Why Voice AI Could Change Software Forever | Mike Droesch

0:00 Of everything changing in AI right now, [music] one thing has the potential to transform software more than anything else. [music] Voice. But here's what excites us. Moving from text to voice drastically increases the bandwidth of data that can flow into AI. [music] And the more and better data in means more and more powerful results out. Today's guest has thought about the potential and future of vertical voice AI perhaps as much as anyone else in measure. Mike Tro is a partner of Bessemer. They're investors in a bridge, Gorilla, and Vappy, which just closes series B at half a billion dollars. In this episode, we go deep on everything voice AI, where it's working and why. Is it a commodity or the source of the biggest potential moat in AI? Where are the blue oceans and the next biggest categories? Where there are too many startups and where do we need more? After all, as we know, there's one main way that humans like to communicate. It's the reason we launched this podcast. [music] Voice. It's the future, and we're just in the first inning. Let's get into everything vertical voice AI with Mike Dro.

1:12 All right, welcome back to Verticals, everyone. Super excited to bring on today Mike Droch. Mike is a partner at Bessemer Venture Partners, which I'm sure you've all heard of. Mike, thanks for joining, man. >> Yeah, happy to be here, Nick. Thanks for having me. >> Of course. This one was in the works for a while, long overdue, and uh excited to dig in. But yeah, why don't why don't uh you share a little bit on you, how you got here, you know, what what what made you a vertical expert you are today?

1:40 >> Yeah. Uh so so to give you a quick intro, I'm Mike. I'm partner at Bessmer. I've been fortunate to be here come up on nine years and uh and really like you know I was an engineer long before this found my way into venture and I've I've just been really fortunate to to join Bessmer you know right in like the heyday of of vertical SAS and so it was hard not to to get sucked into vertical investing when I joined in the you know 2017 time frame and then obviously as uh as the world's changed a bunch since then we we've parlayed that into being focused on vertical AI applications and agents much like much like you have. So um so that's a big part of my focus today also you know vast majority of my efforts are early stage seed through series B and you know we try to be pretty methodical about having these different vertical specific road maps that we that we build over time and then and then go investig >> you've been building those road maps for what eight eight coming on nine years now.

2:35 >> Yeah. Yeah. We've done a bunch of weird things over the years. years. I mean, I've led uh road maps going back to vertical marketplaces, B2B marketplaces, supply chain and logistics. We've done deep dives in, you know, in financial services, deep dives in voice AI, and and then across the firm, my partner are covering every other major vertical you could think of from, you know, healthcare to fintech and and everything in between. >> Yep. Well, we're going to squeeze out a few of those insights today. Uh, but one that I saw and listeners maybe you ran across this too, but Vappy hit a half a billion dollar valuation in their series B announced what just a couple days ago, right?

3:15 >> Yeah. Yeah, that was just um good timing. >> That's super cool one. Uh, and we want to dig into that. But I think just to kind of like give listeners an overarching sense of where we want to start today, voice AI, right? It's it's it's been top of mind for everyone. It's a wedge that's kind of uniquely accelerated and enabled by the current era of LLMs and um you know I I think it's it's produced some tremendous growth. I think there also big questions around how defensible is that and you know what verticals it makes sense in what verticals it doesn't but why don't you tell us a little bit about Vappy and and what you're learning from that one.

3:55 >> Yeah. Uh I mean I think the voice overall has been incredibly exciting. So I I started leading our voice AI road mapap growth two years ago. And at the time actually a lot of the impetus for it was the fact that we had started seeing some very early stage vertical voice agents in industries like home services or restaurants, healthcare that were just starting to get some early traction. And so like our initial inclination was we're going to go after you know all these vertical voice agents of which we have we've backed a handful.

4:23 We've got a bridge in healthcare. We've got real in home services. But but really the it was through that effort that I got to meet the team at Vappy. And just to kind of zoom out a little bit, one of the one of the learnings we had was when we started talking to teams who were building their own voice infrastructure stacks before platforms like Vapia and others existed. It was amazing how much effort was required. This was like a 10 plus person engineering team full-time managing voice infrastructure if you wanted to have a really sophisticated voice agent.

4:55 And uh and we just started seeing the the writing on the wall that like it was inevitable that over time vast majority of software that has some consumerf facing touch point is going to have a voice component to it. And as we've seen with many developer platforms like if it's not your core business, if voice is not your core product, you're likely not going to want to staff an entire engineering team around that. You're going to look for you know some vest of breed third party platform to help uh manage that for you. And we had the chance to meet Jordan and Nquille Avapi I guess probably a year and a half ago now right when they were raising their series A and honestly just blown away by in particular like the two of them how fast they were executing and learning and iterating. I think that's you know one of the the biggest leading indicators that we look for in the teams that we back at the early stage. And then they also just had this really interesting organic developer flywheel.

5:48 So they had a ton of organic uh inbound and they had people building all sorts of products on it, many of which were vertical. And so yeah, we were uh we were fortunate enough to lead the series A back then and it's been a really fun, you know, ride with the team since. I think they're, you know, we can go into some of the the the use cases in particular, but they've while the platform is rather horizontal, they've really found specific resonance in industries that are particularly highly regulated actually, healthcare, insurance, financial services, which I think like I'm sure um I'm sure it's true for you like this been one of the biggest surprises of of this whole AI era to me is just how fast industries like healthcare are suddenly adopting, you know, AI legal you know, it's just kind of amazing some of the industries that have grown the fastest here.

6:39 >> Yeah. No, I mean, it's it's been pretty wild to see just the incredible overlap between what LM do well and the problems in healthcare. But I just I did want to double click on on Vap for a second because I think it's maybe instructive into a big question in vertical AI for kind of voice first startups, which is what parts of the stack do we build? what parts of the stack, you know, do we work with partners on? So would I guess was that part of the original thesis of hey, you know, voice is going to be a critical part of, you know, the future of vertical application layer. We want to be that info provider to all of them.

7:18 >> Yeah, exactly. I think there's I think there's a couple of use cases where this voice infrastructure makes a lot of sense. So like you know one of the examples I would say service titan for instance like obviously a vertical software incumbent in home services and and one of the hallmarks of that era vertical SAS they launched a um a voice you know AI customer customer service agent on top of that and so like that was a natural extension for them. they already have all the customers. They already have the system of record there a natural extension to to launch that product. And I think for many vertical SAS incumbents if voice is a natural extension, you know, a platform like Vappy or or another one is is just like the easiest jumping off point to get started with that and abstract away of a lot of the complexity of managing the underlying infrastructure which is like surprisingly complex to keep up at real time at the scale that these companies are using it. And so I think that's one obvious use case. And then you've also got the use cases of you know the basically end enterprises where if you are you know a large bank insurance company health system that is is looking at building either external facing voice agents or internal we actually see a lot of large enterprises using these platforms to build internal training services things like that. Um, you know, in many cases, you're probably a very specific need. You probably have an engineering team that's going to need to hook up your the whatever the boy system you're using is to all of your internal, you know, homegrown systems. And so, in many cases, like they a platform like Vappy that's super developer friendly and extensible will be a better fit than, you know, more of a like off-the-shelf application. So, those are those are the biggest use cases where we see platforms like like theirs get started. And then I'd say for you know if you are a voice voice agent startup we see lots of them get started on these platforms. I think there's always a question at some point of like if voice is your core business if it's like the only thing you're going to be building like do you want to you may want to control every little detail of that in which case we see plenty of voice startups also building you know truly from the ground up to manage their own infrastructure which obviously makes sense.

9:33 >> Yep. And why why do you call out specifically more regulated environments? Is that because there are, you know, you Flappy has built-in HIPPA compliance, you know, considerations or >> Yeah, I mean, the biggest So, the the biggest um difference is really the wanting to have very tight control between each leg of the conversation. And so I think there are some like more generic customer support agents, let's say, that can be very good with, you know, handling customer requests and and like sounding really uh natural. But the the thing that a lot of regulated industries are solving for is is more so ensuring that you've got really specific gating at every stage of a conversation.

10:17 So like if you're a healthcare organization, ensuring that you've validated a c you know the end customer's birthday before you are releasing any sort of health information to them. like those kinds of use cases require the ability to create like very discreet gates throughout a conversation that the agent has to follow and then you want to have really strict evalance around it and so I think that's one of the areas where where we see the regulatory you know companies with high regulatory scrutiny really gravitating towards these platforms where like they can control every detail of it and they have you know the the support of a platform that was built to have really strong eval suites, which I I still think that's like both the the biggest unlock and like and the area where we need the most development going forward is like the, you know, the eval process is is getting better and better, but there's still tons of room to go there if we want to unlock more and more of these like really increasingly sensitive use cases.

11:15 >> Yep. And what what do you see as the what do you see as the moat for for Vappy? like is this hey we just have the best infrastructure the best quality you know the best kind of we we're applying the best models in order to be able to pull out what they're actually saying like is it is it that more infrastructure layer mode or are there application layer aspects to this >> yeah I think there you know unfortunately like the infrastructure stuff I think is table stakes you just if you're going to serve enterprise customers the infrastructure has to be great so you have to have you know you have to have really great uptime and everything there candidly I think the two modes that they've really built over time are one um you know the developer community they've got a massive community of developers that come in self-s serve and like any great developer platform we've ever backed from a Twilio or an Oz zero they you know a lot of those users are going to be using it for hobbyist projects and like just trying stuff out but you know some small percentage of them bring that product then into their you know enterprise job and uh and end up adopting you for a much bigger project and I think you know the the developer platform is really that's ultimately a big distribution advantage which I think as we've seen in like any of these AI products distribution matters more than it ever has because it's it's so easy to get to you know to feature parry with other products. So I think certainly the just the developer love community they've built there is is a huge advantage they have and then increasingly I think we'll see them and others like them really start to build modes around quality through this kind of reinforcing flywheel where the more conversations you have with a really well-instrumented suite of evals will actually enable them to create these kind of selfimproving agents where you know increasingly they're going to autonomously have the agents run multiple different experiments in parallel, see what the results are of their agent evalu, you know, you to refine from there. And so I think that, you know, we're in the early innings of this, but to me, that's like the direction that this goes where the companies that already have broad distribution and are having the most conversations and are really good at this flywheel will increasingly be able to differentiate on on aging quality and reliability, which is sounds easy like when I say it as a VC. I think in practice like actually doing this with a really high level quality having the emails that really work and represent what the customer cares about is actually much much more nuanced and so that's ultimately excited to see these kinds of industries play out.

13:48 Yeah, I mean I think that's um I think that's a really good good segue because you know one thing we've been thinking about writing about a lot at Uklid is just you know defensibility modes, right? And I think the interesting thing about voice is that there's just a ton of a ton of metadata you are collecting about whatever that workflow is. Maybe one question before we because there's tons tons to go down in the defensibility world that eval kind of um architecture is is that something that lives in Vappy or you know is that something you're finding their clients are handling on their own or using other elements of the stack to do like let's say I'm you know I don't know I'm selling I'm selling flowers and I want to determine hey do I want the agent to talk about roses or tulips like do am I building my own stack for that is is vap happy enable that >> supports eval testing real-time observability like that whole component of it is a is a really core focus of their platform you know whether or not customers show up with their own suite of eval that they built internally and load that into vapy vap team will help help you build those or you've got a third party platform like you know vap is pretty pretty extensible but I think the majority of the you know large customers that we're working with are specifically working with the vap team to looking for that expert guidance as to what what does best-in-class what are best practices look like for establishing a voice eval suite and so increasingly that lives in in that and and the results of seeing all of these real-time evals and tests running in production is is where I think the compounding mode comes from over time >> and what about the interface and kind of handoff to humans uh we had on the pod had I mean a couple founders that spoke to this but the one that's that's that's standing out in my mind right now is um Aier over at Slice and he was mentioning that you know voice is incredible but it's about 95% right and and that you know in in a world where that could mean a lot of pizza sold in the case of Slice right uh they still have significant humans in the loop in fact he had some skepticism of like we got to wait until it's 99%.

16:13 So like are you seeing that play out in any companies you're working with? >> A lot of almost all of these you know either voice platforms or a vertical specific voice agent will support you know some sort of you know failover to a human in the event that you you hit an exception that the agent can't handle. And I think that like having that graceful failover is actually like critical to making these things work. Ultimately, I the way I view it is I don't think it's surprising because most of us only call something. You'd only call your bank if you couldn't already do it online for the most part. Like you probably tried everything else. You're like, I really I have to call the bank now. This is going to be horrible. And so most of the time it's you're already dealing with a bit of an edge case. And my take on this is like voice agents two the two big unlocks that we've had over the last two years was one like latency finally got low enough where these things sound you know you could have an end to end turn of a conversation that's in the range of 500 milliseconds. It's like snappy enough that it's not frustrating to the end user and it sounds almost like a natural human conversation. That that was one big unlock and the other was just when the LM started really supporting tool calling well because ultimately like I don't want to just talk to a voice agent. I wanted to do something for me.

17:27 I wanted to take some action. So that's the I think that's still where we end up in a lot of these edge cases that the agent can't support is somebody called in about a really kind of like esoteric or complicated issue and the voice agent might understand exactly what the issue is and what to do about it but they don't have the capability to actually take the action because the end customer is probably still like in the early days of rolling it out. I think all of these voice implementations we have right now are in like very much the first inning.

17:56 And so I do think I think you're what you're saying is totally right. In many cases, you know, we're still going to have that, you know, 5 10% of calls that that need to be escalated to a human. But increasingly I think the gap is not can agents handle it. It's do we trust the agents enough to give them all the tools to take action for us. So like you know a lot of the you know the use cases that are enabled are still pretty simple but increasingly and I think this all this all comes full circle like if we have the evals to feel comfortable giving the agents more autonomy because we know they're not going to go wildly off the rails then we can hook them up to more systems and we can let them like actually take action on more complicated tasks that typically would get escalated to a human. So, I do think it's just like a bit of a, you know, a maturation curve that we're on where increasingly we'll see people hook up these agents not just to like their scheduling system, but they'll hook it up to like core systems and be able to allow you to move funds and do more complicated mission critical tasks via via voice, which today I think everyone's still like tiptoeing into into this new world.

19:01 >> Yeah. No, you're you're totally right. I mean, even still, I feel like the vast majority of use cases are customer serviceoriented and very like blocking, tackling, like, hey, we're we're going to help you get a meeting scheduled or reschedule a meeting or, you know, or whatever it is, right? Or or or put in a very b like a like slang, right? Put in a put in an order. The next step seems to be what you mentioned, which is almost like training, right? because it's already so voice oriented, right? So obviously you guys have a lot of experience, you know, with that from you, you know, but I mean what's another like maybe VOCA does a little bit of this. I I don't know.

19:45 >> And there's a lot of sales training use cases that you see where it's, you know, a sales rep has to pass a certain amount of, you know, hours with a with an AI trainer to before they're allowed to to speak to real customers. like we see lots of those kinds of internal training use cases. The thing that also gets me excited are just I'm I feel like we're just starting to scratch the surface of some of these use cases that just flat out were not possible before voice and so like I this is a maybe a niche one to investors have always resonated with me but there's a >> we we love niches here. This is um there's a company in New York called Qualitate which I I think is a really cool business kind of like a AI meets expert network and you know one of the use cases that they support is basically using voice agents to go out and do expert network calls on your behalf. Uh and so you know they shared an example with us of they had a corporate customer that was running through an M&A opportunity. they wanted to do, you know, as many calls as they could on a potential target that they were trying to acquire. And like, you know, I'm sure you've been here, like this would take an analyst or associate, you know, an hour per call and then you're transcribing all your notes and it's like a super manual intensive process.

20:58 like this company uses voice AI agents and goes and does 200 expert network calls over the course of a weekend and you show up on a Monday and you've got like all of this perfectly compiled feedback in a use case where like voice really matters. I think like doing interviews is one where you really want to be able to listen to the tone and the feedback that somebody gave you and figure out, you know, that prompts you to ask the next follow-up question. You wouldn't get the same results doing the entire thing over chat or email. And so, so I think like these kinds of use cases where you can now scale exponentially, you know, infinitely, whether it's expert network calls, whether it's interviewing for recruiting, obviously that's a use case that's just taken off.

21:37 I think those are like awesome use cases of the the scale that we get with voice agents which you couldn't have before. And then we're also starting to see people do interesting like mixed modality stuff where you know you'll have the combination of I am watching your cursor navigate my app on a screen and I'm talking to you at the same time and so I can kind of coach you through no no move a little to the right click that button like from a customer support standpoint let me help you navigate this app with a combination of voice and like a computer you know user monitoring app.

22:08 Uh and I I think that's also like this is all kind of foreign stuff. we're still just starting to see people scratch the surface of of what's possible. And similarly, like bring it into the real world, like, you know, I think industrial inspections, you know, we've seen this in roofing. There's a lot of roofing use cases or construction inspection use cases where somebody's going to be talking to an app and taking pictures and stitching together all of that data to to bring AI into the real world.

22:34 >> We wanted to take 10 seconds to call out our partner and supporter on Vertical, Paraffin. You already know why embedded finance is important. You want to power all the ways your customers make money and grow their business. But you don't want to be a bank. You don't want to be a credit card company. You don't want to be a lender. Herafin can help you with that. They help portfolio companies of ours, friends of ours, hundreds of other vertical founders increase retention, increase revenue, grow their TAM. So if you don't know Paraffin, please take a second to check them out. You can go to paraffin.com.

23:07 That's 1F or the link is below this video. Thanks for listening and back to the show. Yeah, I mean I I we're we're super bullish on this, too. And I I feel like um in a lot of ways, I'd love to see more businesses get creative with voice because I I think a natural assumption kind of maybe, you know, springboarding off of past experiences with uh voice AI is that people aren't going to like the interaction.

23:37 But I don't think that's inherently true. Like a lot of the anxiety people have around talking on the phone are like, "Oh, there's another human over there judging me, you know, I've I've talked a little bit about, you know, some of these voice consumer products like turns out like people can kind of enjoy the conversations sometimes and because they have perfect context, there's a lot you can do with that data." So, you know, one one interesting thing I'd love to get your thoughts on is like as we think about what makes vertical AI or vertical application layer defensible, frankly, I think any application layer, I think we're moving away from what's important is the system of record in the sense of like the database itself or the UI itself or the BI itself like not kind of fungeible now, right?

24:27 what what is is like what that system of record can you know tried to contain in the past which is like the context around the business like is this a hot lead is this not a hot lead what should I do if it's hot or not right and like that abstraction from the human brain like that's the only way to do it is to like mind that process keyboards like super low bandwidth voice much bigger opportunity so like I don't know like with Uh if you're understanding how the best sales reps sell, yeah, you can train them. But theoretically, that's an asset. There's a lot more you can do with if you learn like, hey, how does contractor business X like successfully sell, you know, new roofs or whatever, right?

25:14 >> Absolutely. I mean, I think there's there's a a ton of rich data in here to your point, like the the sales use cases are great because it's it's not just about what exactly you said. It's about how you said it. like, you know, how much time you spent listening versus talking. You know, how how well your sales rep interpreted the sentiment and the tone of the other party. So, we've certainly seen that in these in these coaching and feedback tools.

25:39 I also think increasingly, this is like a kind of a a little bit of a tangent, but I just think increasingly we're seeing passive voiced, you know, think every conversation that gets recorded by Granola, like it's just like a wildly valuable data set that we didn't have traditionally. It's just that captures everything about how we all like really do work and and how decisions really get made. So like in terms of you know your comment about the systems of record like your system of record might only capture you know the end state of this is the decision we made the purchase we made this is how our business process you know terminated uh and or where it ended up but all the discussion about how do we get there obviously isn't captured.

26:23 So that's another we we have a company called Axiomatic which is helping to do this in in large enterprise uh transformation. So think like enterpris is going through a system of record transformation SAP migration. They're in they're bringing a you know a consultant in a system integrator to to help migrate um and one of the biggest tasks they're doing is like workshops where you all discuss how does the business process work so that we can then map that to the software specs for the next version of the the system we're implementing. and transcription is just like such a good use case in in these um these situations where we've got lots of people in a meeting. Somebody traditionally was just taking notes uh and now we're capturing like the full context of all those meetings that might be happening parallel. No one person could sit in on all these meetings. Suddenly transcribe them all and you know you feed that as context to an agent that now has perfect clarity on all of the edge cases that we need to plan for the way the business works and and can put together a much more comprehensive migration plan. So I think that like this version of always on uh you know I've got one CEO that wears like an always on his wrist um you know recorder transcribes his entire day puts into claude at the end of the day like I think we're start seeing more and more of these patterns where it's just that that context is increasingly valuable and uh I haven't been able to catch >> I mean do you do you use granola?

27:49 >> We do. Yeah. So, >> I mean, as as do we. And I don't know if you ever have this feeling if you like forgot to turn it on and I'm just like, "Oh, man. I'm so screwed now." Like, there you become so reliant on it so quickly. It's unbelievable. So, I wanted to hit really quick like I I was just thinking across, you know, verticals specifically like where we've seen it kind of take shape and where we haven't. So, you know, we've talked about uh you know, we've talked about a bridge. I think those are two nodes where we've seen a good amount, right? Like so kind of blue collar, you know, we've got Aoka and Broccoli and you know, maybe you could say NetIC and then we've got obviously like you know a bridge in in Freed and in that whole healthcare world which which everyone knows well. I think we've seen a little bit in insurance right um or like call it related financial services uh liberate salient uh but oh and and I guess Elise is maybe the big like you know kind of leader in in in real estate but what I guess what other categories have you seen what do you think is like underdone overdone >> uh salient was the one you mentioned that I was going to mention in terms of just I think like they found a you know a really interesting use case within within financial services where I I expect they'll be able to know expand over time because we we haven't seen as many you know we've seen lots of things like maybe slightly overdone I say like lots of um debt collection use cases scheduling at large we've seen lots of scheduling recruiting obviously like I think those are some of the you know the most high when you think high volume calls that you don't necessarily you know want to be stabbing humans to do I those those are the ones that that come up the most. Uh in terms of other ones that I think are less, you know, where there's still more opportunity.

29:59 We've got trying to think what else. I I mean, so we're starting to see actually some examples in legal where you're doing like intake for personal injury law. you know, these are like complex conversations where yeah, you traditionally wouldn't have thought of using agents to do this, but I think we'll increasingly start to see um agents tap into more and more of of these kinds of like higher higher dollar value, more more valuable workflows. So, legal healthcare I you know the we've got the assort health world doing you know booking and scheduling things like that. I think they'll continue to do more and more within voice. Um, ultimately like really the the sky's is the limit in terms of what we can apply these things to. I I do think you touched on something earlier which is there's still a human stigma around it.

30:52 You know, I think everyone is just so used to the fact that when you got an agent on the phone, you know, a traditional IVR um kind of agent, the first reaction would be like hit zero and and you know, start yelling for a human. I you know, we're we're going to be I don't know, I'd say a year or two from now, we'll be in a period where, you know, hopefully all of us, like hopefully your parents and like, you know, people who don't live in this world every day are starting to rave about some amazing experience they have with like I was I think I talked to an AI today and I'm not sure. And it was like this delightfully helpful person on the other end of and and ultimately the world I imagine we'll get to is a place where we're actually actively trying to get to the AI agent because it's it's less stressful. But we're not necessarily, you know, dealing with the uh, you know, the wide variety of quality that you might get talking to a human agent. And as long as they're snappy and able to resolve your issue quickly, I think that's ultimately what everybody wants. And so I think we cross the stigma bar, we get the eval high enough quality, then I think there's no reason why this doesn't become a much more normalized thing that you can imagine talking to, you know, as the, you know, kind of first interface for any professional service. um for much more higher dollar value, you know, transactions, interactions, but there's still still a little ways to go there.

32:11 >> Yeah, I totally agree with all that. I mean, I think it's like look, if if a if an interaction is higher dollar value, then inherently folks are worried that that 1% of bad experiences like that's going to kill the deal. And so, you know, I think as we get better human in the loop, better like real-time monitoring, like also just gets behaviorally better, um, that will change. And then I think there's huge huge runway for internal use cases as you were saying like there's just there's no reason why every process shouldn't involve it because like at the end of the day, that process mining like that is where I think defensibility at the application layer goes is that sort of data. Um just you know >> interesting anecdote the dollar value use case. This reminds me two years ago I was looking at the very first iteration of some of these home services voice agents and I spoke to some uh you know some kind of I think it was a roofing company where their average ACV was like 30k and for them they're like there's no way we'd ever trust an agent like our dollar value you know every lead is potentially worth $30,000 like we're going to staff a human every time and lower ACV home services were very happy to use this stuff and so like there was a clear delineation even there Today I think that you know that world has really changed and I think you talk to a lot of roofing agencies even or roofing companies they'd be happy to to adopt this stuff. So I think we've even seen even the last two years just like total change in the guard there.

33:41 >> Oh totally. I think there's also an inherent like faulty assumption in that that higher dollar value is necessarily going to turn because like in a multi-touch sale, you're used to having some bad touches and you're assuming the humans are good, right? There's probably 20% of your reps who like just suck honestly and create bad experiences. Now, if I'm trying to order a pizza and I just literally want to do one thing for 30 seconds and your voice say I can't do it, then I'm going to get frustrated. So, I understand like Aar's point there, but I think what's going to be much more normal, like I'd rather talk to an AI than some like like literally like Ctier rep who like just sends me in the wrong direction. So, I know we're we're coming up on time here, but I wanted to I wanted to maybe run through a few like other kind of questions for you and we can just do these maybe a little bit more, you know, kind of rapid fire style, but just just want to like hear, you know, what what you're seeing in your seat over at Bessmer. So I guess the first kind of more rapid fire here would be like just monetization like what are you seeing working in the world of of AI native modernization and not and like how's how's everything changing?

34:55 >> Yeah, I mean I think the very quick hit which I think probably lots of people have talked on is now we're seeing in cases where you can replace labor we're seeing companies capture like dramatically larger ACVs than their SAS predecessor. So, you know, I've looked closely at an accounting company where like, you know, the the um the predecessor for like a closed management tool might get capture a 30k ACV and the AI agent version that's actually doing the entire accounting workflow end is 150k ACV like just doing the labor obviously you know tapping into the same category and you're dramatically growing the software budget by uh by taking on some of the opex budget versus just it budget. So I think that broadly we've seen across lots of industries that historically were really hard to sell software too. Legal one of the best examples in terms of like pricing. I know it's uh lots of people are excited about outcome based pricing and I think it's like totally makes sense. We have I think there's frankly seen few companies that have really gotten to a crisp outcomebased value metric. Like I'm sure you know the customer support platforms like Intercom's Finn being one of the best examples like we can charge a dollar per resolve ticket because the tickets are relatively homogeneous. We have a pretty good sense of how much effort it takes for to to resolve those.

36:15 I just think you know maybe we're we're still in the first kind of early innings of this but uh I haven't seen that many other industries where they're able to like pretty clearly nail down this is the outcome and this is what we're going to charge for it. That said, I think everyone at the moment is in some some version of like platform fee plus a usage bundle which you know feels a lot like the SAS era where we had usage based pricing in in tiers and I think that's kind of where most of the market is right now.

36:43 >> Totally. I I think at the end of the day both both sellers and buyers want some visibility on what they're going to pay. And so like usage based makes sense, outcome based makes sense, but like maybe it's hey you got to buy this like pack of you know credits or you know some creditled model. What maybe just like quick follow up there like have you have you seen any sort of transactional plays here like hey we're going to take a percentage of revenue savings or like any kind of novel transactional approaches.

37:18 >> Uh I think our healthcare colleagues have for sure on the revenue cycle management side of things like that's obviously was a kind of pretty established norm there. So I think that's where we've seen it primarily. I can't point to a ton of other examples. Yeah. >> No, I think you're right. That's that's been the biggest one. I mean, I think um you know, may maybe also around some financial services like collections or something like you said. Okay. So, so maybe one or two others here. So, just h how how has this era of AI change what you look for in a series A?

37:53 >> Yeah. I mean I think increasingly the hardest thing to distinguish is is not uh you know just like is the product magical because they almost all are. I my my going in expectation is that the product is going to do something that's magical. The things that we're looking for increasingly are trying to get ourselves to some conviction on how will this you know how will the company establish defensibility in the future. It doesn't have to, you know, be day one, but we're very much looking for a team that can articulate how they will get to some sort of flywheel that builds defensibility. Now, some sometimes that is truly just velocity. Like we are going to outsprint everyone else in this market and it's not the it's not the deepest mode, but it uh you know it it is a proven mode. I think the bar is way higher than it ever used to be. like what the fast what fast looks like today is um is crazy relative to what it used to look like 5 10 years ago. So I think the teams that are that are just you know using velocity as as their primary remote have to be prepared to move at an insane pace and so and sometimes that's that's enough. We've we've got some really quickly growing companies our portfolio that I think you know are doing just that. And so I I think increasingly what we look for are like teams that can iterate, learn really quickly, ideally have, you know, a compelling thesis on how they're going to build defensibility over time because increasingly the, you know, the amount of time you have once you've demonstrated a product works to to when you see fast followers is is just compressed so much.

39:23 So really looking for speed and um and then just like this this uh really fast learning trajectory I think. >> Yeah. >> Yeah. It's it's it's interesting because I mean I've we've had a couple other folks um who was this I think it might have been Jake at Emergence and and others have made this point that like along with this fast growth is maybe less you know history to understand what retention will look like. I don't know if you guys have have seen that but it's like man if this is growing like 5x year over year like company's been around for two years like I mean we just hope it holds up.

40:03 >> Yeah. The retention thing is uh it's a unique animal. I mean in some way like we see a lot of these companies particularly in the companies that are more proumer where you've got this this dynamic where it's just it's hard to know what the the retention will look like particularly if it's if it's truly kind of bottoms up proumer. That's I think you know we've got a lot of with within the c the category of companies that I primarily focus on like they're selling more true B2B more enterprise of you know the I think in many cases the retention the retention still looks pretty similar or we'd expect it to still look pretty similar to to B2B you know or historical B2B SAS benchmarks you know at least once you got through this like self-s serve funnel assuming there's like some sort self-s serve funnel you might have a ton of drop off of growth but once you graduate to a wheel, you know, B2B contract. We expect the a lot of these companies to look pretty similar.

40:55 >> Yeah. No, totally. I I think it's probably more relevant to consumer, proumer, you know, kind of first wave like app builders and such, but um maybe last question for you. So, you know, we've largely covered voice AI today like on everything you've learned, you know, across Vaporilla, etc. So like if you're you know talking to founders listeners here that are thinking about building in voice AI are building in voice AI like any any tips pros cons like one or two things to look out for you would you would share with them? I think ultimately what what I recommend for anyone that's building voice agents, I'd like really increasingly try to challenge yourself to ensure that you're solving a problem where you're where you can actually solve the the user's entire journey end to end. This is a little bit of what I was touching on at the beginning, which is like I I think it's very easy for people to build voice agents that will have a delightful conversation with an end customer and then hand it off to a human. And so really think log art about are you in a use case where the end customer is going to trust you to fully resolve the the customer's um query and I think like scheduling is a great example like it's very low stakes of course you know a restaurant will trust you to schedule reservations whether or not you know you can handle more complex sales conversations more sensitive conversations I I think that's just going to be like this this bar that I wouldn't want to be you know spending all my time building a voice agent company if I didn't think I could do an entire job end to end. Um, and ideally that's a really economically valuable job. And so I would I would you may have to convince the customers a little bit to to to trust you in the beginning, but I would really, you know, pressure test whether or not you'll be able to get customers over the over the hump there on ultimately trusting you to to own that workflow and end.

42:54 >> Um, I think that's an awesome point. Because ultimately if you don't like that incredible like step function scalability might not be there. You might actually be delivering customers more work at the end of the day. >> Uh well excellent insight. Uh per usual Mike um thanks so much for joining man. This was super fun and uh you know hey we we'll have to uh let me know if I can share any thoughts for your next road map because I think we're all excited to see it.

43:24 >> Awesome. Yeah, I think we're all trying to figure out what comes next. It's it's a super exciting time and also a little confusing. So, we'd love to uh we'd love to figure that out together and appreciate you having me out. Of >> course, man. All right, that's a wrap. Um