Mike Droesch of Bessemer Venture Partners joins Nick to discuss The Path from Aspiring VC to Partner at Bessemer, Building Non-Consensus Conviction, The Most Compelling Vertical Applications of AI, and Why Fintech is the New Frontier in Supply Chain. In this episode we cover:
- AI Applications in Various Industries, Advancements in AI-Powered Voice Transcription Technology,
- AI’s Potential to Disrupt Vertical SaaS Industries
- Business Models for AI Software, Including Copilots, Agents, and Tech-Enabled Services
- AI Applications, Defensibility, and Market Potential
- Open Source vs. Closed Source AI Models
- Supply Chain Inefficiencies and the Role of FinTech in Improving Payment and Invoicing Processes
- Investing in Non-Consensus Supply Chain Software Categories, Focusing on Procurement and Freight Audit
Guest Links:
The hosts of The Full Ratchet are Nick Moran and Nate Pierotti of New Stack Ventures, a venture capital firm committed to investing in founders outside of the Bay Area.
Want to keep up to date with The Full Ratchet? Follow us on social.
You can learn more about New Stack Ventures by visiting our LinkedIn and Twitter.
Are you a founder looking for your next investor? Visit our free tool VC-Rank and we’ll send a list of potential investors right to your inbox!
0:18
Mike Droesch joins us today from Boston, Massachusetts. He’s a Partner at Bessemer, an early-stage venture firm with $20B+ Assets Under Management. There Mike focuses on AI applications, cybersecurity, supply chain software, and B2B marketplaces. He has invested in companies including Tackle.io, Rillavoice, Optimal Dynamics, Raft, and Goodship. Prior to Bessemer, Mike worked with venture-backed startups, advising on sales operations and analytics efforts. Mike also happens to be one of my oldest friends in venture, and Mike it is such a thrill to see how far you’ve come and finally having the opportunity to interview you here on TFR.
1:02
Thanks, Nick. It’s such a pleasure to be here. And it’s it is wild that I’m actually on the show now after I think you were probably one of the first venture podcasts I ever listened to way back years ago.
1:13
Amazing. Well take us through your journey, you know, how did you? How did you chart your path to venture?
1:20
Yeah, happy to. So I started my career as an engineer. And then I quickly took on a consulting role where I was working with a bunch of large companies helping them evaluate emerging energy technologies. And this was during the time of the first clean tech boom. So 2010 2012 timeframe, which, for anyone that was involved in that boom, there was a lot of pretty wild kind of deep science, deep tech, clean tech companies that were venture backed at the time, ultimately, a bunch of great projects and companies that were that were being found at the time, but probably not a great fit for venture funding. And so while that wave of clean tech didn’t necessarily end so well, it was a great opportunity for me to actually learn what venture was was my first exposure, exposure to venture capital as a as an industry. And so I decided to actually spend more time trying to work with with with venture backed startups, and ultimately decided to go back to business school to to work with a handful of b2b enterprise software companies that were backed by Bessemer at the time, and I just continue to be super interested in venture, did a bunch of my work and listened to a lot of full ratchet podcasts and, and was fortunate to also have an opportunity to intern with Bessemer while I was in school, and, and got really lucky that they had an opening when I graduated and was able to try and full time. And I’ve been there since 2017. So it’s been it’s been a great ride. Amazing.
2:45
And what year did you go back to to business school at MIT? Was that 15? Yeah.
2:50
2015 Yeah.
2:50
So the backstory for the show here is Mike actually reached out to me as a listener. And the two of us collaborated on like a weekly newsletter, venture weekly, with a number of your fellow students that at MIT, some of which are still active and partners in the industry. And that was a lot of fun. You’re the first person I actually worked together with on various projects and venture.
3:16
Yeah, I know, it was awesome recipes venture weekly. But it was a really fun project that we worked on for a few years. And I was thinking back to it. I mean, I learned a lot about just the industry, you know, you were really helping to like demystify the way the industry works. Before there were a lot of other venture podcasts. I don’t know, you might have been one of a handful at the time. Like now everyone in venture has a podcast. But at that time, that was the best way for me to learn from amazing folks that you had on the show early on. So yeah, it is kind of surreal to be back doing it full circle. I
3:49
was super impressed. Because I remember you’re you’re gonna matriculate and fall and we’re working on venture weekly. And then even before you started school, you got internships, working in the industry, right with family offices, or small firms. And once you got onto campus, you were like, you were like the VC guy. So shout out to Mike on that. And everyone in the audience is trying to break in, you know, if you really, if you put in the effort and do some creative things, and, you know, find your way into firms that need help, it can really set you up to catch on with, you know, a much more established firm, like, like Mike happened to do with Bessemer. So congrats on that, sir. So let’s get on to the interview. Tell us about your thesis. Tell us about your investment approach at Bessemer.
4:36
Sure, so yeah, for a little bit of context on on Bessemer. I mean, the firm’s been around for 40 plus years and in its current form, and I’d say a few things that just define like the firm’s approach overall, and then I can share more about my specific theses. But you know, firm is a it’s a global it’s a multi billion dollar multistage fund. That said the vast majority of our focus investment already checks that we write our seed through Series B, early stage deals. And we really want to engage workflows with the founders in very early stages. And then having a large fund enables us to invest a lot more capital as companies scale. But one of the things that’s kind of unique about the approach for us is we really take this what we call a roadmap driven approach. It’s essentially just a thesis driven approach. And it’s not just about like, you know, roadmap for the sake of having the blog post that we put on the website. But but very much this is how we organize ourselves how we organize teams internally, and it’s how we organize all of our investment activity. We spend a lot of time trying to go deep on a thesis, we present these internally at our offsides, we debate them, we tear them apart. And ultimately, we then go out into the world. And we try to find the best companies anywhere in the world that are executing against these theses and try to back the best founders we can find. So that’s truly how the firm operates. You know, we have a very decentralized model where we’ve got 20 Plus partners and our investment team have called roughly 50 people. And we give everyone a ton of autonomy to pick their own DC, run with them. And ultimately try to try to find things that are aligned with those theses. But But But ideally, under the radar and allow us to see things a little bit earlier than they might be, you know, totally obvious to somebody that’s not spending as much time in these industries and doing all the work on building the thesis. And so for me, personally, I’ve worked on a bunch of roadmaps over the years. I mean, at the moment, the ones that I’m most active on have been vertical AI applications, which I have, no, it’s not going to shock anybody, I think, I think a lot of the industry is focused on that. And then I’ve also gotten doing deep dives and supply chains offer in cybersecurity, some of the other ones that you mentioned.
6:44
And what is your most common entry point and check says like, yeah,
6:48
most common would be Series A, and you know, the cheque size really depends on the, on the stage of the company, how much they’re raising, but call it anywhere from like eight to 15 million, kind of roughly in that range.
6:59
Perfect. So my, my colleague here at new stack likes to say, you know, everyone thinks it’s, it’s really hard to break into the industry. But it’s much harder. Once you get in to have success and grow in this industry. You’ve seen that firsthand, talk us through like, you know, what were the key factors that helped you differentiate, find success, obviously, you work at one of the iconic firms. And I think your entry point was associate but you know, you’ve leveled up to partner, right, you’re writing checks for one of the most storied franchises out there. Talk us through how you elevated, you know, and found success?
7:37
Yeah, it’s a great point. I mean, breaking in is really hard and doing a good job staying in the industry, it may be even harder. For me just on the kind of breaking inside. I mean, one of the I don’t know, one of the things that I think really helped was just the realization that, you know, VCs, generally speaking, not not all of them. But I think generally VCs are not great people, managers, they’re not thinking about, I’ve got to hire somebody, or I need to hire an intern. So when we come up with a project, they’re not really doing any of that. And so if you’re trying to approach an investor to find a way and find something to work on with them, I just found you had to be super self directed, essentially, like tee up a project that you created, that gives them very little downside. There’s making it really easy for them to say yes to having you come on. And so for me, I you know, at that time, when I was looking for internships, I showed up to anyone who adopted me, and I had this whole pitch a couple slides on a summer project that I was going to do for my, you know, that I created myself and really just asking for permission to do this with with another firm. And funnily enough, this was like, eight or nine years ago, I guess, now, but I was pitching a deep dive I was gonna do on enterprise applications of machine learning, which is like, just funny that it’s all come full circle in that way, it’s still very much the same thing. I’m working on this a little early then. But I had pitched my now partner can benefit in our investments Cambridge office on this idea, and he like quietly sat through my pitch about AI. And at the end, he said, I have no interest in this at all. This is like far too sexy for me. But I’ve been thinking about Govtech. And do you have any thoughts on Govtech? To which I, I had absolutely no thoughts I had done, really no thinking about it. But I said, let me get back to you real quick. And he, you know, he hadn’t promised me a job or anything like that, but, but I kind of quickly hustled, went heads down for a week and work nonstop trying to figure out what the heck was going on and Govtech. And what might be interesting there and I showed up a week later with my presentation, my pitch on Govtech. And you know how I think about investing in the category and well, we didn’t end up investing in gift. To be honest, to my knowledge. I don’t think we’ve made many Govtech investments since it was enough, either. either. He felt guilty about all the work I’d done or he was he was impressed enough to bring me on as an intern. And so that was my that was how I initially broken and I think that I do think that You know, that self directedness, the ability to just be comfortable figuring out your own plan amongst ambiguity continues to be, you know, an important part of being successful in the industry. Like even, even if you’re fortunate to get into a big firm, no one there is going to tell you where all the good deals are. And no one’s going to tell you how to spend your time. You just got to figure it out for yourself. And so I do think it’s a good for people that are curious about getting into the industry, it’s a really good exercise just to figure out like, are you comfortable doing all that work before you get the job, because when you have the job is still, it’s still a grind, and it still takes a lot of self direction to find conviction in an area that may not be obvious to everyone else. Well,
10:41
love it. Perfect. So. So Mike, it came across best summers, state of cloud 2020. For you all published it on June 20. It looks like cloud infrastructure spending is projected to surpass 1 trillion by 2026. And the report mentions this Cambrian explosion of AI applications. Talk to us about what some of the most innovative AI applications are that you’re tracking, and you’ve observed.
11:10
Yeah, I mean, for me, I think about the AI application landscape as we saw this explosion of kind of the first wave of large of text based applications powered by the text based models. And they targeted all of the obvious industries that are largely text first things like legal Customer Support Chat, content generation, of course. But the thing that has been most excited today are all these multimodal applications that we’re seeing powered by the multimodal LLM that are quickly emerging. And for me, it just so much more of the work that we as humans do is not purely techspace involve some of our other senses. And so if you want to have a big impact on industries, where a majority of human work is done, you need to be able to incorporate these capabilities. And so I could just give you a couple of examples, but one that I think we’re not investors in, but one that I think is a cool product, nonetheless, is a company called Flux AI. They’ve built this cloud based printed circuit board design software. And one of the big pain points that their users have are electrical engineers, if they want to import a component to incorporate in a circuit board design, largely all the information about those components is contained in like a 50 page, Texas Instruments PDF catalog, and they have highly paid electrical engineers kind of going through the catalog, transcribing information and recreating the component in their design suite. And so they released an AI feature a couple of months back that allows you just upload a PDF catalog, it uses a combination of vision models and text models to actually extract all the relevant information, generate a 2d schematic in their design software, as well as a 3d rendering of the component. Unless you just like plug that into your design. And it also like understands all the context about how that component works. So you can ask it like, you know, what, what is pin 32 on this component do and it will answer you. And so I think those are some of the cooler use cases. To me, it’s highly verticalized. It’s using multiple models. And and it’s impacting an end user that highly paid labor force, but it stopped doing some kind of manually drudgery work. So like, those are the kinds of things that I think are super interesting. I mean, I’ve got, we’re seeing other use cases that are also plugged in together, vision and audio models that companies that are trying to help generate a police report based on body camera footage and pulling out images and narrative voice transcription or in construction generating inspection report based on images, someone takes on their phone and a narration from from somebody walking through the construction site. It’s interesting, we’re gonna see all these kinds of use cases that help bring AI into the real world.
13:47
Love it. So you mentioned audio a couple of times, you and I have exchanged some thoughts on audio and voice. And we’ve had, you know, voice automation bots in the past for things like customer support lines, which have generally been pretty underwhelming. You know, why are you so excited about the next wave of conversational voice apps?
14:09
Yeah, this is an error. I’ve been spending a bunch of time the last six, nine months because and I think it’s where i Far and away where I’ve been most impressed with the demos we’ve seen from the early stage companies. For me, I would just summarize it as you know, all of the voice applications that we’re used to dealing with largely have like a static kind of logic tree approach, you know, if the user asked for reservations, you know, route them here, but they’re very brittle. And so, as the first step forward is just LLM comes as the conversational logic interface are far more flexible. They can understand way more context or you know, and are much better at routing users to the actual the intent that they might have had depend regardless of how they ask it. So today, we see a whole suite of you know, really like an explosion of voice applications that are using what’s called this cascading architecture, you kind of take audio in, transcribe it to text, feed that text into a, pick your LLVM off the shelf. And then they feed the text response back into a generative voice model to complete the loop. And even those are, like, dramatically better than the prior generation of applications we’ve seen. But they’re still not the most elegant architecture. I mean, these because there are many steps involved, you have a lot of latency that gets introduced, you actually lose a ton of context, when you transcribe the voice from from voice to tax, you lose the tone, and the sentiment and sarcasm, all that stuff gets lost. And so while they are dramatically better than prior generations of voice automation, I think there’s still leave something to be desired. I think the most exciting thing is kind of right around the corner, which is, for anyone that’s seen the GPT, four o demos, that’s as far as we know, like the first demonstration of a truly speech native model is taking raw audio in and raw audio out, doesn’t have the transcription. And so the latency is dramatically lower this like, it’s like 300 millisecond latency, which is basically human level. And it’s got the ability to understand your context, it can understand if you’re upset, or if you’re being sarcastic, and it can mirror that back in the voice generates itself. While while the API is not available yet on the voice side for people to build applications on top of I think that’s one of those things is just around the corner. And we’ll we’ll see more companies released similar speech, hate of models, but that to me, it feels like just the, you know, the 10x better unlock in terms of providing a truly human like audio experience. Love
16:45
it, love it, correct me if I’m wrong, but I believe Bessemer as an investor in a bridge, which is one of the fastest growing companies in healthcare, and they capture that audio from the patient provider visit, which is kind of the first step in the chain. And then it’s got all these workflows behind it, like coding and prior offs and revenue cycle management, etc, etc. But if you’re getting that audio correct first, and getting, you know, every subsequent step, you’re taking so much time out of the provider, ie the doctor’s, you know, workflow. And in theory, you know, you’re optimizing, you know, much better capture of what happened, potentially comparing that against a database of other conditions, you know, to get suggestions about what could be going on. I mean, this is kind of a cool example. And is a bridge are they using? Are they using open AI and chat? A four? Oh, because I know, previous versions used a lot of Dragon Dictation. And that was kind of a nightmare. Not for a bridge, but other companies. Yeah,
17:51
totally. Yeah. So bridge is we’re fortunate to be investors. phenomenal company. And, well, I am not, I don’t work directly with the company. And so I can’t tell you exactly which models they’re using. I know, they do use a combination of, of open AI models and other providers, as well as what’s really helped them separate them, their performance is they’ve trained a bunch of their own models as well. So they’ve got a unique team of Carnegie Mellon, computer scientists, professors, and, and physicians, actually, the CEOs is also a physician himself. So they’ve got this really unique background that they’ve been able to meld. And they’ve trained their own models specific to the medical vocabulary and conversation, which is a real differentiator. And it’s pretty rare, frankly, we actually don’t see that many application companies that are training their own models or fine tuning their own models, but it’s been in healthcare, it’s, there’s an obvious, you know, need to understand the vocabulary, the words that are used, where the traditional, you know, generic models fall really short. So a bridge has, yes, they’re using both kinds of models. And, and it’s such a good application of, of AI for this transcription use case, even if you’re just doing this transcription part, you’re taking a what today is a highly paid workforce, who has to spend a ton of their time on, you know, there’s manual note note writing at the end of the day, and to your point, like, you know, they may forget things throughout the day, like it was this, this provides not only dramatic cost savings, but ideally a much more accurate note and capture the encounter. And then there’s all these valuable things you can do downstream. Yeah, writing, incorporating billing codes for insurance, and more and more on the on the, you know, on the on the doctor workflow, as they expand from there.
19:36
It’s it’s such a fascinating area, because I feel like healthcare and biology, you know, requires such a broad in deep a knowledge base and analysis, critical thinking, you know, and to be able to leverage AI is is, you know, tremendous gains can be can be had in that space. But Mike back to The state of clouds. So, you know, I’m reading through this report, and it states that quote, vertical SAS proved to be a sleeping giant that transformed industries during the first cloud revolution. Today, the top 20, US publicly traded vertical SAS companies represent a combined market cap of 300 billion, with more than half of these companies having IPOs in the last 10 years. So the you know, the incumbents are large and have some clear advantages, but the report is bullish on AI and LLM native startups, you know, making a very bold prediction that, quote, vertical AI is market cap will be at least 10x, the size of the legacy vertical SAS as vertical AI takes on the services economy, and unleashes new business models. So my question for you, Mike is, you know, that’s a bold statement 10x The size of these, you know, massive public vertical SAS companies? So the question is, you know, over what time period do you expect this to happen? And how do you justify such a bold overtake of these large established incumbents?
21:04
Yeah, it’s a great question. I think on the timing, you know, we were thinking about this in the span of vertical SAS, in many ways, has become more mature at this point. So think about comparing vertical AI to vertical SAS, both at steady state, like at their maturity. So this is a longer term prediction. But we we’ve haven’t been having a lot of discussion internally, my partner, Samir is really fond of saying that, you know, AI is just converting services spend into software spend. And so one of our one of my colleagues did a analysis, just looking at the size of the professional services market and business services market in the US relative to the total software market. And it’s, I mean, no one’s gonna be surprised to hear it’s massive, it is much more than 10x, the size of the software market today. And so that’s where a big part of what is leading to our bullish belief that over time, AI will be able to take a much bigger chunk of convert a much larger portion of this, of this services market into software. And particularly as we’re seeing the impact that these multimodal models have in industries, every major industry from health care, public safety, Home Services, engineering, you name it, we’re seeing it across the board. And so I think our assumption is that this will take time, and we’ll certainly see some of these vertical AI companies like displace the existing vertical SAS incumbents, but even without that, I think we’re gonna see just the TAM for software grow dramatically. And that’s what’s going to drive the opportunity for vertical AI to be so much better, or something so much larger. I mean, we’ve seen this just within some of our portfolio companies, we’ve got portfolio companies, you know, for instance, two portfolio companies that service the same customer base, let’s say and one’s a SaaS product and one today I product and, you know, before the AI product came along, this customer might have only been paying for one software product, it’s, you know, it’s their their SAS system of record, and that was their whole IT budget. Now, all of a sudden, an AI service comes along. In this case, I’m thinking of even up which is, which is a really, really cool product for personal injury lawyers that helps my automates, and fast majority of the work they do upfront in generating a demand letter, it’s they’re the primary deliverables for that industry. Even UPS able to command a, an ACV, as effectively as large as the other core software products that they were these firms were buying. So one way to think about is like they’re doubling the TAM of software expand in this industry, because this AI service just has a dramatic ROI. It’s you know, it’s helping them to, to serve as far more business with the same amount of, of staff. And thus, I think it’s, you know, in a way, it’s offsetting their services banned, and thus growing them the TAM for software.
23:57
So question for you on that, right, like a lot of these professional services providers, whether it’s a law firm, or let’s say it’s KPMG, IBMC. A lot of them their business model is hourly rates, right. And, you know, if if you introduce automation, sometimes the law profession is resistant, because it makes them more efficient, and they can’t bill as much. We just had a company that did rpa, out of Boston called pliant, a sell to IBM, you know, IBM was a channel partner that loved working with them. And they, you know, they do things like provisioning servers very easily. They’re incorporating AI to automate a lot of this, and IBM buys them. So how do you square sort of the service providers having a business model based on hourly rates with software, which does not?
24:48
Yeah, that’s a great question. And it’s actually watched our team spent a decent amount of time trying to understand like, what is the way that the incumbents make money? You know, for example, an even UPS case selling the personal injury law is Not, that’s an industry that is not operating on, you know, on the per hourly model is much more on a contingency basis. And so there it is a really natural fit for them to be able to help their customers service service more business. You know, I think we’ll see, we’ll see a mix of outcomes, we’ll see some, you know, some customers who are operating on the on the hourly model and are adverse to improvements in efficiency. But I think ultimately, like the industry will have to move that direction, you know, leaders are going to do it to adopt this software, the bar is just going to continue to rise in terms of like, how efficient you have to be to stay competitive. And I also think we may see, we’ve been looking at other industries, like systems integrators and and other sorts of like professional service providers that increasingly are trying to move to more of a, you know, an outcome based pricing model, the idea that if they can power, more efficiency under the hood, but keep the same kind of outcome based pricing, they can drive larger margins. So I think it’s good question. I think it’s one more we’ll just see over time. But the underlying pressure from competition is it’s going to drive more folks to have to adopt technology.
26:14
So this is a good time to talk business models, I really enjoyed the part of the report that talked about the three primary business models that are emerging with AI, native vertical software. Can you walk us through each of those?
26:26
Yeah, so I think the three that we’d outline would be copilots, which I think everyone’s familiar with AI agents, and then tech enabled services or AI enabled services, businesses. And you know, very quickly, I think most people are familiar with the copilot kind of model. This is software that works alongside a human is often priced like like a SaaS based product.
26:48
And like a fee per month per user. Yeah,
26:51
exactly. And it’s, you know, it’s helping to provide more efficiency to that user. And more in this, you know, the AI agent, kind of model is much more focused on actually kind of paying for a for an outcome in this case. And I think this is some people call this the pilot versus the copilot, like, you’re actually giving an entire task to somebody, this would be like, an example would be like this company, even up that I mentioned, where a law firm will give them all of the documents pertaining to to a case and their AI will go off prepare a full deliverable, they might have some human QA involved in the process. But effectively, it is a fully outsourced deliverable that has been provided back rather than helping the human with their work. And so I think they’re the pricing model shifts much more to pay per outcomes. And I’m sure lots of your listeners have read Sarah capital’s piece on, on selling the work, which I think is a great analogy for this, like, rather than selling software, you’re selling an outcome here. And then on the tech enabled, or enabled services, these, this starts to look more like, you know, industries that are much more services heavy, but that have the opportunity for AI to provide dramatically higher operating margins. And so I think, you know, the use cases where this becomes pretty obvious are things where you’ve got a high labor, high cost, labor pool, and AI can actually have a really meaningful lift on their efficiency. So, you know, we’re seeing folks try this across accounting firms, professional, other professional services, legal, for instance, and
28:25
even tech designers, right, yeah, a deck designer, five to seven grand for a deck, and they flip it back to you in six hours. And it’s pretty good. Yeah,
28:33
yeah, I think we’re gonna see more and more of this across lots of industries, particularly with like, a, an engineering or, you know, highly paid workforce.
28:42
And so that for that middle group, for the agents, you know, do you find that it, it, at least thus far, it, it usually is priced, based on outcomes, you know, based on performance versus like cost plus, like the number of cycles the LLM has to go through and the number of things that need to be analyzed? Yeah,
29:01
I think much more on the on the outcome side versus the cost plus side. And I think that’s, this is also maybe the most nascent of the categories in a way where people are still trying to figure out you know, there’s there’s often a combination of some sort of highly constrained agents that are working on a, you know, taking a task from you know, give given given an end goal. Agents trying to create a plan the steps that they want to take execute against those steps, pull them tools as necessary, like we’re starting to see that work, but a lot of them are still pretty constrained in terms of the the, the freedom that they have, and then they’re pulling in humans in some cases to provide quality assurance. So I think those are the, that may be where the biggest potential is though, to actually, people do not want to like the old saying, like, people don’t want to drill they want the hole like that’s exactly I think what what the potential is here is just to deliver whatever the outcome somebody wants. We’ve seen other companies trying to sell AI agents for software development. is where, you know, they don’t want another a copilot to help with autocomplete or the code, they want the app to be delivered. And so those are the kinds of use cases that I think are probably most compelling. And and in many cases, still, in the early days, it
30:14
seems like the market has finally come around on vertical AI applications. However, you know, there are some pundits that say, you know, these are just wrappers on MLMs. And they’re not defensible. Like, how do you think about the moat and defensibility?
30:31
Yeah, this is I think this is the biggest thing that we spend time debating internally, to be honest, because we see, we see so many companies that are growing at the early stages growing really quickly have a really compelling product. I mean, that’s almost become the status quo, given the power of all the technology that we have, at the moment. And the biggest question is often like, what’s who’s going to win out? There’s so many companies going after each, each category? And so for me, I think the areas where we’ve been most focused when it comes to try to figure out defensibility are, you know, first, like, Are you are you solving some sort of hard technical problem that helps to take you from a demo to production. I mean, this is one of the things we see is like, it’s pretty easy in most of these applications to spin up a demo over a weekend, that’s pretty mind blowing, you know, you’re like stitching together, some off the shelf, MLMs and some other features, and you like put together a demo that works really well. But actually being able to put that in production with with a large customer base and handle all the edge cases often requires a lot more engineering. That’s not It’s not defensible in the way that you’re training your own model, or you have some really unique dataset, but it’s defensible in the way that building software has always been defensible to some extent, like, you know, building vertical SAS apps alone were never super proprietary. But the people that built them faster, better understood the customer needs, were able to, you know, to build the most intuitive UI like those things, always traditionally one out. And I think we’re seeing the same things here, like the people that can solve all of these edge cases, make the the products super reliable, and robust. That’s one of the biggest, one of the biggest advantages. I think, anyone that’s got having really deep vertical, specific workflows, in some cases, they may be fine tuning their models on the vocabulary of that industry, they’re integrating into all the systems of record, that’s a big one that is again, not like an insurmountable moat, but in a near term advantage to get to get people first mover advantage. And then the final thing about defensibility, is, I just think now more than ever, that teams execution matters a ton. You know, it’s always been true of startups. But because the the time to get from, you know, zero to one is condensed, so much with with the LLM that we have available today. Teams that can move super quickly and just be really agile, I think are even that much more of a premium today. When we when we think about defensibility
33:02
there’s so much nuance in each of these verticals, you know, the behavioral, the behavioral nuances, the market dynamics, the different types of systems that they currently use, I was chatting with a founder of ours who runs a construction tech startup, and they use a lot of AI on images, but he’s like, I can’t mention the word AI with this customer set, they will shut down. And so, you know, that’s the whole relationship layer two, which is really important in enterprise, like, you can’t just, you know, build the most powerful motor and expect that like, oh, it’s bigger, faster, cheaper, like you’re gonna sell it, there’s so many dynamics to each of these verticals that need to be navigated.
33:45
I totally agree this is I’m somebody who spends a lot of time in these kinds of under digitized verticals. And I’m always conscious, that’s a blind spot of myself. And a lot of VCs, like, we think we know, this is the best technology, it’s obviously better, cheaper, faster for you. So naturally, everyone’s gonna be a rational actor and adopted. Like, I’ve seen this firsthand in logistics, for instance, where I know we’re selling software that everyone should want, but change management is real and convincing people you know about about why to adopt it. And how it works is is a whole different challenge. So, yeah, I think it’s there’s a real risk of naively assuming that because, you know, we spend all day in technology, we know what’s best for every industry, right? often not the case, often
34:32
not the case. So Mike, I’d love to get your take on the battle between open source and closed source AI models. You know, I guess, where are you spending most of your time? And how do you see this playing out? You know, as well as maybe implications for the broader tech ecosystem.
34:48
Yeah, and so you know, immediately I’m personally I’m not spending much my time investing in the model there, but we do have folks at Bessemer that are would be better speak to that. But as I think about it, in terms of the impact For the broader ecosystem, I’m super excited that we have these open source models, I think it’s, you know, it does a ton of good for the, for the ecosystem. And I’m already seeing all sorts of interesting use cases for at least the open source models where you’ve got people that either want the ability to, you know, to fine tune the model on their, you know, their vertical, vertical specific data set, or many cases where people have really data sensitive, and they want to host their own model, run it on their own infrastructure, I think that’s a role where we’re seeing open source models have a really big impact. And increasingly, there’s, you know, this is movement to experimentation with small language models, you know, much more smaller models that are highly constrained to a specific domain, where I think, you know, being able to pull from a handful of small language models, enables you to do some other interesting things operate a much, much lower cost operate on more traditional infrastructure, that’s not super expensive GPUs. So I do think open source models are continually going to continue to play a really important role for all those reasons. My, my assumption, I don’t know that this is any very novel, but I think we will continue to see most sophisticated companies use a combination of models, they’ll use the GPT, four or five or whatever, you know, the the highest and most sophisticated models are for the tasks where they really are needed reasoning, you know, that really kind of complex cognitive tasks, but then they’ll shift to cheaper models, lower latency models for other tasks, where, where that level of sophistication is not required. And so yeah, I think it’s the good news is for people building applications, like the better the models get, the more capability we have at the high end, and the cheaper, the current capabilities, get a flow. And I think, you know, it’s a good time to be building at the application layer, because every day things get better and cheaper. Is
36:54
there a big tech player that you think benefits most from open source success? Like I know, Amazon historically has done a pretty good job monetizing that much to the chagrin of many traditionalists in the open source world.
37:08
Yeah, I mean, I’ve been it’s really good question. I don’t have a view on, on who’s going to benefit most on the open source side, other than to say, I’ve been incredibly impressed with, with the amount that that meta has invested into their open source in their fair research organization. I mean, like the, the models that they’re putting out in terms of the llama series of models are incredibly high quality. And I think they’re figuring out a bunch of ways to monetize that internally. And I don’t have a view on how you know, how valuable how valuable to be to Meadows business model long term, but I think I know, it’s of huge value to the rest of the tech ecosystem. So I think, if nothing else, they’re, they’re unlocking a ton of potential for, for anyone else that wants to build on top of open source models. Yeah.
37:57
So to transition a bit here, you’re one of the folks with the deepest knowledge in supply chain in the industry that I know. And I was reading through a bunch of the content from Bessemer, from you. And the question I have for you on supply chain is, why is FinTech the new frontier of of supply chain?
38:18
Yeah, so we so I started looking at supply chain in probably late 2020 2021. And we’ve made five or six investments, early stage investments in companies that are largely providing either, you know, AI to automate what today is a manual workflow, or optimize the network in some way of supply chain. And most of those were focused on how can we improve the physical movement of goods, make supply chains more resilient, make them more efficient. And I still think there’s a ton of potential there and work that needs to be done. But that was like, in my mind, that was the first challenge was how do you make supply chains operate more efficiently. And through all that work, we it really opened our eyes to like, there’s also a second challenge that supply chain operators face, which is how to get paid efficiently and how to collect the right amount in a timely and efficient manner. And it’s for people that don’t spend a lot of time in this supply chain is unique, and that there’s a ton that happens in the real world when you’re shipping something, you know, overseas or across the country. And that ends up you know, that results in a big difference in the you know, the price that you have may have been quoted to move something and the eventual invoice you get, you know, often those are very different because lots of stuff happened along the way fuel charges increase, there were other fees that the carrier incurred. And as a result, shippers, you know, the people paying to move something and the carriers people are actually moving it end up in these months long negotiations about who actually owes who what, in the meantime, you know, nobody’s collected anything. There’s just a ton of wasted effort and capital that’s locked up, because people can’t agree on what the right thing to pay is. And so we’re looking at a handful of companies that are helping to provide book better data on what actually happened in the slag and and that’s what does that mean for for what’s the right amount to pay a customer is and then definitely help them to just resolve that, you know that transaction faster unlock capital that today is just locked up in, in these auditing workflows,
40:15
you know what the average float is across the industry? Are we looking at like 60 days, 90 days? 120? Yeah,
40:21
I think for for a lot of you know, if it’s, I don’t know, off the top my head how much is outstanding at any given time, I’d say if the average is 60 days, net SIX Payment terms, they may still have, you know, there may be 2030 40% of invoices that are that the shipper and carrier still can’t agree on. And so it may be another 90 days where they’re just going back and forth. And the way this works in practice today is there’s a whole cottage industry of freight audit and pay companies which are services businesses that shippers pay to just go through their invoices manually and like try to negotiate back and forth and they keep a percentage of the savings they find it’s like, in my mind a totally unnecessary industry because we don’t have better data and we can’t agree on what we what, what the real price should be for this transaction. And so that’s where I think we’re we’re gonna see a ton of potential for for AI to deal with all this unstructured data, all the invoices, all the receipts that came back and help actually, you know, agree on a on a common price, and thus allow folks to settle much faster.
41:29
So in this supply chain and fintech article, I was reading the market map shows various categories of FinTech enabled supply chain software, and which which areas do you find the most promising for investment right now? Yeah,
41:42
in my mind, the two areas where we’ve been most focused have been procurement. So actually, you know, initially procuring and agreeing to contracts with between shippers and carriers. I mean, there’s, you know, in the US alone, in the US trucking market, there’s some 800 billion that’s spent every year on trucking. And you’d be really surprised to hear like, even the largest, sophisticated shippers in the world, in large enterprises that are spending hundreds of you know, hundreds of millions, sometimes billions of dollars a year on truckload freight, they’re largely doing that through spreadsheets, emails, and you know, kind of maybe consultants to help run their RFP process. So like, I think there’s, there’s a ton of potential to streamline and automate that workflow. And so that’s we recently invested in Goodship, which I know you’re familiar with, it’s helping to just take on that portion of the workflow and bring more efficiency to what are the prices were agreeing to at the beginning of the contract. And then at the opposite end, is this this freight audit and invoice reconciliation problem. And so there’s many companies going after this. There’s company called loop, which may be one of the more notable ones that we are not investors in but they’re big fans of that is helping to just apply AI to helping trucking companies and enterprise shippers agree to a helping them automate the invoice reconciliation process. So those are two of the areas where I’m most bullish, we’ve got, we’ve got another investment in a company called raft AI in London, which is helping to do the invoice reconciliation workflow for freight forwarder. So actually doing the same thing, but for the International leg of a shipment, which also has all sorts of complexities to it. And so we’re in the early days in a lot of these categories, but I think they have huge potential.
43:25
So something you and I have chatted about before is a sort of non consensus bets. And the tagline here at new stack is investing in outsiders, right. So we don’t invest in typically your Ivy League, you know, Bay Area located Google train entrepreneurs. But something we have to talk about a lot is like, you know, how much outside of the norm is the founder, because if it’s too non consensus, that destroys the potential for downstream funding, even with traction, so I’m curious for your take, you know, there’s this cliche that you got to be non consensus and right to be successful, but not consensus is a risky way to invest for a variety of reasons. So so how do you balance you know, being a contrarian with, you know, the need to attract additional funding?
44:13
Yeah, I think this is a great question. And I think particularly whether you’re, whether you’re a new fund that’s trying to, you know, find an edge or you’re a younger, you know, investor trying to break into the industry, like you have to find areas that are that are non consensus, I mean, that, that’s just the, I think, really the only opportunity to gain some edge. And to your point, it’s, it’s really hard, they’re non consensus for a reason. For me, I’ve tried to do it in ways where I feel like, you know, this is just my personal take on more focus on things that are, you know, traditional peer software. So there’s obviously non consensus ways to invest in technology, deep science, Frontier Tech, I don’t feel like I have any advantage there. So I’ve tried to stick to software industries that are non obvious, which tends to be things that are Let’s digitized, somewhat legacy industries. And the big question there, I think, in most cases is eventually, you know, all of these industries will adopt technology. It’s really about the timing. Yeah. And I think that’s like you can be, you can get the timing wrong by a few years get the industry, right. And you’re still just as wrong as if you had picked the wrong industry. You know, like the timing is, is just so critical here. And so I try to spend a lot of time on figuring out the why now, like, why is all the sudden this the time that supply chain operators going to care about technology or construction companies are going to care? I mean, I think back to like some of our best vertical SAS investments like toast and restaurants at the time of the initial investment pembo. The consensus internally was that restaurants are terrible buyers of software, like why would you ever want to sell to them? And that’s true up until the point that it’s not, you know, and eventually these markets flip and being able to get that timing right is so, so critical. So I don’t know if I’m doing it right. But I do spend a lot of time trying to talk to the the end buyers in these categories and figure out does it feel like both on the technology front things have changed and on the buyer front, their receptiveness to the to the opportunity to leverage technology is, is fine, that changed enough for there to be a big outcome?
46:13
Mike, if we can feature anyone here on the show? Who do you think we should interview and what topic would you like to hear them speak about?
46:19
I would say, candidly, I’d love to, if you can get my partner Jeremy Levine. I, you know, I’ve worked with him for a long time. And I still would love to hear him interviews. He doesn’t do a lot of these, but he’s just one of the best first principles, thinkers that I’ve ever come across. And so I’d love to hear you interview him about frankly, anything but the other one that I thought of recently is Mike noop. He’s the one of the cofounders of Zapier and I’ve just been I’ve been incredibly impressed. I got to interact with Mike a little bit recently. We were early investors in Zapier, but they, you know, notoriously raised very little capital. And Mike, you know, around the time that MLMs became popular, he made the conscious decision to pivot from I think he was running like half the company on, you know, to being an individual contributor, and going back to just going really deep on all things AI. And they were an early mover in figuring out how do we integrate l AMS and into our product, and it’s been a huge accelerator for them. And I think they’re like at the cutting edge of how do we meld LM with locode applications. So I think he’s both of he’s got a fascinating story, both from first founding, Zapier and then like the second life in a way of like bleeding all of the AI initiatives internally, and so I think he’d be an awesome person to speak to about the AI application landscape. Perfect,
47:39
Mike, what book or article or video would you recommend to listeners?
47:45
Yeah, I admittedly, I can’t take credit for this myself. But I’d recommend this book, Mr. China, this was I think, one I heard Bill Gurley recommend somewhere. So that’s like the credit to him. But it’s a book about some private equity investors in like the 1980s, when China was first opening up to foreign investment. And this team of guys that goes in to essentially try to buy up a bunch of businesses in China and run the traditional private equity playbook. And they have this grand illusions of how they’re going to make a ton of money because there’s this huge market. And they think they know everything about how business works, and how to be investors. And like, the short story is it goes horribly wrong. Totally underestimate how little they know about the markets they’re investing in. It’s both it’s wildly entertaining. And I think it’s a good reminder, like, as we were speaking to earlier, like, we go into these legacy verticals where we think we know what’s best for the end user, because we’ve spent a lot of time and technology. This is just a good reminder of how much like having empathy and understanding of the end user or the market you’re investing in really matters.
48:49
Love it. Love it. Mike, do you have any habits, tactics or behaviors that are a force multiplier,
48:54
I don’t know if this is like a real if this is a habit, or if this is just kind of me being lazy. But I would say my only, you know, my only tactic here is I’m really not on social media. And I’m not like active on Twitter. I’m not, you know, I’m not doing a lot of other, spending a lot of time on social media. And I know I miss a lot there. I know, there’s, there’s a lot of good learnings that could happen. But I also save a ton of time, you know, I’d say like I just it’s I have no mental bandwidth that’s devoted to like checking, you know, what people think about my posts and stuff. So, but definitely both a pro and a con, but at least it does free up a lot of my time to just try to think about like the deeper content and thesis driven work that that I tend to gravitate towards.
49:37
It’s funny, you mentioned that I was talking with an old old friend of mine who’s built a huge Twitter following. And he’s been more silent for the past, I don’t know, four or five months. And I asked him I’m like, Well, you know what happened? You were you’re building this huge audience and he said, well, the quality, my primary goal was deal flow. And the quality of my sourcing through that channel was just it didn’t turn out to be good. And so I’m putting my effort into into other things. So I thought that was kind of interesting. Yeah,
50:06
it’s interesting. I know people that like are really good at it and it works like real impressive results. It’s just like, it’s it’s not my natural inclination. And it’s, you know, in the spirit of like trying to kind of ruthlessly prioritize what are the things you’re good at and feel like they’re a natural fit for you. That’s just, that’s just one of couplers.
50:26
Well, if you want to revive venture weekly, Sunday, it’s all yours. And then finally, here, what is the best way for listeners to connect with you? Yeah,
50:34
feel free to reach out to me on LinkedIn or email me directly. We can put it in the show notes. But my emails first initial last name@bbb.com I’m Drush.
50:45
He is my Dros, the firm. Of course, you all know well is Bessemer Venture Partners, Mike, this was long overdue. Thanks so much for doing it. I’ve really learned a lot and super enjoyed this interview. So thank you.
50:56
Thanks so much, Nick. It’s so crazy to me that you know, I learned a lot about this industry from listening to your podcast 10 years ago, and it was a true pleasure to be back on. Awesome.
51:11
All right, that’ll wrap up today’s interview. If you enjoyed the episode or a previous one, let the guests know about it. Share your thoughts on social or shoot them an email, let them know what particularly resonated with you. I can’t tell you how much I appreciate that some of the smartest folks in venture are willing to take the time and share their insights with us. If you feel the same, a compliment goes a long way. Okay, that’s a wrap for today. Until next time, remember to over prepare, choose carefully and invest confidently thanks so much for listening
