6 Questions Founders Should Ask Themselves to Drive Value from Generative AI with Base10 Partners

Thank you for joining us. I'm TJ Nahigian and I'm joined by my partner Lucy, and we're going to be chatting with you for the next 30 minutes on 6 things that founders should be asking themselves to drive value from generative AI. So let's get started.

Hi everyone. I'm Lucy Fonseca. I'm a partner at Base10. Joined TJ and the team about 2.5 years ago to lead our growth investing efforts. And I've spent the last 18 months or so with TJ going as deep as we possibly can in generative AI and super excited to share what we learned with you guys.

Just to give a little bit of background, Base10 is a research-driven investment firm focusing on companies that are automating the largest sectors of the real economy. And like most venture firms, we've been trying to figure out what is going to happen with generative AI, and we're going to share at least what we're observing today and why we think it's a platform shift that both incumbents and upstarts can truly take advantage of to build significant value.

So last year we were one of the first VCs to actually share some of our research publicly on what we were seeing in generative AI. The first piece we published was one that looked at the platform layer and the application layer and essentially set out some questions and some hypotheses that we were starting to toy with as we saw the impact of this technology. We then very quickly followed that up with a piece on essentially the infrastructure tooling or the middleware layer that was helping platforms and applications essentially use each other to be able to build. We learned a bunch from sharing some of that research, and essentially we have a Friday meeting where we go through our research every Friday, and we are always asking ourselves, okay, what have we learned? What did we get wrong last time? And what are we seeing in front of us today? Now, as VCs, our job isn't really to try to predict the future, even though sometimes it seems like that's what we're trying to do. As VCs, we are really just trying to look at what is in front of us and see what's there. And what's eminently clear is that generative AI is a platform shift, and it's fundamentally changing a lot about technology today, similar to cloud, social, and mobile.

And so I've been fortunate in my career to see a few different platform shifts, or originally cloud, then mobile and social, both happening at the same time. We think generative AI is going to be a platform shift. We didn't always think that, and I can chat through why. So before starting Base10 7 years ago, I was a founder. I founded a, a mobile business called Jobber. It was a mobile job marketplace. And even though we were a mobile business, we really viewed ourselves as a data business. Most of our team was engineers and data scientists. And we would use data basically to make the job application process move much more smoothly on your mobile device. And we ended up using data a lot. We built out our own taxonomies. We used the latest machine learning models to train and run our recommendation engine. But I was always so frustrated. Even though we had an amazing team, we had a great dataset, we got very little value out of machine learning and AI back in 2013, 2014. It was just, it felt really inaccessible. But that all changed. So that all changed when OpenAI came out with GPT-3 about 2 and a half years ago. And for the first time, generative AI actually became accessible, not just to the huge companies and mega businesses out there, à la the Facebooks, the Googles of the world. It actually became usable for developers. And since then, we've seen significant advancements up and down the stack, both from upstarts and incumbents. And our view is that not every platform shift looks alike, but they all tend to have different similar characteristics. And so we're going to take a lot of the learnings that we've seen from cloud or from mobile and apply them to generative AI. One thing that is true, and Lucy's going to talk about this in a second, is when you started building in cloud in the early 2000s, you were a cloud business. Now every business is a cloud business. Jobber, the business I started, was a mobile business. Now every business is a mobile business. Same thing's gonna happen with generative AI.

Yeah. And you might be thinking AI is not new, and we certainly agree with you. And obviously the quickest way to become a skeptic is to try to build a business using AI, which we've definitely, many of us on the team have. But what we can actually see if we just take a look at the last 5 years is that something pretty ground shifting is happening. The most powerful models ever made in AI were all formed in the last 5 years, and you can see some of them on this page here. But if, even if you take a look at just the pace of innovation, the number of patents filed in 2021 in AI was essentially 30 times the number that was published 6 years earlier. And if we just project out and keep that pace of innovation, think of what's going to happen in the next 5 years. We think essentially we're at the very cusp of a golden age in AI. And if we take one of the lessons that we learned from cloud is that cloud really sped up the pace of development by a lot. If we take that and take a look at what might happen with generative AI, what we expect to see at Base10 is that the speed of both development and deployment is just going to accelerate so dramatically. It's going to make our heads spin in the next few years, and we're really excited about what that's going to create for, for innovation and for opportunities for investment.

The other way we look at this is just by looking at how much venture funding is going into the sector. For those of you founders, the last 18 to 24 months has not been the easiest or most fun time to fundraise. In fact, oftentimes impossible, but that's with one exception. That is if you are an AI business, and we're seeing that happen across different parts of the application and infrastructure layer. And that is already a very large number. I think the number last year is over $4 billion raised in, in even actually even the first half of last year. This year we're expecting that to be, it's already 5 times more than it was last year, which is 5 times more than the year before. Venture funding doesn't necessarily mean things are perfectly going to work, but oftentimes that is a good indicator of where there is a lot of innovation happening.

So as investors, it's really helpful for us to have some framework by which we can understand the universe and who's building where. So we're gonna walk you guys through the framework that we use at Base10, and then we'll walk through what we've observed in every one of these segments, and therefore how that's impacting our own investment thesis. Maybe there's some helpful lessons in there for anyone building in these segments. So essentially there's two big categories. Incumbents and upstarts are both leveraging this technology since it's like electricity, but within those categories we see platforms, infrastructure, and applications. So the platforms you can essentially think about as the model layer, that's your Googles and your Metas and incumbents, and your upstarts are obviously your OpenAIs, Hugging Face, Coheres, those businesses. At the infrastructure layer, it's anyone essentially allowing you to use these largely generalizable models to essentially create something with a more specific use case. So those are your picks and shovels. And infra— we'll have more on this a little bit later, but NVIDIA is the huge winner here on the incumbent side. And on the upstart side, we've spent time with Great businesses like Lambda, Pinecone, and LangChain might also ring a bell. And finally, our last bucket is the applications, and we could probably even split this into both horizontal and vertical, uh, applications. We've chosen to keep it simple, uh, but on the incumbent side, you have your Notions, Gorjus. You could probably take Google, Meta, Microsoft, and run them through every single bucket here. And on the upstart side of applications, names like Jasper, Copy, and Harvey should probably be familiar to the audience. So that's how we think of the world.

So at the platform layer, we think of this really as the model layer. OpenAI is the clear giant in the room. They've originally started as a fully open source business. We all know there was some drama, led to Elon Musk actually no longer working with the business, and now are closed source. But they were really the first ones to come out with a scalable large language model. Ironically, they were enabled to do that because Google created the transformer model and open-sourced that a few years prior, which they needed in order to actually create their original LLMs. The first popular one, GPT-3, came out 2, 2 and a half years ago. And since then, we've seen dramatic acceleration happen there. But it's not just OpenAI who's now raised $10 billion plus, most recent round led by Microsoft. There's a host of others that are innovating here from businesses like Hugging Face, Cohere, AI21 just announced another very large fundraise last week. Stability AI, Anthropic, and others. So we're seeing a ton of innovation happen here. This is creating new LLMs and other models that folks are using to actually get value out of generative AI. And moving on, how are they actually doing this? They're building their own tooling, but also they're leveraging third-party tooling, as are other upstarts and incumbents. And the infrastructure layer is a really interesting one. So this is tooling that enables you to actually tweak different models, integrate them, scale them in a cost-effective way, observe them, and do a whole host of others. LangChain's a really popular framework in order to actually integrate and get value out of generative AI, but there's a whole host of different areas here of startups that are actually innovating at the infrastructure layer to enable folks to actually have more control and get more value and cut down costs. As they scale out generative AI for their specific applications.

And now the applications at the upstart layer. So we actually started here in our research, and there's a lot of fascinating innovation happening here. And the way that we think about it is essentially changing mediums like text and writing, image, video and audio, and even code. So we've now gotten to spend time with great businesses like Jasper and Copy and Writesonic that are fundamentally changing how text and writing can happen, but we've also met businesses changing how you create music or essentially recreating and disrupting how films are made or dubbed, or even companies like Replit that are disrupting the way that programmers code, or, and that's all on the upstart side. So these businesses essentially have scaled incredibly fast, and the way that they've done that is they've taken these large foundational models. And they've wrapped it in what sometimes feels like a really magical experience for the end user that has allowed them to scale remarkably quickly. What we have seen though in the applications in upstarts is that a lot of them are now facing challenges, and those challenges essentially boil down to one word, and that's retention. Retention has been tough for some of these, and there's a couple of reasons why that might be, and think about it for the, those that are selling to enterprises, often they're selling to experimental budgets and sometimes those budgets just run out, right? The other reason might be that some are actually selling to more of a prosumer base, but we know that prosumers are actually a very churny, inherently very churny base, and that can be really tough as well. And then I think that the last thing is with this wave of innovation, we've actually seen a lot of copycats. So even if you just take a look at the list of founders building in the last YC batch, there's probably 5 or 6 doing the same e-commerce customer support effort. And so then it becomes really difficult, not only for, for the founder to be able to tell a very differentiated story, the user also might not be able to see where the differentiation is. So remember that word retention is a challenge, and if you're building here, you gotta be a little bit more strategic about this. So you've heard what we're seeing at the upstart level. We're now going to tell you a little bit about what we're seeing at the incumbent level, which is also very exciting.

And starting with incumbents, we'll start first really at the platform layer. Again, think of this as the foundational model. And I think it's been really interesting to study how each of these three large incumbents is playing it. They're all actually taking a very different strategy. Google, I think, as I mentioned before, we'll start there. They actually came out with the transformer model and then eventually open sourced it. That enabled LLMs and other models to be built leveraging generative AI. I think they were caught a bit on their heels when they saw how fast OpenAI was able to commercialize their offering. And since then, Google has been actually rethinking of themselves again, truly as an AI company. Coming out with Bard and trying to innovate in other areas up and down the stack. Meta is another really interesting one to study. They've had incredible technology that they've been leveraging primarily for their ad model, for their recommendation engine across their ads, and monetizing that in a very material way over the last 7 to 8 years, but also their feed model and content model as well. They came out and open sourced Llama 2. It's actually an unbelievably powerful model. And it'll be interesting to see what happens with that. I know a lot of entrepreneurs that we work with are now actually starting to leverage that versus OpenAI's models. And lastly, Microsoft. Microsoft, I actually think, has done a masterclass in strategy here for us, where they've actually opted really to partner and somewhat outsource innovation. They did that with OpenAI early on, initially gave them a billion dollars, most recently $10 billion. And actually have a lot of, I would say, preferred access as well as rights and control over OpenAI. So that was not free money by any means. And what's interesting is if you look at the market caps of these businesses, the share price of these businesses, they've all done incredibly well this year. If you read any of their earnings reports, 2 years ago, the term AI came up maybe once, twice in an earnings call. On average now it's coming up 100+ times in an earning call, in an earnings call from these businesses. And anyways, I think this will be fascinating. I imagine business schools are going to be studying the strategy of these 3 incumbents at the platform layer going forward. We'll see how it plays out. Next, we'll talk about the infrastructure layer. How do you actually enable folks to do this? And perfect example here is NVIDIA. NVIDIA is the creator, popularizer of GPUs. GPUs necessary to run LLMs, and they've seen insane demand for GPUs. Just this last quarter, they announced at their earnings revenue was up over 100% in the quarter, which is insane. NVIDIA is now a top 5 company in the world by market cap and truly propelled by this. It's probably the best pure play incumbent infrastructure company out there. But again, those other companies I mentioned, à la the Facebooks, Amazons, Googles of the world also benefit tremendously from this via their cloud offerings as well. We're seeing a ton of value actually be created and captured by these incumbents at the infrastructure layer.

Yeah, over or under NVIDIA gets to $700. Finally, we're going to talk a little bit about the incumbents at the application level, which have by no means been sleeping at the wheel. And the reason why NVIDIA is benefiting so much isn't just because you have applications at the upstart level, it's because every single company is essentially developing an AI strategy and no executive essentially wants to be caught without an answer as to what they're doing with generative AI. And you can see that just in the sheer number of instances in which these companies are talking about generative AI to their investors. That's increased over 80% year over year if we look at earnings calls. But something that's actually an even more dramatic shift that we're very excited about is All the engineers we're talking to essentially are infinitely more productive, and that's because GitHub CodePilot is now writing 50% of code, and that's every other line of code, to put that in perspective. And that is not only incredible when you think about the pace at which these applications can innovate, if you just think about the access and the democratization of who can actually create, that can get you really excited about what's happening here. I think even us lonely VCs will probably be able to create these slides with maybe Microsoft Copilot next year, which would be great. But ServiceNow, for example, has also added generative AI capabilities into their product. So you can actually do summarization and text-to-code in there. And TJ and I spent a bunch of time with our sort of growth stage portfolio companies, understanding how they're thinking about generative AI, companies like Notion. And what we are seeing is that these businesses are actually being absolute ninjas about experimenting really quickly with generative AI. And some of the lessons that we've seen is that one, they're often doing it in-house. They're just experimenting with open-source tools and just building as they go. The second thing we've seen is, in addition to just using their own engineers, is they're often going out and trying to buy or partner the way you might see in some other waves. Sometimes they're acquihiring, but that I think is a fundamental lesson if you're trying to sell to some of these businesses, because a lot of them are actually building in-house and it's important to keep that in mind. So now you guys have a view of how we see the landscape. Now we're gonna take you into a little bit of our investment thesis and where we're focusing a lot of our time. And what we think are some lessons for the folks in the room.

And as we're doing that, one of the things that we're spending time on is understanding where generative AI actually is going to have true product market fit. And one lesson that we've learned in the past when we've observed platform shifts is companies that weren't necessarily built for XYZ platform shift for cloud, for mobile, et cetera, happen to be perfectly positioned for that. So there's a few early categories that we've been drilling into. These are areas like customer support. At Base10, we work with 6 or 7 businesses in the customer support space. And I can tell you that the expectations from purchasers of that software is incredibly high about what generative AI can do for their end user experience. The reality is a bit far off still, although it's starting to converge. And we're seeing this happen not just in customer support with businesses like Intercom, which has Fin, some of you guys may use it. They're saying it automates about 50% of chats. We're gorgeous as well. We're seeing that happen across a whole host of other areas. So in DevOps, when we were getting started at Jobber, the business I started, I used to do QA. Brutal. You had to do it manually. Now actually you can automate a lot of that using generative AI. We're seeing that happen in legal documents. We're seeing it happen across low and no-code applications and getting more value out of that. We're seeing that happen in medical billing, actually reducing again, human workers, and we're seeing it happen in, in data providers. Clearbit's a good example of that, which essentially just relaunched their entire platform 3 or 4 days ago, leveraging generative AI. So Lucy, maybe you can chat through what our original hypothesis was and what it's changed to.

Yeah, when we first started, we were throwing darts at the wall in terms of what would actually drive defensibility. Where were, where would the real moats be found? And we had 5 areas that we thought might be those. So we saw businesses like Jasper really leveraging the community and brand they'd built over many years. We saw others like Copy really leaning into the user education. Everyone essentially said that they would have access to proprietary data or that they got the workflows of their customers the best, or maybe they would be able to hyper-customize in a way where they would win. But now TJ and I are pretty sure that only 3 things matter, and it's whether you can distribute. So it's distribution, it's your data, and it's the workflows. Only these 3 things matter. And because only these three things matter, what that has meant for us as investors is we actually are developing a thesis that a lot of the value, not all, but a lot of the value in this early platform shift is going to be captured with incu— by the incumbents. And naturally it's because they already have huge advantages on distribution, data, and workflows, uh, over their, uh, upstart competitors. So we're gonna walk through some case studies where we've seen that and then open to any challenges from the audience.

So one great case study is a business called Casetext. Uh, for those of you unfamiliar, uh, it's a legal software platform. They actually sell into mid-market and enterprise law firms. Uh, if any of you have ever tried to sell to a law firm or interacted with a law firm, you know, it can be quite challenging. Uh, getting through legal review at a place like Oric is no easy task. Casetext has been doing that for the last 11 years and they built actually a pretty strong platform doing so. But the business got up to about $10 million or so of recurring run rate revenue and stalled out over the last few years. So that's about 10 years in. If you're sitting there as the founder of that business, you're thinking, oh man, this is tough. What am I going to do? Been flat for a while here. Then generative AI comes out. They were able to easily plug into initially GPT-3, then 3.5, now 4, and launch their version of Copilot called Co-Counsel. Which actually will go out and do a lot of the tasks that a lawyer needs to do routinely in their work in terms of surfacing information, summarizing that, and then actually creating responses to that. They launched this at the end of last year and that product took off. So they launched it as an add-on, went from $0 to $9 million of ARR in 7 months, created a ton of interest both from VCs, but also from strategics and a few weeks ago, they announced they were getting acquired by Thomson Reuters for about $650 million of cash. Thomson Reuters, historically incredibly price-sensitive buyer, right? But they were able to do this because they already had distribution into these law firms and those law firms really wanted to use AI. Turns out if you were an upstart, you actually had a hard time getting through legal review with these law firms, right? They also had data as well. And so they were able to leverage that into huge groundswell that reignited their business and drove a bunch of growth and great outcome for them.

Yeah.

Lucy, maybe you can talk about our portfolio.

Yeah, I love the Casetext, uh, example, but I love the Notion example even more. In our portfolio, we've gotten now to spend a lot of time with this team. I also use it every day, so I've been able to leverage the benefits of Notion AI in my own workflow, which has been great. But so they launched a first of many upcoming AI features. And they've been able to price that in such a way that our guess is it'll probably be $100 million plus ARR business in the near term for them. And they're charging essentially an upsell of $7 to $10 to be able to turn on Notion AI. And there's, I think, some fundamentally really important lessons here in terms of how Notion has executed. So the first thing is they turned this feature on to a user base of 30 million people. Including their very much growing enterprise practice. So think about just the expansion opportunity there and the embedded distribution is just a fundamentally huge advantage, right? Secondly, when I'm using Notion AI, I'm using it in my workflow. I'm using it in a tool that I use day to day, and it has now access to my data, to my workflows. And as you guys know, these models are incredibly generalizable, but what makes them really useful is when you feed them very unique and often private data, right? So that's a huge advantage that they have access to that. And then finally, now that we've spent a bunch of time talking to Akshay and Ivan and the team and really understanding kind of their philosophy around how they're using generative AI, it's very clear that this one Notion AI that you guys maybe have access to now is a single feature, but they think about generative AI as core technology, as electricity. Which means that they have now gone back to their original strategy that they had 8, 9 years ago, which was making toolmaking ubiquitous. That's the core mission. And they're thinking about all the ways in which generative AI can now power that, which is how everyone we think should be thinking about generative AI. Let's share one more example and then go from there.

So lastly, I'll end on another great example, Gorgeous. We thought we'd bring this up because Jason M. Lemkin, the founder of Saastr, is actually an investor in this company, and we've gotten close to Roman and the team over time. For those of you unfamiliar with Gorgias, it is a ticketing platform for customer support focused just on e-commerce and really started primarily with Shopify. So I have this really constrained use case of how do e-commerce merchants actually manage and interact with their end customers. They started the business in 2015. The original thesis was actually to automate that. Turns out, just like me at Jobber, it was pretty hard. They actually couldn't do it. And so they ended up building kind of the workflows and software support system to enable individuals to go in and do that and respond to customers. At the end of last year, they ended up launching automations. Turns out their customers actually don't know what AI is, nor do they care about it. But what they do care about is not having to respond to their customers to give a better customer experience to their end customers. So they launched automations. They launched it as an add-on product priced at about 50% of the core offering. And initially it started to automate 7% of tickets. It's gotten up to 18% today. They think within the next 6 to 12 months, they'll be able to automate 50% of the back and forth interactions between e-commerce merchants and their end customers. Again, have significant amounts of data that they can train this on. They already have distribution. They are, I think, the largest ticketing platform for Shopify and already have distribution across 10,000+ merchants that leverage this and they were able to actually innovate in a way that makes their business model much more attractive and a higher retention product for their end customers. And so moving on, we'll talk through a little bit of what we've been seeing from the business model perspective. So we're seeing a few things happen. Number one, we're seeing ARPU increase, ACVs increase, and LTV go up. Some of that is by launching essentially add-on products. Those can be priced as a subscription fee, or it can be priced transactionally, i.e., per conversation in the example of Gorgeous, which they're actually switching to after starting with a subscription model. And we're seeing that price typically at 50 to 100% of the core offering. On average, about half their customers are adopting it. So you get 30% increase in ARPU, an increase in LTV. Secondly, we're seeing a drop in costs. So In the case of medical billing or in the case of QA testing for software, you used to have to do those things with humans. You're still going to have to do them with humans, but you need a lot less of them when you can leverage generative AI to do that. And so you're— those businesses are not dropping their prices, by the way, they're keeping them the same. And so you actually get a significant margin increase that increases your LTV. Lastly, we're seeing retention increase pretty dramatically across businesses with data distribution and the right use case. And so that's increasing LTV as well. The other thing we've been noticing is some forward-thinking companies are pulling it into their onboarding and conversion funnels. Better product offering can actually increase conversion anywhere from 20 to 50%. And so it's another area that we're paying attention to. Overall, we think for the right types of businesses, this could have a really dramatic impact on increased revenue growth, cutting costs, increasing profit margins, ultimately creating a lot more enterprise value. And then lastly, returning to our hypothesis, as we think back through cloud, through mobile, what can we learn from them? One thing that's interesting is Amazon actually wasn't built as a cloud business. It wasn't built as a storage company. It was an incumbent e-commerce business that just happened to launch AWS and They ended up capturing way more value, just them in and of themselves, than the entire on-prem storage market before them. We're also seeing Google capture a ton of value through cloud, even entering much later, right? We're seeing Microsoft do the same with Azure. Meta was not a mobile company. It wasn't built as a mobile company. They captured more value from mobile than AT&T, Verizon, et cetera combined, right? Just Meta. We think the same thing could happen in generative AI. That doesn't mean there aren't going to be upstarts that win here. There definitely will be. But where is most value going to be captured? Our view again is incumbents are going to capture a lot. We're already starting to see that, right? Meta, Google, etc. have traded up dramatically this year. NVIDIA has traded up dramatically. They're capturing a ton of value already. But we're also going to see a lot of upstarts as well. Just have to be really careful where you play. And so Lucy, maybe with that, I can pass it over to you in terms of some of our key takeaways we'd like to leave you guys with, and then we'll open it up for questions.

Yeah. Absolutely. So any upstarts building with generative AI in the audience? Okay. There, so you, maybe you're thinking we're telling you to stand down. That is not what we're trying to say. What we're saying is actually in order to beat the incumbents at anything, you're ac— you're gonna have to be even more strategic. And so here are, if TJ and I were building an upstart business in using generative AI, here's what we would be asking ourselves. The first question would be, what niche market or vertical can you absolutely dominate? Not rejigger a slight solution to, but actually dominate in a way that's not completely not adjacent to particular incumbents. The second thing, and I hope that the presentation really nailed this particular point, but retention is absolutely key. And generative AI has actually fundamentally shifted the way that consumers and users actually expect to interact with your product. For example, in a more conversational-like manner. And so what can you change about your product to really drive product-market fit in a way that tackles all of the retention issues that some of these upstarts are facing? And then finally, and this is core to everyone in the room building around this, but how do you make generative AI core and foundational to your product and not just a feature? And I will say that some of the most interesting generative AI companies that I've met don't even talk about generative AI. They don't even mention it, but it's powering everything in the background, and what they're selling is just a value prop. Obviously, maybe you wouldn't do that because the multiples are so high with generative AI now, but you're building a business for the long term, and it's really fundamental to keep this point in mind. Those building in the incumbents, any incumbents in the room building? Okay, we've got a few lonely folks. So here are the questions we would be asking ourselves if we were incumbents building here. The first is, let's go back to our original mission and the core problem you were solving for your customer, and what are the 1, 2, or 3 problems you can now solve that you just couldn't before when you were starting because this technology did not exist? And how do you tackle those and throw resources against those? The second thing we want you to do is essentially take out your P&L, take out whatever metrics dashboard you use, and go through every metric and ask yourself, How can generative AI actually help me optimize my COGS, optimize my LTV, optimize my conversion, whatever it might be? And the last question is, we talked a lot about the pace of AI innovation, and you guys can see this, it's dramatically fast. So how do you actually experiment and test in a way that meets the pace required for this platform shift? And there may be cultural organizational changes you might need to make. In order to be able to do that. Thank you for your time. Thank you guys so much. Really appreciate you all coming.

All right. Thanks, Hoff.