Airbnb’s radical AI future (Part 2)

BRIAN CHESKY: Modern corporate America and the modern way a tech company is designed is suboptimal. So what if we could design it differently and better? My approach isn’t to push decision-making down, it’s to pull decision-making in, to pull people in. So if I am like the sun of the solar system, my job is to pull the planets close to me. Now, it gets hot close to the sun. And as long as you’re okay with a little sun tan, you’ll do just fine.

BOB SAFIAN: That’s Airbnb CEO Brian Chesky, and this is Part 2 of a special edition of Rapid Response. If you haven’t listened to Part 1, I encourage you to go back in the show feed. You’ll hear Brian’s theory behind earnings releases, how Airbnb’s branding taps into culture and celebrity, and why what he learns from longtime friends can be more important than what CEOs tell him.

In Part 2 here, we get an inside peek at Brian’s partnership with legendary Apple designer Jony Ive and how they think AI will spawn a new interface for engaging with tech. Along the way, Brian shares his opinions on the hotel industry, on ChatGPT, and his own place in the sun. I’m Bob Safian. This is Rapid Response. 

Copy LinkHow Airbnb is implementing AI

You alluded to AI and the discussions you’re having with some colleagues in Silicon Valley about it. So how is AI being used at Airbnb now? Like, does AI apply less to your business because it’s a physical experience company or more because it’s a digital platform?

CHESKY: It’s probably, okay, like, let’s create a framework. AI is going to affect, in the near term, digital things more than physical things. Let’s start there. Like there’s robotics and autonomy, but it’s mostly affecting the digital world, and there’s not a near term where like the way you go to a restaurant or you’re walking down the street’s affected by AI. Not in the next few years. So in that sense, we may be a little less affected than a purely digital company. And also, therefore a little less at risk of disruption. 

At the same time, compared to hotels, Airbnb’s way more digital. Like, if I was, I don’t think a hotel is going to change like almost at all in the next five years. I know that they’ll have some AI software to like, blah, blah, blah, blah, blah. But like you go to a Hilton or Marriott five years from now, it is not different because of AI. And if it is different, you’re not discerning the difference. So AI is going to affect us much more than hotels, much more than anyone else. 

And the reason why is because of a couple reasons. Number one, you go to a hotel, every room’s the same. If you don’t like the room, you go downstairs to the front desk and that’s a reasonably solved problem within that framework, and AI is not really gonna like, it can only affect that in margins. Airbnb has no skews. Every single thing is essentially one of a kind. There’s no product catalog. There’s no attribute catalog. We only know if the person in Istanbul told us this is their bedroom.

SAFIAN: Right, there are ways that I can sort on Airbnb that try to find things but—

CHESKY: But every stay is one of a kind, like a person, right? Every person is one of a kind. Every house is basically one of a kind. No two homes are the same. And even if they were the same, they’re not furnished the same, and they’re not maintained the same. So no two homes are the same. So we have a long tail of skews. That’s the first problem. 

The second problem is: we have very, very complicated issue types. They’re much more complicated than a hotel, right? You manage a hotel, like, there are only so many problems that can go wrong. It’s on one premise. We’re in 2020 countries and regions. Almost half of our business is cross-border. That means that almost half of our business, you have somebody likely speaking a different language than the person they’re staying —  so you might have a traveler from Japan, staying with a host in Italy that calls a call center somewhere else that are not in either of those two countries about an issue at 11pm at night that they need to fix at 11:30pm at night. And the host isn’t on-prem. You see I’m saying that the issue might be like the heater broke, and there’s no way to know what photo evidence the heater broke. You gotta’ take their word for it. It could also be a much more serious issue. And so, customer service in our business is like, it’s not impossible, but if there is a nearly impossible job, it’s Airbnb customer service. It’s very, very difficult. 

And so imagine we have agents and they’re trained, but there’s like a thousand pages of documentation that no one can ever know or remember, right? Imagine AI can read all that documentation. AI can also speak every language. AI can then do a regression and look at the last thousand times this happened. What was the most likely course of action that led to a resolution? It doesn’t mean AI will do all the customer service. These bots, these models are not autonomous. They’re not even close, they don’t have agency, they cannot really reason, but you can create rules, and they become a very powerful tool for the agents. So, the near term is they’re going to be the front line for some easier tickets, and it will really allow agents to be actually more personal, not less personal. So that’s kind of phase one of AI. 

Copy LinkWorking with legendary Apple designer Jony Ive

So then, we say, “well, where do we take it from there?” One of the designers that I work with is Jony Ive, and Jony Ive, for those who don’t know, most people probably know him, was the person who designed every Apple hardware device from the iMac through till he left before the Vision Pro came out. And he left Apple in 2019. He started this firm called LoveFrom, where us and Ferrari were his two anchor clients, we’re the main. And I’m very close to him. 

And one of the observations he made to me, like this is even before ChatGPT launched, he said, “it’s kind of weird that I open all these apps and none of them know me.” And even our app, like there’s a search box and you have to enter a destination. We’re all stuck in this, like Amazon, eBay, 1997 paradigm. It’s like this internet-era interface. You had the graphical user interface in the ‘70s and ‘80s, then you had the internet, then you had basically multi-touch in 2007, 2008, and everyone’s waiting for a new interface paradigm. And I don’t think that interface paradigm in the next 10 years is the Vision Pro. I think that’s like a world’s fair thing. And maybe in 10-plus years, if the form factors or maybe eight years, I don’t know, but not in the next three to five years, that’s not the platform. 

So I actually think, there’s an interface paradigm in the next five years that’s going to be powered by AI. And it’s not a chatbot. And there’s two problems with OpenAI. I mean, there’s more than two, but there’s two that I’ll call out in the interface layer. The first problem is if you and I search, if you and I type a query in OpenAI, we’re going to get the same answer, or we’ll get a different answer that’s randomly different, not cause you’re Bob and I’m Brian and you— 

SAFIAN: Only because the AI is run— 

CHESKY: It doesn’t know you. It doesn’t know me. And the second problem is a chat interface, and a number of people in … When ChatGPT came out, said “chat’s gonna be the interface of AI.” It’s not, and Jony Ives doesn’t think so. I don’t think so. And no great designer I respect thinks a chat interface is the best interface. You want a chat interface for chatting. But like, you don’t want a chat interface for like, a calculator, or to find the weather. Every application wants a different interface. 

And so what we are working on with our own internal team and with Jony and others is what is the interface after multi-touch? What is a new interface paradigm for this new age? And I can’t say much more than that. Well, I’ll say a couple things: One, I think it’ll be more dimensional. I think flat design is over. I think we lived in skeuomorphism design in the 2000s. We went to flat design. I think we’re entering a more three-dimensional design again, but not so literal representation like skeuomorphism. Because the more time you spend with a screen, the more humans want to see dimension. It’s like, we don’t want to live in flat worlds. 

But then the second thing is that the application layer is going to be so much more personal. So you’re going to open an app, and it’s going to learn about you. It’s going to understand you. And we had this vision before ChatGPT launched. We just didn’t know how to do it. We had this concept of this app that said, ‘Hey Bob, tell me more about you’. “Okay. I am X years old. I live in Brooklyn. I grew up here. I like this food, but also like: I know I never told anyone this before, but before I die, I really want to do this. And I’m kind of a little freaked out about this.” And you know, whatever. I know you might not trust an app to tell it that, but the more you trust the app, the more you’ll tell it. And the more you tell it, the more it could almost be like the ultimate concierge. 

So this was kind of our concept, and it could, like, really understand you and recommend like where to travel, what to do, who to meet. But we didn’t have the technology available. We’re like, “this is like some futuristic thing.” And then when ChatGPT launched, we’re like, “Oh my God.” And you know, that model is not strong enough. GPT4 is not strong enough. Even GPT5, which is coming out soon, isn’t strong enough. But it will be very strong. But probably by the GPT6 era, the technology will be there.

And what I’m excited about is there’s been a lot of innovation in the models. There’s not been a lot of innovation on the application layer. Anyone watching this or listening to this, open your phone, and I want you to look at your home screen. I want to ask yourself, ‘how has generative AI changed any of those apps in the last 12 months?’ And maybe they changed, like, the recommendation algorithm a little bit, right? Like, maybe TikTok’s blah, blah, blah. Not really. Every app’s basically the same as it was before ChatGPT launched. So we haven’t yet figured out what the application layer is for AI. I think it’s dimensional. It’s personal. It’s more conversational. But it’s not chat.

SAFIAN: But it’s still physical as opposed to, like you said, it’ll talk to you? It’s more physical than it is oral or verbal? 

CHESKY: The future is absolutely not voice. Not voice alone. And this is another thing that I talked to Jony about and others. And what we’ve all talked about is like, there’s a lot of problems of voice. There’s a lot of good things about voice. Okay. Twitter’s got hundreds of millions of users and it’s primarily a word-based platform. But TikTok and Instagram are a lot more popular. Words are not the most comfortable way for the average person to express their preferences and desires. If it was, Twitter would have billions of users, and no one would use TikTok or Instagram. So I think a lot of people are very visual, and I think people in Silicon Valley who aren’t as visual, they tend to build the things they want and like, well, you’re not everyone. You like to just consume a lot of content in the written word. 

The second thing is that you’re not going to walk down the street talking like in the movie, Her —  like whispering on the subway or in the bathroom stall. So there’s a big privacy concern there. 

The third thing is most people aren’t articulate enough to summon the words to express exactly their desire. And it just takes a long time. It’s like words are low bandwidth. And so what we think it’s not, the voice isn’t part of the future. But that’s multimodal: it’s voice, it’s photo, it’s visual …  It’s like this. It’s a little hard. I don’t even see, I’m proving my point. I can’t with words express quite what I’m trying to communicate to you.

SAFIAN: But when I see it—  

CHESKY: You’ll know. There’s no doubt you’re gonna have a conversation like Her, it’s just that’s like 20/30% of the interface. That’s not 90%.

Copy LinkCan AI make Airbnb your personal concierge?

SAFIAN: And this new way of interacting, does this connect back to your vision of the way Airbnb’s business is going to change and grow? 

CHESKY: Yeah, a hundred percent. I think there’s like a few big transformations are going to happen. One is that we’re going to go from selling homes to selling a lot of other things. I want Airbnb to feel more like a concierge. Those are my words. People are calling them agents. They’re all the same thing. Everyone acknowledges this agent/concierge type thing is where it’s going, but no one, it’s like a fuzzy idea, and everyone’s describing this thing, and it’s not a chatbot, it’s not pure voice…  And so what is it? And that’s where everyone’s waiting for the ChatGPT of interface. And I’m not saying we’re going to do it. I’m just saying there needs to be a ChatGPT moment for interface that creates the language because otherwise you’re going to have GPT5, GPT6 from your phone, and the apps are still the same. Very few apps have actually fundamentally changed that we use day to day.

SAFIAN: I’ll be honest, this is the first time I’ve heard someone talking with such passion about AI enabling a transformational interface for technology. But as a previous guest on this show told me, when tech titans have decided to invest big into something – and Brian and Jony Ive definitely count as tech titans – you should expect something to come from it. 

What I’m also hearing, first in part one of this interview and now as Brian talks about AI, is how Brian plans to expand Airbnb’s footprint, as he puts it, “selling a lot of other things” beyond just home stays. We’ll get some more clues about how that will happen in the next section, where Brian reveals choices he’s made in how Airbnb operates and the risks that go along with them. Stay with us.

Before the break, we heard Airbnb’s Brian Chesky talk about his partnership with Jony Ive on a new AI-based interface paradigm. Now, Brian shares how he’s trying to keep a start-up spirit at Airbnb, even as it grows. Let’s jump back in.

Copy LinkOperating like a start-up

SAFIAN: You talked about the difference between a start-up and a mature company and that, at least for you, running the mature company is more fun. But you also mentioned that there are things about being a start-up that you want to keep in the Airbnb sort of ethos. So what is that balance? Do you still think of Airbnb as a start-up, even if you’re no longer a start-up CEO?

CHESKY: Yeah, I do. I’ve had a whole conversation with people about at what point does a company cease to be a start-up? Some people say it’s when you go public. Some people say it’s some arbitrary, like, scale threshold you reach. Like, you kind of have a full exec, you’re a quote, big company. But I have an ambition till the day I’m no longer running this company, which I think is hopefully decades from now. Let’s just start that. I’m 42. I’ve been doing this since I was 26. I hope to do this for a couple more decades. That I hope when I’m done doing this and retire, that I retire as a start-up CEO and that Airbnb is the biggest start-up in the world. 

And so what do I mean by that? Well, first of all, we’re one of the only companies in the Fortune 500 organized like a start-up, meaning we’re functional. Almost every start-up starts with an engineering department, a design department, and a marketing department. So they start with functional departments. And at some point, almost every single company, they divisionalize. Because there’s like the engineering department has too many requests and there’s too many decisions that they create separate engineering teams. There’s separate teams subdivided, and we call this a divisional structure. They quite literally divide the divisional structure. 

My recollection was created by DuPont after World War Two. I think they were creating gunpowder and they’re like, “wait a second, powder can turn into paint.” But you don’t sell paint the way you sell gunpowder, and they created separate divisions. So a lot of companies divisionalize. We’ve chosen not to do that. It was crazy. Everyone says, “if you don’t divisionalize, decision-making is going to screech to a halt.” And we had a theory that if we don’t divisionalize and we stay functional, that we’re going to have as few employees as possible. We’re not going to be the Navy. We’re going to be the Navy SEALs. We don’t have a lot of junior people at the company. So almost everyone joins experienced. People don’t have their own goals. People don’t have their own priorities. All priorities are shared. 

SAFIAN: How do you keep people who are in different functions communicating? Because sometimes they become silos. 

CHESKY: Oh, this is such a fun conversation. Okay, so we want to be the biggest start-up in the world by continuing to run like a start-up. And so, what are the characteristics to run like a start-up? The first is you’re functional. The second thing is that you’re in the details.

I’m in all the details of the company. That also means we can’t do too many disparate things. The main downside to the model I’m describing is not that it’s slower, it’s actually faster. You can ship faster, people feel more empowered, not less empowered, mostly. This is very, the people don’t understand this. The primary downside, and you can do a lot of things. You can’t do disparate things, so long as you do many things that all require the same functional capability, you can do 10, 20, 50, 100 things. It’s endless. But this model prevents us from going to some things. 

SAFIAN: So it enforces a certain kind of discipline too. 

Copy LinkWhat Brian Chesky’s day-to-day looks like

CHESKY: And focus. And we can’t be in more details than I can be in. Most of what I do all day, every day, is I review the work. And so, everything in the company is reviewed every one week, two weeks, four weeks, eight weeks, twelve weeks, six months, annually. It’s like, everything’s on a rhythm. 

And when I review it, I review it with the full chain of command, so the head of like… The functional people will be there and then the people doing the work are there because I need them to be able to answer the questions. I’m not deciding on everything. I’m mostly signing off on decisions. But what I often do is I’m pushing. I’m getting everyone to talk. Most CEOs don’t do this. They think, like, “I don’t, first of all, have time to do all the work. And I’m meddling. And you’re closer to the customer. You know better than I do.” The CEO, you have to be close to the customer. And you have to be. So I’m like an orchestra conductor. 

So we don’t have a product management function in the classic sense. We have product marketing and program management. And what people call product management, we call product marketing. It’s product management plus outbound marketing. It’s one function. It’s smaller. It’s leaner. It’s central. And then we have a really robust central program management function, which is what most people call product management, but they’re purely program management.

And so the entire company is kept on one rhythm. I can quite literally feel the work of an individual engineer. And almost no other CEO in the world can do that. The reason I can feel the work of an individual engineer on a Friday is because I can see the assembly of the entire product. And as I see the assembly, if the paint is off or this tire is off, I can identify, and I can ask, well, “who did that? What happened?” But also, I can unblock people.

So there’s not a lot of politics. Not a lot of infighting because everyone’s aligned with the same priority. And what I’ve created is a shared consciousness of the top 30 or 40 people in the company. We’re in all these meetings together. We’re constantly debating. Constantly talking, and my approach isn’t to push decision-making down, it’s to pull decision-making in, to pull people in. So if I am like the sun of the solar system, my job is to pull the planets close to me. Now, it gets hot close to the sun. And as long as you’re okay with a little sun tan, you’ll do just fine. And, I mean, the downsides, it’s very intense to work this way. 

And then we do a top leadership offsite twice a year where we have like 70, 80 people in the company, and we do a two-day roadmap review where we go through everything that we’re shipping over the next two years. And it’s one keynote that we just update. It’s a rolling keynote and that keynote is the same keynote we show the leadership team and the board. It’s a living two-year roadmap. We don’t have an annual plan. We don’t do an annual planning process. We do a two-year plan that’s updated every six months. I don’t think you should do an annual plan annually because the problem is that, if you make the plan in November for the following year, by August, you’re not working on a 12-month plan, you’re working on a 4-month plan or a 3-month plan. 

And so we weigh the annual budget. The annual budget is like, we don’t hire a lot of people. We’re going to spend basically the same as a percent of revenue on marketing. We’re going to make some efficiency targets and then we will incrementally invest throughout the year, not necessarily on a purely annualized basis. If we see an opportunity, the ROI is great. We’ll tell Wall Street, and we’ll make the incremental investment. 

And so that also means we’re always longer term. It’s always a rolling two-year plan. It’s more up-to-date. It’s updated every six months, and I think it goes back to the fact that I’m a designer by training, and I try to even think about how you can design a company differently. I just looked at all the problems, and the CEOs are always like reacting, and they don’t know what’s going on, and they have to therefore create all these processes and these rules. And I just thought — modern corporate America and the modern way a tech company is designed is suboptimal. So what if we could design it differently and better? And not to say it’s better, I don’t want to come across like we solved the better problem. One will find out in five years if it’s better based on our results.

Copy LinkHow Brian Chesky learned to make better decisions

SAFIAN: You’re very confident. And a lot of these decisions come down to you making them. How do you check yourself? 

CHESKY: Before you wreck yourself. 

SAFIAN: Well, to see when you’re wrong. We’re all wrong sometimes.

CHESKY: Gotta check yourself before you wreck yourself. Yeah, so a couple things…  

SAFIAN: You know, because some of those folks close to the sun, they might get burned 

CHESKY: So, let me preface the answer to your question by saying: I’ve wrecked myself more than the times I would wanna admit. Between 2014 and 2020, I greenlit and authorized, and pursued a lot of things that did not work out. We can list the things, but maybe we don’t have to right now. But those can do their research, and they’ll find out. Like I made a lot of mistakes and I grew fairly just detached and disconnected. So I studied Steve Jobs, Walt Disney, people who also made the decision, but the decisions were almost always right. And it’s like, how does that happen? And it turns out, there’s a couple of things… 

I don’t make the decisions alone. I go back to the shared consciousness, that if I’m in a meeting, I often speak last. 

SAFIAN: Speak last?

CHESKY: Speak last. And I’ll make sure that all the other people last. If I am in a meeting, I mean some, okay — I don’t always do that. Sometimes I’m certain, like, “this is like, we should not do this” and I’ll say it. But a lot of times, at least half the time, I’ll turn to Hiroki, my head of design, marketing, and product marketing report, to him —- so whatever you call that, he’s a consolidation of three functions. And I’ll ask him, “what do you think?” And they don’t know what I think. So at that point, they have to tell me what they really think.

And tip number two: I change my mind. Sometimes I genuinely change my mind. But sometimes I’ll wear a different hat on. So everyone’s like, “you’re right, Brian, let’s do this.” And the next day I’m like, “actually, maybe we should try to do this.” “You’re right, Brian, let’s do this.” I’m like, “well, which one? I can’t tell.” And so—  

SAFIAN: So you’ll change your mind almost as a tool? 

CHESKY: As a tool. I mean, it can almost sound manipulative, but I’ll tell people I’m doing this. It’s almost like you try on a shoe, and you’re like, “this fits really well.” You’re like, “well, let me at least try on the other shoe.” So we call it trying on ideas. We try on one idea, and it feels great. We try on the other idea.

We’ll also often try to argue both sides of the point. Not like a red team, blue team. But like, from a first principle. If you do something from first principles, you’re usually right. You only get wrong if your principles are wrong. 

We try to also look at things in their most native form. So I’ll give you an example: In the past, I used to like design features, I’d ship them, and it would never do what I wanted to do. It was like I can never predict. So now we have a rule that we prototype things. So instead of showing me wireframes in Figma, you have to show me a working prototype. And then we sit with the working prototype for a long time. We put it in the hands of the employees. Then we do lots of beta. We don’t do a lot of A-B testing. We do some. But we do a lot more like the way you do in an operating system. We do a lot of private access.

SAFIAN: You said you’re faster, but it sounds like doing all this testing — does it make it slower?

CHESKY: It’s slower, and then it’s faster. So the way I’m describing it, Initially, everything is slower. I mean, first of all, to set a company up like this is incredibly slow. You’ve got to reorganize the company. At least a third of your people are going to quit if you do it this way. So let me 401 you, cause not everyone wants to work this way.

All the decisions are going to come through you. It’s going to really, really slow down. But eventually what’s going to happen is it’s like everyone, like on a crew team rowing in the same direction, they’re going to pick up steam. 

SAFIAN: And then it moves. 

CHESKY: And then it really picks up. And then, I review all the work and I’m changing it. “No, that’s not right. It’s not right. It’s not right.” I mean, three years ago we worked on a press release, and I did 70 revisions to the press release. Like, we did 70 drafts. And at some point somebody said like, “if this is how it’s going to be forever, it’s never going to work.” And I said, “the good news is, it’s not.” It’s like a golf instructor practicing your swing 70 times before you get the swing. And now, we only do 20 revisions. And then, eventually we’ll do two or three. But you’re building a muscle of excellence, and now I don’t review a lot of the work because people know what I’d say —  it’s ingrained in them.

So this process is initially slower, but then it speeds up. Like fill-in-the-blank tech company, you can ship something quickly, but the number of things that don’t work, like 90 percent of things don’t work the first time. So now, unless you’ve got to ship multiple versions. You were A-B testing, you got to design A and B, so if you can design it once, right the first time, and it works 90 percent of the time, you have a lot less waste. And then, if everything is just one foot in front of the other, you’re not swinging in different directions, it accumulates speed, it accumulates mass.

You’re also not delegating understanding. Charles Eames said, “never delegate understanding.” I think this is a really key lesson. And so Robin Oppenheimer, I studied him. He basically kept the whole bomb in his head. I think it’s important that a leader pulls everything into their head, not pushes everything out of their head. And so if you do it this way, it’s like a bicycle riding down the hill. It eventually gains speed. Now again, all designs must have design flaws acknowledged. This momentum can be a runaway train. if you’re going down the wrong direction.

And again, this is not good for disparate things. One of those disparate things might be the next big opportunity we miss. So that’s a big risk. This is also the risk of Apple. Like the reason they don’t do a bunch of things they don’t do is because they’ve also wanted to be functional. And so they can’t do disparate things. So all designs have to acknowledge the trade-offs, but we’ve just chosen that knowing those trade-offs, we’re going to go this direction.

SAFIAN: Well, you have pulled lots of things into your head, and I appreciate you allowing us to go into that head for a little while today.

CHESKY: Thank you very much, Bob. 

SAFIAN: Thanks, Brian. 

Brian is a bubbling cauldron of energy. Just nonstop. And in the years I’ve known him, it seems like that energy just keeps growing. So what do I take away from this most recent, intense, in-depth chat with Brian this time? Brian’s designing Airbnb to move fast and revolve around him, some great insights, and as he admits some clear risks. 

Brian also dropped a bunch of hints about his future product plans for Airbnb. He was careful not to give too many details –- maybe because he knows that the plans themselves are still taking shape. But the aspiration is clear: to move Airbnb beyond lodging —- and to do so via an AI-enabled tech interface paradigm that we haven’t seen before. It’s an ambitious vision, but then Brian’s an ambitious guy. 

Next time on Rapid Response, I’ll share a very different conversation, with two business leaders unafraid to talk about the most controversial, polarizing challenges facing America’s executives. I hope you’ll join us. I’m Bob Safian. Thanks for listening.