The Path to Lean Runs Through Bloat

Asking Claude to “do some deep research on Subject X. summarize it in a table and make no mistakes” is cool, but have you built a thing that builds a thing recently? Have you built a thing that builds a thing that gets better building that thing the more things it builds?

If you believe you haven’t, I know a guy and a studio that can help you with that!

If you want insight into what’s happening with the job market right now as it relates to AI, look no further than your personal setup for a hint. There’s no leaner startup than you managing your AI spend, tinkering around with these tools today trying to automate to-do lists, buy groceries, track stocks, draft emails, and book vacations to deliver value to the shareholder that is yourself. a16z isn’t likely to write you a check for your hard work, but if it improves your life in some way, ship it.

Unless you’re a seasoned engineer (almost all of whom stopped reading the moment they saw a Product guy lecturing about development), you notice two things fast. This stuff is powerful, and it gets messy without proper controls and guardrails. The phrase “one-shotting” describes using code-assist tools to build an artifact, app, or service in a single breath. It’s the loose technology equivalent of those viral sheet pan meals where you throw every ingredient on one tray and let the oven sort it out. They can be delicious. So can a one-shot app. But you don’t need a Michelin star to know that not everything is sheet-pan-able. Ingredients have different cook times. Depth of flavor comes from layering seasoning across stages. Brines and glazes get prepared separately and applied at their own moments. The good stuff rarely survives being dumped on one pan at one temperature.

You can cook a simple meal with a dull knife and a tiny cutting board, the same way you can one-shot a simple idea with a process that leans entirely on Claude “making no mistakes.” But as the complexity of the dish and the idea climbs, you’re going to get burned.

Most enterprises get this part. What I find fascinating is how much very smart people disagree about how it all gets implemented and what it means for jobs. Call me a cynic, but the never-ending AI-complex CEO press tour reads to me less like honest opinion and more like an effort to prop up stock prices long enough to keep a highly levered, circular financing model from collapsing before mass adoption makes everyone whole. I appreciate the stamina it takes for Jensen Huang to be on TV every other day telling everyone to stay calm. I also find it exhausting. Not exhausting enough to stop me from adding to the pile, apparently!

The layers beneath that press tour are what actually interest me. I care more about the blueprint for execution over the next eighteen months than about whether NVIDIA’s next move is twenty percent up or down. This isn’t really a prediction post. It’s more of a ramble through two things I don’t hear many people connecting. The first is the thing almost everyone is quietly racing toward, even the ones who couldn’t name it if you asked. The second is why the jobs numbers still look fine while something underneath them is clearly shifting.

Why should you care what I think? I spent a chunk of my career building talent acquisition software. I sat inside how hiring actually works, the reqs, the pipelines, the headcount math, the signals everyone reads and nobody questions. Now I spend my days deploying AI for my own businesses and in industry-agnostic corporate settings. I’ve watched both sides of this, the demand for labor and the thing now compressing it.

The unlock everyone targets

As I mentioned, there’s more to AI than asking Claude to edit a spreadsheet, draft an email, or run some research. Those are the first order. They make you faster at a task. The second order is using AI to build the thing that does the task for you, a skill or a small service you can invoke on demand without re-explaining the process every time. Useful, but still not the unlock.

The unlock is what happens on the next turn. Recursion is when the tools you build start improving how you build the next ones. Every skill you stand up teaches you the pattern, leaves behind scaffolding, and shortens the path to the one after it. You stop using AI to do the work and start using AI to get better at using AI, and that capability feeds itself. It’s the thing that builds a thing that gets better at building the more things it builds. If you remember the shape of an exponential graph from calculus, you know it looks almost flat at the start and then like a hockey stick rocketing towards infinity. Companies all want to sit on that curve, where internal AI usage compounds into sharper usage and the product itself throws off flywheels for them and their customers, whether their strategy says so out loud or not.

Most takes on AI come from people describing what they did with it yesterday, and for almost everyone, yesterday was a one-off (the email, a summary, a cleaned-up spreadsheet, etc.). You can’t describe the loop from the outside. You have to have built one, watched it start improving itself, and felt how different that is from a good autocomplete. Most people haven’t gotten there yet, so the conversation stays stuck one level down.

Stay in the kitchen a second. First order is a skill that generates a single recipe tuned to your dietary needs. Second order is a meal-planning app with that skill baked in, planning a week at a time instead of a plate. Recursion is wrapping the whole thing in a service your agent plugs into, one that schedules those meals around your calendar and rewrites the recipe skill itself as your bloodwork changes. Next month’s plans get built by a better planner than this month’s, and you didn’t lift a finger to improve it.

One step back to go two forward

I’ll name the tension rather than dance around it. I don’t want people to lose their jobs to AI, and I spend my days pushing AI acceleration. Call it hypocrisy. My read is that it’s coming either way, so the real question is whether you meet it on the front foot or get caught flat. Everything I argue from here points forward, into the storm rather than away from it. Retreating is the one move I’m fairly confident is wrong. Forward isn’t the same as reckless, though. There’s a way through and it’s governed more through defining outcomes we say we want, not the process we try to control. More on that another time.

Here’s what I don’t hear enough people saying. Companies can get lean while implementing AI. It’s just genuinely hard. And the average company can’t get lean and hit business objectives at the same time, especially now that AI opens up a pile of opportunities they’d never have touched in the manual world.

Picture one person wearing the hats most operators actually wear. Say they’re in product or customer success. They’ve got their customer-facing work. They’ve got their syncs with sales, marketing, and engineering. They’ve got go-to-market they can’t let slip. Now, on top of all of it, ask them to pull their own process apart, understand it from first principles, and model an AI version at exactly the right level of abstraction to replace the parts worth replacing. That doesn’t happen while the other plates are still spinning. Most people can run the business or rebuild it, not both at once. If you can genuinely do both, strangler-figging your own job in real time while hitting your numbers, more power to you.

So to actually implement an AI strategy, you need people. You need them to keep the existing business running and to stand up the recursive structure that eventually makes parts of that business cheaper. The path to a leaner company runs straight through a temporarily bloated one, filled with automation engineers, AI platform and integration folks, QA, and the people who can synthesize everything that comes out of it. One step back to take two forward. That’s my read on why the numbers haven’t fallen off a cliff. It’s not that AI isn’t displacing work. It’s that the displacement needs a build-out first, and the build-out needs hands.

I’d expect that gap to close, not hold. The build-out demand is real, but it’s a bridge, not the destination. When I read the hiring signals from my old vantage point, the cracks are already there if you know where to look. Headcount is getting harder to justify. The new gate in a lot of these conversations is “prove a model can’t do this first,” and that gate didn’t exist eighteen months ago. Gartner is now framing 2026 talent acquisition around the AI revolution and pulling cost out of the business, with high-volume, lower-complexity roles first in line for an AI-first approach. That’s just the analyst translation of what I’m describing. The reqs that survive are the ones a model can’t yet eat.

And the labor numbers themselves are noisier than the headlines suggest. Some of the apparent stability is real build-out demand, the temporary bloat I described. Some of it is just lag. Be careful what you conclude from a print or two while the whole system is mid-reconfiguration.

Where to stand

So what happens when the build-out ends? The comfortable answer is that it nets out. Old jobs go, new jobs arrive, the ladder rebuilds itself the way it always has. I’m passing on that idea for structural reasons, not sentimental ones. The jobs the build-out creates are senior, the people who design the recursive systems. The jobs it removes are the ones that automate cleanly, and those aren’t only at the bottom. A lot of middle-management work is coordination, translation, and status-chasing, and that automates cleanly too. Senior work created, junior and middle work removed. Those don’t trade one-for-one, and they don’t happen on the same timeline.

The in-between is where I’m worried, because it’s already starting. Companies are cutting entry-level roles while chasing AI-certified senior hires, which is a short-term, profit-driven way to build a future workforce given that’s where the future workforce comes from. The same logic is creeping up a level as the coordination layer thins. Do enough of that and you don’t get a shorter ladder. You get one with the middle rungs missing, and no obvious way to climb from the bottom to the top.

I expect the build-out to continue for several more quarters. The hiring propping the number up and the automation quietly eating the work beneath it are the same wave, moving at two speeds, and the build-out is the faster half. Right now you’re mostly seeing the fast half. That’s why the print will keep looking fine, right up until the slower half catches up.

So here’s the one piece of advice I’d actually stand behind. Keep refining your core domain skills while getting as close as you can to where these recursive loops get designed. The roles that hold up aren’t the ones doing the task, or even the ones managing the people doing the task. They’re the ones deciding which loops get built, how they’re wired, and what “good” looks like when they run. Make yourself responsible for that. It’s the same front foot I keep coming back to. You don’t out-run the automation. You go stand where it gets authored.