Redpoint Ventures Playbook: How Top VCs Are Really Investing in AI with Jacob Effron

So it's really exciting times, obviously, in trying to determine what the market size of this all is. A bit of a fool's errand. And as famously people have whiffed on, been wildly off before. I think Morgan Stanley has estimated in the next few years, 25% of global software spend could be towards these AI use cases. And so all this is like incredibly exciting, obviously makes the best period to, to be a venture investor. But certainly comes with its challenges as well. Right. And I think there's two main challenges that we see day to day. The first is there's a lot of AI companies being started. And what that means on our end is an interesting pattern that we have as investors is anytime we get excited or interested in a category, there's 10+ startups in that category that pop up. And so that is really interesting market dynamic. The second thing we're seeing is And I'm sure folks have seen headlines and whatnot. The prices for these companies are definitely increasing. There's a lot of investor excitement.

Hey everybody, Saastr Annual will be back May 2026, the world's largest SaaS and AI gathering for executives. Just this last May, we hosted 10,000 attendees with 68 VP-level and above attendees, 36% CEOs and founders, and 25% were AI-first professionals. It's the very best of S-tier attendees and decision makers that come to Saastr Annual and AI Summit each and every year. But here's the reality, folks. The longer you wait, the higher ticket prices get. They're cheap now. They're cheap. So just get them early. Lock in your spot today. Use my code JASON100 for exclusive savings. Get your tickets at podcast.saastrannual.com or just use code JASON100 when you check out. See you there. Saastr Annual and AI Summit 2026.

It will rock. So maybe give a bit of background first on myself and then we can dive into it. I'm Jacob Efron. I'm a managing director over at Redpoint. Redpoint is an $8 billion venture capital firm based in Silicon Valley. We've invested in companies like Snowflake, Stripe, Twilio, Ramp, HashiCorp, and I help lead a lot of our AI investing. I also run an AI podcast on supervised learning. So if you find any of this interesting, definitely check that out. But today I'll talk a little bit about what we're seeing in the AI world, how it informs our investments today. And maybe I'll start with some backdrop that I'm I imagine familiar to most of you, but maybe help set the context for what we're seeing as VCs. And so to kick it off, we're seeing a massive decline in model costs today. So what does this slide show? Basically, for a given benchmark of model capabilities, the cost of— or, you know, per token of those models is going way down, way faster than it was in the cloud era. And so what I think that means for AI application companies is a familiar pattern where basically we get a set of frontier models that have new capabilities. We discover the wild things they can do, and then those capabilities become cheaper and cheaper until they can be wrapped into the vast majority of products. And even in the past year, we've seen some wild new capabilities from, from models. Obviously, they've gotten way better at coding. I think voice models have gotten a lot better. And that's then spawned a ton of different startups in these spaces. And so I think the implications of that are, as we think about things, we're super focused on what's the end use case and how powerful is that with AI, and a lot less focused on what are the gross margins of today, because these models just get cheaper and cheaper over time. I think the natural byproduct of these people playing around with these models for longer periods of time is we continue to just find more and more use cases and get more and more usage of these underlying models. And so just the insane increase in, in usage of AI models, which is a lot faster than happened in the cloud era. And we'll talk a little bit later about some of the use cases that are driving this. But I think there's this pervasive feeling in the investment community today that even if you froze the model capabilities, there's trillions and trillions of dollars of applications just waiting to be discovered here. And obviously for some of the most cutting-edge reasoning models, we've only had 6, 9 months to experiment with what these models can do. And so I think over time there's going to be more and more. And what that means is basically the power of these models, the increased usage, the fact that they're pretty cheap, is that when these startups find product-market fit, they're just scaling way faster. Than traditional SaaS counterparts. And so this is some great data from Stripe that shows how fast AI companies are scaling. And this pace of adoption breaks a lot of rules about traditional startups. As I think about some of the classic rules of SaaS investing, it was maybe don't go against some of these horizontal incumbents or don't sell into hospitals or law firms or banks. Those are those customers that take a long time to buy. But we're seeing a lot of these rules be broken by just the 10x better product capabilities that a lot of these models are able to provide and like insanely fast scaling. And what does that mean is these budgets are only, only going to increase. I think maybe in the early days post-ChatGPT, there was a little bit of skepticism from enterprises about what some of these models would actually be able to do. But there's the combination of kind of actual real-world impact these models have had, plus just everyone watching the pace at which they've gotten better, I think has led to real deep interest from enterprise. So I think these budgets are only going to get bigger. And at the same time, the models are only— are going to get better. And I think we're— with each generation of model improvement, we unlock a whole new set of capabilities. And obviously a lot of focus now is around models that can do tasks for longer periods of time. And so this slide is really around how long can you let a model go and act and actually execute something, or what folks often call agents. And I think as models are more and more able to do this, the set of applications that they're able to address and the set of industries they're able to address is only going to increase. So it's really exciting times, obviously, in trying to determine what the market size of this all is a bit of a fool's errand. And as famously, people have whiffed on or been wildly off before. But I think there's all sorts of estimates of just, hey, this is a really large opportunity. I think Morgan Stanley has estimated in the next few years, 25% of global software spend is— it could be towards these AI use cases. And so all this is like incredibly exciting, obviously makes the best period to to be a venture investor, but certainly comes with its challenges as well. Right. And I think there's two main challenges that we see day to day. The first is there's a lot of AI companies being started, which makes sense if you're an ambitious founder today. What more exciting space to go build in than applying these models to a bunch of end industries? I think YC is a great proxy for what types of companies are being founded. And so you're seeing the proportion of AI companies go way up And what that means on our end is an interesting pattern that we have as investors is anytime we get excited or interested in a category, there's 10+ startups in that category that pop up. And so that, that is a really interesting market dynamic. The second thing we're seeing is, and I'm sure folks have seen headlines and whatnot, the prices for these companies are definitely increasing. There's a lot of investor excitement. And so we looked at data of all the kind of Series B and C companies we looked at that were non-AI and then AI. And as you can see, the AI companies, they're raising larger rounds. They're raising those rounds at higher prices. Now, admittedly, a lot of that is justified by the higher growth rates that these companies have, the kind of step change in adoption. But these prices and the sheer number of companies definitely make for a, a challenging investing landscape to navigate today. And so maybe I'll share a few observations on the broader market before I talk about how we think about navigating all this. The first thing I'd say is these products are hard to build. Yes, there's a lot of companies that we see in each space that 10+ in each category. But I would say a common trend is that in any category we look at, there's really 2 to 3 of them that, that rise to the top and they're providing a better user experience. They're scaling faster. And I think the reason is there's not a ton of tooling that exists for a lot of these problems. Below the, below the line here. And so companies are building their own scaffolding around the models. And with the pace of model progress, that scaffolding has to change really quickly. And it's actually quite challenging to build effective AI applications. And we see a real difference in what some of the best companies are able to do. The second point I'd make about investing in the space is our thinking on what makes an interesting application company and what application companies need to do evolves over time. I feel like there's just a firehose of new progress in model advances, and it's important for us to have strong opinions on, on what makes for interesting applications, but we have to loosely hold them in this just period of rapid change. And I think maybe one example of this that, that could be interesting is the idea of kind of building and fine-tuning your own models for a specific domain. So immediately post-J ChatGPT, there was all this buzz around, hey, enterprises sit on tons of data. What if they pre-trained a really large model that was specific to finance or law or healthcare? And so famously, I think Bloomberg trained a model, BloombergGPT, that outperformed the, the latest OpenAI model of the time. And then 3 months later, OpenAI shipped the new model and it was way better. And I think that consensus has been, hey, massive pre-training for an individual domain doesn't make a ton of sense. And for a while, even fine-tuning was somewhat effective, not super effective. And it felt like the game was all about just building really good workflows on top of the model APIs. But the market does evolve. And in the last few months, OpenAI shipped reinforcement fine-tuning, which is kind of a new way of fine-tuning that they offer. And it's actually quite good. And so it actually is starting to seem again, being able at least to not pre-train but fine-tune data on a specific domain could be really helpful for a lot of these application companies. And so it's just a constantly changing and evolving landscape and one we're always trying to stay on top of what's happening. I think one of the most important things as we think about what's happening is what's actually working. And so maybe to hit on what we've seen really work today, I'd say broadly there's 4 types of applications that have real fit, and then we see them applied in a ton of different ways. The first is chat, but it's a great use case for these LLMs, works really well in the customer support space. The second is the ability to basically search over documents and summarize those documents. And so you've seen that in the horizontal space with companies like Bleem, but also in verticals. Legal AI companies like Legora or construction companies like Trunk Tools. And then these models are really good at speech-to-text and text-to-speech. And so you've seen that in the healthcare space, companies like Abridge that are transcribing medical visits, as well as companies like Liberate and Assort that are doing the front door of insurance brokers or medical clinics and being able to field more calls than folks could before. And then the last bucket is coding, right? These models are unbelievable at coding. They're only getting better. You've seen the explosion of companies like Cursor and Cognizant admission, Poolside, Sourcegraph, a whole host of, of companies in that world. And so I'd say to date, the categories that feel like they've worked best are coding, customer support, legal, and healthcare. But I think we're still just in the very early innings of applying this amazing set of capabilities to a ton of different end industries. And so there's inevitably going to be a bunch more applications or end industries that emerge over time that have great use cases with these model capabilities. And then the thing that we're always thinking about is every few months it feels like the models get better. So what's the set of use cases that's going to get unlocked as these models get more reliable? They have longer-term reasoning and memory, and we're constantly trying to figure out what, what that set of applications is. And I guess the theme of this was our investment framework and how we think about investing in the AI world. And thought I'd put together a, a simple 3-part framework. As a former consultant, it always has to be 3, but 3 questions we're always asking ourselves as we're evaluating these applications. So the first is, is there a really effective wedge for AI? And I would say that generally the bar for what product-market fit and what an effective wedge looks like has gone up in AI because you've just seen explosive growth for companies when they do have them. And I think there's this interesting thing we see in AI today where there's clearly a huge top-down mandate for companies to invest in AI. Boards are telling their CEOs, what are we doing in AI? So there's a lot of willingness to experiment with AI products today. And I think one of the key things we have to do is when we see a company that's growing really well, it's— is this just Hey, a company's messing around with what AI can do and maybe this is something that they adopt, maybe it's something they don't adopt, but they're willing to pay some money just to see what the product does. Or is it something that the end users really love? And so we're always trying to talk to the folks who are actually using the product and figuring out like, is this a game changer or not? And you can really tell there's just some products that people absolutely love. And as you think about this kind of race to be, I think there's a race right now to be the strategic partner to hospitals, to law firms, to banks, to insert any kind of company. And I think a key part of that race is finding the right wedge where AI can just deliver a 10x better experience, a killer use case that is— that everyone just loves using. So we really look for that, and that's a prerequisite to us looking at any kind of company. The next thing we do is we think about like how much more a company can do from this wedge. And so there's a few things behind that. One of them is, I think in larger industries, there's generally more that, that AI can do over time. And so we like healthcare, legal, finance, large industries versus some of the more niche ones. And then I think also there's a question of are there 10 other use cases within this industry that this set of models is really well poised to, to go automate? And so I think that one way you could think about this is there's lots of great wedges that exist, but maybe in smaller markets. So you could imagine one thing I would love is I'm sure you could make a great AI product to write sourcing emails to reach out to a bunch of companies for venture capital firms. And this would probably be a great wedge that could get you some initial scale pretty quickly. The issue with that obviously is there's only so many venture capital firms, that industry is only so many as so big. And at least today, it's not totally clear what else models could do within a venture capital firm. Maybe they could do a bunch. I'm sure over time they'll replace plenty of what we do. But at least today, I think that's a great example of a potentially very compelling wedge that doesn't actually translate into a really interesting company to invest in. And then the last bucket I'd say we think a lot about is the extent to which quality will matter. And the reason this is important is, like I said before, in any category we look in, there's always going to be a bunch of players building there and there'll always be somebody who's going to say, hey, for 10% of the price, we'll sell something that's 50% as good. And the markets that are most interesting are ones where you would rather have the higher quality product. And so maybe an interesting observation has been that in some ways some of the easiest things to sell AI products to today or automate are things that maybe have already been outsourced, that business process organizations are doing already. And the interesting thing about that is they may actually have a real race to the bottom in pricing. And so we're really focused on things where quality of the end application really matters and people will pay for kind of best in breed. And so I figured it'd be fun to take these 3 questions and take a look at how they apply to a few investments that we've made in the AI application space. So there's 2 that I'll talk about that I'm on the board of. One is a healthcare AI company called Abridge, and the other one is a legal AI company called Labora. So let's start with Abridge. Abridge, we co-led their Series C. They recently raised a Series E at $5.3 billion. It's an incredible company. What they basically do is they offer a product that records a visit between the doctor and patient and then transcribes that visit into the electronic health record in a way that the doctor doesn't need to spend a ton of time writing the note themselves. The insurance company gets the information on what happened in a way that is, is effective for billing, and the patient gets the doctor's attention as well as can get a summary of what happened in the visit. And This is, if you think about that 3-question framework on the wedge, you're not going to find a better wedge than this. It is like truly an incredible AI use case because doctors used to have this thing called pajama time where they would spend 3 hours at the end of a really busy day just typing up a bunch of notes late at night in their house, in their pajamas. Uh, and it was one of the things that was just leading to massive doctor burnout. And so for the doctors, this is really a total lifesaver of not having to do this extra work and being able to focus on the patients in front of them. Them. For the patients, have you ever been to a doctor's office and they're typing away at their computer the whole time, not looking at you? It's a way better experience when the doctor is able to focus on you. And then for the hospitals themselves, the ability to correctly document what's going on and then get paid for the services they provide is also really interesting. And so this is a use case that has just spread like wildfire. And if you talk to any doctors that use a bridge, I've never heard people speak so positively about a product. And so incredible wedge. What they— what also makes Abridge a really interesting company is there's tons to build on top of that down the line. And Abridge's CEO Shiv Rao always says that everything in healthcare flows from the doctor-patient conversation, and it's totally true. And Abridge is basically creating this net new data, and they have an amazing opportunity to just build a ton of interesting products on top of that data. And I think what Abridge does so effectively is they sell this Scribe product that's here today that people love, that works, but also paint this really compelling future vision of all the different things that can do down the line. And then that third bucket on quality, I think healthcare is totally a space where quality really matters. Who would want to go to a health system where they said, hey, we're using the cheap AI that's 50% as good but 10% the price, right? Like healthcare is really a space where quality matters. And I think Abridge has done a really great job of kind of leading the science of what does it mean to have a high-quality medical AI product and how does that, how does that basically, how do you evaluate whether these products are hallucinating, whether they're recalling everything that's happened in the visit correctly? These things have major consequences if they're done wrong. And so I think they've done a really great job doing that. And basically in being early to build these products, getting cutting-edge science, they've really gotten an incredible distribution flywheel going, which has been fascinating to watch. And I think it's a great example of a really successful AI company. The second company I want to talk about is a legal AI company called Legora. We led the Series A here. They recently raised a great Series B. What Legora basically does is it's an end-to-end platform, collaboration platform for lawyers where they can do they can review all of their kind of documents and work they're doing. They can draft and compose net new legal documents as well as do research on whatever it is they're digging into. And so I think, again, if you go back to the three-part framework, what makes Legora so interesting is this wedge is really compelling. There's some things that models do really well. We talked earlier about search and summarization. Bulk document review is a really good fit for models today. Research over a bunch of different cases or documents, as well as drafting and writing. So a great wedge that lawyers find a lot of value in today. And then I think over time there's clearly a feeling that, hey, legal is this text-in, text-out space. There's going to be more and more that models can do over time. And then it feels like this is an industry that's fundamentally going to change with AI. And there's a lot more that a company like Legora can do. And then similar to Abridge, that third question, quality really does matter. The top companies in the world, the leading private equity firms or tech companies, They use law firms and they want their law firms to be using the highest quality tools, right? And so again, whoever builds the best product in this space, I think really has a right to win that category. And it may be an interesting data point for our listeners. One thing I think is fascinating about Legora is they've been a really successful second mover in the space. So Harvey was really the first company in the legal AI space and Legora started afterwards, but has really caught up. And built an incredible product. And I think what's enabled that is a few different things. One of them is we're still in such early days of what these legal AI products can do. I think today whatever these products can do is probably 5% of what they'll be able to do down the line, that there really is an opportunity for, for a leapfrog as these capabilities continue to improve. And then the second thing that was interesting is one of the challenges of coming to market later is sometimes people want to pigeonhole you into, hey, we already have an end-to-end solution. You just do this one thing or that one thing. And it's hard to build an end-to-end product when you do that. Legora is a team out of Sweden, and because it was started in the Nordics, actually had a real opportunity to early on in the company's history work with these leading Nordic law firms and build out an end-to-end product there. And so they were able to build this really effective end-to-end product and then bring it to the rest of the world. And I think as we think about what makes for really effective teams in the AI space, The Lagora team is, you know, about the fastest team you can find in terms of just shipping new things, getting the capability of these models into the hands of their customers and building just really compelling workflows on top of the models. And so these are two amazing AI companies we've had the privilege to partner with that really, I think, fit into this, to this overall framework. And so before opening it up to questions, Maybe I'll just end with some, some basic thoughts on some of the things we've learned so far and seen in investing in AI applications. The first is there is just— it's amazing how fast people can build brands and become almost incumbents in the AI application space and just the compounding advantages that come to that. And if you ask someone on the street, name an AI coding company, they'll probably say Cursor or Support Sierra or Health Care Bridge. And that's just amazing to see how fast these companies can move and build those brands. The second thing I'd say is velocity is probably the most important thing we look for. The market just changes so fast and it's both a race to build the sheer breadth of all the different things these models can do for end industries, but then also a race to, hey, a new model or new capabilities come out. How do we figure out how to translate whatever GPT-5 can do to an end industry. And teams that are just shipping incredibly fast have a real advantage in doing that. It really compounds over time. And then the last thing I'd say is a lot of what differentiates these products is really the thousand little things that make one product delightful to use over another. So it's the UX that's way easier to use, it's latency that's way faster. So every time you're doing a review, it comes back quicker. And the interesting thing we've commented on as a firm is that in many ways a lot of what makes AI applications really strong is similar to what we looked for in SaaS businesses in the past. And I think over time, a lot of these companies have only been building on these models for a year or two as we're built, as they're building out for 4 or 5 years. I think a lot of these differentiation that we've seen in the SaaS world of really products that are just delightful to use and are really industry favorites, I think that will only continue to compound. As folks build longer. And so with that, I think that's how we think about things generally. Would love to open up any questions that folks have.

Thank you for that, Judy. Super helpful. There's a few questions in the chat and I'll help, I think, frame them in a way that is more helpful for broader categories as well. A few things you mentioned here. How are you guys now thinking about the balance of investing between, let's say, maybe more traditional B2B be SaaS companies versus ones that are like AI-first? Like I would argue Looker and Bridge are AI-first.

Yeah, yeah, yeah. No, it's an interesting, it's an interesting question. I think one, one maybe funny anecdote is that for a while in our CRM, we used to distinguish between, we're like AI companies and this one's a SaaS company. And we've now collapsed that distinction. And so it's almost meaningless at this point. Like any SaaS company is building, well, they're all kind of building their own AI features, right? And certainly You have companies that have come of age in this AI era and maybe are moving faster or leaning into that more. But I think there's tons of any kind of more traditional SaaS company we look at is still figuring out ways to, to give AI to their end customers. And I think that on the application side, almost any opportunity we look at, whether it was founded post-ChatGPT and is purely like an AI wedge or maybe is a company from 5, 10 years ago that has incorporated this stuff in, AI is definitely a part of the story. And so I think that I'd say almost universally that is true of the application companies we invest in.

And to put it maybe more high level, do you think it's a red flag if you get pitches that don't mention AI at all?

You're like, I'm not even going to look at this one. I never hold rules that, you know, that, that solidly. But I would say I think there's— I think it would be a red flag if the team hadn't deeply thought about and tried, hey, what can these models do in our domain? Now, like I said, it's very possible that you have a really sophisticated vertical SaaS company. They've looked at a bunch of different applications of AI. They said, hey, some of this stuff is, is 10% helpful or 20% helpful, but it's not actually— maybe there isn't a 10x use case yet in our domain. And that's— that might be the case. I think despite all the hype around AI, it's not like in every industry already there's some AI application that scaled incredibly fast and found this amazing wedge. But I think the risk of not thinking about that stuff is over time, as the models get better, there will be a 10x wedge from AI in pretty much every industry. And these companies would risk being disrupted by someone that was doing that if they're not keeping their fingers on the pulse. But there's— I think it's totally legitimate that maybe that 10x opportunity doesn't yet exist in a certain industry.

Yeah, makes sense. Next question from the chat. How important would you say a moat is today?

It's a great question. I think the, the reality is we may have like somewhat unrealistic expectations around moats. Like I was saying earlier, if you think about the reasoning models, they've been out since what, like September? People have been building on them for 9, 10 months. In SaaS world, nobody would say, hey, after 9, 10 months of building a SaaS product, you really should have this defined moat. I think it's really important that there's a path to a moat and that this isn't— and I think that's why we talked about like quality and some of these other things that might help you. There's all sorts of other creative ways folks are building moats. And I think it's important that you can see that on the horizon, but I think it's unlikely in such a short period of time that there's going to be— it's— and like these moats are often at this stage, the thousand small things that they don't always sound the smartest. Like I'd much rather come and be like, The reason there's a Moat is they train the best model and no one else can do that, or they have data nobody else has. And I think in the early days of AI, everybody wanted that to be the story. But it turns out, again, training your own model, certainly pre-training it, not helpful. Yes, you need data to fine-tune, but it's not actually a ton of data. And so a lot of people have access to that. And so in many ways, the, the differentiation that I think will continue to compound is like these thousand little things. And then what you can do as you get scale partnerships, other things like that. All right.

So for folks who want to reach you, what's the best way to reach out with questions to you and your team?

Yeah. So my email address is just jacob@redpoint.com. Really easy one. And I'll, I'll plug it if you found any of this interesting. I do host an AI podcast, Unsupervised Learning, which you can find if you just Google Unsupervised Learning Redpoint, and we talk about a lot of this stuff. Yeah, this was, this was a ton of fun.

This was so fun. We'll get an update in a few months, see where we're on. Maybe like at the end of the year we'll check in with you again.

Yeah, sounds great.

See where we are heading into 2026. Thanks so much, Jacob. This was super fun.

Thank you. Thank you so much. And thanks everyone for listening.