How Startups Can Crush Enterprise AI Sales | E2050

A startup should be looking at like, well, what's different? And it's like, well, part of what's different is we have reasoning as a utility. Think of every role in the economy as a bundle of tasks. And those tasks, some of them fall to AI and some of them fall to, to humans. So I, I do think what startups should be looking at is like, how can I change the way, like, if you look at workflow, the way that it's been built over the last 10 years, it's been built with humans at the center and that's the actual workflow. So if you were to zoom all the way out and say, oh, okay, well, I'm going to have a workflow that's going to be shared, some of it's going to be through an algorithm and some of it's going to be through a human. What would that allow me to do?

This Week in Startups is brought to you by LinkedIn ads. To redeem a $100 LinkedIn ad credit and launch your first campaign, go to LinkedIn.com thisweekinstartups Zendesk. The best customer experiences are built with Zendesk. Qualifying startups can join their startup program and get Zendesk products for free for six months. Visit zendesk.com twisttoday to get started and beehive power your newsletters with AI tools, referral programs and ad network features all in one platform. Get 30 days free and 20% off your first three months at Beyve.com twist Hey everybody.

Welcome back to this Week in Startups. This is Alex and I have a special show for you today. There's a theme going around the world of startups which is AI hype, AI fundraising, AI applications, AI agents, AI chatbots, AI job loss, AI this, AI that. What I've always been very curious about is behind the headlines, how is enterprise generative AI spend going? And thankfully for me, a venture capital firm that I've known for a long time, Menlo, dropped a very long and detailed report digging into precisely that. How is generative AI doing in the enterprise? So I decided to have a couple of the authors of that report come on the show to answer my questions. And then in a stroke of fortuitous luck, Menlo is perhaps best known in the AI game for being one of the backers of Anthropic. And this morning the company announced that it raised $4 billion more from Amazon. So now we have the VCs from the report trapped under our thumb so we can hit them with all sorts of questions as we would like, but just to welcome them to the show, first of all, we have Joff Redfern. Joff, how are you?

I'm fantastic, Alex. Thanks for having me on the show.

Absolutely. And what I like the most about you, apart from the fact that you're part of this report, is that you were at Yahoo back from 2003 to 2009. Right.

I was, I was. I was early Yahoo prior to. I've been a product leader my entire career. Prior to joining benlo about a year ago, I was the chief product officer at Atlassian. Was there for six years. And then before that I was an early vice president of product at LinkedIn when it was a tiny company pre IPO and then helped grow that up into 10,000 employees. So I was there for seven years. Yeah, I love building things. Started companies, sold a company, helped take a company public. That's a bit about me. Yeah, yeah.

But the coolest thing is that you worked at Yahoo, where I've also worked. And it's.

Yeah, it was like of all the things, Alex, that we point out, it's like, I don't know if Yahoo was the one, but. Yeah, no, Yahoo in the early days. And you'll remember this was just super fun. It was like, it was really the Google of the day as the Internet was emerging back in, in the late 90s. So it's always the tricky one because I think younger folks that are listening won't really know Yahoo as the amazing company that, that we were a part of in the early days would be like if an amazing institute like Harvard was all of a sudden, you know, not Harvard, and it was at a different tier. So, yeah, it's changed quite a bit over the years.

It's like if Harvard became San Jose State, was purchased by private equity and.

I was going to, I was going to go with Cape Cod Community College, but.

Cape Cod Community College. Yeah, That's a deep cut. I'm in Rhode island, which is why all Northeast jokes work. But we also have Derek Shaw with us. Derek. Hey, you my friend. I was prepping for the show and the thing about your background that I love the most is that you were president of the Harvard Crimson.

That's right. I started in journalism. A lot of overlap, I think, with venture capital, but trying to. Yeah, it's an interesting start and a lot of admire a lot of the work that you do and excited to be on.

I appreciate that. So also, Derek, on your Menlo page, you are tagged in on the anthropic side of things. Now I know it's Matt Murphy who's the lead partner over at Menlo for the Anthropic relationship. But today the company announced that it raised 4 billion more from Amazon, bringing it to a total of $8 billion. And I, I guess the question that I want to know is why? Why is it still so expensive for these AI companies? Because that's a lot more money not that far afterwards. And to me sitting here where I am, apart from the numbers and so forth, it feels a little terrifying. But I was hoping you could assuage my, my, oh my God, at that number.

Yeah, I mean, I think Anthropic was, has a special relationship with Amazon. I feel like they have a very close partnership and this is just a furtherance of that. But the other side of the coin is that Anthropic is one of the foundation models. Right. This is when we made our first investment back in 2023, the thesis was that this was going to be one of the, you know, companies that will matter the most in the AI revolution. And we've seen that play out. That's why we have been getting closer and closer to the company. And you know, the announcement today is probably just a furtherance of that and deepening of the relationship with Amazon.

So Joff, you guys took part in the Series C and then if I recall correctly, led the series D, is that right?

Yeah, we, through an SPV, we were lead on 500 million of the billion dollar raise that happened on the D side.

And I take it now feeling pretty smart given how AI has gone since that deal was put together.

Yeah, well, we're, we're excited to share with you some of the findings that we've, we've learned on the LLM side as it pertains to the enterprise. Right. There's obviously two large markets for LLMs. We have the consumer based market and then we have the enterprise based market. And it, it's, you know, it's starting to shake out that these are different markets and, and how competition is approaching and attacking. Those are quite different. So very, very happy to see some of the data behind Anthropic's progress on the, on the, on the enterprise side.

I know that I'm slightly harassing you with the Anthropic news when I asked you to come on to talk about the actual enterprise AI report, but I'm just curious, one more question about that. Does having a company like Anthropic in the broader, you know, family portfolio, does that really bring in a lot of information to the investment team that you guys can then I don't learn from and apply directly to making new investment decisions or is that information segregated from the firm? And so you guys can't go fishing in that pond to get for.

We are, we are very, you know, church and state when it comes to that. We're, we're anthropic and in both menlo are, you know, we're not in there deep looking at people's usage metrics or anything like that. But I will say one benefit of the many that comes from it is just being able to see what actually is coming down the pipeline when it comes to new models. We forged a relationship, a special relationship with anthropic. Five months ago we announced our anthology fund. That's $100 billion fund and that's really looking for who are some of the most pioneering AI founders out there. And being part of that fund gives gives founders a number of benefit. Early access to models, $25,000 in anthropic credits, access to some of the devrel teams over there and some of the expertise. You'd be part of a network of fellow founders and builders in the AI space. And then every so often, once or twice a year, we run a builder's day. Our first builder day with Anthropic was the first of November and that was just, that was super fun. So it was at the Anthropic offices, we had a bunch of the experts from Anthropic coming out, meeting with companies that were building on the, building on the platform and yeah, so we were able to have a really successful builder day. So I would say, you know, that part of the relationship is really helpful for both of us. So we really cherish the team over there and they're, they're just amazing. I mean the team and the caliber of the talent that has been brought to Anthropic I just, I admire every day it's really quite talented.

Well, with $4 billion more, they can certainly keep hiring. And I know one of your predictions in the report was continuance of the AI talent drop. So we'll get to that in a minute. Let's start with what everyone wants to know, which is the high level numbers. So you guys wrote in this 2024, the state of generative AI in the enterprise, that spending surged to 13.8 billion this year, up 6x from 2023. Now, before we get into categories, Derek, did that match your expectations? Is it a faster pace of growth than anticipated? To me, big number, big jump. No idea if it was bullish or bearish compared to your projections.

Yeah, I mean, I think that when we did this report last year, we thought that this would take a little while to ramp up. And that's what you saw in previous technological transformations like this, right? You benchmarked cloud, you'd benchmark to mobile. And these are all, you know, trillion dollar kind of markets. Now that took a little while to get started. And so I think when one of our predictions from last year was that this would similarly take a little longer to to ramp up despite all the spending that's going to it.

Have you built something awesome? Are you ready to bring it to market? Well, that's great. But you're probably asking yourself, how do I get to my target customers? Well, here's an easy solution for you. It's called LinkedIn ads. And with LinkedIn you can precisely target the professionals most likely to care about your product. You can use LinkedIn to reach people by job title, industry, company location and more. You know, all the stuff that you have on your profile page. Imagine you're looking to build your B2B audience. Maybe you want mid sized companies, maybe you want large ones, maybe you want emerging small businesses. Right? Small and medium sized businesses, SMBs. Well, you got to build those key relationships and you want to drive results. Well, you just do that with LinkedIn ads and you're going to gain access to over a billion members of LinkedIn of which 13%, 130 million of them are decision makers. That means they can pull the trigger on buying your product or service. And 10 million of them, 1% of the LinkedIn audience, are those elite C level executives. You know what I'm talking about? Chief Technology Officer, Chief Strategy Officer, Chief Revenue Officer, Chief Executive Officer, Chief Financial Officer. They're all there on LinkedIn. And you're going to get two to five times the return on your ad spend. And 79% of B2B marketers say LinkedIn delivers the best paid media results for a reason. So here's your call to action. Start converting your B2B audience into high quality leads today. And we're going to give you $100 credit to start your next campaign. LinkedIn.com thisweekinstartups to claim that credit. That's LinkedIn.com this weekinstartups terms and conditions do apply.

I think what we've seen this year actually is a surprise in terms of how fast that the revenue is actually materializing. And so you look at areas like healthcare or legal which are Actually traditional laggards in terms of technological adoption. And these are actually the verticals that are leading the AI revolution in many ways on the app layer. There's $100 million plus companies now in both of these verticals. And I think that that part was surprising is just like how much that app layer has really taken off.

The reason why I'm actually less shocked by that than I thought I would be is that my spouse works in medicine. So I'm actually viscerally aware of that world and its inefficiencies. But if there are places where you could apply AI to a voluminous amount of written information, legal and healthcare are gotta be two of the ripest apples on that tree. So even if they are historical laggards, doesn't that actually mean that they have more accumulated equivalent of technical debt in their operations and therefore they're the best place to deploy? So actually I can kind of see that, Derek. I guess I'm less surprised by that than I thought that I would be.

Yeah, absolutely. And the other way that we look at it too is the comparison between the size of the total market versus the size of technology spent there. So if you think about some more. Yeah, I mean this, this goes towards, you know, the Silicon Valley kind of phrase now is services software. And I think that it rings true. It's a cliche because of a reason, right. Because these are massive markets that tech used to not be able to touch. And now with AI, they can now automate a lot of that.

You know, one thing that does surprise me, and going back to this comment that Derek made, is just the steepness of the curve on AI. Right. Having lived through a number of these different waves over the years, you know, the Internet came and then cloud computing came and then mobile came. They tend to be much slower. You lose fact that at the enterprise level, it was really, this is the Gen AI movement started two years ago, and I would peg that at January 2023. It was January 2023 that ChatGPT came out and said, hey, we got to 100 million MAO faster than any other application out there. But it was a really telling story. I was still at Atlassian at that point. I was the Chief Product Officer. And in January, like that first earnings call, there was like one mention of AI, but if you looked at all of 2022, there wasn't a single of generative AI. So it went from the first earnings call by the second earning call, there were 18 mentions in the transcript. So like a fun activity is go grab Any company, right? And pull all of the transcripts from 2022 to now and load them up into your favorite LLM and just say, build me a chart of the number of mentions on AI or artificial intelligence. And what you see is, it's like, it goes like crickets to hey, something's happening here. To all of a sudden that is the conversation.

Yes.

So I would say that is. Is definitely one of the things that played out and that's one of the findings that we have in the report. If you look at 2023, that would be the year really of the pilot. So the CEO gets off the earnings call with the, goes over to the R and D group and says to the CTO and the CPO like, hey, what are we doing with generative AI? And very quickly what happens is teams get pulled together. It's like, how are we going to use this new technology? And that's really one of the findings from the survey, is that last year was a lot about experimentation. This year it's really about moving stuff into production. And that's where we get the. The 6x increase in the overall spend at the enterprise level coming up to close to 14 billion by, by our marks.

I want to talk about basically data sourcing for this stuff because I love this particular chart. But to me, when I see this, I go, how confident are you guys in these numbers? Because you could easily categorize things in different buckets. I presume there's some bleed between them. And also it's a growing industry. So can you just talk me through how this chart was put together? And if you're listening to the audio, this shows year over year changes in generative AI spend for foundation models, training, deployment data, vertical, departmental and horizontal.

Let me start with a high level and then I'm going to let Derek talk through the how we've actually calculated numbers in this chart. Really there's two big buckets that you should be focusing on. One is around the LLM and the infrastructure needed to bring AI into the organization. And then the second bucket, these four are actually three. The three bars on the right are really talking about the application layer. So what we can see is that two thirds of the spend that goes on at the enterprise 9.2-ish billion, is sitting in the LLMs in the infrastructure bucket. And for those that can't see, by far the largest spend inside of the enterprise is against the foundation models themselves. 6.5 billion of that is spent there. So two thirds is happening at the infrastructure layer. And then we have another third which is being spent at the application layer. The main thing about the application layer, the big story is there is that that's an 8x increase from the prior year. So that's sitting at 4.6 billion. Then what Alex is, is asking, one of the questions he's asking is like you can categorize your application layers a lot of different ways. In this case we've broken it down into vertical, departmental and horizontal AI. And certainly we can squabble about like what belongs into what bucket. But I would say the higher order story is that applications are coming, they're alive and well. And then for a variety of reasons we've chosen to break it out this.

Way in terms of data sources. We went out and asked 600 IT decision makers. So these are budget owners that have purview over their organization's generative AI spend. This is very lengthy survey, but basically how much are you spending on generative AI? What is that relative to your overall spend and then very specifically what tools are you spending it on? I think that one of the motivations behind the survey was that there's just a lack of good data out there of what are actually IT decision makers really kind of looking at what are the use cases that they have and where is their money going towards. And so this is, you know, we spoke to 600, we didn't go get to all, you know, the Fortune 2000. And so there's obviously, you know, this is directionally our answer, but I think that it's pretty informative to see some of the insights of like what folks are actually, you know, not just playing around with anymore, but actually adopting.

So just to summarize that the 4.6 billion that we see for the different types of AI applications you guys surveyed on, I should probably pay a little bit more attention to the year over year growth in that number than to sit here and go why is it 4.6 and not 4.7 billion?

Yeah, exactly. And then the other thing, you know, the next question you're going to ask, I was like okay, great. What are they doing at that application layer? Or you know, in. The thing that I would say is fascinating about that is that there's a real broad set of gen use cases in the enterprise. So sometimes you see things get very concentrated around one department or one use case. But I think it speaks to the usefulness of the technology. So when we break down, if on average an enterprise has 10 genius cases amongst it, which ones are most popular? Yeah, and that would be, that'd be an obvious question. So when we look at it, code Generation was number one. 51% of the folks on the survey said that they're using AI for code gen, they're building software for it.

Jelf, just to be clear, not 51% of the 600 who are paying for generative AI services, but of all the 651% companies are paying for code completion tools.

Yes.

Okay, Just want to make sure. So this is as big of a number of the 600 as we could imagine. Okay, I appreciate that. Keep going.

That's not surprising. We've seen GitHub Copilot is one of the, you know, fastest growing revenue products out there. And in the application layer of AI, they're north of 300 million in annual revenue. We've seen players like Cognition, Codeum, All Hands come about after CodeGen is support chatbots. Probably not surprising there, 31% are saying they're deploying chatbots. We've got products like Sierra, Decagon, Acera, which is from an ITSM IT service management use case is a, is a big one there.

Hey, startups, when you're a business, you gotta treat your customers right. Unreasonable hospitality is the standard today, but you're going to need tools. You're going to need a platform to help you do this. And that platform is the Zendesk Suite. The Zendesk Suite is going to give your startup all the tools you need to deliver exceptional customer experiences so you can build stronger relationships without growing your headcount. That's key, right? Every dollar matters. You gotta control headcount, you gotta control spend. So use the tool that Shopify, Squarespace, Uber and Instacart all rely on. It's called Zendesk. Let's take a look at another customer Unity. Very famous company. They saved $1.3 million with Zendesk automations and Self Service and they saw an 83% increase in their first response time. These companies love Zendesk because it's so easy to set up and it scales with you as you grow. They'll also give you all the metrics to make your reporting easy, keeping you and your business agile and investor ready. And that's because you're their esteemed customer. And they've created the Zendesk for Startups program just for you, where you get unlimited access to all the Zendesk products, expert insights, all the best practices and entry into their amazing community of founders, all at no cost for the first six months. That's right. They want to support you. Zendesk.com twist get ready to scale with the best in customer support with six months free, nothing to lose.

Then you move to Enterprise Search and Retrieval. 28%. You know, these products like a product like Glean has been around for a while, but with the emergence of gen AI, what they've been able to actually accomplish on the search front has been dramatically better than maybe what they were able to do in their early years. Another product in there would be something like Sana. Yeah, then we move to, you know, things like meaning summarization. You know, so many of us are remote workers and we're so used to seeing that AI agent has been added to either pull a transcript or summarization of our conversation. Things like Firefly and Otter. And then the last one I would mention to kind of round out the top five would be copywriting. And this goes back to the comment that you were making, Alex. LLMs are like calculators for words. They're like really good at authoring content. So you could certainly see a world in which everyone would be using or having a copywriter, an editor, grammar checker, just part of whatever they're communicating to help them be more precise and concise. Products like Writer Typeface, where we have an investment copy AI, they tend to be, you know, some of the leaders in that space. But that gives you a feel for like, okay, what is going on at the application layer of the enterprise.

So a lot of words is my read of it. I think the calculator for words is a great way to think about this. Code generation is creating characters on a screen, support chatbots create characters on a screen, you know, meeting summarization, characters on a screen, copywriting characters on a screen. The thing that surprised me is that workflow automation was so low in this chart to date because if you go back five years, we're thinking about UiPath and you know, RPA, robotic process automation. And there was a lot of enthusiasm pre generative AI, pre LLMs that RPA was going to remove a lot of the human drudgery from digital work. Then we got much better tooling. But now that we're, you know, looking back at 2024, I'm not seeing as much of that show up as I kind of expected. And this is a long way of saying how much progress have we actually made on agents? I suppose that can take some of the work from us and do it versus being more assistive in our day to day work. And Derek, I see you blinking at me, so go ahead.

Yeah, I think that we think about it in terms of waves. And so the first waves of gen AI apps were what we called rag apps or retrieval augmented gener. And so usually, you know, you, you have an external knowledge store and you, you use it for things like synthesis. So Eve is one of our portfolio companies. It's a legal copilot and what it does is it takes a lot of long, dense legal text and makes, it generates reports, it generates legal briefings, things like that, so that the lawyer on the other end doesn't have to go through the drudgery of all of that work. And so that's kind of the first generation. And then as you get more advanced, we move into agentic architectures which, you know, today our survey found that it is a minority, but if you ask a year ago, if you look at our previous report, it didn't exist a year ago. And this idea of you had like baby AGI and auto gen and some of these like, you know, open source projects back then, but it didn't exist as an enterprise idea of something that, you know, when I think about traditional RPA like UiPath, the idea of applying it to the enterprise hadn't really existed and now it does.

So Derek, it sounds like what I was doing was just being impatient and now it is showing up. And so what I expected to happen is I just had my timelines off mentally compared to the market.

I think the technology is now there. And if you look at things that we're really excited for in 2025, this is one of the things that we think will explode, is moving from retrieval based architectures to more agents and things that can automate workflows across horizontal areas. So like you think uipath but also verticals, you know, healthcare, there's solutions like Tenor that are doing kind of ingest automation and you know, a lot of different domain specific applications. I think you'll, you'll also see and.

I'm, I'm literally right now pulling up the OpenAI01 preview blog post because I forgot the term that I need here. It's like time series thinking. When models take a little more time before they make a decision or return a prompt. Is that underpinning the improvement in the technology that is making the agentic approach more feasible today? Derek?

Yeah, test time inference. And so I think that there's a couple layers to this, right? So you can have stuff like O1 which is kind of formalizing a design kind of pattern at the model layer. And so as the model gets smarter, it makes fewer errors. I think the Problem with agents that you had traditionally is that it's kind of run in a recursive loop. So if you think about it, you know, the agent will. The agent, which is the LLM, will think, okay, in order to accomplish this like task that requires 10 steps, what are those 10 steps? And then I'll go out and do it and then I'll think I'm at step one. Okay, check what is step two. And then if you think about error rates there, right, LLMs hallucinate that. That's now well established. If you have a 99% error rate or a 99% accuracy rate in step one, if you haven't that for all 10 steps, by the time you get to step 10, you know, the error rate is something that is unacceptable for enterprises. And so that's traditionally been the problem. And so LLMs getting smarter is part of the answer. But also when you apply it to specific domains, you have data scaffolding around it. We like to use the term agent on Rails, which is basically you need to hard code it or harness kind of the domain of like all the actions that the agent can take. You need to kind of set guardrails on it with hard code, with code in order to point it and get, you know, higher levels of accuracy so that that recursive error rate doesn't compound too much.

Do those guardrails have to be programmed on a per use case, per industry or per company basis? I'm trying to figure out how hard it is to hard code those because to me that could be incredibly complicated or relatively easy. I just don't know where it lands.

Yeah, I think people are trying to figure that out. Right. And I think that obviously the more specific to a particular use case, the higher accuracy and more robust it is. And. But the thing you're trading off really is degrees of freedom, which like, you know, all the way at this very end is AGI. You know, no guardrails, just the model just like put it in a for loop and it runs. And then on the other end is what we have today, which is computers that are 100% hard coded. Application logic is determined by the computer or by, you know, first of all, some programmer who sat down and was like, okay, I'm thinking in the shoes of the user, what do I need? And so I think the answer will probably be somewhere in between. We're trying to move towards AGI eventually, but I think right today we need stricter guardrails.

I think, you know, there's, as I'm hearing You talk, Derek. There's a, there's a good example in the software agentic space, right? Okay. So let's take code generation and we can say, well, how much progress are we making in code generation? And one way we can look at progress is we can look at this score called the SWE bench, right? And we can look at that over the last, what's happened there in the last 11 months. So Sui bench, for those that don't know, it's it's testing real world tasks faced by software developers. And the benchmark was based on things like pull requests and issues from open source GitHub repositories. And I think there's something like 2200 tests in there. So if we go back to January of this year, about 4% of those tests were completed by the best software agentic system out there. In March there was a company called Cognition which got a lot of traction. You might remember that by March it was 14%. Now if you look at it, the number one score on Swebench belongs to one of our portfolio companies. It's an open source software company called All Hands and they can solve 53% of those cases.

Hey founders, let's talk about a super weapon that your startup might be missing. Killer newsletters. Newsletters aren't just a great way for you to stay relevant. They're also an incredible growth engine for your startup. Social platforms and search traffic, those things are losing their edge. But your email list, that remains your direct line of communication to your customers. You don't have to pay a middleman, you don't have to get caught up and try to trick the algorithm into getting your message through. But in order to have a successful newsletter, you need a world class platform. And that's why we moved the this Week in Startups newsletter to Beehive. Beehive is a newsletter platform that's built to help you grow your community and turn your newsletter into an absolute powerhouse. It's packed with amazing features like an AI driven post builder. It's also got this really clever built in referral program and a top notch ad network if you want to turn on monetization. Plus, they have really easy to use tools if you want to charge access fees like subscriptions. So here's your call to action. Beehive plans start at just $39 a month. Super affordable. But we've got an exclusive deal for you here at this Week in startups. A 30 day free trial followed by 20% off your first three months. That's right, go to beehive.com/to claim that 20% off beehive.com/twist to start growing your newsletter today.

So if you kind of look over the 11 months, right, it's gone from 4% to 53%. So that's, that's like, you know, 13% improvement from the beginning of the year.

It's dramatic and I guess, you know, Joff, how confident are you personally, not speaking for any other company, that that rate of progress can kept up for the next 12 or 18 months? Because if you run the numbers out long enough, eventually we get Damn close to 100%.

Yeah, look, I think it's going to be invariant. It will depend on what department and what use case that you're moving against. Will it move from 53% to 100% next year on the software agentic? You know, I do think it attenuates out, you know, the stuff is fairly unoptimized right now. So I think even with the models that we have, let alone new advancements in the underlying foundation model, I think there's tons that can still be done from an accuracy perspective.

Okay, so thinking about the categories of enterprise gen AI spend that we started with, discussions about improvements to agentic AI and the approach thereof, and also improvements to existing leading categories, this all sounds very bullish on the technology side. My question is how much of the revenue that we're talking about is spread out amongst the startups? Because, you know, Jeff, you said that, you know, six and a half billion. Other 13.8 is foundation model spend. OpenAI, anthropic, a couple of names. When we get to the 4.6 billion from the app side, the number of companies that are nibbling at that spend is enormous. So is there enough enterprise generative AI spend for apps today to support the number of startups that are going after those dollars? Or are we like, is there half as much spend as we need? I just don't know if this is too much bread, not enough butter, I.

Guess, you know, it's a merchant. If we look at the software budgets at the enterprise for AI in 2023, they were going towards something called an innovation budget, which a lot of times that's like, hey, we just need to be investing in here. We don't have a permanent budget to pull from. So what I do see happen is you're going to pull from the permanent budgets in time as the ROI improves. Another point that I would make is that historically a lot of the budgets that were being spent by both the business unit and it were around tooling. And back to this comment that Derek had made earlier. Now we have services as software. So it's not just the tooling budget, but it's the human capital budget that we start to move into as well.

So once, can I, can I explain that to people? Because I think it's a good point, but I want to double click on it. So essentially, if you can reduce headcount and replace that with software spend, you can often get quite a lot of bang for your buck. And so budgets for software for AI that might replace some human activity can be relatively rich because humans are costly and need health insurance and travel stipends and office space.

And.

Yeah, and I don't, I don't think it's just about human reduction. Right. Like, I do believe that there's. Their companies are also faced with like a lack of having enough people. Like I don't have enough software engineers. So it allows me to, you know, expand and continue to grow that department, but I can do that in a different way. So it's not just about reducing, it's also about expansion. But, but that sure is part of the story. So I think there's plenty of wood to, to chop in there. When in the report itself, you'll see the AI spend by department, and you'll see departments like the legal department is historically not spending money on technology. They're much more of a late adopter, and here they are being more of an early adopter. So I think the budgets are very promising and I think they'll continue to grow as the ROI continues to prove itself out. So from an ROI perspective, last year it was less clear what the ROI was. This year, as we move from our pilots into production, there, there's greater clarity, but still some question marks, right? Like it's not all figured out, the number of folks that were in the survey, I want to say it was something like, you know, a third of the survey respondents are saying, well, we're still figuring out our exact implementation on AI strategy.

But Derek, picking up on that point, when we think about AI generative AI spend in the enterprise, we're essentially asking, what is the TAM for that today? And I was just talking about on the show the other day about how Uber, when it was very young people were comping it against the taxi market. Turns out that was BS because the market got much larger. It seems that if we can unlock spend from departments that didn't have technology budgets to begin with, like Legal, for example, the TAM for AI, generative AI in the enterprise is huge, but also it makes the overall TAM for software itself larger. And I don't think I actually thought that was going to be the case, but it does seem very bullish, if I'm understanding this correctly.

Yeah, I think that generative AI as far as adoption today is more expansive than replacement. And so you have innovation budgets last year. But it's actually really interesting. We asked where are you pulling budget from? And of the permanent budget types, a lot of it or over half of the permanent budget for generative AI is coming from new budget. So it is not, you know, I am either, you know, replacing spend for my system of record or some other software. It is, I am creating a new line item for this. And I think that you're seeing that across departments, which is, it is both expansive as well as across a lot of different places that used to not spend a lot on technology. And we're seeing, you know, from the startup side, which is where Josh and I spend a lot of our time, you know, we're seeing companies pop up all over the map, whether it's verticals. We mentioned health care, legal, financial services, or departments, Sales and marketing, data science, human resources. Like, it's, it's really all over the map, which is why we're quite excited about, you know, generative AI as a category.

Okay.

While I would love to just vamp about AI for another like three hours, I want to narrow this down to some startup focused stuff. So you guys had a great chart that's called selection criteria for Generative AI Tools. And to my surprise, the highest line item wasn't cost, instead it was roi. And so it seems that when people are approaching AI software, what they want is not to spend as little money as possible, but to have the biggest bang as possible. And so I'm curious, what should startups take away from this particular chart job as they approach the market so that way they can see the most successful.

Build things that are incredibly useful and valuable.

No one's ever said that before. There's no accelerator that says build things people want.

Yeah, you know, well, it's interesting and just sticking with, like, what would be good advice on the startup front. I do find that, you know, you want to look at what can you do now that you couldn't do yesterday? And that's really a thing to focus on. I remember when I was at LinkedIn, one of the things that I did was I started the mobile group. There was a lot of my peers were saying, hey, look, we got 2,000 pages on the website. Let's like jam it into this mobile device. And my point was like, no, that's exactly what we shouldn't do. Like the mobile device has something that is unique and it's special and therefore what we should be doing is focusing on highlighting those use cases so that it, you know, in time I believe that LinkedIn would become a mobile company and indeed it, it did wind up coming to become a mobile company. And the product that we built was quite different. There was only screens compared to the, the 2000 that existed on, on the website. So I, I think in this moment, you know, you have to look at a startup should be looking at like, well, what's different? And it's like, well, part of what's different is we have reasoning as a utility. So you know, if I look at, there's a, there's a guy, Daniel Rock, who's a professor at upenny. He, I've had a number of conversations with him and he was telling me, he's like, look, think of every role in the economy as a bundle of tasks. And those tasks, some of them fall to AI and some of them fall to humans. So I do think what startups should be looking at is like, how can I change the way, like if you look at workflow, the way that it's been built over the last 10 years, it's been built with humans at the center and that's the actual workflow. So if you were to zoom all the way out and say, oh, okay, well I'm going to have a workflow that's going to be shared, some of it's going to be through an algorithm and some of it's going to be through a human. What would that allow me to do? Because a lot of these workflows that we see at the enterprise level are very complicated and very tool rich. You know, if you look at an average sales team, they have like 12 different tools hanging off of Salesforce or HubSpot. If you look at software development, there's over 12 different tools that are hanging off of the software development flow. So you know, you have an opportunity to reimagine what that workflow should be. So you're, you're either going to enter in and say, hey, I'm just going to try to AI ify the existing workflow or I'm going to think about it at a more first principles level and think about what could be possible given what we know now. And that would probably be my number one, my number one suggestion is really get back to the first principles of it.

Okay, so I want to pick up on that because you guys wrote last year incumbents dominated the enterprise market with bolt on strategies that layered gen AI capabilities onto existing products. And jaff, it sounds much more like you're saying, look, don't do that. Rip the page out, start blank and build from the ground up. So Derek, I think we call this an AI first approach just to building software. What fraction of software that exists today needs to be ripped out and started over again? Because people joke like, I don't know what Salesforce does. We all kind of do, but not really. And you know, not to make fun of Atlassian, but Jira as hell. So I'm curious, Derek, you know how much I will not, sir, I have filed tickets and I. You owe me lunch for the pain you put me through. Don't worry, people, the people who make concur owe me a house. So it's fine. I'm just curious Derek, like how much of software needs to be rethought in this way?

Yeah, I don't know that I have a very clean answer for, you know, x percent of software needs to be completely rethought. But it's very across the board. I think that one of the things that we realized this year versus last year is that it's actually not that easy to build AI that works. You know, a lot of people thought last year like, oh, if I'm a Salesforce, if I'm an Adobe, I can just tack on something AI on top of my existing system of record and that will be my.

Because you have the data already. Because if you are a system of record, you have the bucket of data to make your own tuned models with.

The data, the distribution, the trust with customers, all of that. Like, and that is why a lot of people thought like, is AI really a net new category or is it just a feature on top of existing software? And I think what we've realized is actually it is its own independent category. And why it's really exciting for us is because that gives the advantage to startups rather than, you know, Salesforce coming out with Agent Force and everybody just being like, this is the greatest thing on earth and so everybody will adopt it. There is opportunity for startups because people, enterprises we talk to try out features like Agent Force and realize, wait, this is not what was promised. And so there's opportunities through domain specific workflows, you know, AI native approach to actually make that promise work.

But if we think about the companies that exist today that might struggle to move from you know, their kind of legacy software offerings that make them all their money to an AI future. That means that there is trillions of dollars in market cap out there for startups that are being born now with the blank sheet of paper and the modern tools that we're discussing to go after. So in a sense, I kind of like think we should short the NASDAQ and like double our investment in vc. That would be the best way to get kind of both sides of the bet.

You know, it's different domain by domain. Right. Like it's hard to predict and that's what makes investing fun. Like Adobe, for example, they as an on prem company and then when the cloud revolution came, they actually took a tremendous bet. It's actually really interesting if you look at their quarterly revenue, when they decided to go from on prem to subscription, they were able to convert that and you know, you could. Adobe is one of the companies that I personally admire a lot. Will I bet for them or against them in this gen AI revolution? I personally think Adobe is, is a really interesting company. Firefly is a great product, but not all companies will do that. Right. And so when you look at the startups, both the startups tackling it and what do they offer, why are they different from the incumbent as well as who is the incumbent in that solution? Are they well positioned to move with the currents versus against them? You really have to take it domain by domain.

And it is like, it's a very visceral thing when you see a company that gets caught up and left behind in that. Right. You can look at a company like Chegg, 85% of their market cap is disappeared with AI or Stack Overflow, a site that I used to use quite a bit like their traffic is halved as people go directly into the LLM to get their, their coding guidance.

Yeah, I wonder how many of those we're going to have by this time next year. Because Chegg made money off helping people cheat on homework and they're going to get mad at me about that. So PR team, please don't email me again. That's what people are using Chegg for. We all know it now. People use OpenAI to cheat. It's great. I wish I had when I was in middle school. What a crushed chemistry. I did not. Because I didn't. But I wonder how many other companies are kind of on that list and if that will be a good barometer for how fast AI first companies can kind of uproot legacy companies that are, I guess, Cloud and SaaS are now legacy and kind of, they almost feel outmoded. Like, do you guys remember when Salesforce invented SaaS and we were like, oh my God, this is the future. Yeah, it's weird now to be sitting here going, oh my God, is that the past?

You know, it's, it's interesting because I, I feel like these are all things that were building blocks that were needed to get us to where we are today. We actually needed the Internet to get the world's information digitized. We needed to have cloud computing in order to unleash vast, vast amounts of, of GPUs and CPUs to do training and inference and things like that. So it's, it feels like it was more of a lineage building up to the moment that we are today.

Yeah, but how much credit do we give Yahoo for inventing the Internet portal? None. Today. We don't even think about it.

Yeah, well, what's that saying or Xerox.

In the gui, right?

I mean, it's like you're big, you know, you're. My company was built on the shoulders of giants, or AI is built on the shoulders of giants. I mean, there's a lot of things that needed to happen to get us to the state that we're in today.

Yeah, I think there's building blocks and there's value capture. And I think that a lot of the new value to be captured will be by startups and new companies, but obviously building on the shoulder giants.

So glad we ended up here because I have a question about this. I know we're talking about how startups can disrupt incumbents, but there are some incumbent AI companies, and I'm thinking about very clearly OpenAI and anthropic being two of the largest players in this kind of neo space. And OpenAI recently put out a search product which is frankly pretty good. I use it on a regular basis as a testing tool. And I also think that Perplexity will see some of its momentum cut because there's now a competing product from a larger, better finance company. And so when you guys are talking to startup founders, how do you help them navigate building something that won't get stomped on by a foundation model company releasing something that could be considered to be competing?

I think one of the things to think about is what is on their near term roadmap. Right. And for a company like OpenAI and Anthropic, their ultimate goal is to achieve AGI, for Anthropic to achieve AGI safely and what are the things that are going to be, you know the things that they'll tackle next. And so Anthropic has already mentioned, you know, we won't pursue, you know, side quests such as image generation or things like that because it is not necessarily the thing that will bring them to AGI. Better reasoning on the other hand, or using tools like web browsers and computer use on other computer use. Yeah, those are right. And so a lot of the, you know, when you are an AI app today, there are two things that you really need to get right. One is can I make the base technology work? And a lot of the things that actually have happened to date are just how do I make the LLM reliable? How do I connect it into enterprise systems? How do I give it a tool such as a web browser? And how do I, you know, when I was talking about the agents before, how do I make sure that the reasoning is reliable? But you can think about these as data scaffolding things that you know, actually are needed today to help act as crutches for the LLM to actually work in an enterprise application. Those I think will one by one fall away over time as Anthropic makes advancements in the intelligence of the base LLM. But what won't change and what Anthropic won't get to, because quite frankly it's not important for AGI is how do I apply this to my, you know, health care domain specific workflow If I need to, you know, do like clinical documentation integrity, if I need to do RCM on the back end revenue cycle management, Anthropic is not really interested in that. And so that's an opportunity for Apple or startups.

Okay, but here's my question because all of that tracks with me. Derek, I agree with you wholeheartedly. But as we get closer to AGI and as we actually maybe reach it in the near ish future, doesn't that obviate a lot of the work that's being done to make AI apply to specific categories or market verticals? Because as the AI brain gets smarter, it probably needs less help to do more. And so I wonder if we'll see a dilution of the power of going vertical in AI as we get closer or reach AGI.

I think the thing that maybe that discounts a little bit or takes for granted a little bit is the training needed to actually make the solution work. Let's say that you have a super intelligent PhD level human. If you take that person and apply them to, let's say the healthcare example of how do you do RCM and work with payers to make sure that your claims get paid. You still have to teach them how to do it. You still have to. You know, there's a learning curve as they get used to the workflows there. And, you know, obviously if you take that person versus somebody, you know, who may be less intelligent, there's a shorter learning curve there, but there's still a learning curve there. And how you actually get that intelligence and apply it to the specific workflow on the other end. That is the area where application layer companies can add value.

Okay, but you're making a jump there that I wouldn't make, which is you're saying that if you took a PhD level AGI, my hope is that by the time we reach AGI, we're no longer using postgraduate benchmarks to determine intelligence or expertise. We're so far above that that those, those analogies don't hold. And then we won't have to have so much verticalization, guardrails built and so forth. Because in theory, this should be able to be a bit like a magic box.

I'm hoping there's a question too, if there's like, well, one algo to rule them all, or if that algo is a collection of millions and millions of algos, Right. You know, think about my phone. Like, the power of my phone is actually sitting in the ecosystem. It's that there's an app for that. So there's millions and millions of apps that make that product more productive. So, you know, one version of the world is like, I got an algo and it does everything. Another version might be that I have an orchestration algo that knows when to ask other algos, like for their expertise in a given area or science. So part of that is still unknown as we move into the future.

Yeah, I think we joke about how everything is bundling and unbundling, but I kind of wonder if when we get to AGI, if we're going to have AGIs orchestrating AGIs in a bundle or kind of a singular brand, mother of all bundles. Right? Kind of like if you got, you know, Netflix and ESPN the same thing. Can you imagine? That would be crazy.

That would be nuts, Alex.

I know, I know. So to summarize, though, just kind of taking all this in one, in one bucket enterprise generative AI spend grew very quickly and perhaps even faster than expected. We are seeing increased spend on foundation models, but also on the application side. And you guys, in the Menlo perspective is that the upstarts are going to knock off More incumbents and probably we're going to see more agentic AI progress next year. That's going to be very exciting. Is there anything else at the highest level from this report that I have not brought up? Because I want to make sure we get all the, all the key bits of meat out, if that makes sense. Jelf?

Yeah, I think, you know, there's one point that I found really fascinating that we didn't punch into too much is that when we went into the enterprise and we said, well, what is. And this is at the infrastructure layer. And we said, well, what is the LLM that you're using? What they came back and said is that they're leveraging multimodal. So I thought that they would try to get behind a single model, whether it's for hedging or for performance or cost reasons. They're typically on average using three different models, which I thought was really fascinating. And then the other thing is just the shift in the enterprise around who is using what from a foundation model perspective. OpenAI clearly had the first mover advantage in the enterprise, but what we saw on the survey is that they moved from 50% down to 34%. So they lost 16 points year over year. You can see that anthropic doubling coming up. Closed source models are clearly, you know, the majority of what's being used. Meta was flat. Mistral, you know, off a little bit.

But you know, poor Europe, look at that. Look at their one AI champion, like fourth in line. We gotta, maybe everyone should spend 10% of their AI spend on Mistral just to help France because, you know, new administration, Ukraine. Okay, just me. All right, keep going. Jeff.

We'Re investors over here.

I love how uncomfortable he got right there. No, don't touch.

Exactly. It's like, stand clear, stand down. You're. This is where like that PR training comes in play. It's like, wait, is Alex baiting us? Is this going to be the one quote that comes out after all this, like wonderful, great conversation. You know, Alex and team are going to pick on that one time that we blinked.

Joff Redfern just endorsed all Trump administration policies. I'm kidding, I'm kidding, I'm kidding.

Exactly.

But I am, I think just very net optimistic. And that's a good place to end a year, I think, a year of so much change. It does all feel to me very exciting, like we are really making progress. And you know, the computer use stuff from Anthropic not to give your portfolio company extra ups, but like when that dropped, I was as excited about that as I was about ChatGPT the first time I used it, you know, and that that's a great place to be going into a new year of a lot of investment in progress. So it's all very encouraging. Now just before I let you guys go, I do have one lightning roundish question. So sorry for this but I, I can't help myself. So Derek, we'll start with you. What percentage of your net worth is in crypto or bitcoin?

I am actually outside of venture a very unsophisticated investor. So on fortunately.

So a little bit via ETFs.

Sounds like all index funds, you know, bet on the US economy, bro.

You should see my family's portfolio. Exactly the same. And now 0.1% Bitcoin via Fidelity ETFs. Joff, over to you. Same Q. Curious how crypto exposed you are going into 25 00. Yeah, this was supposed to be the freebie, not the, not the. Alex does pick out the quote at the very end and make that into the headline. I mean Jesus.

Yeah, no, I, you know I, I focus on things where I think I have like better understanding or a competitive advantage. So similar to Derek, it's like passive on the economy and then I am way over indexed on, on. On venture investing through a variety of different funds and angel investing. So it's like yeah, if you, and if you look at that it's like why is he doing that? It's because I feel like I have a. Like that's all I've done. I built software. My like startups. That's like, that's the thing I know super well. It has no. And it's like not a reflection at all on crypto or my beliefs there. It's mostly just about what I mean like I have so much interest in startup land.

I appreciate the honesty there because I feel like there's always a pressure whenever crypto does crypto things to appear sophisticated about it for 18 months until it goes away for two years again. Yeah but it has been quite loud and the, the, the crypto fans have been making noise. But anyways guys, thank you so much and we'll come back and do this again next year for the 2025 generative AI and the enterprise report because I'm sure by then we'll have even bigger numbers to dig through. But before you go, drop your Twitter handles and then we'll say goodbye.

Yeah, I might use my LinkedIn handle. Go to. Go to LinkedIn, man. That's where I'm posting. That's my activity. That's the thing I built, so I stick with it.

Fair enough. Derek, what is your preferred social media handle? I apologize.

I am on both, but on Twitter, I'm at Derek Gshiao.

All right, thank you guys, both very much. And Twist is back next week for a couple of new shows, and then it's Thanksgiving time, y'.

All.

We'll see you then. Bye.

Thanks, talent. Super fun.