AI 代理的未来｜Jesse Zhang 访谈

1. Every Brand Will Talk to You Through AI

What we are building toward is this concept of a new UI for the product, right? Just think about how much these large companies invest in their mobile app or their website. This is literally how everyone communicates with them. So this could be a way where, eventually, the way any user interacts with any brand is through an AI agent.

Perhaps an interesting place to begin—I bet you don't expect this—is for you to tell us a little bit about the phrase you've talked about that's written on your wall in the office.

“There's no challenge that can't be overcome, and there's no enemy that can't be defeated.”

We just like it. I think it really fits our culture. We have a very competitive team, and everyone wants to win. We have a lot of energy when we're trying to go out and build because it feels like, hey, there's all this stuff happening in the industry, and we have such a strong team. Let's just go and win.

My dad told me this—I don't know if it's actually validated or not—but at Huawei in China, which is obviously a massive company, they're known for having a killer culture. They have some version of this written in Chinese, of course, in big red letters across the back of the big hall. So, I really like that, and it's kind of interesting anytime someone walks in. It stands out. The quote stands out.

I asked about it to begin because I'd love to spend some time on what this environment is like when you're building a company like yours against formidable competitors in probably the most exciting era of technology that any of us will ever live through.

2. Competing in the AI Era

Words like “defeated”—I've heard the word “violence” recently in reference to a company culture, and “aggression.” These are not words that were being used 3 or 4 years ago. In fact, if you used them, it was a big problem for you. Obviously, that has completely shifted, and not only have founders started using these words, but I think talent and senior people have rallied around them. They want to be in a culture that is defeating enemies, or violent, or aggressive.

Maybe just riff for a while on what it's like to be building and competing in one of the big areas—in your case, conversational AI—at this moment in time, because it's so different from what it was a couple of years ago.

Our whole view is that any space that's worth going after, any large, hot market, is going to be competitive. This is not really specific to AI, right? If you think about Databricks versus Snowflake, or Ramp versus Brex, anytime there are these big, massive growth opportunities, people are rationally going to go after them.

It is exciting, but of course everyone knows it is competitive, because if there is a market, then everyone's trying to build it. It's quite easy to start a company these days, and you can raise funding really easily. So when you go out there, you have to have a pretty deep understanding of what your competitive advantages are.

One of those could be the culture. If you build a good culture, that's pretty hard to replicate, and it also lasts quite a long time. To your point, if you have a culture where everyone is really geared toward succeeding, working hard, and having some level of intensity, it can go a long way.

I do think, from my observation, this generation of company building has attracted a demographic like mine: people who grew up in fairly competitive environments, whether academics or whatever. A lot of these founders have done well. They did a lot of math contests and coding contests growing up, and a lot of the people around my age are doing startups now and doing quite well.

I think there is some element of, hey, if you really embrace this sort of hardcore lifestyle or environment growing up, it lends itself pretty well to the current situation because there are a lot of parallels.

I was with Scott Wu from Cognition recently, who is well known to be one of these math champions—you as well. Talk about that environment. What was that competitive market like? What was it like? Let people into that world, because obviously it shaped you a lot.

When you look back now that we're all grown up, it is kind of a pretty intense way to grow up. But I look back on my childhood with very fond memories. You get a really nice community, with a lot of people who are doing the same thing.

I grew up in Boulder, Colorado, and Boulder is a pretty academic town, but there aren't very many people who are super gunning for these contests. A lot of my friends grew up in places you would think, like California, Texas, and New York. It gives you some level of community where you realize there are a lot of other people out there doing the same things as you, and you get to meet a lot of them.

Now that we're all grown up, those relationships have lasted for a long time. The environment is similar to company building: how well you're doing is fairly objective. There's a lot that has to go into it, a lot of preparation, and a feeling that it's a long-term thing. But because your results are fairly objective, there's constant motivation to improve, and I think that's quite nice.

One of my big theses is that there's a lot of untapped potential there in actually turning them into people who want to do business or build companies and things like that. A lot of them have historically gone into trading or academia. Those are perfectly awesome jobs, but a lot of the folks with this background are correlated to a path that's a little bit more risk-averse: get your good grades and follow a track.

If you can divert some of that talent into company building, I think there's just a lot that can be done.

What are the parallels? What is it about the people, the training, or the specific aptitude in the competitions that makes you and others good at company building? This seems like an obvious, true trend. There's enough of a sample size now of people who had this background and are doing extremely well in this environment.

Maybe it's the best. If you could somehow index that group of people, you'd have fantastic performance right now. What are the parallels that make that true?

One is the competitive nature that we just talked about. The other is problem-solving.

One thing that I believe very strongly is that one of the best things you can do when you're building a company, or anything, is to have good introspection about where your strengths are. At least for this sort of group, it's more around the problem-solving side: really figuring things out.

The problem could be very vague, right? The problem could be, “How do I build a successful company?” You break down the problem, think really critically, and use first principles to analyze a lot of these markets.

Of course, that's not the only way to build a company. Some people just have very good intuition, especially around PLG-type things. In general, I think the problem-solving capability is important. What these contests really teach you is that you're solving problems. Those problems, of course, are not real problems in real life, but the sort of thinking is—

Does it feel more emotional now, company building, than it did in the contests?

I'm a bit older now, so it's not really that emotional. I started a company before this that was, I would say, way tougher than the current one. It just mentally makes you feel a lot calmer and, honestly, appreciative of when things are going well.

Timing matters a lot, right? I graduated a year early to try to start it, and I would say we had a very fortunate outcome. But the whole journey was very bumpy.

What was it? What did it do?

The company was likely called Lowkey. We basically built high-performance video-capture software for video games. When people are playing games, they can easily capture video clips, edit them, and share them. The whole goal of the product was to get as many users as possible.

We exited at a very fortunate time. It was 2021. At the beginning, we didn't really know what we were building. We were just trying a bunch of things, and as a new grad, you don't really have good intuition about what idea is good or not, so you just work really hard for 3 months.

Then you realize that what you're building obviously had no market, and that's tough. About a year and a half in, 2 of my good friends from school, who were my co-founders, both got burned out and decided to leave. It was just me after that, trying to figure out what to do.

I think that’s way tougher than anything we’re doing right now because it’s an emotional thing, the way you put it. You don’t really know if there’s any future or not. There’s a lot of pressure because you don’t want the first thing you work on to be a failure, essentially. So I think that was much tougher from a psychological point of view.

Nowadays, I think it’s tough just because of the sheer amount of stuff to do. We’re not getting much sleep. But again, in the grand scheme of things, you’ve got to be really grateful to even be in this position—to have enough work to do, to have enough interesting problems, and to have a team that you’re really excited to work with every day.

How much do you think you’re sleeping?

It really varies. This week in New York, it’s probably 4 to 5 hours a day. And it’s not good because I’m not someone whose brain functions that well on less than 8 hours.

Yeah. So if you think about the difference between the first and the second company, obviously now you have a company that’s working extremely well. It’s growing really, really fast. What did you do at the beginning of Decagon that was informed by your prior experience to make this one go better?

I think this is broadly true of starting companies. I think the first stage is by far the hardest because you’re just finding direction, right? And finding direction is very difficult because, by definition, it’s not something where you just have a goal and you’re grinding toward it and can get there. I mean, your goal could be finding a direction, but it’s more exploratory.

I think what we’re actually quite good at now—and likely Ashwin, my co-founder, who’s amazing, also has a similar background—he started a company before that got acquired. When you start a company the first time, you often share the same experiences, which is that finding a direction is very difficult. So I think this time we were a lot more thoughtful about it, and again, it goes back to what your strengths are.

I think we view our strengths as, “Hey, we’re very good at problem-solving. We’re very rational about things. We’re good at execution.” So if that’s the case, then I think we just try to systematize the ideation process. It comes down to this: Whatever you work on, it has to be something that people will really invest in. How do you tell if that’s the case? You can go really deep asking them.

I think people are usually a little bit embarrassed or not comfortable going super deep on these questions. But when you talk to potential customers, they actually don’t mind answering questions such as, “Okay, if we built this for you, exactly how much would you pay for it? Would your boss need to approve it, or your boss’s boss? Who needs to approve it? How would the entire organization think about ROI? How would you present ROI to leadership to protect yourselves and also make you look good?”

3. Finding Product-Market Fit

If you really go deep there, it’s almost like you’re asking classic sales-qualification questions, but in founder form. Because you’re a founder, it feels a lot less salesy for you to go deeper, and that process gives you a lot more signal.

When we first started, we were fortunately able to get in front of a lot of large companies, mostly digital-native ones. We just asked these questions, and we kept digging in. We had a bunch of different ideas, right? At the time, we weren’t tied to any idea, and it was just open exploration. We were looking at things ranging from data analysis to security to pre-sales to operations stuff.

That process was very helpful because it showed you that there’s a lot more signal to gain than just talking to customers, which is the general advice, or building something customers want—building something people want. Yes, that is true, but it’s very hard to just know what they’re like.

They’ll tell you what they want, but it turns out that that’s—

So what was the literal process? Would you go to a single person and ask them about multiple ideas at once, or would you target it more like one idea to one person?

You have to target it based on what that person owns. But if that person is very senior, you can ask them about multiple things.

Yeah. So if you talk to a COO, for example, you can talk about a bunch of different use cases, and that gives you some signal, too. If you’re talking to more of a VP of a certain area, you’re probably focusing on one use case.

So maybe go in detail through one of these conversations so that others could benefit from what you’ve learned. What is the order of the questions? How would you structure those conversations to get the most information possible?

You get into the call and start by doing very high-level discovery: “Hey, what are the sorts of projects that are ongoing right now? How do you spend your time? What is stressful for you right now?” Things like that. Then you can get a sense for the types of use cases.

Very quickly, you can form a hypothesis, literally on the fly, of what a product would be that makes sense here. Then you’re explaining, “Okay, what if something like an AI agent could do XYZ? Would that be helpful?” Most likely, they will say yes, because I think there’s this thing that happens where, if someone’s on a call with you, they almost feel like they owe you something good that you can take away.

So they’re like, “Oh, yeah, yeah, that’d be great.” Now you’ve solidified at least the potential product ideas, right? They might adjust it a little bit. They’ll be like, “Yeah, no, actually, it should work like this,” and so on.

I remember we talked to a lot of operations leaders. Matt McInness from Ripling is a great friend of Decagon, and he was telling us about all the different things that happen on his team because there are so many. It was similar with other operations leaders we talked to, and it was a range of companies as well, including people like Oura Ring.

When you get down to it, you’re like, “Okay, great. Now we have these use cases. How much would you pay for it?” That forces them to think, because most people are not thinking about that as they’re talking about ideas. They’re like, “Oh, yeah, this would be cool.”

4. Anatomy of a Great Customer Interview

As soon as you force them to think about how much they would pay for something, it becomes a forcing function for finding some order of magnitude, some level of scale. Then they’re like, “Okay, well, yeah, we have 5 people doing this full-time. If the AI agent could do this, maybe we’d be able to get rid of 1 of them, assign 1 to another role, and then, as a result, I’d pay you $20K a year or something like that.”

In your head, you’re like, “Okay, cool. So now at least you have a general order of magnitude.” $20K, of course, isn’t amazing. But if I could quickly turn out a ton of these, maybe that’s interesting.

Most of the time, that’s not the case because there aren’t that many good ideas out there, honestly. Most of the time, at the end of this exercise, you’re like, “Okay, great. Glad I didn’t pursue this further, because that would have been a waste of time.” At the end, people are paying you a $100 subscription per month, and it’s a big company.

You just do this exercise, and the nice thing about it as well is that it puts the customer in the same frame of mind as you. Then they can tell you other things. One thing that essentially happened with our company is that we were talking about all these use cases, and they were like, “Okay, great. If you did this, we have 5 people over here. But by the way, we have a 500-person support organization, and there’s a lot of opportunity there.”

We’d be like, “Okay, great. Tell us more,” and then you dig into it. As a founder, you have to build your own conviction and do this process yourself. At the time, what everyone told us—including very smart older founders that we knew—was that this use case was super obvious. There were probably just going to be incumbents tacking onto the product, and because it was so obvious, there was probably a reason why no one was super big right now, or there was going to be someone who was ahead. No one really knows.

Even for me right now, when other founders talk to me about other spaces, I have my own opinions, but I don’t really know the details of the space. The only way you can really know is by talking to customers and getting that signal.

In hindsight, it turns out that in any sort of wave, at any time, there’s a very small number of good ideas, I would say, and your job is to ideally find 1 of those at the right time. By definition, it’s going to be pretty non-obvious and pretty difficult. So if you can do this process well, it’ll give you the most signal.

Do you remember the highest number anyone said for how much they’d be willing to pay for something in one of these ideation sessions?

Yeah.

So, at the time, it was probably on the order of low- to mid-six figures.

As you neared the end of that process and settled on what Decagon does—which is probably the right time to ask you to describe in detail what it is—what was the final closing like? Where did the conviction come from? What made you think, “Oh, this is clearly the thing,” after this discovery process?

It was clearly the thing because if you just tallied up the amounts that people said and added them together per idea, this was probably an order of magnitude more than anything else.

5. Everyone’s Wrong Until You Talk to Customers

Very specifically, what Decagon does is an AI customer service agent. That’s the simplest way to think about it. You’re building a conversational AI that can be almost like a front end for a brand. Anytime someone wants to talk to the brand, or anytime the brand wants to talk to them, you can initiate these conversations.

Of course, long term, this is not specific to customer service. But, again, going back to this exercise, customer service is where we felt the most urgent need. In hindsight, I can dissect why we think that is, but that’s basically what we felt.

What gave us conviction is that we had all these folks lined up who were very willing to invest six figures in a random two-person team that they didn’t even know that well, because it was a very top-of-mind initiative for them. Everything else was just a struggle: “How much would you pay for that? I don’t know. This is really exciting, but our budgets are tight right now, and it’d be hard to measure how well this is doing,” and so on. That gave us enough conviction, and then you just take it step by step from there.

Say more about the comment you made that, in hindsight, it’s clear why this was the key problem. Customer service has a bunch of nice properties that I think are very hard to reason through ahead of time, which is why I feel so strongly about this process of discovery and just really staying customer-centric.

6. AI Agents Replacing Human Labor

One of the properties is that the ROI is really easy to justify internally. You have these numbers already tracked, right? It’s like, “Hey, we have so much conversation volume right now. We have a simple chatbot or a simple IVR phone tree. It’s kind of resolving, so to speak, 15% to 20% of that.” If you’re able to take that to 50%, 60%, 70%, 80%, that’s huge ROI. It’s very easy to quantify that. It’s like, “Okay, well, I’m going to take the total cost, chop off 60% of it, and that’s what I’m saving,” right?

The other property, which I think is a little underrated, is that it’s very easy to go live. A lot of GenAI use cases are struggling with that right now, especially at the enterprise level, because there’s risk involved. People don’t want to feel like something could go wrong for them when they release your product and their leadership is going to get mad at them. It’s like, “Why’d you do this?”

That is a big deal with GenAI. I think that’s one of the reasons why it’s been difficult for a lot of use cases to really take off, because at the end of the day, there is always going to be risk. The models are nondeterministic, and something could always happen. But the nice thing with customer service is that you have an escalation path naturally built into the way the product works. The agent’s having the conversation, and if, for any reason, it needs to exit, it’ll just escalate to a human. That infrastructure is already set up, right? You already have your call center and your telephone stack or whatever, so you just connect to it.

I think that property alone makes things way easier. These big enterprises that we work with are like, “Okay, great. We test it internally, and then to go live, we’re just going to choose this one surface area and release it to 5% of the user base.” Even for that 5%, if anything goes wrong, it just escalates. That gives people enough comfort to go for it.

Those 2 things are some of the big reasons why it’s probably the use case with the most traction at the enterprise. Coding is another one. Coding is a little bit different. It’s a lot more bottoms-up.

Can you compare it to coding? It seems like these are the 2 areas where it’s blindingly obvious that it’s useful to customers and you can build great businesses around it, just by looking at the revenue curves. Yours, Sierra’s, obviously—like Cursor, Cognition, et cetera. Compare and contrast coding and customer service.

Yeah, they’re very different. Maybe one framework to think about this is that, at the end of the day, what AI agents are there for is to essentially replace human labor, and that’s why it’s exciting. That’s why everyone’s so focused on it.

7. Why Customer Service

One thing you can do, then, is just map out the spectrum of how much that human labor currently costs. With customer service, it’s generally outsourced already. Especially for the tier-one and tier-two inquiries that AI is now handling, it’s generally not people who are super highly paid. On the other end, you have engineers, who are the most highly paid people.

I actually think one way to think about it is that AI use cases will start eating the spectrum from both ends. The reason why is that, because engineers are the highest-paid people, they have the sophistication to really leverage it well, and AI just gives them so much leverage. There are other factors as well. It just happens that coding is tokenizable, and the models are really good at it.

That’s one way to think about it. I don’t know any company that’s like, “Hey, I would like to just let go of a bunch of my engineers because now I have coding agents.” There’s infinite engineering work to do, so you’re just augmenting them.

On the other end, it’s more of a replacement sense. You have a large BPO, and that’s costing you a ton of money. It’s also really operationally intensive to maintain, because you have to hire people all the time, there’s a ton of churn, you need to train them, you need to QA them, and you need to make sure that nothing goes wrong.

AI is really valuable there as well because, since the work is easier for the AI to do, it can fully replace the human labor. I think the idea that AI is replacing jobs is a little bit overblown. Most of the BPOs we talk to aren’t really that concerned, because there’s already very high turnover in these BPOs. People are just hopping around doing all sorts of different things.

8. Coding vs. Customer Service

What most enterprises will do is not massively cut their workforce. They’ll just naturally let it decrease, and then people go on to do the next level of task that the AI can’t do yet. Maybe that’s data labeling or something. That’s generally what we’re seeing from the BPOs.

Yeah, anyway, I think the spectrum is pretty real. So then the question is: What is the next thing that happens?

That’s a really interesting conclusion: try to augment the very highest talent or replace the most replaceable end of the spectrum. That’s a really interesting conclusion.

I want to talk about what you’ve begun to learn about how to do that second thing well. If others out there wanted to start a company or invest in a company that was doing that end of the spectrum—eating its way in, as you described—what have you learned are the key things in the setup process with a given company to increase the likelihood that you can replace a lot of the low-hanging fruit, whether for types of customer service calls or types of what used to be human-to-human interaction and can now be handled by AI on one end of the spectrum?

I would say the biggest learning we had is that oftentimes a long pole in the tent—and, again, by definition, if you can solve this well, it just makes things go a lot easier—is aligning on what good looks like.

You would think that in our space it’s pretty easy because it’s like, “Okay, maybe you just have a bunch of questions and answers, and that’s what good looks like.” But unfortunately, it’s a lot more nuanced than that. What good looks like could be in terms of tone and brand guidelines, or how conversational you are. Even for the actual answers, we work with a lot of enterprises where the scope is broad, and so one of the things you need to set up beforehand is what this could look like.

In our product, we essentially have a testing or simulation suite. We’re going to build out 10,000 tests, and each one is going to be constantly running, say, 5 times, so you can get a sense of how well things are performing. That’s actually pretty difficult, and I do think that is broadly true for anyone trying to build in this style of company. If you’re going to be replacing human labor, you need to know what good human labor is.

What we found is that you might ask, “Can someone just tell us what are all the answers to all these questions?” Most people don’t actually know. No one is the person who knows how to answer all these questions, because these are large, complex organizations. You have to design a process where it’s very easy to extract these answers from all the people who do know.

Maybe it’s all the CX leaders and people who lead different areas of the product. You have to get them all together and get them to align on, “Okay, here’s what the eval is, essentially.” If you can do that well, then it makes everything a lot easier, because now you’re just building, building, building. You have this quantifiable score that's like, “Hey, here's how well the AI is performing.” Then, once you're done building and the score is high, you can go live.

Is the right way to think about this that you just created a captive reinforcement-learning process within an organization? Is that the simplified version?

Yeah. That's an interesting way to think about it. It doesn't have to be reinforcement learning in the pure sense of training a model. It can just be reinforcement learning in the sense of making the agent improve, and that could be compiling more evals or compiling more guardrails and guidelines around what it can and can't do.

How fast does the spread happen? If I'm a customer and I've got a 500-person customer-service call center or whatever, I'm actually curious. I don't know what the volumes are like—how much call volume or interaction volume a center like that handles for a given company—but if I give you 5% of my workload and I'm satisfied with the AI's performance, it performs well and there aren't a lot of problems, how fast are people willing to go from 5% to 10% to 15% to 20%?

Very fast. I would say even for large enterprises, within weeks. Everyone wants to just go live with everything, right? But the reason why you stage it out is so you can make sure nothing's going wrong, and you can tell if something's going wrong almost immediately because you have these metrics.

So, even within a week, if you have a 500-person organization, I would probably estimate mid- to high-six figures of conversations a year, maybe slightly higher. What you're doing there is making sure that things are going well. Within a week, you can see, “Okay, what is the resolution rate? Is that what we expect? Okay, great. What is the customer satisfaction?” People have those scores as well. They'll probably have some sort of accuracy metric based on human review.

If those all check out, there's really no reason why you shouldn't roll it out. Again, the business case is so obvious there, right? It's like, “Hey, we're both generating a ton of operational efficiency and our customers are happier. So, yeah, let's just send it out to everything.”

What goes most wrong when something bad happens? I'm sure this is happening less and less as the product's gotten better, but even in the early days, what sort of thing would go wrong in one of these customer-to-AI interactions?

It ranges across all sorts of different things, and it's kind of funny. Sometimes, obviously, both sides—us and the customer—take everything very seriously, but there are just a lot of things you wouldn't expect.

In the early days, we had a customer that was essentially a large ticketing platform. Someone came in because they couldn't find their ticket. The AI looked into their account and said, “Hey, there are no tickets here.” Then they were just like, “Okay, well, what I'm going to do is show up to the event, find 8 homeless people from the city, and bring them with me.”

The agent was like, “Oh my God, that's so awesome that you're thinking about doing something nice for the community.” It's just things like that where you're like, “Okay, I would not expect that to happen.” Those are the sorts of things you have to tune over time.

Yeah.

Generally, it's in the spirit of that. You're trying to find the right level of guardrails and flexibility, right? I think that's the name of the game in our space, at least, and that's the way we've designed our product. It's probably the number-one reason we've been successful so far.

What generative AI really unlocks is this super-flexible, super-personalized experience. One way you can think about it is that, in the old days, to map out a conversation, you would build a gigantic tree of decisions. That's very hard because no one likes that experience. You ask something that's not quite one of the branches, and it just forces you down that branch, with no way to go back.

What an LLM does is abstract a lot of that tree into the neurons of a language model. That's really powerful. On one end of the spectrum, you're looking for flexibility, power, and the ability to sound really human-like. But with the enterprise, there are a lot of things where we don't necessarily want that as much. You want full rigor, right? If it's a regulated use case, you cannot afford for it to ever deviate. You want these 3 steps to always be followed in this order, and you can't go to step 3 until something has already happened.

You need to design a system that can be anywhere along that spectrum. Back to your question of what could go wrong: the worst thing that could happen would be that it just says something it's not supposed to say. You need to design an AI that's really robust to that, and you can choose, “Okay, for this use case, we really need it to be over here”—a lot more robust. But for other use cases, where you're just asking basic questions about an account, we don't want it to be like that. We want it to be super free-form, and that's how we get customer satisfaction up.

I think most people are still focused on things that could go wrong because it's nondeterministic. What about the total other end of the spectrum? What have been the things that have gone way more right than you expected? Where has the potential of agents outperformed your expectations in terms of what they can handle or what they can do?

It's really just about elevating the experience. One metric—which is usually a secondary metric that folks think about later but is quite important to us—is how often people come in and just say, “Agent, agent, agent. Get me to a representative. I want to talk to a human.”

I've done that, and you probably have as well, where you're calling into some sort of customer service and pressing 0 the whole time. That's because people are used to bad experiences, so they've already lost trust in these systems.

What surprised us was that if you make it really clear off the bat that this is a different experience, people are willing to give it a chance. Then the outcomes are just way different. We have a customer, likely Oura Ring, the wearable ring. We did a case study with them where, before having any sort of generative-AI system, 1 in 3 customers who came in would not bother saying anything and would just keep saying “agent” until they got to a representative. Now it's 1 in 20.

We spent a lot of time making the beginning of the process feel very different, and folks are willing to give it a chance. I think that's been exciting.

Where do you think that can go? How good can the experience get in ways that it's not yet that good, with the subsequent evolution of your product but also of the underlying capabilities of the models?

The biggest frontier right now is voice. I know there are a lot of interesting startups working on voice as well. It's exciting because it's still definitely not solved. There's a lot to be done there, and the bar is very high.

If you think about how humans communicate, for literally the entirety of humanity—150,000 years or something—the user interface for every human has been language, specifically spoken language. You listen, you speak, and that's how our brains have evolved. That's the most natural way for us to communicate.

Only in the last 60 years of that entire time did we have keyboards and phones, and start communicating through typing. Fundamentally, any sort of agent that communicates with humans has to treat voice as a critical factor, because that's just how we communicate. Our brains are so evolved for this that it's very easy to tell when something feels not quite right. The bar is very high; that uncanny valley is quite large. That's why there's a lot of effort going into making voice experiences good.

Even likely ChatGPT Voice, or Sesame, or these voice-to-voice models, are starting to feel very impressive. But if you talk to them long enough, you can definitely tell they're not human. There's that element of it.

For enterprise use cases like ours, there's still a ton of hurdles to cross because even though those models are good, the hallucination rate is really high. You can't necessarily use them as-is in current systems. A lot of people now go from voice to text and then back to voice, so you can run a lot more checks to make sure things are accurate.

There are a lot of cool ideas to explore around how you make it both human-like and accurate, and how you tie everything together. That's where most of the work is going these days.

At the risk of getting too technical, why is voice-to-voice interesting and worth pursuing versus just always going back to text and being able to manage it that way?

Yeah.

The fundamental difference is that if you're just going to text, then, no matter what, the final audio is just a narration of the text. Voice-to-voice is powerful because it takes in the entire audio of what you said. So it knows your cadence, how upset you are, the tone, and everything. The latency is a lot less as well because you're going straight from voice to voice.

Latency matters so much when we're talking, right? When we're talking right now, our brains are constantly going, “Okay, when’s he done talking? When should I start talking?” If someone interrupts someone else, in a polite way, people adjust very naturally. That is the biggest component of voice-to-voice, and ultimately the prevailing view is that, for whatever the final experience is, if you really want to make it indistinguishable from a human, you have to do voice-to-voice, or at least take into account the voice.

The issue with voice-to-voice, though, is that voice fundamentally has a lot more dimensions to it. The amount of tokens you generate per sentence is just a lot higher than when you generate text, and the more tokens you have, the easier it is for something to go wrong. So the hallucination rate has so far just been a lot higher.

9. The Future of Voice and the Human UI

How much higher? Give us a sense of how far we are from these being really good.

I think probably 8x higher or something like that.

Wow.

So it is quite a bit higher. Yeah, of course, you want to leverage that technology. So maybe there are creative ways to make a hybrid of the two, right? Maybe you can have a text model generate the content, but take into account the audio from before as well, and that makes something very realistic.

At the same time, latency is still the hard problem because, at the enterprise, you're doing a lot before you can start responding, right? You have to figure out what they're asking about, which materials you need to collect, whether you need to hit any APIs and get that data back. You have to do it in a way that feels very natural, and sometimes, if you think about how a human does it, you might have to say something like, “Give me a sec to look that up,” because it genuinely takes 10 seconds for the API to come back. So these are all interesting problems to think through.

So give us a sense today: if you add up all the interactions, some idea of how long they are, what type they are—voice versus text versus some other modality—what does the entire corpus of interactions between a Decagon agent and a customer look like today?

It's pretty balanced at this point between chat and voice. On a raw customer basis, there are more people in chat, at least for us. But if you just think about the large enterprises, like the Fortune 100, they've been around for so long and everyone just calls them. So voice is disproportionately higher there. A lot of them are 90–95% voice and 5% chat.

In terms of the types of conversations, you start with the easier ones, of course. These are things where they're kind of question-and-answer based, right? So tier 1 is, “Hey, you can answer their question based on what you statically know.” That could be, “I have questions about how your loyalty system works,” or, “If I bought something, would I still be able to refund it?” Things like that.

The next level is still question-and-answer based, but you're leveraging a lot of real-time data. That could be, “I got 2x points on this transaction, but I should have gotten 5x. Why is that?” Then it would actually go and look into your account and reason through things like, “Okay, I see the account says this type. Let me go find all the documentation on this type.” And, like, “Okay, actually, it's because you booked through a travel agency. If you had booked directly with the airline, you would have gotten 5x, but this thing doesn't apply to a travel agency,” or whatever.

The third tier is that you're actually taking action. “I lost my credit card. I need a new one.” It's actually walking through a pretty large flow, and that's where AI agents have been really excellent because you wouldn't expect them to be able to do that. It's able to go in, and it can be a pretty complicated system where it's like, “Okay, well, first I need to figure out what your address is and confirm whether the address is correct. Then I need to look and see, ‘Hey, do you want me to lock the old card?’ ‘Okay, great. I'll do that.’”

10. Why Voice-to-Voice Is So Hard

I might need to check for fraud to make sure this person isn't just constantly asking for new cards. It's all these things stitched together. That's really what makes it agentic, and that's why there's been such a step-function improvement with LLMs.

In the spirit of that question of what could go right, what could go right for the company based on the data they're gathering from these interactions that they probably historically have been doing nothing with? What new things can they do for their customer because, on a one-to-one basis, they're just learning more about a person, and on an aggregate basis, they understand the behavior patterns of their customer base or something?

Oh yeah, that's a huge topic. I think that's a huge part of our product. We have a viewpoint that this data, of course, is super valuable because it's literally what your customers are saying, but it's very underutilized because historically it's been very unstructured data.

What people typically would do is, “Okay, well, every month we have 1 million conversations. We'll have a full-time team of 20 people, and they're just sampling these conversations, trying to check on a rubric and compile topics and things like that,” right? And that will only get you so far. But now what you can do is literally have a language model that reads every conversation and extracts whatever information you want from it.

That allows you to do things like, over time, identify topics that people probably didn't even know about, because these organizations are big. The people in leadership positions can only have such granular insight into what's happening, but it will just flag, “Hey, there's this 2% of conversations where things are not really going that well, and it's because we don't have context on this topic.” So let's flag that. Let's draft a suggestion for what could go better here based on how the human agents are handling it or based on the other procedures that we have, and here's a suggestion for how we should adjust the agent.

That allows the agent to improve automatically over time, and that's really critical. So when you think about moats in the agentic world, a lot of it is around, well, if you've been working with a client for a year, has your agent just continuously gotten better by learning from the data? That's a different concept than just training on the data, but has it continuously gotten better to the point where it's very difficult for another agent to come in and perform at the same level?

There's this funny situation today where I think, especially CEOs, really want AI in their businesses. They want it now, but they don't know what they want. They don't have a good framework for thinking about, “Okay, I understand my business. I don't know where to go.”

It seems like I would be missing a major boat if I don't deploy this in my business, but I don't know what to do. I don't know where to go first or who to call. Have you developed any framework for those company leaders who desperately want to use this technology in their business but simply don't know where to start, apart from automating customer service or something? I don't mean specific use cases; I mean a framework for thinking about what kinds of problems might be addressable by agents or by LLMs?

I mean, one framework is similar to the framework we've talked about before: two ends of the spectrum. I would say most leaders we talk to are focused on the more bottom-up end of the spectrum, which is, “Where are the areas where we should just not have humans doing this because it's so mundane and repeatable, and there are tons of cost efficiencies there?” I would say that's where folks are typically thinking.

When we talk to leaders, I think there are a couple of observations. One, pretty much all AI initiatives are very top-down at this point because it is such a board-level mandate. So the C-suite is very invested in, “Okay, where do we deploy AI?” That almost means that if you want to get things going at a larger organization, you have to have buy-in from the top level because it's going to get up there anyway, and they have to make the decision at the end of the day. So that's one.

Two, the way they think about the use cases, to your point, is back to ROI, right? It's, “Where can we either save a bunch of money or make a lot of new revenue?” If you cannot, in basically half a sentence, explain that, then it's just not going to work right now because they're under a lot of pressure. They need to show quick wins, and if this isn't going to be a quick win that they can point to—“I saved $10 million”—then it's not going to be something that's prioritized.

Do you think coding answers that well? Do you think the ROI is clear in coding?

11. Learning from Customer Data

Yeah, it is. I know the coding agents quite well, and the way they do it generally is, one, it’s very easy to test. So it’s like, let’s just deploy out to the engineers, and then you just do a poll of the engineers: “Hey, how much more productive do you think you are?”

Engineers are often the most valuable resource in these organizations, and so those answers are treated with very high importance. An engineer tells you that they’re 50% more productive, and it’s like, “Okay, great.”

Are you confident that that’s true? That they can self-report and be accurate? There was that unnamed study or whatever that came out that actually—I don’t know if the study is good or not—but it found that productivity was down or flat, or something like this. The reported, self-reported productivity was at odds with actual measured productivity, or something like this.

Oh, yeah. I actually don’t have an opinion. I don’t know nearly enough about that. But I’m just saying that it doesn’t really matter, right? If their entire engineering team is like, “Hey, we love this. This is making us 50% more productive,” it’s like building your thing.

Exactly. And then you think about it from the CEO’s perspective, and they’re reporting this to the board or something. It’s like, “Hey, my entire engineering org said that they’re 50% more productive. This is a worthy investment because these people are all paid this much, and now we can accelerate the product. We can accelerate everything.”

12. How CEOs Should Think About AI

Do you think the future is that every time I talk about brands and how a given company might want its agent to feel and sound—that an agent might be a way to express brand culture, style, tone, whatever? Do you want to talk about that? Do you think the end state here is that each company has almost a named, personified representative that you just come to expect to interact with? It’s not just customer service issues, but sales issues. You ask it for advice on what shoe to buy, or whatever it might be, and that’s all integrated. Or will you have different agents for different parts of the company?

I guess what I’m trying to ask is, what does the future of a company’s agent or agents look like in the natural end state, 5 years from now or something?

I would say that in the natural end state, it is more unified, for the exact reason you listed, which is that people want a unified brand out there, and the brand is very important to them. For some businesses, it’s more important than others, but eventually this becomes the front end for the business, right?

That means it’s both how you gain new customers, but also how you support the existing ones and make them retain more, and so on. In the limit, if you’re working with a bank, airline, telecom company, or whatever, the agent could be the only thing that most users interact with. They don’t even have to touch your mobile app. They don’t even have to go to your website ever.

They just have this agent where they’re authenticated. It knows everything about them, has all the context of your previous conversations, has memory, and can just solve your issue. It can take actions for you. You need to book a flight, upgrade a seat, or ask questions about this or that. I think that’s the exciting vision that we’re building toward.

In some ways, we often call this a concierge. It’s just a digital concierge that can do everything for you. But you also have to be pragmatic on both sides. You have to start with a clear use case. I think that is where folks are building toward.

In the near term, I do think that different teams, because the reality is that at these large companies, different teams have different budgets and make different decisions, might have different agents. But long term, you either need to have a unified system that ties them together—a unified framework—or they could just be literally the same agent.

Is the right analogy here a company’s website? They’ll think about their agent like they think about their website. A lot of work will go into it. It’ll look and feel a certain way. It’ll be kind of a unified interface with the world.

Yeah.

Is that a clean analogy?

I think that’s a good analogy. It’s just like a front end, right? It’s like a UI. But instead of a visual UI, it’s a conversational UI.

13. The ROI of Coding and Productivity Tools

How do you feel brands pulling personality requests in their agent out of you? Do they care about what they care about? Do they want it to be nice? Do they want it to be concise? Do they want it to be funny? How has that dimension evolved since you started?

People almost always already have brand guidelines because they need to show brand guidelines to their human agents to train them. The nice part is that they already have all this training process for the humans, and you should ideally be able to apply it in the same ways to the AI.

They’ll have guidelines like, “Hey, you need to do this. You need to always be confident.” Sometimes folks don’t want the agent to apologize. Sometimes people really want it to be apologetic. It’s just a bunch of different preferences, and that needs to be taught to the AI in an efficient way.

But that’s also a lower-throughput way of communicating. The other way that you can show the agent is just to give it a lot of examples: “Here are examples of what great looks like from our top agents,” and just have it learn from that.

What is the very biggest version of what Decagon could be—the version you’re almost afraid to admit because it feels so big?

At the end of the day, what we are building toward is this concept of a new UI for the product, right? Just think about how much these large companies invest in their mobile app or their website. This is literally how everyone communicates with them.

This could eventually be a way where any user interacts with any brand through an AI agent, for all sorts of different use cases. It benefits our customers for the AI to have all this context because it can seamlessly flow between things. A lot of them want to do sales-type use cases at the end of a support flow, or vice versa, and that’s exciting.

What internal context or things do the best customers have that make this better? You mentioned brand guidelines. Maybe there’s a write-up on what the brand guidelines are, or whatever. What are the internal assets that companies have or don’t have that, if they have them, make the Decagon experience way, way better?

The number one thing is just APIs for the AI to use: APIs to take action, APIs to look up data, and APIs to reformat things. If you have those, you already know that within the first month it’s going to be a great experience. If you don’t have them yet, then we should try to build toward that as soon as possible so that the AI can actually achieve that elevated experience. That’s the main thing.

Most people already have documentation. It might not be up to date, but we can help with that. Most people already have SOPs, I guess, and we’re able to use those and generate our format. We call them AOPs, agent operating procedures. They’re just SOPs, but for AI. Brand guidelines, as I said, people usually have those as well, so you just ingest them.

We were talking last time about, in any business, but certainly in yours, the 3 key stakeholders being your team—the talent that you have to recruit, and I want to talk about that in detail—capital and investors, and customers. Maybe suppliers, too, is a fourth category, but I’m especially interested in the first 3.

Last time we were together, I asked you which one you have trouble with, and we laughed because you said, “Well, definitely not investors.” Talk a little bit about the demand from investors to invest in companies like yours and how that feels. What are they doing to try to give you more money and get on the cap table? How competitive does it feel? What does that feel like right now?

It does seem like there are a handful of companies like yours that are in one of these white-hot areas, have key traction, have great teams, and basically every investor wants to be involved in those companies. What has that felt like?

Yeah, it definitely feels like there’s maybe a little bit too much excitement right now on the AI side. It just seems way too easy to raise money. There are so many companies out there.

For us, we’ve been very fortunate. We don’t take it for granted. In my first company, we were also fortunate that it was easy to raise, but it was for different reasons, like 2021. Nowadays, I think there aren’t that many AI companies that have real, real traction on the revenue side, especially in the enterprise. I think that’s attractive to investors because they want to deploy their capital into AI, and that’s the biggest trend.

I think we’ve had a great relationship with all our investors.

I think we just really selected for folks that we get along with at a personal level and who we feel will be very helpful for our go-to-market, normally. That’s kind of been an interesting process. Pretty much after every single time we’ve raised a round, we’ve almost immediately gotten preempted.

That alone can’t be right, you know? If you’re thinking from first principles about making an investment, the previous valuation should not be a super big factor in that. It should be how well the business is doing and what I believe the potential is. So it feels like there’s a little bit of mania, but we’re definitely more indexed on the other 2 things: talent and customers.

What’s the craziest thing that an investor has done to try to invest in the business?

The main thing that we do, and I would actually encourage more founders to do this, is that during the stage where people want to invest but haven’t yet, that’s when they’re most willing to be helpful. It’s a great way for you to use that as a proxy for how helpful they’ll be afterward. If they’re not that helpful in that stage, when they really want to invest and are willing to do anything, they’re for sure not going to be helpful afterward, right?

They’ll still be friendly, and hopefully they won’t be detrimental, but that’s your opportunity to really test folks and see how helpful someone will be. We obviously know a lot of investors very well, and I think no one has an issue with that. They know that they’re going back to competition. They’re also in a competitive sport, and they know that they need to earn the ability to invest in the best companies.

From their perspective, I think they’re happy to work for it. If you just give them the opportunity to, they will.

I’m going to ask about this from both the founder and investor perspective. What advice would you give investors during that window? When there’s a fundraising happening, there’s sort of an open process or a window or whatever, what have you seen the best of them do well?

I’m sure helping you—here are 10 customers you can talk to—is great. But also on the underwriting side, them making sure they understand your business extremely well in a short period of time, which is usually what I’m expecting. What have the very best done in that window?

Number 1, I think the best investors get this dynamic. We’ve definitely talked to a lot of high-profile investors where they’re just not willing to help much until they’re invested. They’re right, right? It’s, “Hey, we have a lot of our current investments. We don’t want to use our social capital or whatever to help with a new investor or a new investment.”

But from the founder’s perspective, if that’s the case, then it’s hard to tell if you’ll actually be useful or not. There’s no difference between you saying that and someone who can’t really help just saying that. I think the best ones are able to give a lot of signal to the founder that, “Hey, I’m really willing to help. I have the ability to help.” It doesn’t have to be that long a period of time—a couple of months before the round actually happens, for example.

The other thing is that if you think about anyone you bring into the world, whether that’s employees, investors, advisers, or anything else, what we really index on a lot is cognitive ability—raw intellectual throughput. You can almost feel that out in an investor in the same way you would feel it out in an employee, just by spending time with them and seeing if they’re thinking about things from first principles in your company.

For example, a lot of AI companies are growing much faster than traditional SaaS companies right now. Because that’s the case, a lot of things are different, and you generally don’t want investors who are just like, “Hey, I’ve seen this so many times. You’ve got to do things XYZ way.” You want people who are intellectually curious, will think about things along with you, and can help problem-solve.

What about on just the pure underwriting side? An investor comes in, they’re really smart—we’ll take that for granted—and they just want to understand your business, the good, the bad, and the ugly, as fast as possible. What have you seen the best do on that side of the investment process, including things like how fast they move, how deliberately they move, or anything like that?

14. Recruiting, Capital, and Demand

The best ones, I would say, get a very deep understanding of our customers. It’s actually funny: our customers have probably made so much money on expert calls because so many investors are hitting them up. I think the best ones, before they even talk to you, have probably already done quite a bit of research and have a pretty full view of your customers.

Unfortunately, what we’ve seen is that there’s a lot of noise there, because a lot of people just lie on customer calls. We’ve had people say that they’ve used us when we literally never heard of them. But generally, if you do enough research and you’re good at it, you can underwrite the business that way, because that gives you the most signal.

The other thing is people who index on culture, because I think culture is quite important. A lot of good investors know how important that is, and if they feel like you’re becoming a place where good talent is congregating, then they’ll index on that more.

Let’s talk about that. Let’s talk about recruiting and culture. Starting with culture, how would you describe it? We talked about the quote on your wall as the first question, so that’s part of it, of course: extremely competitive, an extreme bias to action, getting things done. What are the other key components of your culture? When you’re sitting with a new recruit, what do you tell them about what kind of place it is?

There’s definitely a level of intensity that’s important, and you definitely want to tell people that upfront because you want them to self-select into this culture. Everyone who has joined Decagon, I would say, wants to work hard. They want to be around other people who are really smart and are like them.

They’re motivated by viewing this as the prime of their career: “Hey, I’m going to work hard, but because of that, I’m going to build lifelong relationships with other amazing people. I’m going to have good financial outcomes. I’m going to be able to leapfrog steps in my career because the growth is happening so quickly.” You’re attracting people like that.

I think that alone creates a pretty strong foundation for the culture, because you have people who are there to work. We have a lot of people who live right next to the office, and part of that is because we’ve selected for people who like being in the office with other people. We’re in the office a lot. Those are the foundations.

On top of that, what you have to be careful about is that we still want our office to be a place where people enjoy coming to work every day, because we spend a lot of time in there. You want people to be happy to be there and feel like everyone there is supporting them. Even between the organizations, you want people to feel like they’re aligned and that folks are working toward the same goal.

That’s an ongoing problem. I don’t think we’ve solved the culture, but it’s something we put a lot of thought into. We want to make sure that people feel like they’ll be very fulfilled by staying here for a long time.

One of the most interesting subplots of this entire business evolution, and around AI, is the talent wars. Obviously, it’s happening in the most extreme cases at the model layer—Meta, Anthropic, and OpenAI—and there are great, riveting stories about the lengths people will go to secure a great engineer or someone really key to the business, and the amount they’ll pay to secure them.

Can you give us your perspective on these talent wars and what it’s like to be in them? Obviously, I’m sure you’re fighting for the best talent, too, versus lots of other great companies that are being formed. Tell us from the inside what this environment feels like.

It definitely feels like talent is a team effort. For anyone you want to hire, you need the whole team to swarm around them to win highly sought-after talent. You have to go and do all the things right.

There are some parallels to sales, of course. You’re trying to convince people that this is the place they want to be. That involves oftentimes getting to know their families, getting to know their partners, really figuring out what they want out of their own careers, and making sure that you can design a role for them that’s like that.

Unfortunately, in the application layer, it’s not as crazy as it is at Meta and OpenAI, where there are only so many top-level researchers.

We have some very strong researchers on our team. It’s a lot more applied for us. It’s the same thing that happens, right? We’re going after a lot of people on our team from Harvard, MIT, and Stanford. There are only so many of these folks in the market at any time who are in San Francisco or going to the office, so it’s a competitive place to be.

15. Investor Frenzy

I think we’re kind of fortunate. Our talent brand has gotten a lot larger than when we were first starting, so it definitely has gotten easier to hire. At the same time, the number of people we need to hire has also gone up, so it’s just constantly finding ideas for how to get new people. We just opened our New York office, as you know, and that’s part of the reason for that: there’s another talent pool over here, and we should leverage that.

One of the questions that I think so many people are interested in for companies like yours is the use of whatever core underlying LLM you’re using versus the development of your own models on top of, or replacing, those underlying LLMs. You’re gathering all this incredible data that’s just yours. You don’t have to share it with anybody else, and these models thrive on good underlying data.

How do you think about that aspect of all of this, where 5 years from now it’s going to be either your own model, your own model plus something else, or just your own data and context on top of the best model? How do you think that will evolve? I’m curious about this in terms of how it impacts the product, but also how it impacts your business model and your power in your business, where you rely or don’t rely on GPT or whatever.

When we first started—this was about 2 years ago—people were still figuring out applications, and almost no one was doing fine-tuning. In fact, for anyone who was doing fine-tuning, a lot of the writing at that time was, quote-unquote, “Fine-tuning doesn’t work,” because it doesn’t really get you that many gains. Another big reason not to do fine-tuning at that time was that the models were changing so fast. Why invest a bunch of time into fine-tuning? It’s not reversible. You’re going to have to throw it out the next time there’s a new model release.

I think now the open-source models, for example, have gotten to the point where they’re definitely not smart enough to do everything, but there are a lot of specific use cases where you just don’t need that much intelligence. An example would be, even with an agent, let’s say the first thing the agent does is think about, “Okay, based on what the user said and everything before, what path do I go down?” Or, “What data do I need?” You can make that a fine-tuned model. That model doesn’t have to be good at math or coding or anything; you just need a smaller model that’s fine-tuned on that.

Nowadays, we’re seeing much more of that because a lot of the applications have gotten more mature. You know how your agent is structured, you know the places where you need models to run, and you can take smaller fine-tuned models. That improves the entire system both in terms of performance and latency, and so on. I think over time there’s going to be more and more of that.

I do think there’s always going to be huge usage of OpenAI, Anthropic, and the rest of the world, because if you just need intelligence, you need the best models. I think there’s going to be a balance, but at least in the short term there’s going to be more and more fine-tuning of small models.

What is your perception of the very biggest companies? The entire market has been focused, rightly so, on the 7, 8, 9, or 10 biggest technology companies. Which of them feel most important to you? I don’t mean OpenAI and Anthropic; I mean Microsoft, Amazon, Apple, and these sorts of companies.

What is your relationship with and thinking about them today? They’ve been the driver of equity markets by a huge margin. They’re important companies. How do you relate to them?

Work-wise, you don’t; it’s not super relevant. But of course, we have our own opinions. I’m very bullish on Google, actually.

Why?

I just think that with AI use cases, having individual consumers is so important, because that’s where all the data comes from anyway, and Google is much stronger there. You could say that Meta—Facebook—also has that element, so maybe their new superintelligence lab will be able to make it work.

But something like Anthropic, for example, where they haven’t done as well on the consumer side compared to ChatGPT, I think long term you do need that consumer buy-in, because that’s where all the new data is going to come from. So, yeah, Google. Google just has an amazing team, and they’ve had a couple of rocky starts, but hopefully they’ll make it work out.

Of course, all the large companies—all the Mag 7, basically—are obviously super strong, so we don’t really have strong views on them.

If I let you build a portfolio tomorrow, where you got 5 slots, 20% each, in 5 private companies building in and around AI, what portfolio would you build? Decagon excluded.

Yeah, Decagon excluded. Let’s see. I just gravitate toward where most of the talent is forming. I mentioned I’m close with the Cognition guys, so Cognition would be in there for sure. Cursor also would be in there. I’m actually kind of interested in where those might run into each other in the future.

For me, I definitely would want to take a bet on the hardware layer, even though it’s a much higher-variance bet. Last time we were talking about Etched—companies like that—I think you probably put one of those in there. It’s still on the earlier side, but obviously has very high potential.

Another friend of mine is building a company called Pika, building video models. I just think very highly of that team as well, so we’d probably put them in there.

I’d probably want some sort of bet on the underlying model side, even though all the large language models are out there. Another friend of mine, whom I think very highly of, is building models, but not the types of language models—still foundation models for healthcare, for example. My friend Josh is building Chai, and they’re building a foundation model. Stuff like that is quite interesting.

I would also put the company likely called Physical Intelligence in there. Those are very exciting.

I mean, it obviously is early to see how those would turn out, but if I’m building a portfolio, I definitely want one of those in there.

What do you think—I’m interested in both sides of the spectrum—what do the people who are in your position, who are both technical and commercially at the center of this wave, overestimate and underestimate about the capabilities of AI today? Where is it further along than the world thinks? Where is it further behind?

It’s a little bit more behind in being able to unlock a lot of the enterprise use cases, I would say, because of the nondeterminism. There are 2 things that need to happen. First, you need to reframe the way that people think about agents, because there’s the likely Waymo effect that often happens, where it’ll be objectively way better than human drivers—and human drivers make a lot of mistakes—but because we’re investing in new technology, the bar is a lot higher. It has to be near-perfect.

There’s that dynamic that needs to adjust in some folks’ minds. Instead of evaluating AI in a way where you’re just trying to find mistakes, you’re evaluating it holistically and looking at the success rate. If you can frame it that way, the success rate is going to be way higher than humans, because, again, humans are not perfect.

I think that needs to happen, and I don’t think that’s fully happened yet in the enterprise. That trend needs to happen. Then, on the AI and use-case side, I think the models still need to get better in a lot of areas. We were just talking about voice-to-voice earlier. Hallucinations are too high there, and as those models get better, I think more enterprise use cases will be unlocked.

I do think the general public will just see a really cool demo and be like, “Okay, wow, voice is solved now.” As a result, they’ll think enterprise should be adopting it left and right, and CEOs will see that too. But when they actually get into it, it’s harder to go live. That’s the piece where it’s not quite fully there yet, so it’s a little bit slower than people think.

On the flip side, what’s underestimated is just that things are growing and improving exponentially, and no one is good at conceptualizing what exponential means in things like this. That could be from a performance and cost perspective. Right now, I would argue that if you’re building an application, your margin shouldn’t really matter that much.

People will often critique the coding agents, saying they’re hemorrhaging money, but things are improving exponentially, and the cost will go down exponentially as well. Your margin doesn’t really matter. What really matters is that you need to get market share and mind share of users.

So it's perfectly fine not to have good margins right now. I think people underestimate how much the models improve; everything improves way faster than people think.

It's slightly different at the enterprise. The same principle applies, but you generally don't want to be hemorrhaging cash with an enterprise deal because those are much longer-term. Even though the cost will go down, their expectations might also change, and so on. So you generally want to be fairly healthy there. But I think that's what's underestimated right now.

How do you know that costs will get way lower? I'm very curious about this margin question, because if you knew for sure, then the argument would be that you actually want to run super-negative gross margins. If you knew that it was going to get 98% cheaper on the cost of goods to serve a coding agent, or something like that, I think the argument would be: get the install base, have the best product, get the users, build the affinity with that user base, and don't care at all.

But that hinges a lot on the confidence that you have that those costs will fall. How do you think about that equation of winning the install base versus demonstrating good unit economics now?

I just think it's quite unlikely that where we're currently at is the best that things will be. There's just so much effort being put into it, and one of the main metrics is efficiency.

The other piece is that even if things don't get better, there are a lot of ways you can rearchitect your system so that it is more cost-efficient. It's just that it's not worth putting time into that right now when you can put that same amount of time into getting new customers, because you know that things will change in the future.

I think that's one thing where people are like, "It's kind of a scheme where you take VC dollars, and then the VC dollars go to the chips, and these companies are losing money." If they wanted to, they could probably just spend a month, or even less, and massively improve their margins. But no one is doing that because it's just not worth doing that work right now. What you're really optimizing for right now is quality and growth. If you can do that, then the optimization can always come later on.

How do you think about your margins? Just putting it on Decagon, do you care at all? Do you set them? What are your guardrails or parameters for what's acceptable, or what you're targeting?

The only principle we have is not to have negative margins.

In general, we have fairly healthy margins. One way you can think about it is to think about the supply chain for any good. Let's say you're buying a croissant at the airport or something. That last step, where you're actually solving someone's problem, is where you generally can capture the most margin.

Every step along the way—whoever's enriching the flour or making the butter, for example—you're generally making a margin on top of the costs of whatever your goods are. That's why it's nice to be in the application layer: what you're building is actually solving the business need, and as a result, you can capture more of that.

The way our customers think about it is, "Great, we're investing in Decagon. We don't care that much about what Decagon's costs are. In fact, we probably don't care about them at all. What we care about is the business ROI that we're getting." We're downsizing our operations by this much. We're actually generating more revenue now because the AI agent can engage people and keep them retained. That is where we probably see the most dynamic here.

I do think this is generally not really a hot take. In the early days, there were ChatGPT wrappers and so on, and a lot of wrappers, if the margins are too thin, aren't going to be valuable. But if you have enough software built around the models, then that's where you can actually capture almost the most value.

That's why I think the OpenAIs of the world will continue to move toward applications, because it's quite hard for them to make money long-term on their API, for example. There's such high competition; all the labs are building, and it's very easy for people to switch. Moving from AWS to GCP is very hard, but moving from an OpenAI model to an Anthropic model is just changing one line of code.

Can you say a little bit more about this ChatGPT-wrapper concept? My sense is that, certainly for what you built, but probably for other companies as well, a lot of the work your engineers are doing is not AI work. It's traditional software work: the ability to hook this system up to enterprise customers, good old-fashioned product and infrastructure building that's very different from what OpenAI or Anthropic are providing you.

That's hard work. Building any software system is hard work. Everyone's very enamored with this idea that, in 5 years, I can just show you a piece of software and tell a coding agent to replicate it, and that's going to mean lower software moats. I don't think you think of it that way. Maybe describe your thinking on, or critique of, this ChatGPT-wrapper concern that people have.

People say "wrapper" in a derogatory way. It's like, "Hey, you're just a wrapper." Again, it's not black and white. There are a lot of apps that are just wrappers that aren't going to become real businesses because there's just not that much value.

I don't know too much about this space, but at least from the outside, it has seemed like copywriting, for example, has been difficult because someone can just log into ChatGPT and say, "Hey, write this for me," and it will just write it. So there are things where, if there's not enough tooling and functionality on top of something for it to be really valuable and needed, then maybe it is easier to just leverage the models.

But most of the time, that's not the case. Especially when you get into agents, an agent is not just a model. You have to design it, put in guardrails, and teach it how to do new things. That's where the software layer on top of the models comes in. If that's valuable, then it's much harder for it to be made obsolete by a model update.

The other side of it is that the labs are now quite interested in building applications. They're building a bunch of cloud coding applications, like Claude Code, and those will end up being competitors with Cognition or Cursor. Maybe for that reason, it is wise for the coding agents to start training stuff as well.

For us, at least right now, the sheer amount of functionality you have to build, because it is a very top-down product, is quite large. That has nothing to do with AI. It's things like: How do you have observability into what the conversations are? How do you alert the team if something spikes? How are you able to QA and have unit tests for the conversations so that, before you push it out to end users, you feel it's ready? There's just all this functionality that doesn't really have anything to do with AI. It's just a lot of stuff.

How would you advise other founders thinking about the ideal customers to go after? What are the most interesting qualities of your best customers? When you're qualifying them, you have limited time; you can only serve so many people. I know you're growing really fast, but you can only serve so many customers at any given time. How do you qualify who you want to work with and who you don't? What are the attributes that you've seen matter the most?

We want people who, at the leadership level, are intellectually curious and really excited about technology. You see a huge spectrum of that in the enterprise, and some of the best leaders—and probably the folks we're most excited to work with—are genuinely like, "Hey, we want to move on AI as fast as possible. We're very interested and curious about how all your systems work." As a result, they're going to help cut through all the cruft and bureaucracy to get something going.

You can tell that pretty clearly in the first conversation. You can tell if someone's legitimate about, "This is something where I'm going to both push aggressively but also give you a ton of feedback, and the feedback is going to be good," versus someone who knows it's a board mandate and AI is just a thing on their to-do list, basically.

How do you mark milestones in the business? How do you motivate the team? What have you learned about how to rally around a given thing? I know you're super aggressive when you have a customer that you want to get. It's not just an old-school sales process; it's an all-hands-on-deck effort. You send engineers, do whatever it takes.

What have you learned about motivating milestones, rallying the team, and organizing around common goals?

Yeah, for us, one thing we always do is have a sort of flagpole within sight that can rally everyone around it. Having things to rally around is quite helpful. I was thinking about this the other day. I think another type of rallying is around competition, right?

When people feel like they're in a battle and there are clear enemies, then it makes sense to rally around that. Again, you don't want to get to the point where there's active animosity, but just a healthy level of competition ties a team together because there's something to focus on.

The same thing applies to milestones. If you give everyone a clear milestone—and this can be pretty insignificant—last year, for our revenue milestone, we told everyone we'd get them super nice jackets. We got them Decagon [likely Arc'teryx] jackets, and everyone was super excited about that.

The cost of the jacket—

The cost of the jacket and just how much people get paid is trivial, but it creates this feeling of, "Hey, we're working toward these jackets," right? It brings the team together because now it feels like everyone's working together toward this common goal. That's been a big part of our culture: finding the next milestone.

Is there anything that we haven't covered about either the business or this exciting AI applications world that you feel especially passionate about?

I think what has become almost a meme or is hyped up a lot in AI startups right now are a couple of things. One, obviously, everyone's in person. It's like 996 or whatever. I actually don't think 996 is that healthy. It happens in China, and everyone's super hardcore, but one of the reasons is that no one has jobs over there, so it's very easy for employers to have leverage.

I think here you generally want to maintain a good balance, because if you're working at super-high intensity, you need time to relax a little bit. But that is one element.

The other element is the forward-deployed engineers. Everyone's talking about forward-deployed engineers. I just think it's kind of funny, because my co-founder came from Palantir, so they actually have forward-deployed engineers.

What a forward-deployed engineer at Palantir means is that you're working on a $10–$25 million deal. You're working almost full-time with either one or a small number of customers and building very specifically for them. I think people are conflating that with what startups do, which is that startups are just very hands-on and do things that don't scale.

For you to actually have an FDE model, you need to have massive clients, and most people do not have massive clients. I'm interested to see how that plays out, because I do think there's an overindexing on this forward-deployed engineering model right now. It's like, "Yeah, I have a forward-deployed engineer," and the deal sizes are like $50,000, you know.

That's something we think about a lot as well. We're very hands-on with our customers, but you have to think about, "Okay, well, how do we scale quickly?" If, for every $50,000 client, you have someone that's fully staffed to them, that's impossible to scale, right? So you need to find that line in the middle. I think the full forward-deployed model only works with the Palantir approach.

Presumably, on this side, you do care a lot about your margins. You're much more open about margins if it's just LLM cost or something like this, but if it's fully baked, like people cost, that's going to be a problem.

Yeah, and it's not even necessarily a problem from pure dollar margins. It just prevents you from scaling. No one can hire good people that fast, right? Good people—it's just hard to hire good people. So if your business is fully constrained by the availability of good people, that's also not a good thing.

What do you think the minimum customer size and revenue is to justify a forward-deployed engineer model?

Probably $1 million.

Yeah. Fascinating. Well, I think you know my traditional closing question for everybody: What is the kindest thing that anyone's ever done for you?

16. The Long-Term Vision for Decagon

I did put a lot of thought into it. When I was little—call it ages 5 to 13, elementary and middle school—I was a pretty lazy kid in general. I think most kids are. There are very few people who are intrinsically self-motivated. I wanted to just play games all the time, hang out with friends, or play sports.

My parents had a very interesting way of raising us, my sister and me. When we were really little, it was an extreme level of discipline. When I was little, I played a lot of piano—essentially 3 or 4 hours a day. For competitions, they would just pull me out of school and go hard at it.

Fortunately for me, my parents decided that math was probably a better way to go, and I was quite talented at math when I was little. Even for that, it was just full force: You were committing everything. We did not have a TV in the house. We didn't have video games. We didn't really take vacations growing up.

You were in this mindset of sacrificing most things to focus on one thing. When you're a kid, you don't have a frame of reference, so you don't know—it doesn't feel hard necessarily, because your parents are setting up the criteria for you.

I'd say, in hindsight, I had a very, very happy childhood. But I think that level of discipline and competitiveness is very hard to gain after your childhood is over, because when you're in your childhood, your brain's still forming, so that kind of forms your personality. I'm very grateful for that.

I think that's also why, when I talk about my generation, there were a lot of immigrant parents from my generation who came over for grad school, and they're all around my age. I think this crop of folks is doing very well, partly because of that upbringing.

What my parents did, I think, which is the more unique side, is that a lot of times, especially with the stereotypical Asian parent, what happens is that it just continues and you have overbearing parents. Even though my parents were very intense about things, they never had any semblance of overbearingness. They wouldn't prevent us from doing things we wanted to do or force us to pick something.

17. Kindest Thing

Towards the end of middle school and into high school, I think we had already established these personalities. My parents had basically pounded a sort of lazy, wanted-to-play-around kid into someone who was very, very driven. Then, to their credit, they completely laid off.

They don't have any opinions on what we do for careers or what we should do. They're very supportive. You basically just need parents who are willing to spend a ton of time crafting this childhood for you to develop this. I think a lot of the things I have in life right now come from that.

So, yeah, that's probably the kindest thing. My sister as well—she was always trying to sacrifice things for me when I was trying to achieve things. Now we try to spend as much time with our parents as possible.

Well, this has been so much fun. I'm so fascinated by the business that you've built and are building. Thanks for explaining it to us and bringing us right to that white-hot center in so many different ways. Thanks for your time.

Well, thanks for having me.