The business case for AI: Brent Hayward of Salesforce, David Carmona of Microsoft & Nagraj Kashyap of Touring Capital

(upbeat music) - What is fundamentally different about this wave is who started it. It was Microsoft as a big incumbent, teaming with OpenAI, which is a startup. So they're innovating at the speed of a startup with the resources of Microsoft. - Welcome to Orbit, the HG Polcar series, where we speak to leaders from across the tech ecosystem to hear how they've built and scaled some of the most successful software companies in the world. - Good morning. It's my pleasure to introduce you to Brent, David and Nagraj, who have a lot to talk about, based on what we heard. A number of you at the pleasure to meet with David last year. David is CTO and VP of strategic incubation for Microsoft, or is in charge of topics such as AI, quantum computing and space. A lot of what he talked about last year actually happened, so I suggest you listen carefully to what he has to say today. Brent is a senior exec at Salesforce, and most recently, CEO of Millsoft, probably the most successful acquisition in Salesforce history, which is $2 billion, run rates, and a 40% revenue cager in the past five years. A number of you in the room actually use Salesforce. I know a vast majority actually do, as part of the core infrastructure to run the customer information and customer business. And so we're very keen to hear your thoughts. And then finally, Nagraj is the general partner at Touring Capital. He's a 20-year experience in venture capital. He was the head of Microsoft Venture Business M12. Previous, these were for quantum ventures on a hardware side. And then he was also a managing partner at the soft bank vision fund, the largest venture capital fund in the world. I suspect a number of Nagraj's previous investment will be part of your daily life, depending on your level of activity, zoom, ways, and for some of us, or some of you, I shall say, fit bit. So welcome. I think we're going to talk about AI today, AI tomorrow. We're also going to talk about the scaffolding of fair beds and how we see those layers moving. Maybe to ground the conversation first. What does the current moment in AI mean for you personally and your company? So I'll start with you, David. I love the comparison by the way of before from the presenters on kids versus adults. I loved it. And I think it's something that we're learning how to manage. So under that, I immediately consider myself a teenager, because my role in Microsoft is really to work closely with those incubations and turn them into businesses. That's my role. And I was starting to die with the cloud. Then I moved to AI eight years ago, where most of my work was with the kids, right, with Microsoft Research, and trying to commercialize that innovation. And most recently, with Gen AI, a few years ago, we called that project internally Microsoft Aristotle, and that was, I mean, the solo that is what you see today on commercializing that innovation, which means that it's very troubling years, this teenage area, because you are neither an adult or a kid, and I feel exactly like that. So that's my point in my career, personally, with AI. Great, friends. Maybe build on the teenager theme. I mean, personally, I think it's been amazing. I use Gen AI every day. I use it primarily to summarize tasks. I feel like the business workplace today is just so noisy. I get literally thousands of Slack messages a day and to be able to sort of take the teams output and be able to summarize that quickly and act on it is a simple task that probably saves hours and hours a day. My youngest daughter uses it to write her papers so that she can go perfect her career as a Fortnite player. So there's wonderful other team and things going on with the technology. I think from a business perspective, though, and I think as CEOs, we all need to appreciate that it's an existential threat every single day. I think the difference though between this one and other existential threats is very much the hype cycle. I mean, I see this somewhat analogous to moving to cloud. I mean, that was just as existential. We built a company MuleSoft that took advantage of the fact that the juggernaut's and IBM and Tipco and others who cornered the market in integration software had not figured out a world where all software was going to run in the cloud. And so here we go, $2 billion later in growing. We built an entire business out of that. And it's generated an entire ecosystem. So we have to look at these things as a threat. But I do think it's a marathon, not a sprint. There's just something about AI. I don't know if it appeals to our romantic nature as human beings, but we love that there might be some catch-all technology that solves the hardest problems that we don't know how to solve. But if you believe that, I've got some ocean front, real estate to sell you in Switzerland and found some magic beans today when it's on a walk. I'll sell those to you for a good deal, too. I think the reality is this is going to take time. The consumer model doesn't necessarily apply in the same way to the business model around UI and data and governance. And it's going to take time to emerge. So I'm not particularly surprised that we're 18 months in. And there's a lot of noise around why it hasn't applied the same in the consumer role of the business world. But there's really exciting innovation, and we're starting to see some of it monetize as well. Great. And garage? Yeah, thanks. I really appreciate Microsoft, of course, with my home for five years. And I have a great respect for Salesforce as well. My career is actually spent what I called the last two platform shifts. So what I was at Qualcomm, we went through the mobile platform shift at Microsoft right at the beginning of the cloud computing platform shift. And I think the sole reason at the tail end of my career decide to start my own VC firm was Chachi Petin. November 22 is like, this is going to be another 10 plus years of wealth creation. I couldn't be sitting on the sidelines of. So I'm personally very excited by Gen AI where it brings to it. I think what many people have said, this consumer is a good example of how to frame it as AI in its sort of some form has been around for multiple years. We remember back in Qualcomm ImageNet, which was the advent of very early machine learning. We didn't call it AI at that time. We can call it AI1.0. Then we had basically a big leap forward with GANS in the 2014, 15 time frame. And then finally, Transform models came in 2017, but we didn't have the Chachi Petit breakthrough, powered by Transformers until 2022. And actually why everybody is so fascinated in the hype is so, so much with Gen AI is because it actually consumers can touch and feel it. Before Gen AI, no consumer had any understanding of what AI can do for them every day. Now, Open AI has Chachi Petit specifically. I think has made that part of everyday life. And that I think is red to this hype cycle. So I'm a very firm believer in the same thing, the long term value of what Gen AI will bring. The short term, we don't see it ready for enterprises. I mean, there's very few use cases we'll go through that. But it's clearly the hype cycle is driven by the fact that everybody, not only in this room, but our kids, everyday consumers, not in tech, can touch and feel what AI can do to them. And they never knew about that in the last two times the AI revolution. So I think this one is real, but the hype cycle is really taken over. So I guess I guess this begs the question. What do you guys see as actually working today? So if we go in the, almost like in the first bucket that Chris describes, what's really today in production? And I'm going to sort of follow on by saying, suspect both in your internal organizations, because you obviously wear two hats, your software vendors, and platform providers, and then within your customers, and maybe within the startups you're seeing. What's working? Yeah, I mean, within Salesforce today, we have over 100 projects running just to implement AI within our own capability across sales, across service. I mean, what's nice about the software we provide is that we provide the same functions to our customers, that customers provide to their customers. So we really feel like we can be customers, zero of the technology. But like, panning back a little bit, I think, I love the concept of it's a broader wave we're in. It started with sort of predictive AI. I think we bought a company relative IQ for about 400 million enterprise value 10 years ago. It now generates over 2.5 billion predictions a day. Stuff that we all see when we go into like Amazon and shopping carts to say, what's the paired product that you would have with that. So predictive AI has been out for a really long time. We're now getting into this generative AI world and very quickly, I love the agent talk earlier. It's not just human in the middle. It's going to be semi-autonomous and then fully autonomous. And then eventually, we'll get to AI around basic human intelligence, which I think those are the doomsday fun scenarios to talk about. But in this world of like generative, moving to semi-autonomous, that's where the majority of the experimentation, at least is happening within Salesforce. We've taken all of our human resources content, all of our medical benefits, all the questions you might ask an HR rep or your manager across 75,000 people and a globally distributed company. And Slack is now answering 375,000 questions that it would have taken a human to answer every single quarter. Our engineers are using our code development tools and not only producing better code, but saving 20,000 engineering hours a month. That is real enterprise value on the human productivity side. I think what's going to be interesting is how do we deploy that more standard with the customer having to do less of the work themselves? I think it's unreasonable that every customer needs to build their own LLM. The UI's are fundamentally going to change. And the underlying data needs to be far more secure and accessible within the enterprise, especially in things like regulated industry. So I might bunk a little bit. Like we are actually seeing tremendous return on that investment. Would I say it's at a place where the cost equation, the margin equation is there yet? And where it's generally releasable? No, I think there's very rapid innovation. But I do think those things are coming. We have lots of customers that have deployed the technology, rocket mortgage, for example, win or low. And simple metrics, like 20% less average handle time. Like if you were in a service industry, who wouldn't want 20% less average handle time on a particular call? Rocket mortgage is actually experimenting with using it to auto loan. So this isn't just work productivity. This is fundamentally changing the way they are going to generate offers and acquire customers around their mortgage products. So we're seeing that today. I think it is very real. Yeah. So in our case, I think it's a little bit different. The difference between first party and third party we have a mantra in Microsoft. You probably remember it was, and it is, first party equals third party. So if we see anybody doing something that is only for internal usage, that's a red flag. And we have even the 1p equal 3p that we bring in in meetings. So for us, and that applies also to AI. So the way that we were able to scale AI in Microsoft is actually creating the platform that then is exactly the same platform that we offer externally. Now, on what use cases we see more real today? I love the way the framework, right? And I think it's something super valuable. And you will see it in different ways in the literature out there. But that concept of knowing where you are in your journey of AI, to then target the things that will be successful, or don't take too much for where you are the other way of looking at that. Super important, right? It wouldn't be just that now the industry is getting more mature. They are, you have a ton of research. You have a ton of even evaluation frameworks. We just ran one in Microsoft, maybe three weeks ago, we released it with 1,500 companies. And we were first identifying in what stage those companies were in their maturity of AI, and then seeing what are the success factors in every stage. And it's usually something very similar to what you saw before. So we see that when you're early in that journey, where we see the majority of the use cases being successful is like that first level. Productivity efficiencies like more than 80% of the use cases are in that stage, or about that. Then the more that you move into a more mature company on AI, then the more you start doing things that are like revenue growing, growth growing, AI first kind of activities. And we've seen those companies more than 50% of their use cases more on that side. Now, nowhere you are on what to do, right? So we call, probably I mentioned this last year, the pilot purgatory, right? That notion that we see many companies that are in this pilot purgatory forever, and you don't get out of that. Maybe you don't even need to be in that pilot purgatory from the beginning. So if you are calling something a pilot, maybe that's already our first step, right? So you should be doing more projects that have a continuity, right? Where the business is heavily involved. So you are starting from the beginning. In those cases, you just skip directly, right? So you just do product development, quick iterations on product development, still of doing pilots, right? So if you feel in that situation, rethink where you are. Yeah, I want to give some examples of that's good. And actually, I really like the first session, I think, Joe and when Joe talked about services, we are seeing this actually work really well when we talk about displacement of services businesses. So good example from one of the companies we did went after essentially digital marketing teams. And digital marketing teams are everywhere when there's an end customer basically as acquiring their customers through Facebook, Google, whatever other channel there is. Historically, it's been very, very siloed. You had a team that would go at one channel for Google, the other channel for Instagram, one for Facebook. And now, using not Jenny, right now, just using a predictive discriminative AI, they've actually built models where you can do cross-channel optimization and what are they taking away from? They are not taking away from software, they're taking away business from agencies. And if you add that workflow, a digital market it gets up in the morning, is able to optimize the ad campaign across multiple things and what else do they want after that? Then they want to be eventually be able to build their ads without going to the agencies. That's where Jennyi comes in. So we're seeing actual production of Jennyi in not necessarily video ads, which are much hyped up. You have much bigger, basically, tam of static ads that you can generate. Everybody talks about Sora and what OpenAI is doing with Sora. You don't even have to go that far. Jennyi using pure static ads and how you can experiment with them and essentially test across different cohorts is a real use case. But that is taking away from services businesses. It is not taking from software. It's fully disruptive. That's a really large time. Another one of our companies is going after a vertical which is, I think a lot of the software there is P-owned. Truly the difference between I think what was talked about is don't disrupt Salesforce. You can't do it there. Like a new incumbent versus a player that is a Salesforce for the auto industry. Call it the DMS software providers. And those are very, very legacy for the most part. I'm sorry if there's one, I'm just saying, at least in the US, if you go talk to anybody in the auto dealership, it is the worst customer experience ever to try to buy a car. Tesla has changed that. But if traditional customer service is also relates back to one of the things here on the tam side. They are not going on there and saying, we have a tam of support. They're saying there's a 65 billion employee tam in the front office that doesn't actually need to be there. What are they doing? Coming up, taking customer calls, most of the times not answering those back, then trying to route to a service advisor where their car is. All of that, whether it's an agent architecture, even if the wrapper can be fully automated. We are seeing this actually working. And it's interesting when some of these companies see these solutions, their jaw drops because their frame of reference is literally software they've been using that is not innovated in the last 15 years. It is not Salesforce. It is not something that comes from Microsoft. It came from providers that would have to re-architect their tech stacks so much. It'd be very, very hard to do. So these are real examples of things that are working. You just have to pick your spots. I just wanted to-- I love that. I think that point was made like two sessions ago as well that when you really look at enterprise spend in technology and software, it's about 15% the average enterprise. The other 80 give or take is just human capital. It is a mistake if you're starting a company today and your only viewfinder is on existing tam that you want to replace or take out. I think that needs to be an angle. What software companies are your competitors? What's unique about the value proposition the workflows you're providing? What's unique about the data you're providing? But I think if you challenge your teams to cast their eye on the human element, that is where we have a massive expansion in the tam is. And who in this room isn't worried about hiring great people, training them to be effective in an organization's retention turnover, like software could be solving a tremendous part of that 80% if we allow this opportunity to take that angle. If we just sort of are fighting over the same 15%, it's a much less interesting problem that we're solving. Can we spend a bit of time on this concept of agents? I think it was described. It would be helpful if perhaps, you know, David, you could describe how you'll seeing those happening, that sort of orchestrate, that sort of basic understanding, orchestrate, a bunch of models to deliver more complex use cases. So, you know, to some, it looks magical. To some others, it looks suspiciously like RPA, which has not worked as well. So, you know, if you could explain on that, what are those, what needs to happen for these to truly be working in the enterprise space? I can't take it. Yeah. Yeah. So, it's funny because we use a very similar one, when I was seeing the frame that you were showing before. God, it was so similar to our thinking, right? We refer to it slightly different, and let me just go through it, because I think the evolution of those give you the answer on what is an agent, right? For us, the first stage where we are right now from the Gen AI point of view is productivity. No need to explanation in there. So, those are applications that are infused with AI, because chatbot, a centric application where they can improve your productivity. Let's keep that. That's where majority of the efforts now, by the way. The second one that we see is called reasoning. So, that is where you can not only do things that you could do before quicker, but things that you couldn't do before, because now you are building on top of the AI that can go much farther than you, that can address, you can address with that push, cognitive capabilities that you couldn't do before. That's simple, right? And we see a lot there already going on. That's for us is that second wave. The third wave, the way that we're referring to, it is discovery. That's the one that is more alien with agent. Why do we call it discovery? Because in that case, things shift. So, in that case, it's not only about that co-reasoning paradigm. It's really the human in the loop of guiding that AI to perform that reasoning. That's a huge shift, if you think about it, right? That's for particular field, for highly cognitive fields. And, for example, research and development, which is our focus for this technology initially, is like really, truly transformative, right? Because now, in that moment, you see a new concept of an autonomous reasoning loop where the human is always in there as well, but the role is different, right? In that autonomous reasoning loop, you have from learning, right, to a hypothesis, to testing, to experimentation, and then all over again, for enabling, not only things that you can make better, but really to uncover new insights, to be proactive instead of reacting, just waiting for the prompt to be written. In that case, you have something that is much different. Yes, it is the starting point of that is an agent, but don't look at an agent like a technology or an architecture or something like that, it's much more powerful and deeper, we're going to see there. I mean, I agree with the framework, and I agree with sort of the promise of fully autonomous agents. I don't know that there is alignment in the industry about how we're going to get there. I think today, I think we should overlook today, I think what's happening today is a technology like an RPA, which is typically a quite rigid architecture to solve a problem. It's extremely brittle to record an existing process across different systems and data. When that process needs to be changed or is dynamic, the bots tend to fall over. But we have these two new things that are really interesting. But generally, if we have this ability to have a conversation, so we're making access to these workflows far easier. We're going to extend a workflow that existed in the contact center to support someone or existed with a field service agent, and we can make that more accessible and more conversational with the end customer. We've just raised a bar for self-service. Or by the time you take first call, we're solving a real problem for you and not a press one for change your reservation, press two, and so on. I think this intersection between conversational side and activating more dynamic workflows is where there's the largest opportunity. And that should replace more traditional brittle RPA architectures over time. I think the question then becomes how do we get to this fully autonomous idea. And we're ignoring some huge elements when we try to make that leap. We're ignoring trust. We're ignoring toxicity. We're ignoring data grounding. What's data grounding? It's actually being able to articulate where this answer came from. Where did this information come from? And we're also ignoring data privacy. This is where the consumer model, where basically a whole bunch of companies went out to crawl the internet effectively. They not like the word, but stole a lot of proprietary data from Time Magazine, from Reddit, and from others. And now they're sort of apologizing for doing that, creating agreements with these companies. But that's how those models were built. And they're fascinating. You cannot do that in the enterprise. Bank of America is not going to just offer their data up for the world to crawl to create these autonomous agents to do headless banking. Not only that, but when you crawl sources of data that are not trusted, it ups the bias. It ups the toxicity. The other big thing that's going on, it's one of the reasons actually Salesforce, I would argue, is a bit late to the market rolling out some of our AI tooling. Because until we had an agreement with OpenAI, that every time we feed that model, a prompt, it guarantees it will not retain the data for its own training purposes. That was essential. I mean, trusted customer data is what our customers own. Not Salesforce. And so I think as we start to add conversational and workflows and also pay attention to the fact that in the enterprise, this data is yours, it has to be trusted, secured, and there's real liabilities to getting the answers wrong, not just reputationally, but legally. I think that creates this interesting sort of human in the middle loop. And I think the best companies will start to evaluate, not just the ROI, that they're delivering customers, but they'll also look at things like the accuracy. And when we start seeing 95%, 96%, 97% accuracy, of course, we will set that free, and that will be semi-autonomous and ultimately autonomous. I think the interesting world is when we start having these capabilities, and it's not gonna be one company that owns them all, when we have them everywhere. If you go on hugging face today, I think this morning I logged in, there's 701,000 LLMs that have been uploaded to hugging face. Right, so massive proliferation. And I would argue, commoditization of the LLM, outside of the equation. But when you start putting all of these things together, I think there's a huge opportunity for orchestration. And we were talking about sort of the legacy automotive vendors. A lot of what's coming is just someone that's attaching into these legacy capabilities, and providing orchestration of these conversational AI capabilities and workflows on top of that. So there's gonna be the great unifier of infrastructure, Microsoft, Amazon, but there's also really a lot of room for the great orchestrator of these B2B processes across technologies as well. >> Yeah, I couldn't agree more. I think one of the things coming to Silicon Valley, I think I'm actually quite, we have a very different view, viewpoint than what you probably heard earlier. We don't think in this wave, you have to be very careful on where to invest, at least for us and where not to, on AI front. Coming from the last two platform shifts, the lesson we learned was in the mobile, if you look at the incumbents that started out, some of them from here, Nokia, they didn't exist, they didn't actually survive the platform shift. Google, Samsung, Apple basically took over the mobile smartphone market. You look at cloud computing, when I went to Microsoft, Microsoft was not the incumbent, AWS was incumbent, Microsoft was quite flat, but Google was nowhere there. So what happened was a lot of the efforts within Microsoft or Google were just catching up to make sure you could become, have the cloud computing efforts going. What they're allowed was a huge amount of innovation in the software side, whether it's a data log, it's a snow, like a lot of companies on the infrastructure tooling side came up, because everybody's ball was somewhere else. What is fundamentally different about this wave is who started it. It was Microsoft as a big incumbent, a teaming with OpenAI, which is a startup. So for us, when we look at companies that are trying to say, we are going to build this horizontal thing, we're going to build an agent, we look at who is actually best placed to do it today. I don't think it will come from the startup space. A, the idea on the enterprise side, how complex it is already was mentioned. In addition to that, the level of integrations it is to do a full agent is off the charts. And to say that somebody who's going to come up with a one startup in the valley with $100 million from Sequoia, anybody else who can sort of disrupt that, actually is oversimplifying the problem too much. So I think the only advice I would give is don't start pilots with Silicon Valley startups on this, because I do think you're not going to see the results here. You're better of waiting for some of the incumbents in this, and it's heresy for me to say this, but better of waiting for the incumbents here to come up with products, because the innovation speed in this platform shifts by the incumbents is off the charts. You wait three months, the next open-air developer day, open-air Microsoft, or plug the gap. If you remember Rag, I mean, I think 12 months back. Rag is just like the new thing. There's 50 startups on Rag. If you think where Rag will come from, will come from traditional database vendors, which have incorporated vector databases into the workflow, because you have data prep, you have so many other parts of the process that bind from a single vendor for a single purpose, you can experiment with it, it will not go into production, in my mind. So I think you'll be just careful of when you spend your time and money, which is very, very important, on things like pilots, who you're trying out with, even if you try it out, I don't think will be a eventual provider here. So I guess this begs the question, as to where, and now I'm checking as an investor, as to where application software starts and ends, right, in this world. So where does a stack, how high do they go, like the platform vendors, et cetera? And where do you invest, I guess, it would be the-- - Yeah, for us, I mean, it's a very small portion of the world we see, but for us, the easiest place to invest right now is what we are calling vertical AI. I mean, the auto example I gave, that is not where the incumbents are focused. And for good reason, those are very small markets for incumbents. They're going after the horizontal layer. If you took up LLMs, we absolutely believe LLMs would be commoditized. That is why that is not a venture or an investable area from a VC perspective. The amount of compute and cost it requires. You raise $100 million, $80 million in video. That's where they're $3 trillion today. That is not a sustainable venture model. It's not shouldn't be backed by venture dollars. So that, I think, will get commoditized. And I think the way to think about LLMs is basically what happened in mobile. There was 50 different mobile operating systems that everybody in the world had an operating system. And who survived, there's like two, really. So they will be open AI. Clearly, that's-- I don't think you can display that. And they'll be one or two more. They'll be at scale, but it'll be a commodity layer at the end of the process. So then it becomes-- there is some tooling, which is one of the layers, and then there is application software. And there is going to be a little bit of where application software starts and where tooling starts. Our view is, when we look at the vertical AI space, the one advantage we have, we think there is. And as a lot of people in the room here is we do believe in the fundamental advantage of proprietary data. And I think a good example I'll give is without-- I can't disclose the names and all that. But one of a European company that is in the space that we expect to be able to mine the data from legal, which is how we can mine all this data, is trying to be bought, or it's got an unsold offer from a quote-unquote Silicon Valley company, that is a Gen AI-based legal company. Because they cannot get it started. They cannot jumpstart their business without the hundreds of millions of data points that have been collected over many years by some of the companies in this room. And they have the scaffolding. They have a wrapper, but they don't know what to do with it. They cannot jumpstart their business. And that data that you're collecting, that is proprietary customer data that is privacy associated with it, that cannot is not available on the wider web, is the mode where an open AI cannot train on that, where a general purpose element cannot train on that. So we do believe if you can build on that, ultimately, you can argue that some of the companies in the resume can take that data and start building domain-specific foundation models that are much easier to train, much less cost, much less compute. And that becomes a proprietary way for you to sort of basically get into Gen AI and not depend on-- I think I'm not saying that general purpose elements are less better at doing domain-specific answers. It's just they don't have the data for it. We do a simple two-by-two, which is value and data. And the upper right is if you have access to or have the privilege of managing your customer's data, like we're given the opportunity to do it Salesforce with our customer's most impressive, you know, precious customer data. And we can develop specific functionality for a task, whether it's to market, sell, to service, to do those things. That the intersection of those two things-- I was actually at the bar last night talking to Ed. We're so in contract unit-- an amazing like telematics company up in-- he's located in Denmark, but they track, you know, heavy equipment at construction sites with this amazing IoT device, this generating just millions and millions and millions and millions of records of data. So they have data, secure proprietary data on behalf of their customers and their customer's equipment. Do they have processes and workflows like when they should replace a part or where something is located, or even geo-coding? We've had horrible weather in an area that's a construction site. So that changed the way we think about servicing and providing for these capabilities. So in the intersection in your enterprise, whether it's internally or externally, where you have high value and high data, that is, I believe, where you're going to get the most ROI. Now, we all don't have that privilege, but I think if you don't own the data, you are going to compete with everyone else that is building products to commoditize data feeds. If you also don't own the data, or have good access to the context behind the data, the metadata, today, general AI can tell you that you should go return a product. The best thing to do because of this error code and issue with the product you have is you should go return it. Great, you saved me a bunch of time at enough to call a call center agent. But it can't do is do the return for me. I want to save return it. That requires the metadata, the data, the context, the business process. So I think it's, again, romantic but folly to think that ERP systems, CRM systems, marketing systems, systems that have all of this data, and the context and the functionality of business process behind that data, that's simply if you take the data out and put it into a general AI that somehow magic happens. It's just not true. I think we're seeing a ton of spend in the LLM. I think surprisingly for a lot of us, that market is that the innovation cycles are rapid, but it's incredible how many very specific LLMs in general LLMs are out there. We've taken a stance of open LLM. Some of that also is for frankly consumption. If you're a service agent, you don't need GPT-4 to run that query. It's bad for the environment, honestly. It takes too much compute. You need a much less sophisticated LLM. So being able to orchestrate calling the right LLM for the right workload for the right function means that the cost of providing that service is sort of economical as well as the value provided. But we tend to always start with sort of value first in answering the question of what products and services that we can provide. I think one of the problem with all these pilots is we've sort of lost our mind and forgot that the reason we funded pilots was we stepped this thing called a business case in ROI. I mean, most of these pilots have no executive sponsor. They have no ROI. They're put in IT lab where we ask people to build things. They shouldn't be building. Your team should not be building LLMs. And we're not focusing on the real issues, which is what is the return on the pilot? And we should be willing to experiment a lot to find out that the return on the pilot is negative and lets throw it away. I think what most CEOs are frustrated by is that 80% of these pilots that are not working had no predetermined outcome. So we are randomly experimenting, which is very different than actually piloting on what we think might be a great idea to take our customers. One thing that I would add to-- because I love that that concept of data being your differentiation, right? So that's something that you need to keep in mind always. I think what we're seeing now with these new models is not only the data, it's the knowledge. So it's the knowledge that you as a company that your people, your company, have created over all these years would make you very patient. And now for the first time with this new generation of AI, you can exploit that knowledge. That's very powerful, right? Because that adds a new layer. So we are all visualizing this stack, right? It was perfect, the slide I was shown before. Where did you want to focus the value? Well, just remember something, by the way, right? So even in Microsoft, right? Which usually we are close to the bottom of that platform. We are not following that approach. We decided as a company. The first thing that we say to every business unity Microsoft, to any developer, to any product manager, is hey, if you can do that problem without creating a model with just an out-of-the-box model, do it. That's the first option. That's what you're seeing M365 is actually out-of-the-box models with maybe some rack on top of that. So if we don't do it, should we do it? Should you do it? So think about that, right? But then the application. But now with this knowledge, you have a very interesting additional layer, which is services. So it was mentioned before today. How is services going to transform? I had a conversation a few weeks ago with a audit company, one of the big fours, right? And they are seeing a huge opportunity to do this. So the way that I visualized in this is that, hey, I've been a force to do like a P times Q business model for all this year. So how many people, how many hours does your price? Now have a new element. Now I can do P times Q times T technology. Now I can take all these knowledge that I being imagine, the audit industry, right? The knowledge that is there, I can put this knowledge to the work of this platform that can then increase the value automatically to all my customers. That's something that I think we'll see more and more. Don't think of, hey, this is going to replace my services. So why should I do it? No, this is going to amplify your services. So take that into account. Actually, that one point of inspiration, I think organizations that are actually getting these pilots out of the technology lab and into the hands of the actual users that are meant to enrich or see more success. So I know we're all a little bit terrified of, are all our call centers going to go away? It's sort of analogous to when Cloud came about. It's like, are there no more people in the data center putting racks on, you know, racking up equipment? And the reality is, that was wildly overhyped. I mean, it's just as companies, we don't have data centers anymore. But the job market for the AWS is the world and other providers is as sure as just, you know, isn't probably better than it's ever been and more sophisticated than it's ever been. But I think you've got to overcome this fear of, we don't want to disrupt our workforce. Actually, most of the time, they're struggling with the tasks that AI would help them be far more proficient at. If you actually give them these technologies at the edge, they will automate their own work. And if they automate their own work in the products that you're building or services that you're building for the market, you can turn around and package that up and offer that as innovation to your customers as well. So I think these pilots need to get out to the edge where the human productivity opportunity is first. That's where you get to that, the V in the equation. And then it does become this life cycle of, wow, we really unlock something for ourselves. Let's go and lock it for our customers. - Just one comment, one David said about the accounting firm. I think this is also a little controversial, but I'll say it because why not? One of the things we are seeing in our early companies and where there is some risk for, I would say, not large incumbents, maybe some not, but even for some of the big incumbents here, is the business model for SaaS. Again, it's heresy, but seed-based models have survived for many years. And I think I could go back to the example of one of my companies which sells into digital marketing teams. So what they found was after year one, the digital marketing team they were selling into went from 10 to two. Think of what happens to a seed-based model there. You basically have no NDR, you're basically going below 100% NDR. So what can you do there? So that's one thing, food for thought, on how you can think of aligning pricing based on value creation and consumption. It's actually very hard. But if you don't start thinking now, at some point you will have less people in seed that you're selling to, and then you'll have to figure out what to do about that. That I think is, the good news in the startup side is actually because they're smaller, they're able to adapt faster to that model. One of the companies I previously founded when I was at M12 had a seed-based model for sales and they've gone to an account-based model. It does not mean you could have multiple sales people there, but it's an account-based model. So I think that is one, I think from a business model perspective, is good to start at least thinking about in your own business because it's going to be very bespoke. And there's no one answer to change it because my company in the additional marketing side was easily able to align with ad spend. Well, if you give me this much ad spend, I save you 30%, that basically becomes my business model. That is not like a one-fit, one-size-fits-all. Gentlemen, thank you very much. And it's now time for coffee. Thank you. Thank you. Thank you, it's great. Thanks for listening to Orbit, the HG Podcast. If you'd like to find out more about HG and our work building and during enterprises, subscribe to our newsletter at htcapital.com/newsletters.