Amazon CEO Andy Jassy on Agility, AI Strategy, and the Changing Role of Managers

Yeah. Well, first of all, thanks for having me on. I appreciate it. We talk about wanting to operate like the world's largest startup. And I actually think Amazon already, based on our DNA and the way that we have become a company over the last 30 years, moves very quickly. But when you get larger, there are all sorts of natural ways that you can get slowed down. And so when we talk about operating like the world's largest startup, we think about a few things. One is I think in whatever we build, whatever we spend resources on, we have to make sure we're solving a real customer problem. And I think a lot of companies, technology companies in particular, fall in love with the technology. And they are building things that they think are cool because they love the technology. But they get to the end, and they haven't really solved anything remarkable. And so you have to be—startups are missionary about trying to solve problems for customers. And that's what we've got to make sure we spend our time on. And then we need a lot of builders, disproportionately so. And I think of builders as people that invent, people that dissect a customer experience and figure out what's wrong with it, even if it's pretty good, and then rebuild it. And then you need owners if you're going to be a startup. You need people to think about, what would I do if this were my money? What would I do if I owned all the resources? Hey, I see that I own this piece of the problem. I don't really know if the rest of it's solved. Should I spend my time on it, or should I just assume someone's got it? You need people who really feel accountable. And it's part of what our effort in trying to flatten our organizations is about, is we want the people doing the work to think like owners. And then you really need speed. And I think speed disproportionately matters in every business at every time. In my old job, when I was managing the AWS business, I had the privilege of speaking to a lot of CEOs of companies and enterprises. And they would often say to me something like, I don't know if you understand, we're big. We have security issues. We have compliance issues. We have lots of organizations that have to be involved. We just can't move fast. And I think, honestly, that speed is a leadership decision. You can decide you want to move fast, but you have to figure out what's slowing you down and knock all those barriers out. And you've got to get the whole organization aligned that you are going to move really fast, even if you make mistakes. And so I think part of that, too, is you're a larger company, is you have more people, is you need to really try to root out bureaucracy. Well-intended people, as you get bigger, especially as you have a lot of managers, they keep layering in processes. And pretty soon you have process upon process upon process that really slows people down so they can't get the real work done. And then I would say the two other things you need if you want to operate like a startup is you've got to be scrappy. You can't think about every new project taking 50 to 100 people to do. We started in our cloud computing business, AWS, our storage service, we had 11 people—or 13 people—when we started. Our compute service, EC2, we had 11 people. You can get going with a small number of people and build something that people actually find resonant, and then keep iterating from there. And then I think you've got to nail—you have to be willing to take risks. And I think that as companies get bigger, they often get very risk-averse. If you hire achievement-oriented type A people, which a lot of companies, including ourselves have, they're not used to failing. And so a lot of times when they want to pursue something very different, they worry that they're going to be ostracized if they get it wrong. And so they'll play not to lose. And the only way to build something unique and different is to do something different from what people have done. And you have to be willing to take risks and be willing to fail sometimes.

Yeah. I think they're actually pretty consistent, though. I mean, there's listening to customers. Customer—actually, if you find the right feedback loops, customers will tell you what's wrong with your product or what's wrong with their experience, and they want to change. So there's listening to customers. And oftentimes, customers can tell you the top 10 things in their mind they wish they were different. But if you ask the right 'why' questions on what you're trying to solve, a lot of times customers will tell you what really is bothering them, what's really constraining them. And they can't sometimes tell you exactly how you should fix it. But if you're listening to them, and you understand the need, then you can invent on their behalf, which is a lot of what we do. A lot of the invention we do is listening to something customers are really struggling with. And they won't tell you how to do it. But we start asking ourselves, why do those constraints have to happen, and start to invent on their behalf.

Well, I think the first thing I would say is if you—most companies that have been successful for any period of time have a culture that's been a key part of their success. And that is for sure true with Amazon. And I think we have a really strong culture. But it is not our birthright to keep having a strong culture. You have to—things change—size of companies change, the scope of businesses you're going after change, the geographic distribution of your people change. And you have to keep working on strengthening the parts of your culture that you see being stretched if you want to keep being successful culturally. And for us, it comes back to this notion of wanting to operate like the world's largest startup. Amazon has hired really smart, really ambitious, really motivated people who we have given a lot of responsibility to as owners. And then we let them make the lion's share of the two-way door decisions. The two-way door decision is one where if you walk through it and you're wrong, you can walk back through it and no harm. A one-way door decision is when you walk through that door and you're wrong, it's really hard to walk it back. But the two-way door decisions, which is the overwhelming majority of the decisions we all make, we want to be handled by the people doing the work and those owners. But as you get larger, as we have, as we've grown so much the last 10 years, you end up logically with a lot of managers and a lot of layers of managers. It's the way you organize yourself. And so then you find that you end up with things like there's a pre-meeting for the pre-meeting for the pre-meeting for the decision meeting. Or you find that owners don't feel like they own the decision, can make the recommendation any way because—anymore because that decision is going to be made three levels up. And so that's what we'd like to try and limit. We really want our owners doing the real work, to own the two-way door decisions, and to be able to move quickly and autonomously. And that's part of this effort we've taken, which is to increase the ratio of individual contributors to managers by at least 15% across the company, which we'll beat by—we have beaten already by the end of the first quarter. But we want to flatten our organizations to move faster and to drive more ownership. The other thing we noticed was that when we brought—like most companies, people were largely working remote. And when we brought people back to the office three days a week in May of '23, we noticed that a lot of things about how we were inventing and how we were collaborating got better. And if you do a lot of inventing like we do—and our style of invention is often very collaborative. We're in meetings together. We're iterating with one another, that when you're together, that invention is stronger. What you find is people riff on top of each other's ideas better if they're together. Turns out, sometimes it's actually useful to interrupt each other because you get to a faster spot more quickly. You feel that energy. A lot of our best inventions have been after really messy, wandering meetings, when we've been trying to invent something where we haven't quite gotten there. And we've resolved ourselves to go have another meeting in a few days to figure it out. And then three people stay behind on a whiteboard and actually map out what it was they couldn't figure out. Or on the way back to their office, they figure it out. Or later in the day, they walk by each other's offices. When you're remote, it just—like, the meeting ends, and you're on to the next jingle and the next meeting, and you just don't find that type of invention together. I also think if you want to teach the culture, it's much harder to see it when you're remote. You have a lot of people who are lurking off camera. Or if they're on camera, they look like they're looking at the camera, but they're also working on a spreadsheet. And when you're in a meeting together, and you're watching the body language, and you're watching people's expressions, you internalize the culture much better. I think also the teaching and the apprenticeship is much better if you can walk over to somebody. Or after a meeting. Sometimes it's a hard meeting. And I may say, Adi, don't be surprised that that was a hard meeting. This is a really difficult topic. And next time when you come back, think about these three things. You just don't end up doing those as much when you're remote. And so we realized that we were much better for customers and for the business if we were together.

I would say that I think it's very hard to measure, truthfully. We've done a lot of measurement ourselves. Our data doesn't suggest what you said. But I also—when I look at that data, it's just very hard to measure. I mean, how do you measure how well you're inventing? You don't actually know how well you're inventing for probably a few years because it takes a while to get through the invention process, and you have to hire a team. Then you've got to go build the product. Then you've got to get the product in the market. And then you've got to see if people respond to it. Oftentimes, it takes several iterations. So it's quite difficult to measure. But I will say I spent a good amount of my own time in product and invention meetings. And there is no comparison to being in person and remote.

Well, first, I would say that I don't believe that we have it nailed or perfect yet. I think you have to keep iterating at it. But I think the first step is to want to address it. And it's actually not simple to address because you get used to operating a certain way, and then it seems impossible. How can you change the organizational structure to lose a lot of that bureaucracy later? So I think the very first thing is the leadership team deciding they want to actually change it and resolving themselves to take action. And then, of course, it matters organization by organization. But in our case, we really felt like we wanted to have more ownership with our individual contributors and the people doing the actual work. So that's why we've taken this action of trying to flatten and have fewer managers. I think some of it is actually getting visibility into what's actually happening. So when we announced that we were going to take this goal to flatten and to increase the ratio of individual contributors to managers, we started this no bureaucracy email alias, where we encourage anybody in the company that thought they were experiencing bureaucracy to email me at this alias. And we tried to explain there is a difference between process and bureaucracy. Sometimes people just not like a process and they'll say it's bureaucracy. And most companies of any scale need process to scale the right way. But there is really a difference. And the difference is often processes that have been layered in that don't really add any real creative value. And I've gotten over 1,000 emails at this point, and not all of them are true bureaucracy. And sometimes it's people saying they don't like this manager. So it's not 1,000 pure bureaucratic examples, but there are plenty. And I've read every single one, and so have the relevant leaders in my org. We have changed already 375 processes because of these emails. And it's really hard to see some of the red tape deep in your organizations if you have large organizations, like a lot of enterprises do. But getting that visibility starts to—you can knock down a bunch of things if you see them and you resolve to change it. And it also starts to motivate you to just look for opportunities when you're in meetings to reinforce with teams that you're not going to tolerate bureaucracy. And so I think it starts with the resolve. Then you've got to figure out where you think the biggest problems are initially. Then you have to have good feedback loops to see where all the issues are, and then you got to keep working on it.

I think AI is probably—it's for sure the biggest technology transformation since the cloud. It's probably the biggest technology transformation since the internet. And so I think it's going to change every experience that we know. We have a very substantial investment in the AI space right now. And I think in the early days of people getting excited about generative AI, people forgot that if you really want to pursue AI in earnest, there are three macro layers of that AI stack, all of which are gigantic, all of which we're investing in. But we invested in a bunch of areas that didn't get as much attention early on. But really, at that bottom layer of the AI stack, is for model builders. And what they care about is two things. They care about the compute to do the training and the inference, which is really the chip. And they care about services to make it easier to build models. And we've built a chip there, our own custom AI called Trainium, that's going to help people save a lot of money relative to what the cost has been to date. And we've built a service called SageMaker, which has really kind of become the standard way for people building their own models to get the data in, to build a model, to experiment, and to deploy into production. And so for model builders, I think we've had services there, some of which have been around for a while on SageMaker that maybe aren't as public but have a huge amount of traction, and then chips that I think people are excited about. At that middle layer is for people that don't want to have to build their own models. They want to leverage an existing frontier model. They want to customize it with their own data. And then they actually want lots of features that make it easier to build a high-quality generative AI application. These are things like guardrails so the model doesn't say things you don't want it to say, or rags that you have up to date information, or agentic capabilities, so you can take a bunch of automated actions in succession. And we've built this service called Bedrock, which has the largest selection of those leading third-party frontier models, including our own. But it also has the best collection of features to help you build a high-quality generative AI app. And Bedrock is another one of those services that has substantial traction, really significant traction, if you talk to enterprises or smaller companies, that also hasn't gotten as much public limelight. But these are services that are part of why we have a multibillion dollar annual revenue run rate in the AI space. And then the top layer really for applications. We're building some applications. We have something called Q that's the best AI-powered coding assistant. But the overwhelming majority of applications are going to be built by companies. We have over 1,000 generative AI applications across Amazon that we've built or are building. But most of them will be built by other companies, hopefully, on those first two layers of the stack that I was mentioning. So while there was a lot of attention early on on ChatGPT, which was kind of the only really large-scale generative AI application, I think people have slept a little bit on the other layers of the stack, where we have big investments and are doing really nicely.