The first worker AI replaced was never hired

The first worker displaced by AI was never hired. Put yourself in the chair. Ten junior headcount approved, a board asking about the “AI dividend”, and no idea whether you need ten people, five, or two in eighteen months. You wait, because waiting is free. Firing an incumbent costs severance, morale, legal exposure, institutional knowledge, and the specific professional humiliation of rehiring that capacity in nine months when you were wrong. Not opening a requisition costs one conversation with a recruiter. The rational response to uncertainty is not to cut, it is to not add.

The unemployment rate has no column for the job that was never posted. Worse, disruption shows up at the edges of an organization first and works inward, and the edges are new entrants, contractors, junior cohorts, small suppliers. Aggregate employment statistics are structurally a measurement of the core. We built an instrument that reads the center of the organism and we are using it to detect something that starts at the fingertips.

They cannot name the workflow, so they cut the headcount. To answer the board properly you would have to say something like: this workflow, currently forty-one steps and six handoffs, becomes eleven steps with an AI-participatory review gate here, and that changes our staffing need in this function by this much. Almost nobody can give that answer, and not because executives are lazy. Their references are off. Most senior leaders running an AI transformation have never used these tools as a daily working instrument, (and chatting with a model is not working with one).

Then add the part that has nothing to do with AI at all. Companies are extremely good at saving money under uncertainty. Delay the hiring, cut the non-essential, trim the discretionary, consolidate, cancel the experiments, wait for clarity. It is a good playbook, it has saved a lot of companies, and it is available on a Tuesday afternoon with no new capability (or even hard thinking) required. Right now essentially every executive team on earth is uncomfortable, and the ones who sound confident are usually selling something.

Deep discomfort plus a very well-practiced response to discomfort. You do not need a conspiracy or a spreadsheet proving AI can do the work. You need a nervous leadership team with muscle memory. Headcount is legible. Workflow redesign is not. A req freeze fits on a slide by Thursday.

The readiness trap.

Read that twice, because both halves are true at once and that is what makes it a trap rather than a mistake. You are not stupid to wait. You are buying something real, and certainty genuinely does accumulate while you sit still. It just happens to be the one thing you can only purchase with the currency you needed for everything else… time.

Freezing hiring preserves optionality on the staffing decision. That part works. But every quarter spent waiting is a quarter you did not spend mapping the work, cleaning the data, building evaluation discipline, or getting your leaders' hands dirty enough to have working references. Those options have expiry dates and no rollover. You are preserving the cheap options and quietly spending the expensive ones, and the spreadsheet will never show you that trade because there is no line for it.

Tactics is the company buying eight thousand licenses with no operating model behind them. But there is a second failure that gets far less attention… strategy without action is daydreaming. That is the company with the beautiful deck, the steering committee, the maturity assessment, and a hiring freeze standing in for a decision. Right now I see considerably more of the second, and it is more dangerous, because it looks responsible.

When a company announces an "AI-related workforce reduction," at least six different things are hiding inside that phrase.

Amazon ran four of these in a single month, which is why I keep coming back to them. Nova foundation model teams cut on July 22. Three weeks earlier, AWS committed $1B to embedding engineers inside customers. Capex guided toward roughly $200B for the year, up more than 50%, alongside 14,000 corporate cuts, with a hundred-plus AGI roles still open. No model replaced those people. The budget moved. That is a rounding error on the capex line and a career for every one of the humans.

So ask the embarrassingly simple question every single time: what work did the AI actually do that caused this job to disappear? If nobody can answer in a sentence, you are looking at one of the other five.

Ronald Coase answered an AI relevant question in 1937 and nobody invited him to the AI panel. He asked why anything gets done inside a company at all, if markets are so efficient. His answer won a Nobel Prize: firms exist because using the market has costs. Finding the right person, negotiating, contracting, verifying, coordinating. The boundary of the firm sits exactly where internal coordination becomes cheaper than external transaction.

With apologies to Mr. Coase, here is what I want to do to his theorem. If the firm's shape is set by the cost of coordination, what happens when coordination costs go dramatically down?

Because AI is not primarily reducing production cost right now. It is demolishing coordination cost. Summarizing the thread. Reconciling the two versions. Writing the status update. Translating between the engineering ask and the finance answer. Chasing the eleven people who owe you something before Thursday.

In my own work I have noticed something I find slightly alarming: building the thing is now often cheaper than the meeting to specify the thing. That inverts a century of organizational logic. We built layers of management precisely because coordinating humans was expensive and specification was cheap relative to construction. Flip the ratio and the layer that exists to coordinate has the weakest defense. (Yes, I am also quietly asserting that construction costs for software are dramatically dropping here.)

So the non-consensus call: the first structurally exposed group is not the people doing the work. It is the people whose day is mostly about the work. Coordination overhead, status reporting upward, capacity arbitration, handoff management.

Now because I do not want to make a call like that and then hide behind vagueness.

One important caveat, April 2022 was the peak of the cheapest money in modern history, and a large chunk of that 42% is the interest-rate hangover and the tech-hiring correction, not AI. Anyone quoting that number at you without saying so is (again,) selling something.

Which is why span of control is the metric I would actually bet on. It is a ratio, so it does not care whether hiring is up or down. It only moves if the cost of coordinating people changes.

So here is the falsifiable version. If I am right, span of control keeps climbing through 2027, past roughly thirteen direct reports per manager on the Gallup and BLS series, and management's share of reductions stays above 30% even after overall hiring recovers. If span of control plateaus while hiring comes back, the mechanism was the rate cycle and not Coase, and I will say so on the channel.

One more thing, and it is the reason this prediction needs its own instrument. Every standard AI-exposure measure, including Anthropic's, is built from ONET task lists plus observed usage. Ask yourself where "chased eleven people for status before Thursday" appears in an ONET task description. It does not. Coordination work is diffuse, invisible, unglamorous, and almost nobody logs it as a task they used AI for. The instruments we are all watching are structurally blind to the exact thing I think is happening. That does not make me right. It does mean the absence of evidence here is unusually weak evidence of absence.

Here is the part I would tattoo on a board, or at least write on a boardroom wall if they let me.

Look at what entry-level work consists of. Research. Drafting. Documentation. Testing. Reconciliation. Basic analysis. Chasing down the answer nobody else had time to chase. Every item on that list looks automatable, and most of it genuinely is, which is why it is going first.

Now look at the same list and ask a different question. How did you learn your judgment?

You learned it doing exactly those things. You learned how the business actually works by reconciling numbers that would not reconcile. You learned to smell an exception because you handled four hundred non-exceptions first. You learned what quality feels like by producing garbage and having someone senior explain why it was garbage.

That grinding, tedious, inefficient work was not just output. It was the friction that manufactured your expertise.

That is the paradox, and I do not think there is a clever way out of it. Remove all the friction from work and you accelerate output while starving development. You get people who can produce senior-looking artifacts without the judgment to know when the artifact is wrong.

The data is almost uncomfortably neat here. Stanford's Digital Economy Lab, working from ADP payroll records across millions of workers, found a 16% relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations, controlling for firm-level shocks. Currently running about 3.8% a year and steepening. Older workers in the same occupations are stable or growing. And the detail that should stop you cold: the damage concentrates in occupations where AI automates rather than augments. Where AI augmented the human, young workers were fine. Where AI replaced the friction, the ladder broke.

My favorite piece of evidence is that somebody is doing something about it. The University of Chicago Law School banned phones, laptops, and tablets from first-year classrooms. The lazy read is "old institution afraid of new technology," and that is not what this is. Their actual strategy is three parts: build AI-resilient teaching and assessment, deliberately elevate the human skills that separate excellent lawyers from adequate ones, and teach responsible AI use after students have the fundamentals. That is not a ban. It is a constrained-AI rotation, run at institutional scale by people who understand precisely what would be lost. The genuinely funny and slightly bleak detail: roughly 85% of their students were already using AI. They are not preventing a problem. They are performing a rescue.

A law school figured out the Friction Paradox before most Fortune 500 boards did.

I owe you this, and I want to give it properly rather than as a token gesture. If I am wrong, here is why.

The New York Fed published an analysis in May using Lightcast posting data and Anthropic's own exposure measure, running an event study around the ChatGPT release.

Two findings, and both hurt.

First, yes, postings for AI-exposed occupations declined relative to less-exposed ones. But that divergence started before ChatGPT existed, and it stabilized after 2023. If AI were the cause you would expect the gap to open after late 2022 and then widen. It does not. It was already opening, and then it stopped. That is far more consistent with the rate cycle and a tech-sector correction than with AI displacement.

Second, and this goes straight at my Friction Paradox: they compared junior versus senior postings within highly exposed occupations. If AI were eating entry-level work specifically, junior postings should fall relative to senior. They move in parallel. No divergence.

So Stanford, using payroll records, finds a 16% relative employment decline for young workers in exposed occupations. The NY Fed, using postings, finds no junior-versus-senior gap. Different instruments measuring adjacent things, and they disagree. Anyone showing you only the one that fits their narrative is doing you a disservice.

It gets worse for me. NY Fed business surveys say firms intend to respond to AI mainly through retraining, with limited hiring effects. Under Anthropic's own measure, fewer than 10% of workers are in occupations with exposure above 0.4, while 40% are in jobs with zero measured exposure. And the newest one, from Indeed's Hiring Lab in July: over the past year, the most AI-exposed occupations have seen the largest rebound in postings. Not the largest decline. The largest rebound.

What do I do with all of that? Three things.

I lower my confidence, and you will see it in the numbers below. I notice that the disagreement is mostly payroll data versus postings data, which measure different moments: a posting is an intention (what you say you are going to do), payroll is what happened (what you actually did), so I lean on payroll for what has occurred and postings for what is about to. And I hold my ground on the coordination call, because every one of these studies is organized around occupational task exposure and none of them measures coordination intensity. My prediction has its own metric and it is not in these results in either direction.

I would rather be the guy who showed you the disconfirming evidence and turned out half right than the guy who curated the charts and got lucky.

These are priors, not forecasts, and they are conditional on general capability not arriving inside the window. They moved while I was building this piece, which is what is supposed to happen.

Augmentation economy, ~45%. Most jobs survive, output and expectations both rise, AI fluency becomes table stakes, experienced people gain more than novices. I moved this up, because the NY Fed retraining finding and the Indeed rebound both point here, and because "firms retrain rather than cut" is the most boring possible outcome and boring outcomes are historically underrated. Tripwire: measurable productivity growth alongside continued hiring into redesigned roles, plus a visible wage premium for AI-capable experts.

Hollowed ladder, ~35%. Seniors do great, junior intake keeps thinning, fewer people ever acquire the experience to become senior, and in a decade you have a succession problem no amount of capital fixes. I moved this down, because the junior-versus-senior posting evidence genuinely does not support it and I am not keeping a number where my instincts want it. Stanford's payroll data still does support it, which is why it stays this high. Tripwire: early-career declines persisting past twenty-four months, experience bars on entry-level postings continuing to climb, graduate cohorts shrinking again next cycle.

Workflow substitution, ~20% by 2029. Agents get reliable across end-to-end processes, organizations restructure around smaller human teams, the coordination layer thins fast, displacement accelerates after this long apparent delay. Unchanged, because nothing in the counter-evidence touches it. Tripwire, and this is the one I actually watch: a credible company reporting a staffing reduction tied to a measured end-to-end workflow gain, with before-and-after numbers. Not a press release. Numbers. The day I see three of those, this number moves up hard. (I root for and help with this one so I want to see this go up)

And separately, on the record and unhedged: span of control past thirteen by end of 2027, management holding above 30% of reductions.

Notice what all of the tripwires have in common. Not one of them is the unemployment rate.

The obvious/easy response is "hire more juniors and give them the old work back." That is wrong, and it is wrong in a way that feels virtuous, which makes it dangerous. Protecting obsolete work to preserve a training mechanism is nostalgia with a headcount budget. The tasks are genuinely going away.

What is worth preserving is not the task list. It is the development of judgment. So design that deliberately.

Constrained-AI rotations, where a new person works with the assistance dialed down on purpose. Not hazing, curriculum. If that sounds unworkable in a real business, a top-five law school just implemented it for an entire class year while 85% of the students were already using the tools. Productive-failure projects, with real stretch work, explicit failure tolerance, and a structured debrief, because the debrief is the product. Cross-domain stretch assignments that deliberately disrupt their references, and recurring reference recalibration, which applies to the executives too. The dirty hands problem is not a junior problem.

Then redesign the entry-level role around what actually pays now: supervising AI, verifying evidence, owning exceptions, understanding a system end to end rather than one function deeply. Those are real jobs and they are more interesting than reconciling spreadsheets ever was. (Though I will cop to a personal bias against spreadsheets)

And stop calling it a ladder. The ladder assumed you got deeper in one thing and climbed. The shape that wins now is a lattice: multiple areas of working-depth competency connected by a core that can learn a new domain fast enough to stay relevant, with lateral movement treated as progress instead of a detour.

If not corrected, there is a second bill coming and it is not sentimental. Lock a generation out for five years and you do not get a patient cohort waiting politely for the ladder to be repaired. You get people who watched the arrangement decide they were surplus, holding the cheapest company-building toolkit in human history, with no legacy ERP to work around, no pre-AI governance committee, and no coordination layer to defend. They get to start AI-native, which is the single largest structural advantage available right now, and incumbents are handing it exclusively to the people with the strongest reason to use it against them. You would rather have that generation inside your building, slightly annoyed, than outside it and extremely motivated.

The job apocalypse has not arrived, and the reason is not that the models underdelivered. Capability does not enter the economy directly. It passes through products, workflows, data, budgets, lawyers, managers, and human trust, and every one of those is slower than a benchmark.

But the absence of mass unemployment is not the absence of damage. The layoffs may be arriving before the automation. The coordination layer is quietly dissolving. And the first rung, the tedious friction-filled work that used to manufacture experts, is being optimized away by people measured on this quarter who will not be in the chair when the bill arrives.

The bill arrives twice. Once in about a decade, when you go looking for senior people and discover nobody made any. And once sooner than that, when the generation you kept outside the building starts shipping.

Do not watch the unemployment rate. Watch the first rung. Watch the requisition that never gets approved. Watch whether your company can name one workflow it has actually redesigned rather than one tool it has actually purchased. Watch whether anybody under thirty is being handed work hard enough to learn from. And notice that waiting for more information is not a neutral act. It is a trade, and you are giving up options to make it.

One more thing, because I think this is where the whole conversation is heading whether we are ready or not. AI is going to force us to define humanity. We cannot keep defining ourselves by our ability to out-produce a machine. That was always a losing bet... we just used to be winning it.

So tell me, and hit reply on this one. In your organization, is anyone actually measured on whether the next generation is learning anything? (If you do, send them this and get them to reply, I want to talk!) Not headcount. Not utilization. Learning. If nobody owns that number, we already know how this ends.

The episode version has the live take, the visual walkthrough, and a few things I only say out loud: https://youtu.be/MU_6Ig3LjoY

Chuck
Chief Technology Troublemaker chuckdevries.ai