problem-domains, mostly don't exist in the public in Shmasht? So like we provide the drill bits used in, you know, deep sea oil drilling. And it turns out there's nothing. - You know, I'm right about that. - Yeah, so out of set. And you know, we work with this like medical device company and it's a deep cut of human knowledge. You know, and you can train it all on that. In fact, we do a lot, one of the things you want to be really careful about if you have a really well-known brand. And we work with, I want to say, a third of our clients over 10 billion revenue over half of over billion revenue. So most of our clients are actually quite well-known. So one of the challenges when you're offering either sales or service or customer service are really well-known brand is it's harder to ground it. It's actually easier when the internet has never heard of you and you want to make a well-grounded agent. It's actually pretty easy because there's no temptation from the LLAMs to go off script. So actually, I would say, ironically, the harder challenge is when it's a very well-known brand, it's like, no, I got this, no, you don't, you got to go look it up. That's actually a harder problem. And so how do you force the LLAMs, like mechanically, how do you force them to not, you know, answer off the top of their heads, but actually look it up. So we use, we call it a constellation of models. So our platform, we call it even studio, you essentially configure the goals and guard rails of a process. And goals and guard rails, not the sequence of steps because you want agency, but you want guard rails around it. And within that, we'll use reasoning, but we use supervisor models to actually inspect that reasoning. And so if you were, you know, an AI agent in Sierra, and you decided to go off script, like, I got this. Like, I don't, you know, well, what ended up happy as a supervisor agent would observe your reasoning, say, you know, I think John should have actually looked up the policy here and sent it back with notes and say, actually, you're not allowed to make that decision. You know, here's the reasons why, you know, go, go redo that decision. It's a really effective technique. The way I think about it, which is a little simplistic, but I think basically right, if you imagine a reasoning system is right 90% of the time, but has some either guard rail, malfunction or hallucination 10% of the time, it's obviously better than that. And then you have a supervisor that's right 90% of the time. If you chain them together, you get 99% effectiveness. And so that methodology of layering, reasoning and intelligence has been really effective. And in general, it's sort of makes sense. You're basically layering compute, you're layering reasoning on top of it. What's neat about it though is we can sort of abstract that complexity from our clients. So, you know, they're sort of expressing the goals and guard rails and we have all these evals and tasks and all these other things. When we can find ways to make it more and more and more robust over time, but it doesn't require you to, you know, prompt engineer writing all caps or whatever, like the hacks that people use to get these things to be conformed. And you started in 22, 23? We launched the company on February 13, two years ago. So I guess a little world like-- 24. Yeah, so we are two year birthday was like a couple weeks ago. There's one other thing as you were saying that is, do you sort of co-evolve chain of toss and RL and some of these things that are now in the models, but did you have to build your own kind of janky version of them before they were in the models? So yes, it also, it's a talk which is the weird part of the building product right now and the company right now because so much of what we write we plan to throw out later. Yes. And it's just a very weird way to build the company. So, Google's chain of thought paper, which preceded '01 and doing reinforcement learning chains of thought, was out roughly when we started the company. It was an earlier paper and effectively provided sort of a substantive basis of why asking a model to explain its reasoning step by step produced more robustness. So we used chain of thought all the time and it was like a methodology we used. And then, you know, OpenAI very innovatively came up with the idea of we could do reinforcement learning on those chains of thought, which is where '01 came from and then most labs are doing that now. So we throw out things like all the time. Do you do it? OK, the model just does this for us now. We work with a lot of financial services for VW Work with one bank that was a large Hong Kong business and they speak Cantonese. And like, OK, well, we need really good Cantonese voice support. And it turns out that that's really hard and there's not an obvious model that does that. So we spend all this time evaluating all these models. What certainty would you ascribe to like every voice model supporting Cantonese well in three years? 100%, 99%. 50%? So we did all this work. In fact, you know, we, I think, have the best Cantonese support on the market. Great for us and it's a huge selling point. And it's a technology that will certainly be commoditized in three years. So a lot of what we think about, I think, is going from essentially technology innovation now. I think a large part of why we work with the largest companies in the world is because our technology works. In three years, the same clients will work with us because we have the best product. And I think, and if you look at the early marketing for like early software's, the service companies, they'll explain why having multiple tenants in the same database is safe. And that was a huge part of their marketing. Nowadays, if you came and you marketed your product that way, if you feel like, what are you talking about? Like, I don't care what database stripe is. Yes. You know, I think we're just at this period where the technology is so immature. It's a very technology-forward conversation just because it's like people are figuring it out. Just like when Netscape's business was like monetized through a web server 100 years ago. And it will evolve from being a technology-forward conversation to a product-forward conversation. So the interesting part about building an applied AI company is you can't have the luxury of waiting for all the models to catch up with your aspirations. You know they will. But you know they will. So you have to have the best technology and have to be comfortable with throwing it out. Yes. And so it's a real momentum and pace of innovation game rather than thinking of this as like precious intellectual property that makes any sense. It absolutely does. But isn't this organizationally hard where, if I'm the head of Cantonese, you know, language as Sierra, my incentive and like not disingenuously. So I'll notice all the corner cases where like the models aren't that good at Cantonese. And obviously we saw this in prior tech waves, right? For the cloud adoption lagards were companies that had their own on-prem stuff. And they had a million reasons. Half real, half fake as the white cloud did not suit their business purposes. But how do you avoid getting stuck in this mode of thinking where like, oh well, their chain of thought doesn't do what we need? Is like the classic thing you hear from someone within the organization. It's a huge shift. I mean, going back to the first thing we were talking about, it's hard for me to not care about like the elegance of the source code, which I think is an impediment to my fully realizing like being a software engineer in this new world. I think teams that start to treat the code that they wrote that as precious that has been of the area. by a general-purpose AI model will fundamentally fall behind. - Public markets deem the software industry 28, 30%, less valuable than they did maybe a three months ago. - Yeah, exactly, very recently. The two sides of the debate are one, the valuations were based on what the businesses will do in 2030 or 2035, like far in the future. And just there's much more uncertainty there. And so this is deserved. And the counter argument is that it's still not the case that the agentic software production is really going to build your workday and need anthropic just installed workday, very famously. Where do you net out on, is this a rational response or not? - I think it's rational, but I think it's a bit over but at the same time. So I think it's rational just in the sense that there's probably been, there hasn't been more uncertainty in this market ever. - Yes, yes. - And so unless you have a strong thesis about an individual company, my guess is like, will these companies be less valuable 10 years from now? And now I think the answer's probably yes. Well, that'd be true for every individual company. I don't think that's true. And so if you're just thinking about portfolio of investments, I think it's sort of an indictment of the sector more than it is an indictment of an individual company. - I don't know of the value of these platforms was who could buy and a weekend ever. Not that we knew what buy-putting was. My point is everyone who's ever built a software as a service application is at a hacker news comment of I could have coded this in a weekend. Like every single one, famously drop back, I'm sure you have as well. Every single product I've ever made. It just happened. It's like a right, in fact, if no one said that I'm your product, like I'm not relevant to you. - Yeah, I'm not relevant to you. And obviously most of those comments were incorrect. But if you think about all the work you've done in compliance or the relationships you have with large financial services institutions or working to unfraud, the things under the surface that aren't the forms and fields in the web browser are actually incredibly valuable. If you think about a large software company, they'll have thousands of quote-a-carrying account executives. There are present sales capacity, which is basically a channel. And distribution turns out to be a very important part of software. There's social proof. There's the old saying no one gets fire for buying IBM, which a few people say right now. The IBM's actually doing really well under Arvand. Do you want to be maybe the first health care insurance coming to adopt something? There's another health care insurance that says I want to be the fifth. I want other people to prove it. There's all these network effects around these businesses and scale and both some sort of Silicon Valley speak around them. I think the big risk is where is value in the software industry and yours from now? One risk is that more people will build than they do now versus buy because the marginal cost of writing software goes down. I think they'll be true for some software, particularly developer platforms and things like that that are already being consumed and purchased by other engineers. Little libraries or-- Is that already were part of the build versus buy? Calculated, it shifts about. Absolutely. The other part of it is systems of record. So I think these systems of record have always been sort of the gravitational center of the relative solar systems. And it roughly breaks down by department. So ERP systems are associated with the finance department, an SAP and Oracle and work day of ERP systems. And you have Adobe in the marketing department. And they had Salesforce in the sales department. And you had ServiceNow in the IT department. And everything sort of rotated around them. And why? Well, first, their database was sort of truly the system of records. So every application that wanted to interact with the data and that had to-- you essentially collect taxes from your ecosystem. And then similarly allowed each of those systems of record company to essentially have revenue expansion opportunities to go to adjacent areas where they're all sold to the same buyer and all of that. The thing that's really interesting is AI agents are actually performing valuable labor. Is the database in the system of record does that continue to be the gravitational center of each of those workflows? So let's take marketing as an example. The database of your customers that you used to drive sending out an email blast on Black Friday has some value. But if you had an AI agent that drove way higher, more leads for your sales team from the marketing blast, you probably-- that's worth more to you than the system of record itself. Similarly, if you imagine-- I'll just take a CRM system. And you think about the AI agent that's carving your territories if you know whenever logs in to actually do it manually. All of those things have a lot of value. And a lot more value than relatively speaking than they did because they're actually performing the action. And so the real question to me is, does it up end this-- I would say something's been true for 30 years, which is all the value is in these systems of record. And the way I think about is agents are to some degree a system of record of a process of generating a lead or auditing your financials or reviewing a contract or whatever it might be. And I don't think we've ever had a piece of software like that. And will those encoded, well optimized processes start to have more value than the databases? I don't know that's the case. For example, if your ERP system is your company's ledger, that'll have a lot of value. But I wonder for all these others. And my theory is, the closer you get to literally the database is the value IE ledger, the more durable it is. The closer you get to be in a system of engagement, the less durable it is. That's a very interesting framing. Yeah. And it kind of gets back to the point you were making about the company that was looking to standardize and not have three different ERPs and stuff like that. And you're like, oh, why? Just try not doing that. And I think maybe the consumer example of this is I think people have probably had the experience of you paste data into an LLM to do something with this. And you know, like the formatting is all messed up and like the tabs and spaces don't come through. And everything like that. So it doesn't matter. It doesn't matter. I don't care. You can just like paste through whatever. And it'll work with this. And so this idea that, as you say, if like the system of record is important because it's your general ledger and it matters to the auditors, that's one thing. But if it was a system of record in this kind of all your data in one place, way-- and because it was easier to build incremental software atop this, maybe that advantage is going away because the agents are fine plucking data from 10 different places. That's roughly my view. But the bulk case was mixed up bare and bold. I need to spend more time in Wall Street. The bulk case though is, I think all these companies sort of have a right to win. They're all big. They still have sales capacity. They have all these advantages. But it's a race. How fast will smaller companies build differentiated, scaled businesses before the incumbents go into this new world? But for a wide variety of well-documented reasons, disrupting their own business model, it's harder. But I think your ask is, is it irrational? I don't think it's irrational. I think there's just more uncertainty now than there's ever been. And I think that's markets are telling you there's a lot of uncertainty. And that's why you see people received from the whole category, basically. I feel like there's also a totally separate thing playing out here, where for a long time, certain companies were criticized for not taking profitability, that's seriously. And at some level, there's just a return to normal valuation levels and like a fully loaded stock based, can't bake in and everything basis. That's kind of independent of the AI thesis, but maybe just some return to a gone a fully loaded gap basis, more normal valuations. Well, essentially, if you look at a traditional software as a service company, the way most people model it is, you have any recurring revenue, which is based in an annuity. And it should throw off that much cash every year. Then you have attrition, which is subtracting from the annuity. And then you have net new ARR, which is adding to the annuity. Your salespeople sell software to add to the ARR. You tip we have a account management team or customer success team to keep churned down. And you grow that annuity and you grow your head count, often just a little bit ahead of that annuity, because you need to grow a new business. And if that annuity is not an annuity-- Yeah. Then that math really changes. It really changes. And so because the whole idea of software as a service is you can just slow down hiring and you become very profitable because the annuity starts starting off. Cash that's been the thesis of every private equity firm who acquires slow growth, softwares and service companies. If you don't assume that that revenue is going to be there two years or three years from now, your discounted cash flow analysis looks pretty different. And I don't think it's actually quite so dire in the time frames of people think. but again, if you're asking for it. like markets are, there's all those great quotes about Wayne and I don't know, you get it. Yeah, I get it. Like there's probably more safer sectors to invest in, but I don't think it's an indictment of individual companies, and that's my point, I actually think, you know, when we first met, I doubt either of us had an extremely positive view of the future of Microsoft. At the time, it felt like a previous generation company. Now you look at Azure, their open-air relationship, all these things, like what an impressive turnaround. So, you know, I think any one of these companies could do it. I think it's just, but it's more of an indictment of the market. Yes, yes. I have a lot more questions. Would you like them to get us? Sure. Brett has been through a few platforms shifts, and one thing he's been pretty consistent about is being mindful of the external forces that are shaping the ecosystem you're in. He talks a lot about building with the broader wave of AI agents in mind. Stripe sessions is our way of helping builders see that wave of clothes. What's changing in the internet economy, what's actually working in production, and what the next era of software looks like when agents are running real commerce workflows. It's not the usual conference fluff, it's insights into what the fastest moving companies are actually after. So, if you want to experience the next chapter of the internet economy firsthand, join us at Stripe Sessions this April. Use the code cheekypind for 50% off a conference pass at sessions.stripe.com. Can you talk about business models? Are you guys usage based, or how are you innovating on the business model front, or are you? We are trying to. We do outcomes based pricing. For customer service context, that means in the AI agent, it resolves the case. No human intervention. There's a pre-negotiated rate for that. We do have to ask a two person. It's free for sales. It would be a sales commission. Wherever possible, there's a way to align our interests with our clients. We choose it. I'm a huge believer in this. I think the analogy of going from impression-based ads to CPC ads is apt. I don't think any ad platform thinks like, "Man, think of all the impressions we're giving away for free." Because when you charge for something closer to a business value, it's actually more valuable. It's a lot more efficient. I think the idea of an agent's outcome is measurable. It's a really compelling way to both for clients, obviously, because it's aligned with their business. It's also quite disruptive because most, I'll say, legacy software companies are not necessarily equipped to do it for a variety of reasons I'm happy to go into, but it's just a very disruptive model. There's a few lens you can have on it. One is that you get more alignment, like usage based just more aligned than other ways of charging. They do say it's more efficient because you're incentivized to drive the right outcomes. People also make the analogies to, it's almost more correct for the labor substitution dynamics that you get. Just because you have real inference costs, you have to do a usage-based model. I mean, do those factor in it all? It would not be possible almost to a fixed price contract because. I would actually argue outcomes-based is pretty different than usage-based. Just think of it this way. If you have an AI agent that is making sales for Stripe to small businesses, and I told you, I will sell one-tenth the number of new Stripe GMV. However you value that, I'll use one-hundredth of the tokens you probably wouldn't care. You care about the value to your top-line-your-business. I would argue there's not a strong correlation between token usage or utilization and value. There may be, but there's not always. There's that infamous, I think it's called folklore, but it's this website where that Apple engineer used to put just all this Apple folklore in the store. Folkloader, yeah. Folkloader, yeah, I love it. It's like, if you're an engineer, it's a fun site to go to. But there's a story about some new Bozo manager asking for lines of code every day. One of the engineers wrote a negative number as a way of saying like, "FU to the man because he refactor the code base," or whatever. I think that is the essence of why tokens are not correlated with value. They may be, but the idea that they definitely are, I don't think, stands for reason. I think usage-based is like charging for storage or something. Outcomes-based is what business outcome is this agent designed to produce and did it produce it effectively. That is really aligning because it creates a whole vertical alignment. As a company reducing your token utilization for the same outcomes is your problem, not your customers. That's a great incentive to just drive more efficiencies over time. It means that to grow your relationship with the client, you actually have to make your product better. Not just theoretically better to stake dinner, better better. How do you have usage based or say, outcome-based when you move beyond customer service for the secure was this Resolve, or not, to product usage where people were shopping and yeah, they didn't like buy a house there, but like they mostly don't buy a house on most web site visits, but it was a successful visit. It's the right question. There's not a great way to do it for every type of agent right now. You can all fall back to usage-based, which is fine. But in that over time, it's like, wouldn't it be interesting? I think AI agents should have memory. I think AI agents should drive relationships, not conversations. It would be really interesting to say, could we make an AI agent that actually drives home ownership over time? I think that's actually, it's hard, but it's not. I think so. Even because it's hard today and we're a pragmatic company, I think it's sort of the right thing to ask, though, because that's fundamentally the value of the software's designed to produce. I think actually it's a really sort of values-aligning thing. It also, though, changes the dynamics of a software company's relationship to its partners, to its clients, because they've you go back ancient history four years ago. There was a really stark separation between software and implementation and usage. It was the client's accountability to use the product well. It was either your IT team or systems integrators responsibility to implement the software and the job of the software can be just to make it and throw it over the wall. Obviously, it's not exactly that, but that was kind of the work we were in. Everyone had good intentions, but what's the same success as a thousand fathers failures, Norfolk? When the software didn't go well, everyone was blaming everyone else. The client was like, I'm using it just fine. It was implemented poorly. The person to implement it was like, no, the platform's broken. The platform people would say, and everyone's pointed at everyone else. What's nice about outcome space, whether or not the client sets it up, you become more accountable to help them be successful because until they do, they can use it. If there is some long sort of last mile of implementation, it creates a strong incentive for the software company to have skin in the game to just help you navigate that last mile. I think so much, so many of the problems in the software industry are due to that lack of accountability. If you talk to any companies ever implemented an ERP system, it's like a multi-year process. It's in Bating, Russia. Yeah. You don't even remember why you're doing it. Midway through. You've gone through two CFOs and three CIOs, but the time is done. We're like, okay with that. That's just the way software works. My view is just like, I think, AdWords sort of change the advertising industry on the internet. Just drove it. I think you're going to be paid for mobile app install now directly and truly pay for outcomes. I think it's a really positive step forward. It's not really possible for everything. You have to have pragmatism. But I think it's the right way to actually have a partnership. You just share in the outcomes. You want to wire the companies to be thinking in this outcome-based way. Like in your main customer service stuff, you can do that in other ways. You might not be able to. Yes, but you want people to be spring loaded to be thinking about. That's right. If the whole company is incentivized towards outcomes, it's a way like a way better partner to work with because of it. We find this a striper. Again, we have outcome-based pricing. Exactly. It's transactional. But we find there's a lot of uplift we can guess on just getting people more revenue and finding ways to, you know, we're sometimes hammering customers where it's like, you should be accepting local payment methods for internationalization or you're crazy not to be turning on this feature. But we really feel it. We have the same incentive as a customer. This will be revenue maximizing for both of us. I'm going to ask a very AGI-brained question. I just can't tell you. We'll go over a second. Exactly. We get to this. You described building stuff that you know you're going to throw away because the model capabilities will get there and you're occasionally developing capabilities that you develop yourself. Isn't Sierra itself short AGI? Sorry, I said it could be this. No, it's the right question. You know, short answer is I don't know. I mean, the fog of war in software industries is pretty thick right now. I really believe in the apply day I market, though. I think most companies don't want to buy models or by software. They want to buy solutions to their problem. And if you just go back to the cloud industry, Why doesn't Amazon and Microsoft do everything for everyone? There's not really like a, sort of by somewhat similar logic, like why should any software as a service company exist when you have bigger scale, all this technology and theory that can just develop all the software. And actually, many of them have tried. There's actually competitors to Salesforce and almost all the above. I think there's so much nuance in how these companies align themselves with different departments that these companies solve. They're very unique problems in very specific ways. There is a mix of product, not technology with product, go to market. It's an ecosystem around it. And I think a lot of that still exists because I'm not sure like coding the software was necessarily the hard part. And then similarly, I actually think especially enterprise software, how you engage with your clients really matters. And I think it turns out that GPT-5 and Clawed whatever version it's on right now, or Opus, excuse me, is sold to a different buyer than like the CFO or the chief customer officer or the chief digital officer. And that seems small, but it's actually big. And so I think you tend to see software companies orient around individual buyers within companies. You tend to see consolidation around departments and around buyers. It's possible that you can go beyond those lines, but it hasn't happened traditionally. And I think the reason for it is most business users want actual solutions to their problems and they want a company that serves their unique problems in a very specific and bespoke way. So I actually am extremely bullish on a platy eye. I actually think we could accelerate, I'll make one statement which is, I think if we paused model development, we'd still have trillions of dollars of economic value. I totally agree. Then if you have to be realized. And I think if we had a mature platy eye market where the CFO could go by that agent to onboard new supply chain vendors that just worked, we could actually accelerate that trillions of dollars of economic value. So I think not only am I somewhat skeptical that there will only be two companies in the world, I actually think one of the main things in competing adoption of AI is the lack of existence of all those other companies. And so many of the startups, particularly on here in San Francisco, are basically doing relatively wrote kind of tools around the AI rather than actually building agents for business processes that are boring but important and valuable. So I'm really bullish on it. And I guess you held companies in sure that they can always have access to the latest models which sounds like a minor thing but like the leading model is always changing. And so that's how the trivial. I agree and I don't know like I'm not sure how much the long-term value is. I think it is. You know, I think your customer. Up to this point, the like the race is led by a matter of months. Well, every single month there's a new frontier model and your customer experience doesn't change that frequently. So you're absolutely right. But I also think there's just a big product right now. Like our clients use it to optimize their sales. And that is a product, not a technology. And it's very particular to the workflows of people building customer experience teams, building sales teams, and that's really what we're focused on. And I think those departments deserve purpose-built software. And I think there will be enduring value there. But it's interesting. It's the right question to ask. I don't think we've ever lived in a world where production of software was easy. And you know, software engineering was the most scarce access asset in a company and now it's the most plentiful. And I don't think we've ever lived in that world. Yes. Yes. Well, that kind of gets to. One of the biggest conundrums in Silicon Valley right now is what will the shape of AI productivity be? And I think that's a strong sense that the AI has gotten really good and it should change the composition of companies and it should change the hiring plans somehow. And you've seen this in some corners, you know, block announced their 45% 50% AI layoff yesterday. And you have some companies not growing as quickly. At the same time, in coding, you see a lot of AI benefits. You can kind of argue that either way, right? You can say, "Engineers have gotten much more productive. Therefore we should hire fewer engineers." Or you could say, "Engineers have gotten much more productive." The ROI on a single engineer is way higher. Like we now have super engineers that we can hire. Therefore we should hire way more of them. And because there isn't like a fixed amount of stuff for Stripe or any other company to do. And then the AI productivity story in other roles is just a bit less clear because as we've discussed, AI is going to uniquely well-suited to coding. And so what do you make of just how does the AI productivity show up? I feel like every company in Silicon Valley is trying to figure this out right now. Well, first, I think I'll go back to my why I believe in applied AI. I think the atomic unit of productivity in AI is a process, not a person. I don't think AI, I don't know if you have an assistant. But if you do, he or she might help you prepare for a podcast, might help you prepare for a meeting, he or she might also get you a cup of coffee. AI will be really good at the first two, but quite poor at the last one. So no matter of age, I sort of robotics, we'll get you a cup of coffee. So I think it's wrong to think about AI as like sort of replacing people in addition to being inhumane. It's just sort of nonsensical because AI operates in the world of digital technologies. And I think if you go to like an example of even a mundane process in your business, like onboarding a new supplier, think about all the departments and people involved in that. There's a legal department to do a contract. There's some finance department procurement to negotiate the relationship. You probably have IT that's involved to sort of onboard them into your core systems. And then there's usually a business that's sort of sponsored right now. Fairly mundane happens all the time. You let's just say you tracked what is the median amount of time it takes to onboard a new supplier and it was 17 days, just for argument to say. I bet you could say as a CEO of a company, I want to use AI to optimize that process and make it 17 hours or one day. And you could go through and if you had a product manager on that and optimize every part of it, I bet you could achieve that. But the hard part isn't like a person's job. It's actually all the systems and people in between it. And so I think part of the reason why I think it's been slow to get the person's management is we sort of ship our org charts as companies naturally. That's the natural state. There's not usually a person responsible for that process. There's the legal team responsible for the contract. There's a procurement team. So I think actually we will end up reimagining our companies with the benefit of AI. We actually think of our companies as a collection of processes, have people responsible with KPI's who can apply AI. And I think I bring it up just because that's my theory of the world. I might be wrong, I might be right. But I'm not sure our companies are set up to, essentially, absorb the benefits of AI officially right now. And we need to do that to really do so. But the bigger point, I think, is that there's the paradox of will you want more software engineers. But on top of that, most of the world isn't just digital technology. And so I think a lot of the people in sort of the AGI community have only ever worked to like a research lab or software company. You look around like, wow, AI is going to do all of this. And as they walk by the flower shop and get their coffee at the coffee shop and you think about like your local flower shop, like if you took all the AI in the world and gave it to that, you gave it super intelligence. Like how much would impact the flower shop's operations? Like maybe a little. I mean, I'm sure it would help. Don't get me wrong. But someone's still clipping the ends of the, you know, stems of the flowers, arranging the bouquets and, you know, thanking you on your way out the door and congratulating you for your daughter's wedding or whatever it is. And so I think if you think about, you know, what parts of the economy can absorb intelligence really efficiently, it's certainly software and we're seeing that already it's finance seems particularly meaningful here because so much of finance today is just digital information. You know, we sort of everything's in digital systems now. Not even just crypto, I mean, just everything's in digital ledgers everywhere. It still doesn't touch a wet lab. It still, you know, can't do a clinical trial. You know, you still need to get, you know, a crate from this country to that country, you know, on a ship. So as a consequence, I think I'm not sure we'll see the productivity enhancement we see in software and every sector as quickly. And then on top of that, I think companies need to stop just giving like co-pilot every employee and be like, we're AI now and start to think about from first principles. What are the parts of your business that have a lot of digital workflows? Where can AI have a real big impact? And how do you actually set up your company to actually have someone accountable to drive that? And that feels like a real big change management opportunity that most companies haven't done. We just push on that. So software engineering, I think we're clearly are seeing a lot of AI productivity gains and software engineers have all those loves tools and the latest tools that are just kind of headlong diving into it. Then you have stuff like you see in the Photoshop where just and stuff that requires really good robotics that we're far away from. That will take a while. What I'm talking about is like, there's like a-- - By the way, I might prefer a flower shop with a florist. - I just talked to you, yeah, yeah. - Just to say, I'm not sure if it solves a problem I have with my flower shop. - Absolutely. - I might be wrong, I might be unique in that. But I think a lot of the economy is actually why color, knowledge work, not coding. Think of finance departments, legal departments, things like that. - Where you should be able to see a lot of AI uplift and a lot of AI productivity improvements. And it just feels like a current course in speeds were not on track to get those productivity-- - Well, I'm not sure I'm right, but I would argue thinking about it by department rather than by processes where it's off. - We can talk about the process. - Well, but hear me out there on this because if you said, I want to make the legal department more productive. So I want to make it easier to do red lines and you optimize that. But why is the contract there? What is it for? You might, if you're, for example, on board in a supply chain vendor and you have hundreds of them, you might actually say, actually making a abstract technology for your legal department to red line contracts more efficient is actually a harder, more general problem than for your supply chain vendors because you might actually have very rigid rules around your supply chain. Let's say you're a CPG company. And you might actually have very specific, like look, if you want to work with us, here's our legal terms, here's the axes of independence. And if you want to make an AI agent to automate that contract, that's actually a much more narrow problem domain that doesn't require general purpose, red lineing technology. In fact, if you sort of reduce it, you could say, well, there are like 10% of our suppliers where we let them negotiate their contract, but only for this spend. Let's have them go through our legal department, the rest, let's do it all with AI. And my point on it is, if you look at it through the lens of like an end-to-end business process, you can turn science into engineering. And I think solving legal AI, that's a science problem. And this is my point though, which is I think, people are going through department by department. Similarly, there's not like a person accountable for that end-to-end process. In the more you can narrow the domain that you're solving with AI, the more you can build a harness or a scaffolding with existing technology to actually fully automate it. And my hypothesis is most companies just aren't set up that way. That's just not how we're organized. And as a consequence, we're all optimizing our satellite. We're all just installing co-pilot and co-pilot's great, by the way. Didn't mean to insult it. But it's not actually like, yeah. >> Yeah, and to be clear, that's kind of a thing we're doing where end-to-end companies, good companies did this before they are, a continuous process improvement. And I feel like that is the best thing to do. And I think what you're saying is like, this message thing is an AI lawyer instead, but it's like improving your commercial contracting. Like that is a thing that you can tend to. >> And even more narrowly, like pick one domain of commercial contracting and solve that. And I actually think those are truly solvable. And I think the companies that really think about their business that way, I think they can see the value. And again, I'll go back to the immaturity of the apply day. I market is probably one of the bigger barriers right now. And my hope is that as the apply day, I market maturity of the next years, we'll see kind of the step change in productivity. >> Yeah, there is a canonical way to build a Silicon Valley company. You have engineering and product and design. You have this number of ratios of engineers. Product managers and engineering managers. And then you go to market organization, and you have these pipeline coverage ratios. And you have the product marketers and all this kind of stuff. But I find it interesting how similar, so many Silicon Valley companies aren't to each other because they've all learned from each other, right? >> Yeah. >> There's like a shared recipe and a shared playbook as to how to build a company. And we get tweaked, but ultimately, it's pretty good IP, like certainly companies are much better off withers than without. How is that canonical template for building a company different post AI than before? >> Yeah, that's a really interesting question. One is, I've always believed in the primacy of tech leads over engineering managers. I, both Google and Facebook, where I spent some of my early career, both of this well, where, you know, in like a product review, you weren't just talking to a manager. You were talking to the tech lead and PM who are a product manager who are building the product. Whereas if you went to companies that produced, we're software, I'd notice you sort of move up the chain of the command, like the military. >> Yes. >> I think that we will end up with individual tech leads who, because of the existence of AI agents, will become even more important, where if you are a, I'll say, a product engineer, trying to find the right word for it, we might invent one who has taste, but didn't necessarily know CSS, who has infrastructure ability, mean that you understand the basics of distributed systems and debugging, and you understand your customer very deeply. With the presence of codecs, you can produce amazing results. Those people are truly worth a thousand X, other people, because it's relatively easy to find someone who's a great infrastructure engineer, not easy, but like relatively, finding someone with good taste, that's relatively easy. Find someone who also understands your customers extremely well, like the nuances of the problem they're solving. Those people who can like combine that, will I think, and being able to actually produce products, like capital P, valuable products, with relative autonomy? And I wonder if it will change our view on generalists broadly. I've always sort of identified myself as a generalist, just because I've been both a software engineer and a suit basically, and I've kind of back and forth in that world. And as companies grow, you tend towards more specialization, just because the person who's sort of the jack or Jill of all trades ends up sort of not fitting in. You know, like there's not only a place for them, because, okay, well, you're not really the deepest engineer, you're not really the best designer, you're not really a product manager. If you've been at the company for a while, we'll give you an honorary something to do. And you have to lead through influence and down on that. Could that person actually endure as one of the most valuable people on these companies? And I think, I don't know whether it's naive optimism or true, but I actually think those people who often exist in early stage startups are often the people who get sideline, but actually in a way that actually harms the company. And I'm hopeful that in a world of AI agents, those generalists who, again, I think the most important part is understanding the customer need, with agency, you know, pun intended, and empowerment can end up more powerful in the Silicon Valley company. - I'm the exact same thing. It's the exact same thing, which is high agency, really caring about customers, just really caring generally, high work ethic people, who maybe weren't the best engineers, you know, previously or now, those people are massively ascended as far as I can tell because they suddenly got the exoskeleton. - Yeah. - And like they always have the ideas as to what we should be doing, and this is the better way to serve the customers and everything like that. But now they have the way to make all their schemes real. I really notice that it's right. - Well, it's interesting you talked about work ethic. It's addictive right now, because you can do so much with the technology. Everyone I know who's really used it works harder, because it's like, wow. - Yes. - I can do so much. You know, like you think like you're about to go to bed, like, should I get an AI drill to do something? Like if I waste it, like the next, you know, eight hours of my life. - That's awesome. - And you know, that might be a novelty that wears off, but I think it's really exciting. And so I'm hopeful on the product engineering design side, you end up with these hyper high agency people who really deeply care. I really like the way you said it actually. It's right, it's not just customer problems. It's like care period. Just care, I can end up more empowered. And I'm curious what that means for organizational structures. You know, it's, we have a new job role we need to invent. It's like what role do these people in? I mean, hyper generalists. - Yeah, like kind of product managers, but like sometimes maybe without a product, like Minister or without a portfolio, they're just doing stuff, but now they can do much more. - Yeah, and it's almost like product design or product management engineer. That's why I said product engineer, but that means something different. But it's interesting because we've talked about this, you end up where the grass is always greater with the organ structures. You go functional organization, okay, we're going to have engineering product design. Let's go to business units. They're like, well, that led to silos and more in fact. - We're going to grow the way you say that. - You sway back again, you know, that's welcome to. - Just one more rehearsal. - Yeah, exactly. I'm a little bit injured now. And I think that it is interesting if these people become extremely important, what does it mean to organize around them? And I think it does feel like like something that will end up flatter just because of the amount of impact and individual can have. And so that feels really exciting to me, but I don't know, it sort of feels like a blurry picture right now. Inhance. I agree. It's very, very. It's so interesting. And you were on the Twitter board during the super interesting takeover battle with Elon Musk. How do your reflections on that experience a few years later? It was really interesting to sort of be in the public spotlight. I hadn't really experienced that in my career before I joke like no one really cares about enterprise software. It worked for Salesforce for six and a half years. I was a joke. And I worked for Salesforce for six and a half years and I don't think my mom does what Salesforce does. Yeah. And so to have something that was not really just like a business issue or a technology issue, but like a sort of in the mainstream, I realized I didn't love that very much. You know, like I don't offer enterprise software. Exactly. I'm like a builder. I like to build things and have people use them. I know it's not sort of funny and reductive. That's what gives me joy. So the one thing I realized is the conflict of it all, however it turned out, like victory defeat, whatever it was. It just didn't, it wasn't something that like filled my bucket very much. What do you make of the fact that in all these kind of head counterbates, Elon is now running Twitter with 80, 85% fewer people? I think Nikita bearded recently this. All of the end product design as Twitter is 50 people and you know, maybe it's really impressive. Yeah, it's been a little shaky in pockets or just at times, but mostly the service works and they have shipped new features. And I think those two statements are undeniable, but just what's your takeaway from that? I don't know. I haven't followed as much as sort of the like, I didn't see that tweet as an example. You call it tweet still. Yeah, sorry. I'm a little fashion. Yeah. So I don't know about that, but I mean, it is interesting right now because obviously a lot of that predated AI. But I mean, any person who's been an individual contributor engineers knows that the size of the team does not produce like linearly greater outcomes. Everyone in the world has experienced that. So, you know, I think the, you know, the idea of can you actually give individuals who could taste more agency, no pun intended. I think it's always been sort of an enduring thing. What was Jeff Bezos, a two-piece box, sort of thing? But then do large tech companies underrate this phenomenon? Like do they pay lip service to small and power teams and two pizza teams, bus extras that maybe they're distributing more? I think their companies largely act somewhat rationally. I can't remember who the CEO was, but it might have been the Ripley and CEO just talking about, you know, there's this idea of being like lean and agile and then there's like you want to capture market share and grow your product and grow your platform. And you know, at the end of the day, you could be clever but not smart. And you know, you might be so clever to think like I'm not going to have anything more than two people in these features. And if you have a competitor who maybe does something a little less elegantly but wins, like who cares that you are clever with your, you know, two-piece box team or two-piece person team or, you know, one AI agent team or whatever it is, when someone said we're going to have like an X billion dollar company with one person, I think they might have been right. But it's not how you could have had a ten-billion dollar company if you'd hired a bit more. That's right. And I would actually argue the more specific is if all of a sudden for some, you know, clever reason you want to prove you can, the idea that like a competitor might have ten people and beat you is probably more likely than you would have in a ten billion dollar company. So I think at the end of the day, you know, when you're building a business, especially one that's in hyper-graph, which, you know, successful businesses and tech tend to be, if you are too clever and austere and going back to your point about Silicon Valley cultures all being the same, there are examples of companies that really innovated in culture, you know, you wouldn't think of this way, but like HP sort of like a lot of the kind of traditional open office floor plan came from them, Facebook. Oh, I didn't know that. That's interesting. And then, you know, Google offered free food to their employees, which a lot of people did, and then, you know, Facebook, you know, a lot of the way the layouts of offices all looked like Facebook for a long time. But then you have other companies like working to innovate in HR. And they spend all this time and energy on it. And in fact, the smart thing to do is just be like, we're not, it's not what we do. Just do the same old things that we're not else because everything is just push button. I don't need to worry about it. And so I think I do think it's the right question asked for every technology company. Yeah. After being on the Twitter board during the Elan Takeover, you were then on the opening iPods when Sam got fired. Have you considered thus you are the problem? You are bringing the drama. I came in after the draw over there. Oh, right. You joined the AF. Oh, sorry. I brought in as the mediator. I see. Yeah. Post the, okay. Yeah. Okay. Your hands are shaking. Well, I mean, I'm a little bit upset. Yeah. I was, I wasn't actually on the other side of it, but I was, I got a phone call with it Saturday or Friday after. Yeah. And, you know, basically, my understanding was I was the person that both the existing board and Sam agreed upon to kind of help mediate the situation. How did you learn from the Open AI board? A lot. I mean, certainly the most interesting part is the AI research. You know, I've never been affiliated with a true research lab before and that's fascinating to me. I, it is very inspiring. I mean, it is, it's very easy to grow not cynical, but like, you know, you can look at, you know, Open AI Google Anthropic and say like who's, you know, whose model scores better on this leaderboard to actually go in and see this company. Where every single researcher trying to make safe AGI and not come out of those board means inspired is impossible. Like, it's amazing. The other thing is it's the first not for profit board I've been affiliated with. And that's really interesting as well, just because there's been things. Well, any myths in the fiduciary duty is you have a duty to submission. Yeah. And that is really clarifying and interesting as well because when you're making decisions and you realize, you know, you have your sole duty is to ensure that artificial general intelligence benefits humanity. That's really different. I've never had a fiduciary duty to a mission before. So that's really interesting to me because I take those duties really seriously and like, reflecting in a board meeting and you're making a decision, you think about it very differently through that context. I mean, the other thing was because I was brought in, you know, after that crisis, there was three people on the board on the other side of that when I agreed to temporarily be the chairman and it's still there. Funny that we had to grow the board essentially from scratch. And so that was really interesting too, you know, just to think about, normally you're at one board member at its time. Yeah, yeah. This one was like, do you have a book right for you? You know, like, we're going to build a board ahead of the other team. So you really think about, you know, spend time with the other two board members just really thinking about like, what is the composition for an open AI board look like, you know, how do you represent the not for profit part of it? How do you represent safety? How do you represent, you know, the economic impact of AI? Oh, we're doing lots of infrastructure investments. Like, how do we find someone with like that specific type of financial expertise? And so I was really rewarding as well. I'm just sort of building a board, not from scratch, but you know, effectively from scratch. Last question. How do your AI predictions for 2020? I think we will have some scientific breakthroughs with AI that positively breakthrough into the mainstream press and awareness. We've already had some interesting math proofs, but I joked with one of my friends that I can tell I can understand what the title is. I'm not sure it's going to make the kind of. I'm excited about the interventional manifold space. And it won't quite be like the Apollo landing, but I remember the Casper of Chess Match. And certainly things like AlphaGo were really meaningful. I, given the progress in math, I'm hopeful we have at least one moment of discovery that is inspiring. Because I think a lot of the dialogue around AI right now is economic opportunities, but also what could go wrong. And I actually think one of the main things that came go right is actually discovery and science that actually can improve the human condition. So I'm really excited for it because I think we'll contextualize why so many of us are excited about this technology in a way that sort of captures attention. So as you said, something beyond and dimensional, manifold, blah blah blah. And I feel not confident in that, but it certainly feels like the ingredients are there for that. I think we'll continue to see mainstream adoption of AI by both consumers and companies that doesn't really feel like a prediction, but I think this will be really a year of adoption of agents. We're certainly seeing that in Syres customer base, but I think we're going to see it more writ large. Then you already see in chat GPT growth really unprecedented levels. Things like OpenClaw, you can see that translate over to agents and the more long-running autonomous tasks. It does feel like by the time we exit this year, can that go from a niche community to something more mainstream? It feels probable to me. One other thing is, I think most companies in Silicon Valley won't write code by hand. That might seem almost. It sort of feels obvious right now. Yeah. Of course. You're just nodding. Yeah, why not? If I had said that four months ago, that would have been a closing prediction. I think that's really interesting just because that's such a fundamental state change. Can I say in Silicon Valley because I do think it takes a while for these tools to sort of diffuse their society? Silicon Valley is insular enough that I think it will here, but I'm not sure it will happen through every company in the world yet. So the Euro-Agence across businesses and just people finally getting their kind of clause to agents and then all code written by you as well. Yeah. It's good to have operations. Right. Thank you. Thanks for having me.