Ep 159: Aaron Cannon Is Building the Leading AI Customer Research Platform; It's About to Upend a $140B Industry

0:00People actually share things with AI that they won't share with the person. Outset is what we call an AI moderated research tool. The research like underlies everything that we interact with in the world. There are billions of surveys done a year and they are the lowest quality data you could imagine. You can have AI get the depth of that interview, but you can scale it infinitely. It's like such an obvious opportunity to be smarter. I think you ate X revenue last year. This is an insane amount of growth. How did you get Microsoft, Nestle, Google, Uber? This is the killer application for insights and research with AI, period.

0:41Aaron Cannon is a founder at Outset. We backed him at 8VC last year. He grew 8X last year. It's one of the highest growth AI companies we're involved in. This is a really fun one because he's doing things that are only possible now thanks to AI. Over 100 major companies are using him to interview tens of thousands of people to learn more about their product. The world over the next decade is going to change. We're going to have real-time feedback from everyone about how everything should work. It changes how all companies are going to work, not just big ones, but small ones as well. I'm planning on using their product after talking to them here today.

1:09Excited for you guys to learn more. Welcome to American Optimist. Really excited to have Aaron Cannon here, the founder of Outset, and my partner, Jack Moskovich. Jack, thank you for being here too. Thanks for having us. All right, Aaron, you spent a big chunk of your career as a builder all over the valley before, and you decided to take a big swing and create Outset. Why do you do this? Yeah, great question. So I spent a bunch of my career kind of building stuff and being in the industry. I came from the world of research and insights and understanding customers. Then I went on to work at a number of tech companies to build stuff, but I never started my own thing until later.

1:41and uh yeah like there's an archetype in silicon valley that i'm not uh there's the 20 year old kind of naive you know dropout founder type and i'm old compared to that and i think the the uh i waited to start a company until actually i had a kid and that was like a i kind of reset my mentality by my career there was something interesting to me about like you know are you uh working for someone else are you going and doing your own thing in the world and i thought that was really, I don't know, I got reinvigorated by that. And it happened that right as I was having that kind of a mental shift, I, uh, large language models started coming out.

2:17So this was back in 2022. And that's where I, it felt like there was going to be a huge change to everything. And we started kind of, I started thinking about it. So eventually I just, I, I, I took the swing, um, which, which I was just going to like, I think the being the older archetype of founder is like, there are some, you know, obviously there's some upsides of being the younger type. Like there is, you have a certain, like you're naive about the world and that can be really valuable. Uh, at the same time, like having a little bit more experience under my belt, like it also, you know, we sell to the biggest enterprises in the world.

2:49There is something like really helpful about walking into a room and being taken seriously, having credibility, being able to speak to executives when I've sat in their shoes. And so there's a real kind of, advantage to that too. You actually know what they need to know about their customers to get their job done really deeply and you can do it from having done it before. That's right. So how'd you and Michael connect to the side of buildings together? He has a pretty unexpected background of his own firefighting. Yeah, yeah, yeah. So Michael and I worked together at a previous company where I was the VP of product.

3:19He was the tech lead. He was the best engineer at the company. And interestingly, after he left that company, and he had done the fire academy when he was younger. he has really interesting background where he had gotten really passionate about that and so after that company he went off to be a professional firefighter yeah and so he did that I think it was during I think COVID during the worst of the fires so he was doing wildfires that's pretty dangerous actually right? absolutely yeah they have to be really strong he's a lot bigger than me so he's your best engineer and also really fits yeah that's right that's right exactly So if we have to outbuild someone or fight them, we're on the right side.

4:00Michael and I actually work out at the same gym. I thought you said he was the angsty one. He's really strong, though. Both. It's actually both. Turns out you can do both. So Michael is the younger, angsty, always paranoid one. And Aaron gets made fun of a lot for being old. You need someone like that at this startup who is nervous about things. Especially if I'm always projecting all this optimism. You kind of want someone around you saying, here's where it can go wrong. That's right. That's right. I think the job of the founder is to be always paranoid and optimistic. Like you kind of have to be extremely both.

4:32But yes, we balance each other in that way. And no, and Michael, like, I do think also like really people who are really good at one thing are often really good at other things. And that's where like he just was a very good firefighter. He figured out the systems. And then he's actually like a taken a bunch of those principles that he used in the fire service and brings them to how we operate the company today. It's pretty interesting. The cool thing, by the way, that I've observed with the two of you is starting a company with someone you've actually known for a long time is beneficial in weird ways.

5:04And maybe sometimes in unexpected ways where you complement each other in weird ways. You have always been more motivated by what is possible and what could go right. Versus Michael, I actually think, is uniquely a person who is motivated by what could go wrong. And actually having both of those things at the company has been really useful. We learned this in the first year working together that we kept butting heads about like we kept coming into like points of tension because he would frame everything against the negative. We cannot lose this. If we lose this, we're screwed. The fire is going to kill us all.

5:36The fire is going to kill. You're right. The fire is going to kill us all. And so whereas I'm like, you know, we could stop all the future fires. Right. And like, that's exciting. Let's let's run towards that. So anyway, so we started working together. He had also built Watch Duty, which is like the most downloaded app to track fires in California. And so he built that as well. And so he had already brought kind of the software and the firefighting together. So let's start simple. What is Outset? What was actually broken about how companies understood their customers before you came along with Outset?

6:05Yeah. Yeah. Okay. So Outset, I'll start with that. Outset is what we call an AI moderated research tool. So what that means is AI is leading interviews, like actual one-on-one interviews with participants who are your customers, your users, people out in the market that you want to learn from. Right? So traditionally, you basically have to – well, actually, I'll take a bigger step back. Like I think a lot of people in tech don't quite realize how big research is. So you think about like – I'll go all the way back to free market capitalism, right? Is like the whole thing, you have to sell something.

6:39The whole thing is you sell something to someone else. and that's how the whole obviously system operates so the question is how do you know what to sell how do you know how do you know how to meet their needs how do you know how to meet their needs how do you know what their needs are how do you know how to position the thing you have this you know this mug how do you know how to position this to them and say well actually it's really nice because of these reasons like everything around you has been researched right you walk down a grocery store aisle you'll actually see a thousand different like packages every one of those has been tested and those were the winner by the way of 20 other ideas that were tested so basically research like underlies everything that we interact with in the world and so i think it's like a massive massive opportunity but like uh unfortunately the tools have been like pretty rudimentary so uh the tools we've we've had like what's been broken is what we have is surveys which are you know there are billions of surveys done a year and they are the lowest quality data you could imagine these are super annoying people always spam me with these things and i won't do that's right and even if you do you're just trying to get through it right it's like it's not even you don't care It's not people will say the voice of the user.

7:40It's not really the voice of anybody. It's just somebody clicking a button. And the alternative is you actually call people up one by one. Right. And that is actually where the most the deepest, most thoughtful, most like powerful insights are. But, you know, you have 24 hours in the day and you're one person. It's not very scalable. No, it's like zero scalable, right? So ultimately, if you can have AI get the depth of that interview, but you can scale it infinitely, you wind up in a much more like, it's like such an obvious opportunity to be smarter. This was obviously not possible until the last few years.

8:13And it's probably even more possible today than it was a couple of years ago. Exactly, exactly. So, you know, we built our prototype on, I think it was on DaVinci, right? Which is the earliest model of GPT. And like, it kind of worked. but of course it was very rudimentary. Now it's like video and screen sharing and multimodality and there's all sorts of awesome things it can do now. But yeah, the why now for this is just it was not possible and now it very much is. I think there is two cool things we talked about during the race. One is we obviously asked the why now question and the why nows on this were pretty obvious and as obvious as we've seen in a long time which is computers can now talk to people and computers can also process large amounts of unstructured information and actually without either of those things, you're still going to be bottlenecked by human time of either going and talking to the people or reviewing what they said.

9:03And then the other cool thing was, in many ways, all of us companies we work with solve some fundamental trade-off in their industry. In this case, you either had large-scale surveys, which were low fidelity, or manual interviews, which were lower scale but higher fidelity. And now that you can have computers talk to people, you can have the scale of the survey with the fidelity of the manual interview. And I'll just add to that. So I started my career doing the low scale, high depth, extremely repetitive work. I mean, we used to I would run an interview. I would then and then, of course, I'd be one on one 90 minutes.

9:42Right. I could do a couple a day. I do a handful a week. It would take two months, three months to do a project. And every transcript I read four times. There was a system we had. Right. You then put Post-it notes on the wall. Everything was manual. That sounds really annoying. Yeah, it was really annoying. The interview itself was actually pretty fun, but everything else around it is extremely difficult. And then we'd be delivering this to some Fortune 500 company that would have to make multi-billion dollar decisions based on me reading the transcript four times and pulling out some good clips and notes.

10:12And that is a hard thing to do. And they're doing it, right? It still happens today. How do you get someone to talk to an AI for a long period of time? It just sounds like a cute girl or something who's interested in them. No, seriously. Are there tricks to this? What do you do? You pay them. But can you make it less boring or less arduous somehow? It's actually – yeah. So obviously paying people is part of research. That's a very common thing. But the reality is this has much lower fatigue rates than surveys. So meaning that people will go for an hour and they won't drop off. And the reason is like it listens to you and it plays back what you're saying.

10:53And the same way like if we were talking and like you didn't nod or listen to anything I say or play it back and you just asked, you know, just deterministic questions. And I know like it'd be really hard. So it makes it a conversation where it's actually you the same way a person would is engaged. That's right. And the really interesting thing about this is that not only does this scale and get this kind of economies of scale that we talk about, but people actually share things with AI that they won't share with the person. It was one of the most unintuitive things when we were talking for the first time.

11:25And Alex Kolesic, a very skeptical Serbian CIO, was like, why would anyone ever talk to a computer extensively? and then you kind of experience it and the realization is like oh people actually really like sharing their opinions but they want to make sure they're not just going into the ether they want to make sure they're actually captured so they know the ai is gonna be better at capturing exactly but it's even more than that there is a uh there's the idea around like social desirability bias so it's the bias of like i'm gonna share something in a way that like i want you to receive me in that way right like yeah like i want to be seen as a person that fill Yeah, not a jerk that's positive on something.

12:04That's right. That's right. It was made famous by a lot of political polling that there was a huge skew. I forget what year it was. Everyone will not admit they're voting for Donald Trump in certain contexts. Things like that. Exactly. Exactly. That people want to be seen as someone who votes for Obama even if they don't. And that was a huge kind of polling error. And that is a microcosm of the reality in research is that when you're talking to someone, they want to impress you. They want to look a certain way. so our first uh like our first study ever our first customer ever our first study ever like in in production was with weight watchers and that was really interesting because you're asking people about their weight loss journey oh so they're gonna be more honest the computer exactly and like weight watchers really cares about this like it was their growth efforts that were it was for their growth efforts and of course if you're marketing to somebody like you want to tap into the emotional drivers and like if you know you ask the first question they i want to be healthy It's like, okay, but like, why is that important to you?

12:55And before you know it, people are talking about how they want to live long enough to see their grandkids graduate high school. And you can immediately imagine the commercial you'd build around that. You know, we have a weird version of this in investing where our research is called expert calls. And we run into this all the time where you're talking to this expert who's actually being paid money to teach you something about a product or an industry. And you have to drill down like five different times to actually get to the truth. Because frequently they'll be like, oh yeah, and I can pay for this.

13:22and then you'll ask them like five more times and you're like but can you really and they're like okay fine it's actually my boss's boss who can pay for this right yeah i thought so that's right that's right and why would people use us for expert interviews too right like there there is a uh a huge advantage it's almost like we've we fixed the trade-off and there is an advantage of getting closer to ground truth at all times how did you get like microsoft nestle google uber like how these are huge companies they've done this a lot like why do they trust you to work with you Yeah. So it's a hard thing selling to the giant enterprises.

13:52Like there's basically, there's both like headwinds and tailwinds into approaching them. So I think there's a, like right now, obviously the technology is moving extremely fast. Humans are moving reasonably fast, right? There is a huge amount of social pressure and people want to keep their jobs. There's a lot of that. And then enterprises move slowly. And so you're dealing with all three speeds at the same time. um it is so so ultimately like what we what we do is like we try to sell very um it's very like consultative where you're we recognize that we're not just replacing something right we're not taking a thing you're doing and just replacing it you're adding something they couldn't have had us that's right that's right which means it's a new behavior and if you're selling a new behavior it's actually a much more complex thing right we actually have an idea of forward deployed researchers at our company we are forward deploying people who are experts that are professional researchers you're Teaching them how to do this you couldn't do before.

14:45That's right. That's right. And the reality is just it is really hard to take a huge division of Microsoft or, you know, the entire team at Uber and like change the way we understand customers. So explain it. So if AI is running the interviews, it might sound to some people like you're just replacing people. Like what's the remaining job for the human researcher? Yeah. Are you even are you creating more jobs in some cases for them to do? Like explain this. Yeah. So I think it's easy to think about this as an automation play. Right. But it very much is not like we have not seen that the case. So like to draw an analogy, kind of an opposite analogy, with customer support, you have a static amount of demand, right?

15:20You have the tickets you get and you're automating N percent of them and you want that to be as high as possible. And that's going to probably cut people. And that's just reducing cost and you make the money on the margin. I think in the other case, like what we're in is there's not static demand. In fact, demand has been pushed down because of the tools that we have. So if we bend the economic curve, we could do more. So if one person they hire can now use you to do a thousand interviews, then each person they hire might even be worth more for some reason. Exactly right. Yeah, yeah. So each person is infinitely more valuable.

15:50And the other thing is like if research, if you can do enough research to drive a meaningful change in sales, let's just say, just to use direct revenue. If you do the right research, figure out the right thing, build the right product and sell it and drive more revenue, you'll invest more money into that research, right? Because that's just a growth engine. It's a growth driver, not a cost center. And so what we see is that future of researchers is certainly different. They're not doing as much of the pure execution. But we kind of see the barbell where they're either playing a big role in infrastructure.

16:19They're laying the pipes for the entire company to be much more in tune with customers, to be listening much better. And on the other end, they are the experts. They're the methodologists that can actually interpret data in a really, really good way. So how does this change the corporate world's relationship with its customers? Because you have this always on dynamic thing now, I guess. You're going to be always hearing how your customers' needs and wants are changing. Because I'd imagine the surveys one year could be very different than a year or two from now in certain categories. What's that look like in the future?

16:51So we talk a lot internally about productivity improvements and how that's obviously a much better aspirational goal for how AI can help with a lot of these corporate use cases. and I think one of the really cool things observing you and the company has been you can almost start asking these like sci-fi-ish questions every company wants to talk to every customer all the time but it wasn't economical and now you can ask the question well what if you could do that well you probably don't want to talk to every customer it'd be too annoying for sure until they're in the way yes yes at the limit yeah at the limit right and like You want to know what they're thinking, I guess, all the time.

17:33That's true. And you can. And that's really cool. And I do think it changes how companies fundamentally operate a lot, actually. Just to paint the picture of how silly the current universe is. So the best always-on understanding of people we have is an NPS program. Exactly. We all use Net Promoter Scores. And it's like nobody likes it. Well, that's not true. Some people like it. But most people don't like it, and yet they are stuck with it. And all that's happening is you're asking somebody a forced choice of 10 numbers and pick one. And then you're pulling that together, building a DEX, giving it to an exec.

18:08They say change something and you go to it's like the most manual and also extremely like low value data. Right. Each person. And so so it's it is like what what it should be is a constant on the constant listening to customers. Right. So a AI moderator talking to everybody that they can. right all the time always on you should know the moment something's going wrong in your business or your product you should not not just know something's wrong you should know like why and what happened and what's the product experience and you have videos of people talking about showing you their lives showing you what they're doing wrong like so you should have this loop all the way to the to the to the point where like uh we should then not only tell you what's wrong but start fixing it for you right so let's let's give an example here because this made me think of something so at a par for example a company i started there's like 1500 firms on some are very big banks big ra some smaller family offices and they're constantly 15 years old they put billions of dollars in the product it's the number one in the market but they're constantly trying to like optimize because there's like infinite demands for different types of things for them to do and in some cases i imagine it feels like they'd want like the people to have the relationship where they're hearing from them because they want a relationship with a person because that's a really important valuable relationship for us but then in other cases there's probably lots of people we're not getting that we could be getting what parts people do what parts the computer do how do you think about this yeah yeah it's um what when people say people often ask me to like judge my credibility they're like when do i not use a i moderated right it's like uh they're testing me like when a buyer will ask that the answer is like if you are want to have a relationship with that person then have a relationship with that person right don't like put ai between you and somebody you want to have a relationship with so like a hundred percent but the question is like what is the kind of cadence at which you want to have relationship and what is the cadence at which you want to understand more.

19:53So I would like, people are all the time mixing. I'm going to have a couple of interviews with people who are like, they'll call it like a red carpet research, right? When you're really talking to somebody with deep relationship and then the rest of the year, you're sending them AI interviews, right? It's, but like, of course, like there's, there's a mixture that like ultimately needs to happen. But the other thing is like, and these are stuff we're working on is you should take that interview with them and then be able to like, that should help, without that help inform what else you should be asking them from AI and even start informing a simulated version of that person to start asking.

20:26I imagine some of these firms, just to do the Adapar example, might have like a thousand people in it and there might be people in the back office we would never talk to that have like certain things they love and certain pet peeves or something like that, right? You probably could talk to people you never talked to before. That's right, exactly right. It's like some amount of common sense is you're not going to be using this to go and talk to the CIO of a firm who is directly responsible for procuring the product. However, there's probably like 50 other stakeholders in that account who might actually want their, their, to be heard because they're using it.

20:56You know, the other interesting thing, by the way, is we live in such a high velocity world. I don't know if you ever noticed this about yourself. Like I'm increasingly more and more upset if a digital thing I'm using is broken and doesn't quite work the way it should. And I expected it to be fixed really fast. This directly is a function of that. You know, we joke that, you know, maybe by the end of the year, the thing that's possible is you see some NPS score go down, you go interview people with the AI moderated researcher, you synthesize the data, you put together a presentation, and then those tickets go into something like cognition, which actually just goes and compute.

21:35That's hilarious, yeah. So the cognition agent goes and like, fixes the thing that's like, Like, maybe a person slides off, yes, go fix that. And then it just does it. So there's two directions here. Like, one direction is, you know, we gather data. And instead of doing that whole loop you just described, we literally just, like, push a PR to cognition and say, like, or a request or prompt cognition to say, go fix this thing. The other version is that you tell cognition to do something, and it hits outset as a research agent to go learn stuff. Double check. That's even higher level. So the AI could be using you to figure something out.

22:07and then coming back to the human, ideally for now, I have you in the loop for now, and saying, I think this is what we want to do. That's right. For now, at least I want to sign off on these things. Eventually, they might run the world, but for now, I think you need to do it. That's right, that's right. It's this whole kind of idea that we're building is self-driving research, right? It's the idea that research should not actually have to be constantly pointed by humans. Humans are critical in the loop, right? Like, we are decision makers. We have judgments. Similar to a boss at a VC fund, we like to think we're critical.

22:34So even long past the time we are, I will still get Jack and other people and I'll be like, I'm the boss signing off on this, even though I've not contributed anything. I'm sure you add some value here. Every once in a while. But you have to pretend that I do. I'm going to not comment on this. Yeah, you can't comment. Okay, the funny thing is, so we've had a couple of APC build companies use the product to do really early research before they start building. Exactly. That's cool. And the CEO obviously is responsible for all the decisions. Yeah. But now they're going to be making decisions with access to information from a thousand people.

23:10A product went out and reached. I'm probably joking about that. I do think people are going to be in a live long. There will be certain circumstances where the AI maybe. I think humans just play a very different role, right? Like we do not need to be doing nearly the manual work I did in my career. No, we can all be the boss like me. We should all take credit for all the work that is being done by all these smart agents like Jack. I want to go back to something you started talking about, which is this notion of, okay, we've gone out and we've interviewed a bunch of people. And now we actually have some amount of information on what people think on different concepts and topics.

23:48And a very hot, buzzy thing in the venture community recently has been this notion of synthetic research. Talk a little bit about that and your thoughts there. Yeah, okay. So I think synthetic research is the idea of simulating someone. And so if you can simulate someone effectively, you can ask them questions and do research without talking to a person. And my take is like, I think there's it is. I don't know if it's overhyped, but it is definitely going to be a kind of side by side with human research. Like humans are extremely weird creatures and we evolve quickly and we change our minds and there's things that are hot and not hot.

24:24I think the best application of synthetic is to do it hand-in-hand with human, where you can do human research that builds this synthetic picture of someone. So what we use are digital twins, right? The idea of like, I have a digital twin of Jack. Like I did two hours of AI moderated with him. I now understand him enough that I could ask the digital twin questions. That's dangerous. I mean, is this really going to work all the time? It's not, right? So here's the thing. And so the more you do the digital twin work, the more it tells you, we don't know stuff, go talk to the real Jack. And so there's a virtuous cycle where I can, rather than having to call up Jack every week to get the latest.