They Automated Everything. Then They Hired More People.
Part Two of a series. Part One: “It Came Back in Ninety Seconds” (June 2)
This week I got a recommendation from an AI-assisted workflow I’ve built to track content trends and my own editorial calendar. It was a good recommendation, well-reasoned, backed by data I’d gathered myself. The essay it flagged had the highest brief score of the week. The research was already done.
I almost ran with it.
Something felt off. Not wrong. Off. A different texture from wrong. I sat with it for a minute and realized: I knew something the workflow didn’t. I knew what readers had just been through in my content. I knew what they’d felt reading the piece I published a week earlier. I knew the emotional thread that needed to come next, not the highest-scoring thread. The right one.
The workflow didn’t fail. It just couldn’t hold the story.
That same morning, I’d been sitting with an essay about a company that had automated everything it possibly could, and then doubled its team. Those two stories clicked together in a way I haven’t stopped thinking about.
The Company That Automated Everything
Dan Shipper runs Every, a media and software company that has automated nearly every workflow it possibly could with AI agents. Last month he published an 8,000-word essay called “After Automation,” the most-read piece in his publication’s history. The thesis sounds counterintuitive until you sit with the mechanism: his company didn’t shrink as he automated more of its work. It grew. From four people to thirty since the GPT era began. Over the past year alone, headcount doubled from fifteen to thirty.
He wasn’t expecting that when they started. But once he saw the mechanism, it made complete sense.
Here’s what’s actually happening. When AI makes a type of expertise cheap and abundant (writing, research, analysis, design), it doesn’t eliminate demand for that expertise. It collapses the value of average output in that domain and creates fierce demand for what’s different from average. Judgment. Framing. Taste. The ability to look at something close-but-not-quite-right and know exactly what’s missing and why it matters.
Models improve on fixed frames. Give a model a benchmark, a defined problem, a structure it can optimize within, and it will get measurably better over time. Humans operate outside the frame. We zoom out and redefine the problem. We make choices that don’t exist anywhere in the training data. We decide which story to tell, and when, and to whom, and why this particular moment is the right one for it.
Shipper calls this the allocation economy: human value shifts from producing the work to directing and stewarding machine output. I’d name it differently for the people I write for. Call it the judgment economy. That’s not a coincidence. It’s the mechanism.
The Steward Premium
When competence becomes cheap and abundant, the scarce appreciating asset is the judgment that decides what matters, what’s wrong, and what’s merely average.
Automation doesn’t remove the human. It promotes the human from maker to steward. And the steward premium, the value of the person who can tell good from almost-good, rises the more you automate.
This is the engine behind Every’s growth. More automation didn’t mean fewer humans. It meant the humans who stayed were doing higher-leverage work. Every agent they built needed a human to direct it, question it, and catch it before it drifted into producing something competent but slightly off. Without a human steward actively in the work, the agent’s output collapsed toward the average.
That’s the same engine that ran through my workflow that morning.
Every agent needs a human who knows enough to feel the drift. Who knows when something is close-but-not-quite-right, not because they ran it through a checklist, but because they’ve been in their work long enough to recognize the difference between something technically correct and something actually right.
That kind of knowing takes a long time to build. It can’t be prompt-engineered into existence. It can’t be replicated by someone who started two years ago. It is, almost by definition, a function of experience.
What I Want to Be Honest About
I’m going to stop here and do something a lot of essays in this space don’t bother with.
Every is a thirty-person AI media company that sells AI tools and subscriptions to other AI enthusiasts. Its growth is partly because it productizes AI, not just because automation creates human work. A mid-size accounting firm, a regional hospital system, a construction operation. They may not see the same dynamic. The sample size is one company, and it’s a self-interested one.
Every’s own head of technology consulting publicly disagreed with parts of Shipper’s conclusions. Shipper shared his colleague’s counterpoint himself, which I respect. Smart people who are close to this work are not all reading the same signals the same way.
And the easy dismissal of AI layoffs as cover stories for bad hiring practices is too convenient. Entry-level erosion is real. AI is genuinely reshaping which rungs of certain career ladders still hold weight. The defensible position isn’t “AI isn’t taking jobs.” It’s something harder: AI is reshaping which human work matters, and the reshaping is not painless or evenly distributed.
So let me be precise about what I’m claiming and what I’m not. The steward premium is real. The structural argument is sound. But it is not automatic. And it is not equally available to everyone.
Most importantly: it does not come from simply having experience. It comes from bringing experience to bear. There is a difference, and the difference is everything.
The Part Most Essays Don’t Say Out Loud
Crystallized intelligence, the pattern recognition, judgment, and contextual wisdom that accumulates over a long career, is structurally the scarce asset in the judgment economy. This isn’t flattery. It’s the mechanism.
Models are trained on what’s been made explicit, on what’s been written down, published, codified, turned into data. But decades of tacit knowledge are different. The accumulated understanding of what actually works in your specific domain, the ability to read a room or a client or a market or a moment, the hard-won instinct that something is off before you can articulate why — none of that was ever made explicit. That is exactly what models cannot replicate.
Your judgment is not in the training data. Which is precisely why it’s valuable.
But here’s where I want to push back on a comfortable story.
Crystallized intelligence is not a passive inheritance. The steward premium isn’t waiting for you because you’ve put in the years. It has to be earned again, in real time, by actually being in the work.
I see a version of The Drift that has appeared specifically in the AI age. It’s subtle enough that you can miss it entirely if you’re not watching for it. It looks like sophisticated engagement. Reading the newsletters. Attending the webinars. Staying informed about what the tools can do. And then letting the agent run while you remain loosely upstream, reviewing outputs, approving summaries, staying out of the details.
That is not stewardship. That is observation.
I’ve been guilty of this. I’ve called it efficiency. I’m not sure that’s what it was.
Observation doesn’t generate the steward premium.
The difference is friction. Real stewardship means being in the work closely enough to feel when something is close but not quite right. It means having enough skin in the process to catch what the agent cannot hold. The story. The thread. The thing you know from being in the room for the last thirty years that doesn’t exist anywhere in the training data.
This morning I caught something my workflow couldn’t see. Not because the workflow was poorly built. Because I was paying close enough attention to notice the gap between what it could know and what I knew from living the work. That gap is not incidental. That gap is the job.
One Thing Before You Close This Tab
Pick one workflow you’ve already handed to AI: a research scan, a first draft, a summary, an analysis, a plan. This week, work through the output closely enough to notice where it delivers something competent but slightly wrong. Not broken. Not bad. Just off.
Write what’s missing in one sentence.
That sentence is your crystallized intelligence inventory. The thing you know that the model doesn’t. The gap between close-enough and right. Do this once and you’ll start to see it everywhere, in every output that almost nails it, every recommendation that scores well but misses the thread.
That’s the difference between supervising an agent and stewarding one. The steward premium lives entirely on the stewarding side of that line.
The judgment economy isn’t coming. It’s already here. The question isn’t whether your experience has value. It does, structurally, in ways that compound as automation spreads. The question is whether you’re bringing it to the keyboard with enough friction to matter.
Think of one person whose name just came to mind while you were reading. Send it to them.