AI 改变一切后，如何重组组织｜Block 的 Owen Jennings 做客 a16z Show

The biggest moat is going to be which companies understand something that's super hard for other people to understand. If your answer to that is, “I don’t know,” then you maybe could get vibe code in the way.

Block was one of the first to make a pretty drastic decision, cutting 40% of the workforce. What led up to that decision?

There’s been this correlation between the number of folks at a company and the output from the company for decades and decades. I think that basically broke the first week of December. What we were seeing is that one or two engineers—or a designer and an engineer—who were on the tools, quote unquote, were able to be 10, 20, or 100x more productive.

Over time, it’s pretty obvious that these systems are just going to be so much better than having 1,000 humans who are doing that work. I do believe that, fundamentally, for a given product or a given roadmap, you’re going to need fewer engineers, fewer designers, and fewer PMs. I think that’s very, very clear.

You show up on Monday, 40% of the company’s gone. What’s the most meaningful difference in how you’re operating?

I think the biggest thing is

What does it actually look like for a large public company to restructure itself around AI?

Owen Jennings is the business lead at Block, where he oversees product, operations, and customer support across Square, Cash App, and Afterpay. Before this role, he was the CEO of Cash App during its critical scaling period. Recently, Block executed a roughly 40% reduction in force, and they’ve been pretty candid about AI being a critical component of that decision.

Owen has gone through the AI transformation at scale across product lines and business units. We’re going to dig into that decision around the RIF, how Block has adapted, and the current and future state of the business. Thank you so much, Owen. Welcome to the stage.

Awesome. Jonathan did an amazing job setting the stage for this conversation, talking about how important it is to be founder-led. Block was one of the first to make a pretty drastic decision, cutting 40% of the workforce. Maybe walk us through what led up to that decision and how you thought about it.

Sure.

I would probably start 2 or 3 years ago. One thing about Jack is that I find Jack to be generally right and generally early—sometimes very early. I think that’s flowed through Twitter, Square, Cash App, Bitcoin, and so on.

We were pretty early on the agentic development side. We actually launched Goose, which was the first agent harness, at least that I know of, in early 2024. That started to augment how we approached software development and how we thought about internal tooling. Over that period, 2024 and 2025, there was meaningful progress.

Then, in late November or the first week of December, there was a binary change. You basically had Opus 4.6 and Codex 5.3, and you got this shift where I think the tools and the foundational models were pretty good at writing code, especially for new ventures and greenfield projects. It became clear, almost overnight—maybe in a couple of weeks—that they were incredibly capable of working with existing, complex codebases.

There was a massive paradigm shift. At least from my perspective, there had been a correlation between the number of folks at a company and the output from the company for decades and decades. I think that basically broke the first week of December. What we were seeing is that one or two engineers—or a designer and an engineer—who were on the tools, quote unquote, were able to be 10, 20, or 100x more productive.

That’s really what led us to make the decision a few weeks ago. We spent Q1 discussing what this fundamentally meant in terms of how we were going to build products, how we were going to build software for customers, and how we were going to run a company. What would it actually mean to run a company? We spent Q1 as an executive team, with Jack, working through that.

Ultimately, that led us to this place where we did a reduction in force that was slightly greater than 40%. That wasn’t even, you know, the conversation we were just having. The tools are flowing through really meaningfully on the development side, and so the cuts were much larger on the development side.

If you think of something as outbound sales or account management, the cuts were fairly de minimis. That was really what we were reacting to. The tools were flowing through meaningfully on the development side, and so the cuts were much larger there.

Can I push you a bit on this? Alex, when you introduced the conference just an hour ago, you talked about the surf period. How much of the RIF was overhang from 2021—kind of overhiring—versus AI and the actual productivity gains that are going to be in the business?

If you look at where we were from a gross profit per full-time employee basis from 2019 through 2024, we were basically right in the middle of the pack with all of our competitors. If you look at last year, I think we were in the second quintile or something like that. I think it’s basically Nvidia and Meta that are ahead of us.

When you look at the composition of what we did, if you thought it was cruft and bloat and so on and so forth, then this RIF would have accrued to the operational teams and that sort of stuff. We made really, really meaningful cuts on the development side. You don’t make really, really significant cuts on the development side if you’re not seeing a technology and a tool that have fundamentally changed how we build.

We’re not writing code by hand anymore. That’s over. That’s done. Everyone has their narrative, but it’s largely not true.

Maybe just walk through, tactically, how did you actually execute this transition culturally and operationally in the business?

The nice part about this RIF, relative to some other things that have happened at Block or at other companies, is that we were coming from a position of strength on the profitability and operating-income side. Sometimes, when it’s really financially motivated, the CFO or the CEO says, “Okay, we need to do a 16% RIF in order to hit this target.” That wasn’t the case at all.

We said, “What should the org look like, given how these AI tools are flowing through now and what we expect to happen in the coming months and quarters?” We had some core principles. The first one was reliability. When you do something this size, the worst-case scenario is that you have an outage or go down. That’s P00—not acceptable at all. Obviously, things have been great over the past several weeks, which is fantastic.

The second was building trust with customers and compliance, and navigating the regulatory environment. We all operate in a super-complex, nuanced regulatory environment. That’s non-negotiable. We have to make sure that we’re doing that right.

For instance, we basically did not touch our compliance team or our compliance technology team. Even if the tools are there, it’s like, “Let’s not take any risks.”

Third was continuing to drive durable growth. There are things on the roadmap that we already know we’re building, and we need to continue to do that. We know that it might be a squad of 3 people instead of a feature team of 14 building that. We’re going to make sure we continue to build those features and make longer-term bets.

Then we built up the org from scratch. In some areas, like the regulatory council team or the SDR/BDR team, the org looked pretty similar to how it looked in January. On the development side, it looks completely different.

From an execution perspective, we thought very deliberately. I’ve been in the company for 12 years, and a number of folks we parted ways with are friends and colleagues of more than a decade. We were in a position where we were able to be generous in terms of the severance packages we gave.

We didn’t cut people’s technology access instantly, which can suck. We chose to have an all-hands with everybody at the company, so Jack and the executive team were looking each other in the eyes and explaining this decision and the drivers behind it.

I think it was on a Thursday. The Friday, Saturday, and Sunday involved a lot of shock and dealing with ambiguity. Since then, we’ve massively reduced the number of meetings we have—probably by 70% or 80%. I now have time to build and work; it’s not back-to-back meetings.

We’re also meeting with the company every week. We have a 1- or 2-hour all-hands with Jack every Monday. It just feels like we’re smaller, leaner, with fewer layers and larger spans. It’s been back to building.

You show up on Monday and 40% of the company is gone. What’s the most meaningful difference in how you’re operating? I don’t know—maybe it’s in the EPD org or elsewhere.

There are a few different components to this. One concern I have with how some of these org changes might flow through the tech industry gets back to the founder-led point. If you’re not founder-led and don’t have the ability to be bold, you’re probably going to take a more incremental approach.

The way that’s going to feel is that you do a 15% RIF and it’s like, “Oh, it’s fine.” Then you do another 15% RIF.

And then culturally, that's just devastating for your team because there's always this pending RIF looming over your shoulder.

This was obviously a decision to go in a different direction. I think one of the benefits that we got from this is that we were already seeing a very meaningful increase in AI tool usage, especially on the development side. This is just a massive forcing function. If we're building Moneybot and we want to roll Moneybot out to 50%, and there used to be a team of 15 people working on it and now there's a team of 4 people plus $2,000 on the tokens, that's unlimited access to tokens and you can use fast mode on Claude Code.

So now you have 4 people plus the tools. It's like, "Okay, well, you need to have 8 instances of Goose up, and you need to shift your workflow from sequentially working through a PR, submitting it, getting a review, and making the change, to: I have 14 agents who are building PRs on my behalf right now, and I'm going to context-switch between all of those."

And it's not just on the software development side. It's for PMs, too. It's for growth marketers, too. The biggest shift, myself included, is that I have countless agents running right now that I have to go check on.

It's less of a linear workflow and more of a situation where, in the background, there are 10 or 20 agents doing a whole bunch of stuff. Then I have to check in on the work, nudge it, change it, and what have you. Then I can commit it to GitHub, get the Markdown file, put it in the source of truth, and move on.

We have a lot of public companies in the audience. We have a lot of founder-led businesses in the audience. Do you expect other companies to follow a similar path? And I guess what conditions need to be in place for that to be successful?

I don't necessarily want to—I talked at the beginning about the groundwork that happened in 2023, 2024, and 2025. We built this agent substrate, Goose, and then we built a lot of tooling at the company on top of it. We have an agentic operating system, internal only, called G2, where anyone can automate any deterministic workflow.

Anyway, I think there's work to do to be successful. I would expect many companies are doing that work. Some of them are incredibly far ahead of others, and so I don't know what to expect.

What I will say is that, to the extent that I do believe that fundamentally, for a given product or for a given roadmap, you're going to need fewer engineers, fewer designers, and fewer PMs, I think that's very clear after December. That doesn't necessarily mean that there are going to be fewer engineers, designers, and PMs in the world.

It's the classic Jevons paradox thing, where I think there's probably now just a superset of things that can be built. So I don't know: a given tech company might be way smaller, but there might be 50 or 100 more tech companies, or you're going to start getting this development work in sectors and areas where that hasn't historically been the case.

But I'm not here to predict the future. I'm focused on Block.

Fair. You talked a bit about the AI infrastructure build. Maybe you can go in a bit more depth, both in how it's impacting the technology organization. I'm also curious about how you're using AI in other parts of the business. You oversee ops and customer support.

I got asked that at an investor conference last week: How is AI flowing through Block? To me, that's like asking, "How are computers flowing through Block?" It's a fundamental, inbuilt thing that has changed in a binary way over the past 18 months, and then it feels like it changed all over again in the past 4 months.

I'll break it down into internal and external—how we're thinking about our products and what we're putting in customers' hands. Then I can talk a little bit about the future and where we think things are going.

On the internal side, I think the biggest difference is the shape of the organization. We used to have a classic hierarchical structure. It was functional, which was great, but it was fairly standard if you averaged across a bunch of medium-sized tech companies. You would have 8 server engineers, 4 client engineers, a PM, and a designer, and you would work linearly through your roadmap.

Now we have small squads, squads of 1 to 6 people, meaningfully smaller than the other teams would be. We have way more flexibility and fluidity, where a given squad can work a few cycles on one product, get it live, and then work a cycle on another product.

That's different from how things worked a year or 2 ago, where it was, "I'm on the banking team. I'm going to be on the banking team forever." We also have way fewer layers. On the development side, I think we probably cut our layers by 50% or 60%. On the product side, I only have 2 layers, maybe 3 layers in a couple of places. Information is flowing way more freely.

In terms of how we actually build, things have changed on the development side. I think everyone's probably seen every CEO out there going on Twitter and showing their green dot on GitHub, but that's real. All of our designers are shipping PRs. All of our product managers are shipping PRs. That's not that interesting anymore.

I think more interesting is that we have internal tools that are similar to Claude Code, but they're more plugged into our infrastructure. We have a tool called Builderbot. Builderbot is autonomously merging PRs and actually building features to 100%. We've had some fairly complex features that are built to 100%. More often than not, it's building them to 85% or 90%, and then a human who has a lot of context does the final 10%.

The ability to go from an idea to "This is in the hands of 100,000 or 1 million customers" has been compressed massively since December.

Outside of development, I would say most of what we're seeing is that anytime there's a deterministic workflow, we're able to automate it. Generally, at a scaled tech company, you have individuals who are working queues. A lot of that is just being completely automated away.

From a customer support perspective, this is not new, but our chatbots, AI phone support, and whatnot are automating a majority of the inquiries that we get. Then it gets into product operations, risk operations, compliance operations, and any sort of decision-making. Generally, the models and the agents are going to do a better job than humans.

Right now, I think it's critical that we have a human in the loop. That's the key buzzword when you talk to partners, regulators, and what have you. But over time, it's pretty obvious that these systems are just going to be so much better than having 1,000 humans doing that work.

So that's on the internal side.

On the product side, maybe just catch people up on the shape of the business. Obviously, you have Square, you have Cash App, and you made a big acquisition in Afterpay. What do those businesses look like? And how are they changing with AI?

Sure. We used to operate in a business-unit structure. Square used to be its own business unit with its own CEO, and Cash App was its own business unit with its own CEO. That wasn't leading to the right outcome.

About 18 months ago, we functionalized the company, meaning that all of engineering rolls up to our head of engineering, all of design rolls up to our head of design, and all of product rolls up to me. We have a financial platform team that spans the entirety of Block. We have a business platform team that's doing a lot of this automation and spans the entirety of Block.

Increasingly, we're building features and products that actually connect the Square side, the Cash App side, and the Afterpay side. Naturally, you're building technology and infrastructure that is not brand-specific. That's central to our overall strategy and thesis.

Cash App went from, when I joined Cash App in 2016, having just started to figure out how to monetize and having our first dollars of gross profit, to now being, I think, probably 60% or so of overall gross profit at the company. Overall, it's been growing at a healthy clip over the past decade, but Cash App and Afterpay have definitely been growing more quickly.

Increasingly, we're trying to think about things from an ecosystem perspective. That's maybe where Goose as a platform comes in. We built Goose internally. The way to think about Goose is that it's a nod to Top Gun, or whatever—the co-pilot thing.

The way to think about Goose is that it's an agent harness, and it's model-agnostic. I can run Goose on an Anthropic model, an OpenAI model, or an open-source model. There are probably 120 models that we have, and depending on what I'm trying to do, I'll swap out the models.

That was useful for a human to use, but we've built the agentic layer on top. Now a lot of the automations at Block are actually routing through the Goose agent harness.

We've been able to leverage this across the products that we're building. So, Moneybot, which we'd like to think of as a CFO in your pocket, is essentially a proactive chatbot that can take actions on your behalf within Cash App. That is built on top of Goose. Managerbot, which is roughly a similar thing on the Square side, is built on top of Goose.

So, it's a lot of this foundational work on agentic systems, and then the triggers and the underlying data and events that you need to power them. That's working across the entirety of the company. On the product side, I think the biggest shift has really been that we're going from a world where, for the past 10 or 15 years, everyone's used to a static UI, a rigid UI. You tap through the UI. Everyone's Uber or Lyft or Cash App, or whatever, looks the same.

That's going to fundamentally change in the next 6 months. Generative UI is here. We're seeing it with Moneybot. We're seeing it with Managerbot. As the models get better—

What is that going to look like in practice? I'm curious.

In the simplest terms, your Cash App should look really different from mine. The reason why is: I get my paycheck in the Cash App and I'm super into Bitcoin. Let's say you don't, and you use Afterpay all the time. Great. When we open up our apps, they should be totally different. You could probably achieve that just through personalization. That's not that interesting.

What we're actually seeing—and Anthropic had some releases this week that are incredible—is that I can go into Moneybot and say, “How have I been spending my money?” It'll show me a bunch of charts and visualizations, where it's actually generating that visualization on the fly. It's not actually in the code itself.

That's really cool. It's also potentially a nightmare from a QA perspective. We need to figure out how you're going to QA all of these nondeterministic outputs for tens of millions of customers. A great example on the Square side is with Manager Bot. Maybe charts aren't that impressive to you, but let's say you own a multi-location quick-service restaurant. You say, “Hey, can you build me an app where I can manage scheduling for these 2 locations and automatically fire off texts via WhatsApp or Signal or whatever to my employees?”

It's actually going to create that app for you. The way that app looks and feels is not in the source code of the actual application that we push to the App Store. I think it gives folks way more control. It's way more personalized, and ultimately, I think it'll lead to higher engagement. I think it'll lead to better product discovery.

I don't think that if we ask customers to prompt these tools themselves, they're necessarily going to know the right prompts and come up with the right answers. So, we've invested massively on the proactive intelligence side. What we've found, especially as it relates to money, is that we need to be prompting our customers with things that we think make sense for them. That's where we're creating a lot of the value.

I think we're all incredibly bullish on the impact of AI, in the way that all these businesses run and the products you can create. How does that flow back to your stock price? The stock has been roughly flat for, I don't know, 6 or 7 years.

Thanks for reminding me.

But the business has grown a lot, to your point. The gross profit per employee has grown massively. How do you reconcile that dimension?

Yeah, I think markets are cyclical and there's all sorts of things that are happening. I remember in 2021, when our stock price was, I don't know, $260, and I was like, “That was a little bit irrational.”

You can take a longer-term, mature view and say, “Markets are voting machines in the near term, but they're weighing machines in the long term.” Just focus on building.

You and Jonathan earlier talked a bit about defensibility. How do you think about your own moats at Block? You talked a bit about the ecosystem. You guys obviously have regulatory infrastructure. How do you think about the business overall in that context?

Yeah, I think in the near term and the medium term, there's a bunch of moats that exist for Block, and we can talk about the industry more broadly. I think distribution and network effects are one of them. I agree on the Citrine piece and DoorDash. I don't think anyone's vibe-coding DoorDash in the next couple of weeks here.

I like to say that any of us can create a peer-to-peer app in probably a week. No one's going to vibe-code 50 or 60 million monthly actives who are actually using that. So, I think that's true. I think licenses and regulatory posture definitely exist. Hardware right now is harder to imagine how some of the AI tools flow through to the hardware side. You can't vibe-code a piece of Square hardware.

But I think longer term, if we look at the rate of the change and the change in the change, the key thing that's going to make a company defensible is the extent to which the company understands something that's pretty hard for other companies to understand. So, we're increasingly building toward a world and talking about Block as an intelligent system itself.

The way that I see this going, if you extrapolate from the past several months, is that ultimately a company is sitting on top of some sort of signal, some sort of rich data and deep insight. For us, it's how sellers and buyers participate in the economy. Most companies, I think, have this thing that they understand deeply. Then the question is going to be how quickly you can iterate to improve that understanding over time.

So, we're building world models internally and externally of understanding who our customers are, but then also understanding how Block operates. You can imagine, for any company, just a Markdown file of who you are. Then you need the feedback loop with 2 things. You need the feedback loop with the signal, which is: What do you deeply understand that's hard for others to understand? And then you need a tool like Builder Bot or Claude Coder, or what have you.

Then you can just iterate through that loop over and over again. It's like: This is what I'm seeing; this is what's happening. Great, this is our Markdown file for Block. These are our values. These are the metrics we're trying to optimize for. This is what we care about; this is what we don't care about. Then you have agentic systems—you can just build stuff.

Right now, you've basically taken what humans used to do, and it used to take a couple of months to build a feature. Now it takes maybe a week or 2, and there are still humans involved. It's pretty clear that in the future, you'll be able to run that loop, I don't know, hundreds or thousands of times a day. Maybe there are some humans involved, maybe not. Maybe the humans are more like editors.

So, I think the biggest moat is going to be which companies understand something that's super hard for other people to understand. If your answer to that is, “I don't know,” then you maybe could get vibe-coded away.

This has been an amazing conversation. Thank you so much for joining us. Appreciate it.

Thanks so much.

Awesome.