E452: Why 95% of AI Projects Fail, Anthropic’s Valuation, and Why AI Is Still in the First Inning

What if the biggest opportunity in AI isn't building better models, but unlocking the 80% of data those models can't access? In this episode, I sit down with Joseph Coyne, Co-Founder and Managing Partner of Studio VC, to explore why he believes enterprise AI is still in its first inning. Drawing on nearly two decades in venture capital, including experience at Bain Capital Ventures and Samsung, Joseph explains why enterprise AI implementations struggle, how proprietary data creates lasting competitive advantages, and where the next generation of AI businesses will emerge.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00David Weisburd?

  So I did invest in Anthropic at a $4 billion valuation. How would you underwrite it today?

• 0:05Speaker 01

  Yeah, but look at the implied forward capture of the market. So how big is the market likely to be? And then what percentage are you expecting to capture as Anthropic vis a vis or relative to OpenAI and some unknown entrants? So I think that's hard to handicap on the who's coming next, right? Conversation.

• 0:24David Weisburd?

  It's unknown unknown. Who's going to have the idea or the resources?

• 0:27Speaker 01

  That's the unknown unknown. But you have other folks that are devoting a lot of resources to the space that have billions of dollars in reasons to be highly defensive. The Googles, the Metas, the Amazons, et cetera.

• 0:39David Weisburd?

  I'll tell you the bull case. Virtue of Lucas Swisher. He runs Growth for a Co2 company. Talked about the three S-curves of adoption. So the first S-curve, the early adopters in terms of LLMs from a revenue standpoint have been developers. What's crazy about that is that there's only 50 million developers in the world out of 8 billion people. And on top of that, only a small percentage of developers have actually implemented. If you look at worldwide kind of engineers and you have all of enterprises, then you have consumers. And then you could argue a fourth S-curve, which is agents. So some smart people would argue that your cam is not actually 8 billion. It's 8 billion people and their agents talking to each other. So in terms of the adoption curve, we started this conversation talking about

• 1:23David Weisburd?

  we're in the first inning of AI. It's reported that pre-seed rounds today averaged $100 million valuation. How do you underwrite?

• 1:33Speaker 01

  I think.

• 1:37David Weisburd?

  So Joe, you've been in and around venture for nearly two decades with firms like Bain Capital Ventures, Samsung. And you believe that AI is in its first inning. Why?

• 1:49Speaker 01

  Well, I think specifically enterprise AI is really just starting to scratch the surface of what it can do within enterprise. What we've seen at large enterprises, i.e. the customers of our portfolio companies, is that they're either scared to adopt or slow to adopt or have high failure rates when they adopt AI. So I think because of all we're seeing from our customer base through the portfolio lens, I think there's so much more to come. And I'm excited to see what it can do.

• 2:14David Weisburd?

  Which one of those is the real issue? Being scared to adopt or being slow to adopt or just being poor at adopting today?

• 2:21Speaker 01

  It's a new muscle for a lot of enterprises. I mean, they've gone through data transformation over the many years. They've built sort of database architecture. And now as they start to use that foundational data to adopt AI best practices, it's all new for them. And so they either need external help. It could be a startup, could be a growth company, could be a consultant. But internally, their resources are generally pretty poor right now from what we've seen. 95% of enterprise AI deployments are failing these days. And I think that's per Deloitte. So that only has a lot of room to scale, a lot of room to grow. And obviously, if you've done something a few times, as in deployed AI across different verticals or different data sets inside of an enterprise, you only get better over time.

• 3:03David Weisburd?

  That number you said from Deloitte, 95% of pilots fail. I haven't heard that number, but I experienced it firsthand. In your seat, you're seeing these hundreds of enterprises implementing AI over and over every day. What are the best practices in terms of the 5% that do succeed?

• 3:19Speaker 01

  If they come from a deeply technical background that's already been doing deep data analysis and has great data infrastructure plus security, then they have the foundation of success, right? So that's like already tipping the scale in the favor of tech companies that are pretty well poised to do that. Some of the legacy companies, legacy industries that are lower tech, we've seen implementations where, once we get under the covers, the data is pretty unruly. And so getting the data to a point where, from an infrastructure standpoint, it's ready to adopt AI is half the battle. Now, I think once we cross that chasm, a lot of these AI best practices can start to proliferate and accelerate.

• 3:57David Weisburd?

  What does it mean to have good data versus bad data when it comes to implementing AI on an enterprise level?

• 4:02Speaker 01

  You see in large enterprises, right? They're, the larger the enterprise, the more complex, generally speaking. And they have lots of data silos. That means data tagging fidelity across different fields and tags can be different. And so when you try to run enterprise-grade AI on top of it, you have all sorts of sometimes conflicting tags and data structures, which have to be resolved before you can really do it in a meaningful way. And so those are some of the data transformation projects that are foundational to then running great AI and analytics on top of enterprise.

• 4:32David Weisburd?

  Something that I hear over and over about AI implementation is that it's mostly a cultural issue. Yeah, sure, there's data, there's processes, but it starts at the culture level. How much does culture play a role in terms of enterprises being able to implement AI within the organization?

• 4:49Speaker 01

  That's right. And it starts from the top within organizations, right? If senior leadership is embracing it, it has to be, in a lot of cases, top-down to reach sort of the most mission-critical parts of the enterprise. And so, you know, some leaders, especially within tech companies, are thinking about this on a daily basis. And I can tell you for a fact that some companies leading huge banks, huge consumer companies, huge airlines are not native to this. And so there's a lot of education and foundational work that needs to happen for them to really reap the benefits of AI's power and possibility.

• 5:21David Weisburd?

  This episode is presented by Juniper Square, the operations partner for private markets. So if I'm the CEO of Delta, let's say I just put my foot in my mouth with a Starlink and saying that I don't want Starlink and don't like SpaceX, all these things. But let's say I want to rehash my image and I want to show that I'm leaning into technology and leaning into AI. Is that the problem set that you would go into, which is elastic pricing for seats? What would you advise the Delta CEO to do? What's the next steps?

• 5:53Speaker 01

  It's looking at a lot of historical data to understand how they've approached it in the past, but then also having a horizontal view of pricing within the industry. One of the ways you can get a data moat, which is very important these days, is to have horizontal insights about an industry. Meaning you see hundreds of airlines. The airline you're talking to probably sees mostly its own data most of the time. And so when you bring horizontal insights around best practices and price elasticity and price per mile, what type of seat across different organizations, you can then bring that to bear within large, high willingness to pay customers. And so I think it's A, looking at their own historical data, but B, gleaning a lot of model insights

• 6:34Speaker 01

  across thousands of other potential opportunities.

• 6:39David Weisburd?

  This connects to what we were talking to a couple of weeks ago, which is you said that the vast majority of data is not accessible to LLMs. Why is that?

• 6:47Speaker 01

  One of our favorite stats, and it actually came through an enterprise that we work very closely with, which is that about 80% or more of the world's data is behind firewalls. LLMs have done an absolutely heroic job, right? Going from infantile phase a few years ago to in the pockets of almost every American and around the world. So would never speak ill of what they've accomplished, but they've accomplished it with the publicly available data, similar to how Google crawls the public web. To get a lens into private data, you have to earn customer trust. You have to bring unique insights. And oftentimes, you need a subject matter expert to be able to have the trust and expertise to enter a market like airlines or banking or insurance. When we see companies that are starting to unlock those types of value,

• 7:32Speaker 01

  they get access in a highly secure way to proprietary, behind-the-firewall data, the 80%. When they do that, they can start to produce new insights for the company and for the industry they operate in. And that's where really high-value opportunities exist. Years ago, we were just pure-play enterprise software investors. Now, we always see those companies come with an AI strategy. The best of those companies come with an enterprise software workflow that systematically helps companies to unlock their data and then run AI best practices on it in a highly secure way.

• 8:05David Weisburd?

  You made a very important but subtle point, which is the reason 80% of the data is behind the enterprise firewalls is because it comes from the customer. And in order to get that customer data, you must provide a good product. So everything is still downstream of having great product. The difference in AI world today is that your data moat starts to be one of the most sustainable forms of competitive advantage. So again, 95% of enterprises, according to Deloitte, fail in terms of their pilots. The ones that are succeeding, what are some early wins that you're seeing in enterprise that are maybe not as obvious to the outside?

• 8:39Speaker 01

  Enterprise software that is creating workflows that unlock new data feels like the oil rig company pulling new oil out of the ground for enterprise. And what we're seeing.

• 8:51David Weisburd?

  Give me an example.

• 8:52Speaker 01

  From inside the enterprise. So one of my favorite ones to talk about is Inspiron, which is a portfolio company, just closed $170 million from Insight and NewView Ventures. And they are in the healthcare space, which is a space that generally speaking, if you asked me 10 years ago, Joe, are you going to invest in the healthcare space? I probably would have said no. But their business model, their workflow, the data that they extract is perfectly aligned with our thesis. And the reason for that is that they have a very low-cost device that they place across senior living facilities. And it uses computer vision and other sensors to extract real-time data 24/7 about residents. That data set prior to this device was primarily anecdotal. So it's what is the nurse saying? What is the doctor saying? What have we logged, right?

• 9:37Speaker 01

  This is a human problem. Humans are not great at this organizational level of detail. Nor can they go as granular as 24/7. So what is 24/7 monitoring unlock? Well, first of all, you can do it in a very anonymized, safe way. So that has to be table stakes, right? And then the second thing is you can unlock insights about patients, doctors, nurses, staff members over time. And so that helps you do things like prevent falls. If you're monitoring patients 24/7 and you see somebody that you has a known fall risk, right, that can be built into the AI, you can monitor them in the middle of the night. If they stand up or they look like they're about to stand up, you can send a staff member minutes before they otherwise would have arrived to help that

• 10:21Speaker 01

  person with whatever they need in the night. Could be a question, could be a run to the restroom. But either way, if that person is standing up under medication in the middle of the night and they're 85 years old, the negative outcome is a fall, a broken hip, a lawsuit. That's a tens of millions of dollars problem eliminated. So Inspiron, one of the first use cases was both in hospital and senior living, was they were able to show they could eliminate 60 to 70% of falls that would happen oftentimes in the night. The second thing is you have longitudinal data now. So you're watching seniors or occupants over time and how their activity changes. What that can do is actually do predictive analytics and start to predict things like

• 11:02Speaker 01

  dementia and Alzheimer's based on activity levels, behavior levels and things like that. We can also predict if somebody starts exhibiting certain behaviors, things like infection, which is incredible. And lastly, we're now creating a sort of time series global view of an entire operation. So, you know, senior living facility may have hundreds of rooms, you know, dozens of staff members, dozens of doctors and nurses. Now they get to see in real time who's doing what and where. So where are the resources being deployed?

• 11:33David Weisburd?

  Logistics essentially.

• 11:35Speaker 01

  Right. And so that was anecdotal before. What, and so, you know, you would ask a staff member how their day was and they'd say, oh, it's just, I was so busy, way understaffed, right? Classic response. Maybe they noted it in their logbook, maybe they put it in EHR, maybe not. Now we can actually say, wow, Joe, the staff member was supposed to visit 10 rooms today, but he got pulled into 25 rooms. That's a major problem. By the way, Joe, as a staff member is likely to leave if we're working two and a half times harder.

• 12:03David Weisburd?

  You could predict burnout.

• 12:03Speaker 01

  Yeah. And so we're also eliminating burnout among staff because we can load balance according to the needs of the patients. And lastly, you know, over time, patients and occupants start to exhibit signs of additional needs, right? Like they tend to degrade, especially in senior living. And so you want to appropriately staff up based on the patient's needs. So if my father went into senior living tomorrow, he might need one visit a day, but in two years, he might need 10 visits a day. That is a different, different value prop. That's a different staffing mechanism. And my father might need to be moved to a different floor to achieve that, where he can get the level of care that he actually needs. So it results in better staff satisfaction, better patient outcomes, better profitability

• 12:44Speaker 01

  at the end of the day, and an all new data set that was previously anecdotal at best is now digitized moment by moment across the entire facility. So that's, that's oil, as I said, being pulled out of the ground. It was sitting there. It's always been there. But we've, Inspiron has pulled that data out of the ground and created a very sophisticated model to help multiple decisions, whether it's tactical decisions like preventing falls or operational decisions like staffing decisions and care level decisions.

• 13:13David Weisburd?

  Spiron is a fascinating case study. It's kind of like Minority Report, but it's actually helping the patients. And also in many ways, it's this gateway drug to physical AI. Jensen Wang predicts that physical AI would be a $50 trillion market. A lot of the top VCs I know right now are trying to invest into physical AI because they see it as a new frontier once kind of all the software and all the data based AI gets integrated. And most of that application is actually building from scratch. So humanoids or something, sensors and factories, things like that. But here it's actually going into existing systems and getting existing insights from systems versus kind of building these brand new technologies from scratch.

• 13:58Speaker 01

  Yeah. And all the new technologies you just mentioned live and exist in the future based on ingesting the data we're capturing now. So imagine that instead of having a staff member, you had a robotic assistant. Well, how is that robotic assistant going to learn over time? Well, wouldn't it be great if you had a 24/7 data set of video capture for thousands of patients over thousands of days? That sounds like the perfect training desk data set for that type of use case. And so obviously the impact right now at Inspiron is tangible, but the data set they're creating is, hasn't even been monetized. And that's one of the things that's exciting for us.

• 14:36David Weisburd?

  I imagine they're going to be monetizing not per seat or not per device, but if you are, I think you mentioned 60 to 75% of falls are prevented with, with the device. If you're able to keep these large, you know, multi hundred thousand dollar lawsuits from happening over and over, that's extremely profitable to the healthcare centers. And if you monetize off of the losses that you're preventing, that, that could be quite a sizable business.

• 14:59Speaker 01

  Yeah, today we actually do have kind of a SaaS pricing model. So like a per room deployment pricing model. But you're right, I think in the selling process, that is one of the most tangible aspects to get in the door in the first place and then to win a pilot. But what keeps us there over the long run is that the staff, the administrators, the owners all become addicted to this new data. Once you've pulled that new oil out of the ground, putting it back would be pretty difficult, right? You've already realized these staff retention benefits, these profitability benefits and these patient improvement benefits in addition to those, those first use cases, which is, hey, if we could prevent falls, would that be helpful, right? That gets us in the door. What keeps people is that, that long-term reliance on this new data set.

• 15:45David Weisburd?

  Fascinating example of physical AI. In terms of physical AI, maybe you could double click on additional use cases you're seeing today in your portfolio.

• 15:53Speaker 01

  There's certainly some things to be fearful of in AI. But when you see tangible impacts helping people's lives, it starts to, it's the other side of the coin, right? It's like, you can't get all the positives without the negatives. The negatives can be mitigated with cybersecurity and hardening and data protection and things like that. Obviously personal, personal protections as well. But when you see the positive impact in something like healthcare or a company in our first portfolio called Bark, which protects kids online using natural language processing. So we're ingesting about seven million kids data, real time across multiple apps and across the OSs to establish emotional baselining for those kids.

• 16:35Speaker 01

  We can then predict with very high accuracy when somebody's experiencing depressive episodes or specifically has flagged, even if they're using tons of slang, you know, self harm, harm to others, or if a child predator has entered the chat. And so the impact coming out of that company, as somebody who has three kids especially, gives me shivers every time I open it, right? The, the VC muscle memory tells you to open up a company, a company update and say, what was the revenue? What was the growth? But in this case, I go straight to the impact and that's what gives me shivers. It's dozens of school shootings, hundreds of suicides prevented, child predators escalated at the FBI, dozens every year.

• 17:15Speaker 01

  And so doing something like protecting kids with a complex natural language processing engine plus predictive analytics and AI is just another use case where I think you can get behind AI if you see use cases like that. But again, there's plenty to be fearful of, and I think there's plenty of companies being built to protect against some of the, the dark side of AI. And that's a thesis that we're leaning into going forward. And then on your physical AI question, I would cite a company called Evercoast, which does 4D, meaning in motion capture of primarily human beings. And they use a camera array, which is pretty much off the shelf cameras with very sophisticated

• 17:58Speaker 01

  processing capabilities in the background to reconstruct human beings or objects in motion. So effectively video in 4D. And those, the early use cases for that business were primarily in the media space. So we worked with the Olympics and captured every Olympian in Paris, for example, doing their specialty with a 4D camera array around them. That created digital humanoid reproducible assets from those gymnasts or from those, from those gymnasts and other athletes that NBC could use in other media generatively. And so that was the initial use case. We've done that now tens of thousands of times with hundreds of customers, including

• 18:39Speaker 01

  Apple, Netflix and others, and have a really nice business doing that capture, processing, asset management and dovetailing it with the production workflows.

• 18:51Speaker 01

  That business is interesting, right? They're already making millions of dollars in revenue. What's really interesting is they have the world's largest data set of human 4D captured assets. And why does that matter? Exactly, we said before, robotics training, right? Generative humans for robotics training, real world applications. And so to do that at scale, you could imagine a data set like Evercoast becoming a pipeline to create generative human beings at scale to further train robots, self-driving cars, you know, the digital nursing example we gave earlier, home care, et cetera. And so that's sort of step two of unlocking the data set there. And, you know, notably, that data set is very valuable to a lot of folks.

• 19:36Speaker 01

  And so we have to be very careful about like, who do we partner with? Do we sell the data? Do we build our own internal models? And so that's a conversation we have almost weekly with the founder there.

• 19:44David Weisburd?

  One thing I've learned from talking to hundreds of investors is that great investment firms aren't built on investment returns alone. The firms under are great at the things most people don't see, their operations, their relationship with LPs, and the quality of information they use to make decisions. And here's what AI has changed. Every firm now has access to the very same models, so the intelligence isn't the edge anymore. The edge is what you could feed it. A firm with its fund operations and data in one connected record can actually put AI to work. A firm running on disconnected systems simply can't. That's why thousands of GPs run their funds on Juniper Square. Juniper Square puts your fund operations, data, and administration together in one connected record.

• 20:25David Weisburd?

  That means less time managing disconnected systems and more time investing, working with LPs, and building your firm. This episode is brought to you by Juniper Square, the operations partner for private market GPs. Learn more at junipersquare.com/HowIInvest. That's junipersquare.com/HowIInvest. A lot of people have trouble thinking second order effects or third order effects, but as the LMs continue to get better and the inference continues to be more and more sophisticated, the moat will continue to be the data.

• 20:57Speaker 01

  Right.

• 20:57David Weisburd?

  Just think about Tesla, for example, with their self-driving car and their self-driving technology. I think it took a decade for them to gather the information, just driving on the streets of Austin, San Francisco. And that was the bottleneck, because the 10s of years, 10 years. And then obviously they had to build the models, which was non-trivial at the time. This was pre-chat GPT. Now I'm sure they're using a lot more LMs, but in a post-LM world where you have all the models, where you have basically genius level physics PhDs, you know, 10,000 of them that could work on your problem set, the data, the moat actually becomes the data.

• 21:33Speaker 01

  We hear this word a lot in our kind of investment committee, which is, you know, either flywheel, which is people use a lot, but we also talk about like pro-cyclical advantage, meaning once you're Tesla and you get a lot of data, then your model becomes better and you get more users and then you get even more data.

• 21:51David Weisburd?

  That goes back to the customers being the source of the data.

• 21:53Speaker 01

  Right. You have to prime the pump. Sometimes that's costly. In Tesla's case, it was building a car that had autonomous driving features years ago. Now it's a pro-cyclical event where every incremental user often chooses Tesla because of its self-driving, and so they further cement their data lead. So that's a pro-cyclical we talk about a lot, where that becomes part of your defensibility over time. A new entrant, starting at zero or even even above zero, will still be so far behind a Tesla or an Evercoast in this example to catch up. And so incremental users choose those companies again and again and further deepen their data moat.

• 22:32David Weisburd?

  I want to go back to this question of business models. I don't think there's been a technological revolution that has not led to significantly new business models, but yet people are continuing to charge in this kind of SaaS, SaaS model. What do you see in terms of business models and how they evolve in a post-AI world?

• 22:49Speaker 01

  Yeah, I think you have to meet your customer where they're comfortable, especially when you're a startup. So will SaaS modeling be the traditional or most frequent use case in the future? Possibly, possibly not. But right now, that's where large enterprises have been trained to purchase and been, you know, when they think about unit economics, going back to the Inspire example, you know, how much is our customer paying us per month? How much does this cost us per month? What's our cost savings? What's our cost savings per user? Unit economics leads you to think about it in terms of a cost per unit. Oftentimes, that rolls up to a SaaS pricing model. Now, in more avant-garde tech, you might see usage-based models, which we're already seeing a lot of. You may see even outcomes-based models, right?

• 23:33Speaker 01

  Where if we can produce this for you, large enterprise, you'll give us this, right? Those are a little bit harder sells. I mean, I think that they're very aligned with outcomes-driven and ROI-driven implementations. So don't get me wrong, I kind of love that, but the buying behavior is new. And so there's some education and some comfort level, particularly with large enterprises, some of whom have been around for 100 years, right? And certainly over the last 20 years, they've been habituated to to price based on SaaS per seat or per month. So getting them out of that will take some time. I don't think it's impossible, but I think we'll see a slow ramp away from it.

• 24:10David Weisburd?

  What comes next?

• 24:12Speaker 01

  I think the usage-based, consumption-based, token-based models are all things that are in the conversation right now. What we've seen is some hybrid approaches, so some SaaS plus utilization. We've seen some pure play utilization or token-based models. Those can be scary for the CEO, CFO of a large enterprise, because you have some opacity around, what am I going to pay? You know, I know what my budget is. We've all bought into the budget. We told our board, our budget's X. What if we exceed that, right? That can create problems. And so understand that and creating visibility is important for enterprises to really get comfortable with that buying behavior. I think you can cap it, you can tier it. That creates more hybrid approaches that look sort of like SaaS plus utilization.

• 24:57Speaker 01

  That may be the on-ramp process that gets them comfortable over time to say, look, we've been doing this for two years now. Our utilization is X. We've paid $1 million a year for the last three years. This year, we'll probably use it 20% more, so we've budgeted 1.2 million. That'll take some time, right? They need to have the data behind that and the comfort level to make that type of decision. But, um, you know, it's how most markets develop over time, and I don't think it'll be different here.

• 25:21David Weisburd?

  Sitting where you're sitting, a couple of weeks ago, I had David Breach, the Vista Equity President. They've deployed over $100 billion in software. So he certainly thought long and hard on this issue, and he believes that the future of software is not just selling software, but selling solutions. It's kind of do-it-yourself versus done-for-you model. Do you see that as the evolution in software itself?

• 25:42Speaker 01

  In the early days, we're seeing a lot of that. Consultative selling. I think we've seen portfolio companies with very powerful technology enter Fortune 500 companies, take the covers back or open the hood and be surprised by how disparate and unruly the data is. And that very quickly turns into a data consulting project. For a VC, that's not what I love to hear, right? We want to hear like-

• 26:07David Weisburd?

  Doesn't sound scalable.

• 26:08Speaker 01

  This, this will work out of the box, right? And so I think you have to have, your customers have to continue to deploy best practices in data management, fidelity, infrastructure, in order to unlock, right? The plug-and-play value that a lot of portfolio companies and software companies will ultimately bring. But the on-ramp in the early days can be consultative. That, that can be a positive to the startup in our world, because they learn a tremendous amount about the problems that the enterprise is feeling. Once they've felt that and solved that, they can then replicate it and even create agentic workflows to replicate it across other use cases, so that the next Fortune 500 they go into takes 10 times, is 10 times faster for the implementation.

• 26:51Speaker 01

  But these are baby steps, I think, in the process of really unlocking the true enterprise value that will, will come in years.

• 26:57David Weisburd?

  Reminds me of the Paul Graham advice, which is do the unscalable. If you want to build a bill, great business, do the unscalable, figure out where the problem is, and then automate. Same with the Elon algorithm. A lot of people think the Elon algorithm starts with automation, actually ends with automation. Because the number one problem you could have in engineering is that you automate something that doesn't actually add value. So you have to figure out, make sure that adds value, run it, kind of make sure that that ROI is done, and then, and only after then, then you automate. People have this, especially venture capitalists have this bias, oh, I want something scalable, I want something automatable from day one, but sometimes it can be counterproductive even to the, to the design of the product itself.

• 27:33Speaker 01

  Yeah, potentially. And, and I think like, when you think about funding startups, you know, we always enter at seed, I think you need to create some realistic benchmarking and understanding that that will happen. And that leads to things like, you know, how much cash flow do you need to achieve a particular thing, a particular outcome? So what does it take to reach a Series A, Series B? I think on one hand, from a product standpoint, you can reach product maturity faster than ever. So like the dollars that we're putting into companies are being used with an efficiency that has been never seen before.

• 28:06David Weisburd?

  Double click on that. Why, why are companies able to scale so much quicker with so much less capital?

• 28:11Speaker 01

  I mean, I think it starts when you're thinking about product and code. I think agentic coding has been massively impactful, especially to people who know how to wield it. So early-stage founders that are technical, that have engineering backgrounds themselves are the people that we've seen in sort of the data research we've done that are benefiting the most. The most efficiency right now is being accrued to the engineers. You know, I think next on the list is something like customer service, which is just a great data set, great use case that you can augment quickly, it's digestible. But when you think about like, the biggest delta in efficiency, it's at engineering. And so talking to large enterprises, you know, we have a number of advisors, technical advisors on our team that work at places like Meta, Amazon, Oscar Health.

• 28:53Speaker 01

  And when we talk to them, we say, what does your job look like now as an engineering or AI manager? And they say, I can accomplish four times as much in half the time. And so when we-

• 29:03David Weisburd?

  ADR improvement.

• 29:04Speaker 01

  Yeah. And so when, when you hear that, you start to think about, and we see this just in practice, right? It's like the founders that we're backing are being way more cash efficient to get to product maturity in a shorter period of time. And so that's part of the reason why we spend our time at seed. I think the delta, the alpha creation opportunity from seed to A is, that delta is greater than ever right now. And so when we think about risk allocation and investment portfolio decisions, we think in sort of sharp ratios, right? We're allocators and buyers of risk. And for every unit of risk, we need to pay the appropriate price. As things go from non-obvious, pre-product to post-product and obvious, meaning some product market

• 29:47Speaker 01

  fit, customer references, revenue, during that curve, that curve is happening shorter and faster than it's ever happened. And as a result, you see series As and series Bs being priced where series Ds and Es were prior. And so I wouldn't like to play that game as much as I'd like to capture the alpha going from non-obvious seed to relatively mature product in a matter of year, short, short number of years.

• 30:09David Weisburd?

  I don't think I've heard a single intelligent investor take the counter on AI disruptive, AI has intrinsic value. I haven't found anyone to say, oh, it's all BS, it's all hype. That being said, AI being valuable and all, it's reported that pre-seed rounds today on average, $100 million valuation. How do you underwrite that?

• 30:32Speaker 01

  I think, again, you're buying risk, you're allocating dollars behind some level of risk. At early-stage venture, these businesses are inherently risky. So if that's true, it's very hard for my team to pay that kind of price. Because even if they're sophisticated, if they have the world's best team, you are still betting on an outcome that is extremely low probability.

• 30:54David Weisburd?

  Support for today's episode comes from Square, the all-in-one way for business owners to take payments, book appointments, manage staff, and keep everything running in one place. Whether you're selling lattes, cutting hair, running a boutique, or managing a service business, Square helps you run your business without running yourself into the ground. I was actually thinking about this the other day when I stopped by a local cafe here. They use Square and everything just works. Checkout is fast, receipts are instant, and sometimes they even get loyalty rewards automatically. There's something about businesses that use Square. They just feel more put together. The experience is smoother for them, and it's smoother for me as a customer. Square makes it easy to sell wherever your customers are, in-store, online, on your phone, or even at pop-ups, and everything stays synced in real time. You could track sales, manage inventory, book appointments, answer your reports instantly,

• 31:39David Weisburd?

  whether you're in your shop or on the go. And when you make a sale, you don't have to wait days to get paid. Square gives you fast access to your earnings through Square Checking. They also have built-in tools like loyalty and marketing so your best customers keep coming back. And right now, you could get up to $200 off Square hardware when you sign up at square.com/go/howinvest. That's S-Q-U-A-R-E.com/go/howinvest. With Square, you get all the tools to run your business with none of the contracts or complexity. Run your business smarter with Square. Get started today. Support for today's episode comes from Square, the all-in-one way for business owners to take payments, book appointments, manage staff, and keep everything running in one place. Whether you're selling lattes, cutting hair, running a boutique, or managing a service

• 32:23David Weisburd?

  business, Square helps you run your business without running yourself into the ground. I was actually thinking about this the other day when I stopped by a local cafe here. They use Square and everything just works. Checkout is fast, receipts are instant, and sometimes they even get loyalty rewards automatically. There's something about businesses that use Square. They just feel more put together. The experience is smoother for them, and it's smoother for me as a customer. Square makes it easy to sell wherever your customers are, in-store, online, on your phone, or even at pop-ups, and everything stays synced in real time. You could track sales, manage inventory, book appointments, answer your reports instantly, whether you're in your shop or on the go. And when you make a sale, you don't have to wait days to get paid. Square gives you fast access to your earnings through Square Checking. They also have built-in tools like loyalty and marketing so your best customers keep

• 33:08David Weisburd?

  coming back. And right now, you could get up to $200 off Square hardware when you sign up at square.com/go/howinvest. That's S-Q-U-A-R-E.com/go/howinvest. With Square, you get all the tools to run your business with none of the contracts or complexity. Run your business smarter with Square. Get started today. One thing that I've learned from talking to hundreds of investors, founders, and business leaders is that planning for the future isn't just about growing assets, it's also about protecting the people who matter most. Life comes with a lot of important milestones. Maybe you're buying a home, starting a family, planning for college, or simply thinking more intentionally about your financial future. As those responsibilities grow, it's worth making sure you protect those you love. Not sure what kind of life insurance you may need? Concerned about how much it may cost?

• 33:53David Weisburd?

  Think life insurance is too complicated? Your local State Farm agent can help answer those questions. They can help you explore your options for life insurance and choose coverage that you can adjust as your family's needs change over time. Financial protection for your loved ones when it matters most. Learn more about life insurance at statefarm.com. Like a good neighbor, State Farm is there.

• 34:12Speaker 01

  And the numbers are sort of the numbers. Now, those numbers have shifted, right? Like, the, the big outcomes are now tens of billions, not a billion. So there is a world where that, that can happen, but it's still very low probability. And so when I think about portfolio construction, those are the nice to haves, not the need to haves in terms of how I underwrite. Like, if we're fortunate enough to get into a company that's a 20, 50, $100 billion company, that's fantastic, obviously. But we are underwriting a more rational, high-probability outcome, and that gives us some price discipline at entry. But I think that's why we also play in the intersection of what would be formerly enterprise software and AI core use cases, not in the sort of lab-based world that requires billions

• 34:55Speaker 01

  of investment into the top talent, which has just become more expensive over time. We'd like to be the beneficiaries of that great work that's being done by those model companies without investing and owning that team, right?

• 35:08David Weisburd?

  I had this various conversation with Matt Harris from Bain Capital Ventures, where I know you used to be. And the heuristic you used is, in evaluating, for example, Anthropic's valuation, it's reported it's going to be $2 trillion valuation, is, one way to think about it is just to take off three zeros. So imagine taking off three zeros from $2 trillion, you get $2 billion. And imagine taking three zeros from $100 billion, reported, kind of 2026, and, and run rate, so 100 million. And a 20x revenue is not that unreasonable, it's just people get, I think, rationally so, or at least evolutionarily so, people get so hung up on the sticker shock that they can't imagine there still being upside at $2 trillion, but then when you look at the revenue, it starts to make more sense.

• 35:53Speaker 01

  I think that's a fair argument. My counterpoint to that would be, the probability of getting that type of outcome is just very, very long tail, right? It's like a Six Sigma-type outcome, or a generational outcome. If you rewind 20 years, it would be like being the investor that found Google, right? Those guys did okay. Similar to that, the folks who found OpenAI, Anthropic, et cetera, will do very, very well. Now, as I think with like my public market investor hat, or my late-stage investor hat, those companies are also facing new threats, both from each other, right? Every model release is a step change in terms of functionality, and they're fighting in sort of a duopolistic market right now, but new entrants are not a fore-drawn conclusion.

• 36:38Speaker 01

  Like, they may come, right? Open weight models may come into, into a higher use. Foreign models that we, that we aren't even in the conscious right now, could come into the, into the purview. And so I think rising competition for a massive, right? They've demonstrated the TAM and the growth of the TAM. What you see in markets like that is that you have new entrants, lower-cost entrants. The consumer willingness to pay, or expectation to pay, is between $0 and $10 per month. Like, and those are some of the early adopters. Over time, that probably goes down, while models and switching costs between models become, switching costs become lower, right? I've, I switch between four to six models in a week, and agents. So switching costs are low.

• 37:23Speaker 01

  I'm always chasing the best and fastest LLM for my use cases. And so what that results in, potentially, is that it's harder to monetize your huge user base because of competition. And we're not even talking about some new entrant that could be the next Anthropic. If we rewind two years ago, we wouldn't have been talking about Anthropic necessarily, we would have been talking about OpenAI. And looking at that graph can be a cautionary tale in the sense that OpenAI was inflecting at a rate that nobody had ever seen, and then here comes Anthropic and actually apexes it. And so seeing that is an interesting thing to watch, let's say. I was not fortunate enough to invest in OpenAI or Anthropic, but some of my LPs were, and they've asked me over the last several years, like, when to sell?

• 38:06Speaker 01

  And I've given a similar argument for them, especially depending on where their basis is.

• 38:10David Weisburd?

  So I did invest in Anthropic at a $4 billion valuation, so maybe you could give me some advice, because I'm trying to get a balanced view in terms of the upside and the downside moving forward. How would you underwrite it today, October 2026?

• 38:23Speaker 01

  Well, I think I would look at the implied forward capture of the market, right? So how big is the market likely to be? And then what percentage are you expecting to capture as Anthropic, vis-à-vis or relative to OpenAI and some unknown entrant, or entrance? So I think that's hard to handicap on the who's coming next, right? Conversation.

• 38:44David Weisburd?

  It's unknown, unknown. Who's going to have that idea or the resources?

• 38:47Speaker 01

  That's the unknown, unknown. But you have other folks that are devoting a lot of resources to the space, that have billions of dollars in reasons to be highly defensive. The Googles, the Metas, the Amazons, et cetera, right? And they, they may take a different approach or a similar approach to building very, very robust models, but that will only up the ante in terms of competitiveness over time. And then if you look at, or make an assumption around whether there'll just be two players in the future, I think it's very unlikely that there'll be two dominant players in the LLM world. So if you start to model, and you can sort of stress-test this, right? A few more entrants, and then you look at the market size, of course, it's going to continue, continue to grow. What do you, what's the implied capture rate of that future revenue in three, five years

• 39:32Speaker 01

  that Anthropic would have to have to justify their current valuation? I think right now, it is converging to a rational sort of pre-IPO price, so you're probably in a, in a good spot, but that's absent, my assumption, around increased competition and lower-cost players. Now, I think the other angle or lens is, what is there, who are the customers of all these LLMs? OpenAI tends to be slightly more consumer, Anthropic tends, especially with Claude Code, tends to be slightly more enterprise. So for that reason, I would value every dollar revenue differently across those two. And so I give a little boost to Anthropic in that regard, and you can see that's probably why they've eclipsed OpenAI in terms of their growth rate.

• 40:07David Weisburd?

  I'll tell you the bull case. Virtue of Lucas Swisher, he runs Growth for Co2, a previous podcast guest, and he talked about the three S-curves of adoption. So the first S-curve, the early adopters in terms of LLMs, from a revenue standpoint, have been developers. What's crazy about that is that there's only 50 million developers in the world out of 8 billion people. And on top of that, only a small percentage of developers have actually implemented, if you look at worldwide kind of engineers. So we're still at really the small end of a small market. Now, sure, enterprises have adopted it, but to your point, 95% of them have failed. There's not that mass adoption. There's not many enterprises that are super

• 40:51David Weisburd?

  users today. So we're still seem to be, and then obviously there's all of enterprises. First, you have engineers, then you have all of enterprises, then you have consumers, and then you could argue a fourth S-curve, which is agents. So some smart people would argue that your TAM is not actually 8 billion, it's 8 billion people and their agents talking to each other. So in terms of adoption curve, we started this conversation talking about we're in the first inning of AI.

• 41:17Speaker 01

  Well, I think you just made my point a little bit there.

• 41:19David Weisburd?

  I think there's an argument.

• 41:20Speaker 01

  Especially on the enterprise side, there's room to run, right? There's definitely room to run. So because you established that this market is growing and the TAM is expanding even from a very large base, it's extremely attractive. And I think that will also attract new entrants, new players, whether it's the FAANG players or some player we haven't even started talking about. But that could erode pricing power, right? So it could erode in two ways, which is over time, you just see switching costs and you see the markets divided among more players. You can see the entrance of where we invest, which is bespoke enterprise software meets AI for specific industries that has foundational

• 41:56David Weisburd?

  Verticalization.

• 41:57Speaker 01

  Yeah, that has more foundational subject matter expertise baked into it, which then leads to trust within specific verticals like airlines, banking, insurance, et cetera. And then over time, I think you also have a pricing compression, which is like enterprises are going to be willing to pay more, but how much more will they be willing to pay? In order to really capture the lion's share of the wall, you need to have much higher ROI than we're seeing right now in the first inning. So it's sort of multi-factor. It'll be fascinating to watch, but I don't think you can lose coming in where you can.

• 42:29David Weisburd?

  Also, investor in vertical AI solutions like Legora. How do you think about horizontal versus vertical? Are there going to be different areas where vertical wins versus horizontal wins? And what's your framework for that?

• 42:43Speaker 01

  When we think about horizontal, we think about more utilities that are foundational to AI. So it could be a company like Monitor, where it's helping any vertical or any one enterprise manage their own models. It could be the data loss prevention example I gave with sort of like, where is my data flowing in and outside my enterprise? That again, would be a horizontal solution to avert potentially a vertical problem. And then when I think about verticals that will win, I think selling into any one of these verticals requires trust. And especially in the early days, building that trust within an industry as esoteric as airlines or insurance underwriting or banking underwriting lending, that requires

• 43:26Speaker 01

  trust, understand the customer process, the legacy process, understand the data, the KPIs, et cetera, can give you an early lead to start building that data flywheel within legacy industries. And so when we think about vertical AI, I think it's companies that have domain expertise plus the ability to build product very, very fast. So engineering-led teams that have AI-native workflows that also have deep understanding of what we call subject matter experts on the team that can earn the trust because they have some sort of asymmetric advantage to penetrate a market. And so we're oftentimes like underwriting teams in that way, which is technical prowess. What have they built today? How fast can

• 44:10Speaker 01

  they build their future roadmap? But then also, what does the go-to-market motion look like? Is it appropriate for their ACV, for example? So are they being rational about how they scale? But then the next thing is, we don't want to fight in a fair fight, right? There's lots of players. People can build companies faster than ever. And so when we're looking at a company, we're looking for an unfair fight, bringing a gun to a knife fight, right? It's who knows this industry better? Who's better connected to this industry? Who has the license to win in this industry? That can often, in addition to having the best technology, you have to have base case very strong, if not the best technology. But having that other tipping point can really tip the scales in your favor in terms of how fast you can penetrate an esoteric market like airlines.

• 44:53David Weisburd?

  I think your answer will be very disappointing to engineers. I think since 2008, when I started in the startup world, engineers have been trying to replace salespeople. And I've liked this high school jocks, jocks versus nerds thing where for now, almost 20 years, engineers have been trying to replace salespeople. The problem with that is they may be ultimately successful. It might take 10, 20 years. Why? Because they're still human beings running the organizations and human beings need solution, they need trust, they need all these things. So this idea that you could come to the market with the best AI solution still needs those salespeople that help with implementation, still needs those people that help actually on the operations level, and all these things that still require real-life human beings, at least for now.

• 45:38Speaker 01

  Yeah. And I would argue, that's all true. That may go away a little bit as you automate sales process, and that's fine. But what still remains is understanding customer needs. People who deeply understand their customer, some of them were formerly the customer they're targeting, right? They bring back unique insights and problem sets that the best engineers thrive with, okay? Generally, the best engineers, unless they're the founder of a company, are not going out and talking to the customer or have lived inside the customer for 20 years. And so the salesperson can also be sort of a customer solutioning specialist that brings back unique insight or understanding to the engineers that informs product roadmap,

• 46:21Speaker 01

  solutioning, and making sure that whatever we're building within a startup has extremely high value day one to the customers because it came from them.

• 46:27David Weisburd?

  Do you know the story of how GoDaddy massively increased their revenue?

• 46:31Speaker 01

  I don't.

• 46:31David Weisburd?

  They put salespeople into their customer service calls. So they realized that the best time to actually upsell somebody and figure out what people need is at the time that they have issues. And it goes back to this kind of paradox in customer services. When you have a problem and a company solves it for you, you may end up actually building affinity with the company more so than if you didn't have that problem. So they actually

• 46:54Speaker 01

  Yeah, it becomes sort of de facto design partners with you.

• 46:56David Weisburd?

  So they figured this out a decade ago, and now we're kind of seeing this reemerge again in AI.

• 47:01Speaker 01

  Fascinating. We invested into another AI company that was doing post-sales data gathering for service delivery companies. So sales makes a sale, right? They capture some data in their CRM about how that process went. And everybody has a steak dinner and high fives and says, great job, everyone. And that's usually the, it's like the day of your wedding, right? It's like the happiest day of everyone's life is the day that you consummate the deal. And then after that, you start actually delivering the service. You could be an investment banker, you could be a consultant, you could be a lawyer. You're delivering service over a long time series. Monitoring those interactions is what our company known well does. And what it's pulling back is unique insights for the management

• 47:42Speaker 01

  to say, it's great that you sold this customer. We're going to monitor minute by minute, interaction by interaction, A, how happy they are, but we're also going to distill down everything the customer's telling you about what they need. You can take those insights and immediately put them back into those enterprises to probabilistically and sort of stochastically lay out, what should we build next? Because you have a time series, multiple customers, and frequency graph of what your customer is telling you. When you digitize that and digest that with AI, it creates unique insights for the service delivery partner, which says, it could say, wow, we just really don't like this team that we're working with, or so-and-so was not really great to work with, or you weren't very timely. But it could also say, hey, the software you sold me does XYZ, but we were really

• 48:27Speaker 01

  looking for tau. And that can be a unique insight that can then recursively get pulled back into the product roadmap.

• 48:34David Weisburd?

  When I talk to engineers at Anthropic, I also have some friends at OpenAI, they're telling me that average engineers are getting a 4 or 5X efficiencies and that the 10X engineers are getting actually 100X. So the 10X engineers, which oftentimes are twice as expensive, they end up being even better returns on investment. Are you seeing the same phenomenon on the sales side where the top salespeople are actually getting these major efficiencies from AI?

• 48:59Speaker 01

  I think it's still pretty early innings on that. The best salespeople generally are still the best salespeople. I think sales is probably slightly slower to adopt, or at least, just like a lot of enterprise that we discussed earlier, is only starting to scratch the surface of the true scalability. I think if you're an AI-native, if you're a founder and not just like an enterprise salesperson, a founder doing sales can do things that they never could have done at a scale they've never done. Even in like my own side projects, my agents are running business development workflows and analysis while I'm sleeping, right? And so that's me being sort of like a hacky founder mind that's just running those in the side. Similarly, for fundraising at Studio VC, lots of agents

• 49:41Speaker 01

  involved in identifying LPs, drafting initial emails, keeping us in the loop in terms of follow-up, all these things, right? So it's almost like we have a chief of staff there. But that's us as AI-native, AI early adopters. When you think about traditional sales, there's some defensibility. There's a, well, how do I matter in a world where AI is doing a part of my job? And so the willingness to adopt some of those technologies is a little slower. And again, that comes back to my point earlier where it's like, if the CEO is native and understands the impact that they could have on sales, that will start to trickle down quicker. But if you're asking that head of sales, right, they're going to try to protect their fiefdom potentially.

• 50:18David Weisburd?

  Yeah, it makes me think about this alignment issue where salespeople within startups where they have equity may be faster to adopt than ones in traditional kind of Fortune 500 companies.

• 50:27Speaker 01

  Yeah, it's a great alignment question, yeah.

• 50:29David Weisburd?

  If you could go back to 2009, when you first started working in venture, and you could give yourself one timeless piece of advice, what would that be?

• 50:36Speaker 01

  Wow.

• 50:38David Weisburd?

  That's one people take a long time to think about.

• 50:40Speaker 01

  Two things coming from the culture of Bain Capital. One is, some of the smartest 10 or 12 partners I've ever been around. So it was an amazing place to start, be a fly on the wall, learn from and listen to. But there was also a lot of humility among those partners and as they interacted with founders. What I think the cultural fit for me, and just to go back before that, when I was leaving investment banking, I thought I was going to go into private equity, just like all my colleagues did. But I got a call from Bain Capital Ventures, took the meeting and just fell in love immediately. It was the best interview I've ever had. I called my parents and said that was the best interview I've ever had. And the culture was such that they were extremely smart but humble. They wanted to drive value in the ecosystem and they understood that this was a long game.

• 51:24Speaker 01

  And so that's how I've tried to operate in my career, which is like, even if you're not investing in a company, drive some value, right? Like, what's the worst that could happen? Maybe they come back a few years later, or they send a friend. But the second thing I learned from Bain Capital was I think that they did still have some consulting DNA in terms of how they looked at deals. Very complete analysis and diligence process at Bain Capital Ventures, which I was impressed with and I still keep in my back pocket to this day, which is, it's a combination of technology, financing, market, strategy, and reasons to win over time in a competitive market. So that's what fascinated me about working there. And as I've kind of progressed in my career, I've gone earlier and earlier

• 52:08Speaker 01

  towards seed stage because I think that's where the problem is most acute with founders. And I love getting in there and helping them or creating the conditions that help them to solve those problems. And hopefully, I'm not the smartest guy in the room, like hopefully they are, right? And I can help create conditions and foundations and connections for them that helps to make that happen in an asymmetric way compared to the average seed investor.

• 52:31David Weisburd?

  Well, Joe, this has been absolute masterclass. Thanks so much for jumping on.

• 52:34Speaker 01

  My pleasure. Thank you so much, David.