“Most coaches aren’t very good.”

I hear this constantly. Founders. Investors. Executives who’ve had a coach. It’s not like they don’t know what I do for work. They volunteer it anyway.

I guess they’re telling me because they think I might be an exception. Maybe they trust me enough to say what they really think. Honestly, I don’t know. What I do know is that I hear it a lot.

I haven’t written about coaching here in a while, and it seems like it’s time, because the demand and the need for good help have never been higher. But before you go “oh, another piece about coaching”, give me a minute. I’m not going to sugarcoat the current state of the field or the people it tries to serve. If anything, it might look more like a takedown. But I promise I’ll guide you somewhere more constructive.

Here’s the reality I see: clients don’t know what they’re buying, and coaches don’t know what they’re selling.

Harsh? Sure. But I believe—deeply—that if this weren’t the case, more coaches would be busy delivering results for their clients, and I believe a lot of those clients would be creating better outcomes in the world with fewer unforced errors.

Seems like everything’s unprecedented these days, so I’ll spare you the “we live in unprecedented times” talk. This particular moment in history, in society, in business, and in our concepts of self is unusual. We’re in the transition between two very different ways of living, and there’s plenty of VUCA to go around.

I also won’t claim here that coaching is the cure for all of society’s ills. That’s nonsense. The practice itself is just one way to reach an outcome. There are others. I am going to make a strong case that good coaching can make a lot of this current moment much more comprehensible and the people do their work a lot more successful.

And I’ll start by saying no one really needs a coach.

A coach by any other name…

Coach.

I’ve never loved the word. It doesn’t describe the work. It doesn’t tell you anything about what it’s for, who needs it, what’s different afterward. It’s the professional equivalent of “other duties as assigned”, containing pretty much anything while describing nothing.

I might have a distaste for the word because of a stunningly foul thing my high school basketball coach yelled at me during a game long ago. It’s too coarse to share here, but send me a note and I’ll tell you the story. It didn’t set me up to love coaches.

During my corporate years, I worked with two coaches and didn’t get a damn thing out of it. It was only years later that I recognized that they were both very good, and I was very unready to look at my patterns. That one’s on me.

So I wasn’t wildly enthusiastic when I accidentally became a coach. My path looks very different than a lot of people’s in the field. My first client pressured me into being their coach, then forced me to take money for it, then insisted on making referrals. Six months in, I was working with more people than I would have if I’d been trying.

I resisted at every turn. But a few years in, I realized that I was shifting something for people, something that made them more effective, more successful, more powerful. My presence more than my knowledge was helping in a way I hadn’t been able to appreciate. And that changed the course of my life.

I still don’t love the word. People recognize it, though, so I use it.

What I’d prefer is for us to give the word meaning, to make clearer what it is and what it does. I’d like for people to understand what it’s for and to think of a coach that truly helped them as closer to the norm rather than the exception.

That starts with us coaches and an essential concept from a wise man.

Why no one needs a coach

Clayton Christensen knows why people seek a coach than most coaches do.

His “Jobs to Be Done” theory does a solid job of explaining why anyone makes a choice about anything. The theory posits that people don’t just buy products. They “hire” them to make progress in a specific situation.

“Somewhere between 75 and 85 percent of all new products launched into the market don’t succeed financially,” Christensen says. “The reason is they don’t target a job that people are trying to get done.”

If you’re a coach, the job to be done isn’t often going to be what you’re selling. You can offer business coaching or somatic awareness or transformation or whatever, but when that connects with someone, it’s not because of your offer. It’s because people want something they don’t yet have, and they can imagine you helping them get it.

If you’re not addressing a job to be done when you talk to them, there’s no movement.

Nobody has that first conversation to examine their inner life. They’re taking the time because the co-founder fight is spilling over into the team. Or the team isn’t leaning forward. Or the round isn’t closing. Or cash keeps tightening and invoices are going unpaid. The problem they walk in with is intense, external, and already known. They probably already got some feedback about it.

And what do they want that they don’t yet have? A more constructive relationship with their cofounder. Or they want to avoid having the same argument again. Or they want investors to lean in. Or they want customers to value them enough to pay their bills. You know that’s an inside job, but they might not. And if they knew it, they might be doing what you do.

More than that, they want to feel. They want to feel ease, or confidence, or relief. If there’s a path to feeling those things, the rest of the conversation is logistics. They want to know that you can help them get the thing they want but don’t yet have.

When you answer with a process, a framework, a methodology, or a brand of transformational work, the founder hears words that don’t address the thing keeping them up at night. The discussion stays pinned to the symptom they already understand, which means it stays optional.

Put differently, they hear, “yeah, that’s a problem.”

It’s that simple. We talk about what we do, but they’re living inside a situation that’s getting worse. Those two things have yet to meet.

No one really needs a coach. They need the situation to improve (the job to be done). And sometimes the most direct way to get that is with a skilled guide. A coach.

You must first become useful

Useful work requires two kinds of expertise at the same time.

The first is the outer terrain. Fundraising. OKRs. Taking a product to market. The ordinary friction of growing something from nothing into a company that can make payroll on Thursday. Without that, I don’t get a serious look, and I shouldn’t. Someone who’s never been in those rooms can ask good questions and still miss the situation.

The second expertise only shows up in what the person does under pressure.

I worked with a founder whose receivables weren’t being received. Work got delivered and invoices went out. But money didn’t come in.

When he got on the phone he felt smaller than the invoice. He was careful. He left room for the other person to set the terms. If I owed him, I would have paid someone else first. I would have assumed he’d wait. But he couldn’t see that.

(By the way, notice what he wanted. It wasn’t money. It was what the money would make him feel. Respected. Valued. Appreciated. Confident. The job he was trying to get done was feeling those things. Because I didn’t have a rigid process, I could listen for those things and find a lot of ways to get to them.)

We didn’t start with stronger contracts. We started with the way he started that call. He took a slightly different posture. He used a phrase he’d tried before and decided didn’t work. Practice, then another call, then another. A couple months later the money started moving. Clients who’d been drifting began to pay. He sounded like someone who expected the relationship to endure.

I don’t need a special vocabulary for what shifted. Money in the door did the talking.

I’ve seen the same thing with a founder who couldn’t stop triangulating the team. Every decision came back through him, and he called it quality. The team called it waiting.

We didn’t change their meetings. We practiced him staying out of one decision all the way through. Then another. The team started moving, then moving faster. He had more cycles than he’d had in a year.

That combination of depth in the inner terrain and outer context is rare. Familiarity with the business, plus the ability to see the pattern the founder keeps repeating that recreates the outer problem, plus enough patience to stay with the practice until the new way sticks. A lot of people have pieces of this. Former operators often know one or two versions of the movie in real detail and get real quiet when the person in front of them isn’t playing the part. People trained only on the inner work can hear the pattern and still not know what a bridge round does to a cap table or how hiring an assistant creates more work. Coaching starts to matter when both are in play at once.

Non-optional

Most coaches get the reputation for not being very good because that combination is uncommon. Real comprehension of the client’s situation plus the ability to introduce someone to a version of themselves they haven’t met. Maybe enough stubbornness to stay with it until the new capability becomes more than a performance.

The language founders (the people I help) already understand is blunter than the language the coaching industry prefers. To get the chance to help, I need to be fluent and fluid in that language.

Are we going to keep getting the same results with the same operating system?

What becomes possible once an imagined boundary isn’t there?

The highest-value problems feel non-optional. Essential. Anything that gets deprioritized the moment something else catches fire was never real work.

A lot of people think most coaches aren’t very good. But I know a lot of good coaches. And I know a lot of people (like that previous version of me) that don’t see the value yet really need the help. They don’t take action because they don’t trust that a coach can even understand what they do, let alone who they are. And in that mindset, you’re a nuisance and possibly a threat.

You get to lead any way you choose to. If you’re curious about what’s driving the words you’re hearing and hold your frameworks lightly enough, you might get the chance to change someone’s life and work. But to get that job, you have to build trust, and that comes from the person who needs help feeling understood.

That’s also your job.

A year ago we published a projection most people still found hard to believe. The consensus put meaningful AI-related job displacement three to five years out. The data, read through the causal chain, said the window would open in Q1 2026 and that the first visible effects would arrive faster and more quietly than the classic layoff story suggested. Our…

Read more

That quote captures how I feel about this moment. W…

Read more

In 1898, the smartest urban planners in the world got together to solve the biggest problem facing every major city, and they gave up.

The problem was horse manure. New York had something like 170,000 horses moving people and goods, and each one produced more than twenty pounds of the stuff a day. The streets were buried. Dead horses, about 15,000 a year…

Read more

If productivity increases in the economy and no one measures it, did it happen?

I began this week with three somewhat counterintuitive propositions:

1. I do not believe there has been any meaningful job displacement by AI.

2. I do not believe there has been any meaningful job creation by AI.

3. I do not believe there have been any meaningful productivity gains from …

Read more

There’s a false debate running through the center of our economic conversation.

On one side, you have the advocates of technological displacement who warn of a jobless future, pointing to the surgical cuts in tech, finance, and professional services as the first signs of a structural collapse. On the other side, you have the cheerleaders of augmentation …

Read more

I’ve been tracking the evolution of the workforce for the past couple of years, and I’ve been getting the feeling that we’re all describing the parts of it we see and understand—not the full picture. The last few weeks have brought some of those parts into sharp relief. “AI is going to take all the jobs” has had a longer run, but “AI will create more jo…

Read more

Why do I write about the macro jobs picture?

A big part of my work is reading signals for what’s coming. The shape of jobs is a critical element in a bigger picture, and I want to tell the story of that bigger picture. That story runs right through the monthly jobs numbers, and if you can read what they say (and don’t say), you can begin to see this nex…

Read more

Read more

A friend works in marketing at a mid-sized software company. Senior role, good pay, been there six years. She told me last month, half-laughing, that she’s not sure what she actually does anymore. The work has been moving under her for about eighteen months. Tools she used to use by hand are now mostly automated. Reports she used to write are mostly drafted by something else. Meetings she used to lead, she’s still in, but the deliverables coming out of them are produced differently than they were two years ago.

She still has the job. The W-2 is intact. The benefits are intact. The title is the same. The salary went up at her last review.

But she said, “Something’s gone. I can’t quite name it. The job feels less like a job.”

I think a lot of people are having that conversation right now, mostly with themselves. The headline numbers say everything’s fine. Unemployment around four percent. Markets up. The S&P printing buybacks. Most people who have a job, still have it. And yet something is gone, and we don’t quite have a name for it yet.

This piece is a try at the name.

What a job actually is

When we say “a job,” we usually mean the work. The thing you do. The deliverables. The expertise. But that’s not actually what a job is. A job is a container.

It’s a container that bundles five things together. Income. Identity. A time-structure that organizes your week, your month, your year. Belonging, in the form of colleagues, a team, a place where people know your name and what you do. And benefits, which in the U.S. context means health insurance, retirement, paid time off, the legal infrastructure that keeps you upright when you’re sick or old or laid off.

Five things, bundled into one form. We hand it to you on day one, you hand it back when you retire or leave, and in between, the container holds your life together in ways that are mostly invisible until they stop working.

This bundle isn’t natural. It isn’t ancient. It isn’t how humans have always organized work. The bundle is about 150 years old in the form we recognize. It got assembled out of factory-era logistics, postwar labor compromise, and a very specific accident: in 1942, the U.S. government froze wages during World War II, and to compete for workers, employers started offering health insurance as a benefit. That’s why your health insurance comes from your job. Not because it makes sense. Because of a wage freeze in 1942.

(My friend Josh Allan Dykstra has written about this history in detail, and it’s worth the read if you want the whole arc.)

The retirement piece got bolted on later. The 401(k) wasn’t a thing until 1978. Defined-benefit pensions, the kind your grandfather had, peaked in the 1980s and have been declining for forty years. The career-arc piece, the idea that you’d work for one company for thirty years and then get a watch, was specific to a postwar window from roughly 1945 to 1980 and has been quietly dissolving ever since.

So the bundle isn’t a law of physics. It’s an arrangement. A specific historical arrangement that got assembled in pieces, that’s been quietly disassembling in pieces, and that we keep calling “a job” as if it were the only possible shape.

This matters because the question we’re trying to answer over these four weeks is what happens to work in the next three years. And the answer depends on whether you think “the job” is the work or the container the work has been kept in. If they’re the same, then any change to one is a change to the other. If they’re different, then the work can change shape inside a container that looks intact, or the container can hollow out around work that’s still happening.

I think we’re in the second case. We have been for a while.

We’ve seen this before

The container has hollowed before. Twice that I want to walk through, because the pattern is more useful than any single example.

The first time was the household. For most of human history and almost all of pre-industrial Europe, the household was the container that bundled work, family, identity, and place. A weaver in 1750 did the work in his home, with his wife and children participating, on raw materials supplied by a merchant who picked up the finished cloth and paid by the piece. This was called the putting-out system, and it organized the bulk of what we’d now call manufacturing.

Then the spinning jenny arrived in 1764, the water frame in 1769, the power loom in 1784. These tools didn’t replace the weaver. At first, they just got added to the household. Some weavers did better with the new tools. Some did worse. The container, the household-as-workplace, looked intact.

The factory was the change in container. By the 1820s, Manchester had textile mills with hundreds of workers under one roof. By 1840, the household as a place where you did paid productive work was effectively gone in industrializing England. The work hadn’t disappeared. It had migrated. The household survived as a container for family life, but the bundling of work, family, identity, and economic activity it used to do was over.

The transition took roughly sixty years. From the spinning jenny to the death of the putting-out system. Two generations. The people who lived through it mostly didn’t know it was happening at the time. They knew their work was harder to find, their bargaining power was worse, their kids were going to the mills instead of staying home. They didn’t know the container was hollowing. They knew something was gone.

The second time was the clerk. In 1880, the clerk was a craftsman of paper. He kept ledgers by hand. He drafted correspondence in longhand. He maintained files. Becoming a clerk was a respectable apprenticeship that took years to master, and a senior clerk in a big firm was a person of standing.

Then came the typewriter, in commercial use by 1880. Then carbon paper, then the filing cabinet, then the adding machine, then the calculator, then the photocopier, then, eventually, the spreadsheet in 1979. Each of these tools mechanized a piece of what the clerk used to do.

The clerk-as-container didn’t collapse. It bifurcated. The judgment piece, what to write, how to organize a deal, how to manage a senior person’s calendar and politics, became the executive assistant. The codifiable piece, what to type, what to file, what to add up, became the data-entry temp. The wage gap between those two halves grew steadily for a hundred years and is now one of the larger wage gaps in the U.S. economy. The judgment half went up. The codifiable half went down or disappeared into software.

The transition took roughly a hundred years. Slower than the factory because the tools were less concentrated and the work was harder to mechanize. But the pattern was the same: the container persisted in name; the contents migrated; the survivors did different work than the people the container was originally built for.

Hold these two examples next to each other. The household and the clerk. Different industries, different centuries, different technologies. Same pattern. The container is what we keep calling the thing. The work is what changes shape inside it. When we talk about “the death of the household economy” or “the decline of the clerk,” what we usually mean is that the bundle came apart, not that the work stopped.

It’s happening right now

So here we are. Spring 2026. And the same pattern is showing up in three numbers that don’t usually get put next to each other.

The first number is the rate at which the U.S. economy is hiring people. Not unemployment. Hiring. The Bureau of Labor Statistics tracks this in a monthly release called JOLTS, and the most recent print, for February, was 3.1 percent. To give that number a feel: 3.1 percent is the floor of the Great Recession in 2010 and the floor of the COVID lockdowns in April 2020. Two of the worst hiring environments in the last twenty years. We are matching both of them right now, with no recession on the official books and an unemployment rate of 4.3 percent that looks fine on the front page.

The hiring engine has stopped. People aren’t being hired, and people aren’t quitting, because they don’t believe a better job is on the other side. The pipe is closed at both ends. The number of Americans unemployed for six months or more crossed 1.9 million in March, one in four of all unemployed people, and that share has been climbing for three straight years.

The second number is what’s getting built while the hiring stops. Hyperscaler AI capital expenditure, the money the largest tech companies are spending on AI infrastructure, was about $256 billion in 2024. In 2025, it was $443 billion. For 2026, the projections range from $602 billion to $750 billion depending on which analyst you trust. That’s not a feature investment. That’s the kind of money you spend when you’ve decided that the work running on top of the new infrastructure is about to look very different from the work running on the old. Microsoft alone spent $72.4 billion in the first months of fiscal 2026. Meta and Alphabet are running capex-to-revenue ratios above 45 percent. These are existential commitments. Boards approved them. The bet is on the page.

The third number is what’s happening to people while the first two are happening. Mercer publishes a global talent report every two years. The 2026 edition, released in February, has employee thriving collapsed from 66 percent to 44 percent over two years. Twenty-two points. AI job-loss anxiety up from 28 percent to 40 percent. Twelve points. These aren’t small movements. These are signals that something has changed in how working life feels, even for people whose jobs are still nominally intact.

Three numbers: the hiring freeze, the substrate buildout, the mood of the people sitting inside both. Put them together and the pattern is the same one we saw with the household and the clerk. The container is still there, in form. The W-2 is still there. The benefits are still there. The title and the salary and the company picnic, mostly still there. But the work is changing shape inside the container, the engine that used to renew the container by hiring new people into it has frozen, and the people inside know something is gone even when the headline numbers say it isn’t.

The job, as a 150-year-old bundle of income, identity, time, belonging, and benefits, is hollowing. The container is doing less of the work it used to do, the contents are migrating to other forms faster than the container is being remade, and the gap between what a job is supposed to bundle and what it actually bundles is the part most people are feeling but haven’t had words for.

The faster part, compared to the household and the clerk, is what makes this disorienting. The factory took sixty years to hollow the household. The spreadsheet took a hundred years to bifurcate the clerk. The hollowing we’re inside of has been measurably underway for about three years and is accelerating. By the time the historical analogues finished their work, three or four generations had moved through the transition. We’re going to do it in one.

Slow hollowings give institutions time to adapt. Fast hollowings ask a lot of individuals before institutions catch up. This one’s real fast.

Where this goes from here

What I’m not doing in this piece, on purpose, is solving it. I haven’t told you what’s being built underneath the hollowing, what survives it for the people inside, or what posture this is going to take. That’s evolving quickly.

For now, let’s stick to the diagnosis. The job is the container, not the work. The container is hollowing. We’ve seen this shape before, and the part we haven’t seen is how fast.

What do you think? Does that match what you’re seeing in your own work, or in the work of people around you? I’d rather hear from you than guess.

I do rigorous signals intelligence for these pieces. They’ve allowed me to see movements before they happen, things that change a company’s position in a meaningful way that usually doesn’t show up in the planning process. I’ve been doing portfolio and company analyses for investors for a while now, and I’m extending the same tools to founders. You can see what investors see here.

A friend asked me a few weeks ago what I thought the next three years would look like. He’s a founder, mid-forties, building something he cares about, and he wanted a real answer, not a hot take. Just somebody he trusts saying, here’s what I think is coming.

So I told him what I’m seeing. And he said, “Why don’t people talk about this?”

So I am.

A brief caveat before I do: nobody knows what the next three years will look like. I want to say that at the top because most of what’s getting published right now skips that step, and the skip is doing a lot of damage. The future is genuinely uncertain. The honest version of any read on it includes “I could be wrong about this,” and the people writing in this space who don’t include that line are, I think, doing a thing that’s more about confidence than clarity.

That said, I do think you and I, if we look at what’s underneath the headlines, can get pretty close. Closer than most. It’s not because we’re smarter. It’s because we’re willing to look at the parts of the data that aren’t being put on the front page, and we’re willing to think about this stuff for longer than a news cycle.

So that’s what this series is. Four pieces, one a week for the next four Thursdays, on what I think is most likely to happen between now and Q2 2029. It’s not a forecast in the “bet on this” sense. It’s a read. A careful one. With receipts where I have them and honest, transparent uncertainty where I don’t.

Two stories on offer, both probably wrong

The two loudest versions of the future right now are pretty far apart.

One says we’re months away from something that looks like a singularity. AI cures disease, replaces work, ends scarcity. The people writing this aren’t making it up. The pace of capability growth is real and the optimism comes from somewhere honest. I just don’t see the leap from “the models are getting impressive” to “abundance is a year away” actually showing up in the data.

The other says we’re months away from something that looks like collapse. Mass unemployment, hollowed middle class, real social rupture. The people writing this are also not making it up. The structural displacement signals are real and the dread comes from somewhere honest too. I just don’t see the leap from “the labor market is freezing” to “civilization breaks” showing up in the data either.

I’m not going to spend this series arguing with either camp. I think highly of people in both, and I believe they’re seeing real things. I think both are taking what’s loudest in the present and projecting it forward in a straight line, and that’s a pretty reliable way to get the next three years wrong. Most people are smart enough to know that. They’re just not seeing a third option that feels honest.

So that’s what I want to give you. A third option that feels honest.

What I think’s actually happening

Let me say what I see.

The hiring engine has stopped, not slowed. The JOLTS hires rate, which is the rate at which the U.S. economy is actually hiring people, is at 3.1 percent. That’s the floor of the Great Recession and the floor of the COVID lockdowns at the same time. Unemployment looks fine because the pipe is closed at both ends. Companies aren’t hiring, and people aren’t quitting because they don’t believe a better job is on the other side. The numbers in the headlines look normal. The numbers underneath don’t.

At the same time, AI infrastructure is being built at a scale that’s hard to grasp. Hyperscaler AI capital expenditure went from about $256 billion in 2024 to a projected $750 billion in 2026. Microsoft, Google, Amazon, Meta, Oracle. That’s not a feature investment. It’s the kind of money you spend when you’re betting that the work running on top of the infrastructure is about to look very different. They’ve made the bet. The receipts are in their 10-Ks.

And the human signal underneath both of those numbers is loud. Mercer’s 2026 talent report has employee thriving down twenty-two points in two years. AI anxiety is up twelve. Stanford’s data on workers ages 22 to 26 in AI-exposed jobs has them down twenty percent on employment. Something is happening to people. It’s already showing up in their mood, their bodies, their trust in institutions. It just isn’t showing up yet in the headline unemployment rate.

Three things, all pointing the same direction. The hiring freeze, the substrate buildout, and the human signal. They’re not opinions. These are numbers anyone can look up.

What I think comes next

Here’s the part I want to under-claim, because the whole point of this is honest reading.

I don’t think we get a singularity (yet). I don’t think we get collapse. I think we get something quieter and stranger and, in some ways, harder to talk about because it doesn’t fit either headline.

I think the job, as a 150-year-old container that bundles income, identity, time-structure, belonging, and benefits, starts coming apart for a chunk of the cognitive middle class over the next three years. Not for everyone. Jobs hold up fine for people who are AI-complementary, where the wage premium is already 56 percent and rising. Jobs hold up fine for the trades, because AI can’t fix your sink. But for the cognitive middle, the people whose work is mostly typing, mostly producing documents, mostly meetings, the container starts to hollow. The work changes shape. A lot of the work routes into freelance, fractional, micro-firm, and small consulting arrangements. By 2029 the freelance count is on track to clear ninety million Americans, which is roughly half of the workforce.

That’s not a textbook collapse. It’s more of a reorganization. And honestly, for a lot of people, it’s already happening. The fractional CMO. The solo consultant with a niche and a Substack. The friend who got laid off and “is figuring out what’s next” and is somehow also doing pretty well. These aren’t anomalies. They’re the leading edge.

The reason this isn’t on the front page is that the aggregate statistics were built for a labor market that doesn’t fully exist anymore. Unemployment can stay around four percent while the cognitive middle leaves the W-2 quietly, one Substack subscription and one fractional engagement at a time. The country can look fine on paper while the shape of work is changing underneath. In some ways, that’s what’s making this transition feel weird. The data on the screen doesn’t match the feeling in the room.

It’s going to be a lot of change. Some of it’s hard. Some of it, for the right people, is genuinely good news. The five things that travel through this transition, the things that hold their value or actually go up in value, are mostly old human capacities. Discernment. Judgment. Taste. Empathy. Relational skill. Most of the people reading this already have them, often in surplus, and have been told for a while that what really matters is technical skill and credentials. The data says the opposite. The premium on the human stuff is going up.

What would change my mind

This is the part I rarely see anyone include, so I want to.

If the JOLTS hires rate climbs back above 3.8 percent for two quarters, the freeze was cyclical and I’m wrong. If hyperscaler capex reverses below $400 billion in 2027, the substrate slows and I’m wrong. If the 22-to-26 cohort employment recovers above minus five percent, the entry pipeline is repairing itself and I’m wrong about how durable this is. If a federal portable-benefits law passes for 1099 cognitive workers, governance moves and the read changes.

None of those is on track right now. All of them are watch items. Two of them firing within twelve months would meaningfully shift what I think.

I want to be wrong-able about this, because the alternative is the kind of confident prediction that’s been making the rest of this conversation exhausting to read.

What’s coming over the next four weeks

Each Thursday for the next four weeks, I’ll publish one piece in this series.

Week one: the job is the container, not the work. The W-2 job is hollowing. The actual work is leaving the container faster than the container is being remade. We’ve seen this shape before. The factory hollowed the household. The spreadsheet hollowed the clerk. Containers persist; contents migrate. This piece is the diagnosis.

Week two: something is being built on top of this. The substrate. $750 billion in capex is a big enough number that it’s worth slowing down and looking at what it’s actually buying. The benchmarks are public. The capability is getting real fast. The earnings calls have started saying “digital labor” out loud, which is a verbal admission worth paying attention to.

Week three: what you already have that survives this. The five attributes that travel. Most of the people reading this are already carrying them. The math on the experience premium is empirical. The takeaway is more hopeful than people are expecting and more specific than the usual upskilling advice.

Week four: how to stand in this. What a working professional, founder, investor, or leader actually does with this read. No five-step plans here. This is a fundamental posture. The posture is the part that matters.

The honest version of all this

I think it’s all going to be okay. I want to say that clearly because I don’t see anyone saying it cleanly.

Not okay in the sense of nothing-changes-everything’s-fine. Okay in the sense of: we’re going through something real, the change is fast, it’s going to feel disorienting for a while, and the people reading this are mostly the people who can come through it well if they see what’s coming and start moving with it instead of against it.

There’s a lot to be honestly worried about. There are people whose lives are getting harder, in measurable ways, because the structural changes are happening faster than the institutions that used to catch them can adapt. That’s real. That’s worth taking seriously. None of this read pretends otherwise.

And. The people reading this Substack are mostly thoughtful, mostly curious, mostly already a little ahead of the consensus. The part I want to keep alive over these next few weeks is that the things that make you good at moving through something like this are mostly things you already have. The transition asks you to trust them more, and to trust the credentials and the comfort of the old container less. Most of the work is a small reorientation, done early, before the reorientation gets forced.

That’s what I want this series to be. A small reorientation, done early. With the receipts I have, the uncertainty I have, and an open invitation to push back if you see it differently. (I’ll keep updating this read as the data moves; subscribe if you want first looks at how it holds or breaks.)

More than anything, I want you to know this:

It’s all going to be okay. Maybe a lot better than okay.

If you’ve been reading for a while, you know that I do more than think about the future. I use what I’ve learned over decades to examine the structure of the systems our world operates on and how those systems change. These analyses focus on macro trends like markets and workforce, specific cuts like sector analyses, and more narrowly focused portfolio and company projections. I identify thresholds and decision points for investors and company leaders that cut the fog and show what’s most likely to happen, and when. You can dig into some of these projections at josephlogan.com. For clear projections on your company or portfolio, contact me directly.

I didn’t set out to predict massive layoffs. I didn’t set out to predict anything.

What I wanted to do was establish a way to pierce the fog of opinion and fear and get a better read on what would likely happen, and when. That led me to create and train two thoughtmodels that could find signal in the noise, back it up with data, and test against the knowledge and wisdom of generations of accumulated insight.

I began with a projection last summer that aimed to be roughly right rather than precisely wrong. It found a years-wide gap in consensus thinking and likely outcomes. Six months later, the conditions it predicted were happening 2-3 years earlier than the consensus.

Today’s analysis is the most information-rich and data-heavy yet. I’m not offering it to stoke fear—much the opposite. I’m showing the direction under the numbers because I believe business and work are changing in profound ways, and I want to share what I’ve learned. What follows is grim news for those who like predictable quarters and full-time W2 employment. But I don’t see it as grim news at all.

I’m robustly optimistic about the direction business and work are headed. Still, transitions can be hard. This one will bring some pain. But it’s creating the conditions for what’s next. I believe we have a unique opportunity if we can keep clear eyes, heads, and hearts.

On April 10, the University of Michigan published its preliminary consumer sentiment reading for April 2026. It came in at 47.6: the lowest number in the 74-year history of the survey. Lower than June 2022. Lower than the worst month of the financial crisis. Lower than any point during COVID.

Six days earlier, the Bureau of Labor Statistics reported that the U.S. economy added 178,000 jobs in March. Markets rallied. The word “resilience” appeared in every headline. FinancialContent ran a piece titled “Labor Market Resilience: March Jobs Report Shatters Forecasts.” Consensus had been 59,000 to 65,000. We nearly tripled it.

So which is it? A resilient labor market or an all-time low in how people feel about the economy?

Both, actually. And the distance between them is the story.

I spend most of my working hours inside economic data, building projection models, tracking the structural layer underneath the headlines. It’s the kind of work that makes you unpopular at dinner parties. But four installments into this series, I keep finding the same thing: the number and the thing the number is supposed to measure are drifting apart. This month, the gap is the widest I’ve seen.

Is unemployment a useful measure…?

The number says one thing. The structure underneath says another.

+178,000 nonfarm payrolls, seasonally adjusted. Total private sector: +186,000. Government: -8,000. Unemployment rate: 4.3%, down a tick from February’s 4.4%.

Good print. Until you pull on the threads.

Thread one: the strike. Roughly 35,000 of the 178,000 were physician office workers returning from a strike. These weren’t new jobs. They were the same jobs, temporarily absent from the count in February, returned in March. Strip them out and organic job creation was approximately 143,000. Still decent. Not the headline.

Thread two: the revision. The same release revised February from -92,000 to -133,000. That makes February the second-worst payroll month since the pandemic. The combined January-February revision was a net -7,000. The two-month picture got slightly worse, not better.

Thread three: the three-month average. January: +160,000 (revised up). February: -133,000 (revised down). March: +178,000. Average those and you get roughly 68,000 jobs per month. That is not a healthy labor market. That is a labor market running below population growth.

Thread four: who left. The civilian labor force shrank by 396,000 people in a single month, from 170,483,000 to 170,087,000. Labor force participation fell to 61.9%. People classified as “not in the labor force” increased by 488,000. Discouraged workers, those who stopped looking because they believe no jobs exist for them, spiked 39% in one month, from 366,000 to 510,000. Marginally attached workers rose by 325,000 to 1.944 million.

The unemployment rate fell partly because the denominator shrank. People didn’t find jobs. They stopped being counted.

Thread five: the other survey. ADP’s National Employment Report showed +62,000 private-sector jobs in March. That’s a 116,000 gap from the BLS private-sector figure of +186,000. ADP and BLS diverge regularly, but this is a wide spread. ADP’s size breakdown tells its own story: small businesses (1-49 employees) added +85,000 jobs. Medium firms (50-499) shed -20,000. Large firms (500+) shed -4,000. The entire positive print rested on the smallest employers.

Thread six: the frozen labor market underneath. The February JOLTS report, released March 31, showed a hires rate of 3.1%, matching the COVID-era low and the lowest since January 2011. Job openings fell to 6.882 million. The quits rate has been locked at 1.9% for eight consecutive months, a decade-low signal that workers feel trapped, unable to risk a job change. Long-term unemployment (27+ weeks) sits at 1.8 million, representing 25.4% of all unemployed and up 322,000 year-over-year. Average unemployment duration: 25.3 weeks, up from 21.4 a year ago. U-6 underemployment: 8.0%.

This is not a resilient labor market. It’s a labor market that generates a resilient headline.

Change is here, but it’s not evenly distributed

The March gains were concentrated. Healthcare and social assistance: +89,900, but remember, 35,000 of that was the strike reversal. Leisure and hospitality: +44,000, driven by seasonal food services and entertainment hiring. Construction: +26,000. Transportation and warehousing: +21,000, almost entirely couriers and messengers at +20,400.

Now look at what’s bleeding.

Financial activities: -15,000. Finance and insurance specifically: -16,200. This sector has now shed 77,000 jobs since its May 2025 peak. Credit intermediation: -8,500. Insurance carriers: -5,700. Month after month, the cuts deepen.

Professional, scientific, and technical services: -13,500. Computer systems design: -13,200. The broader professional and business services category showed +2,000, essentially flat, only because administrative and support services (+18,500) masked the professional-services decline.

Information: -3,000. This sector has been declining for months. Computing infrastructure, data processing, telecommunications: all contracting.

Federal government: -18,000. Federal employment is now 355,000 below its October 2024 peak, a decline of 11.8%. This is an unprecedented peacetime reduction in the federal workforce.

The pattern is worth stating plainly. The sectors most exposed to AI, finance, professional services, information technology, computer systems design, are the ones bleeding consistently. The sectors holding the headline up, healthcare (with a strike bounce), leisure and hospitality (seasonal), construction (physical labor), are the ones least exposed to the current wave of AI displacement.

The headline aggregates them. The structure separates them.

Wall Street rewards layoffs

Something shifted in March. For the first time since Challenger, Gray & Christmas began tracking the metric in 2023, AI was the number-one cited reason for layoffs in a single month. Of the 60,620 job cuts announced in March, 15,341, or 25%, were attributed to AI. For Q1 2026, tech-sector cuts totaled 52,050, up 40% year-over-year. According to Nikkei Asia data, 78,557 tech workers were cut globally in Q1, with 47.9% of those cuts explicitly attributed to AI and workflow automation.

The individual announcements tell the story at company scale:

• Oracle: An estimated 30,000 cuts, roughly 18% of its workforce, explicitly tied to AI infrastructure spending. Stock down 27% year-to-date. But it spiked nearly 6% the day of the announcement and continues to climb.

• Atlassian: 1,600 cuts, 10% of the company. CEO Mike Cannon-Brookes: “It would be disingenuous to pretend AI doesn’t change the mix of skills we need.” Severance cost: $225-236 million. Stock rose 2% on the announcement.

• Block: 4,000 cuts, taking the workforce from roughly 10,000 to under 6,000. CEO Jack Dorsey: “Intelligence tools have changed what it means to build and run a company.” Then, the line that will echo: “Most companies will reach the same conclusion within 12 months.” Stock rose 24%.

• Citigroup: Targeting approximately 20,000 reductions, about 13% of workforce. CFO Mark Mason cited AI-enabled systems for middle-office functions.

• Salesforce: CEO Marc Benioff, on AI agents now handling roughly 50% of customer interactions: “I need less heads.”

Read those stock reactions again. Oracle cuts 18% and the stock spikes. Atlassian cuts 10% and goes up. Block cuts 40% and goes up 24%. The market is not treating these as distress signals. It is rewarding them. The numbers are lying to you, but not on purpose. That’s almost worse.

This is the profit-per-employee thesis becoming operational. Block’s COO Amrita Ahuja has been tracking gross profit per employee as the company’s core efficiency metric: $500,000 in 2019, $750,000 in 2024, $1 million in 2025, targeting roughly $2 million in 2026. Their internal AI agent “Goose” has been in production for 18 months. Developer productivity is up 40% per engineer since September 2025.

The AI-native companies have already arrived at the destination. Midjourney generates approximately $18 million in revenue per employee with a team of roughly 11 people. Cursor hit $1 billion ARR with around 300 employees, $3.3 million per head. The old SaaS benchmark of $200,000 ARR per employee is now baseline; “good” starts at $300,000; “great” at $500,000-$700,000. Gartner forecasts that new unicorns by 2030 will target $2 million ARR per employee.

Most CEOs see these numbers. They’ll pursue this ratio.

The KPMG Q1 2026 AI Pulse survey found 54% of organizations are now actively deploying AI agents, up from 12% in early 2024. The Federal Reserve’s April 3 FEDS Note estimates that 78% of the American workforce now works at a firm that has adopted AI. McKinsey is running 25,000 AI agents alongside 40,000 human employees. CEO Bob Sternfels says every human employee will be “enabled by at least one agent” by year-end 2026.

The infrastructure is not a speculation. It is an expenditure. Hyperscaler capital expenditure for 2026 is projected at $660-690 billion, nearly double 2025 levels. Amazon alone is spending $200 billion. That money is being spent to build the capacity that replaces tasks currently done by humans. That’s what capital allocation has looked like for a couple of years now.

Layoffs, regret, rehiring, and…?

If you’ve been following this series, you know I refuse to flatten complexity. The regret-rehire cycle is real, and ignoring it would make the analysis dishonest.

The pattern looks like this: companies announce AI-driven layoffs. Then they discover the AI doesn’t fully work, or that they need more human oversight than expected, or that customer satisfaction drops. Then they rehire, partially.

The data is substantial:

• Forrester’s 2026 predictions report: 55% of employers who made AI-attributed layoffs already regret the decision. More than one in three spent more on restaffing than they saved from the cuts.

• Careerminds survey of 600 HR professionals: 35.6% rehired more than half the roles they cut. “More than half of HR leaders said AI required more human insight than anticipated.”

• Gartner predicts that by 2027, half of companies that attributed cuts to AI will reinstate employees for similar roles, often under different titles.

• Harvard Business Review reported that 60% of global executives cut headcount in anticipation of AI. Only 2% because AI was actually performing in production.

Klarna is the canonical case. Workforce from 7,000 (2022) to 3,400 (2024), driven by a chatbot that claimed to do the work of 700 agents. Then repeat contacts jumped 25%. CEO Sebastian Siemiatkowski admitted: “We went too far.” They rehired. And now? Siemiatkowski expects headcount to fall further to roughly 2,000 by 2030, not through another round of layoffs, but through natural attrition at 20% annual turnover.

People watching this cycle and concluding “see, they need humans after all” are reading a correction as a reversal. It is not a reversal.

This is a damped oscillation within a reinforcing loop. Each cycle, cut, regret, partial rehire, learn, cut again with more precision, settles at a lower equilibrium of human headcount. Klarna’s trajectory is not 7,000 → 3,400 → 7,000. It is 7,000 → 3,400 → partial rehire → 2,000 target. The oscillation dampens toward displacement, not recovery. The companies in their second or third cycle are not making the same mistakes. They’re cutting with better information about which tasks the AI can actually do.

The 2% figure from HBR, only 2% of cuts driven by AI that was actually working, is itself a time-stamped number. It was true when measured. It is less true each quarter as the tools improve. The gap between “cut in anticipation” and “cut because it’s working” is closing from both directions: models improve, and organizations learn what to automate.

The evaporation thesis

Everything I’ve described so far is top-down. Companies announce layoffs. The numbers show up in Challenger data. This is the visible displacement.

There is a second dynamic running simultaneously. I explored it in depth in The Evaporation Thesis, and it still doesn’t appear in any government report.

Bottom-up displacement is when workers automate their own workflows, but not because they’re told to. More because the tools are available, cheap, and obviously useful. The Fed’s FEDS Note found that 41% of the labor force now uses generative AI at work, up 31% year-over-year. The KPMG survey shows 63% of organizations now require human validation of AI agent outputs, up from 22% just a year ago. Think about what that means: agents are increasingly doing primary work and humans are reviewing it, not the other way around.

If that math ain’t mathing, try it this way: GPT-4-level intelligence cost $30 per million tokens in 2023. It costs roughly $0.06 per million tokens today—a 500x collapse in 16 months. That’s more than a price drop. It’s is a phase change in the economics of cognition. At these prices, the math on automating a $60,000/year knowledge-worker task is trivially positive. (I’m tracking these cost curves monthly now; subscribe if you want first looks at where they’re heading.)

Meanwhile, business applications are running at roughly double pre-pandemic rates: 491,941 new business applications in March 2026 alone. What these formations represent is another compelling thread to pull. Many are likely necessity entrepreneurship: displaced workers who can’t find a corporate role starting consulting practices, automation agencies, and AI-augmented solopreneurships. The gig economy has swelled to $674 billion globally with 70 million U.S. participants, projected to cross 50% of the workforce by 2027.

But AI is compressing freelance value too, and that picture’s more complicated than I realized last summer. Ramp’s research tracking actual firm-level spending found that freelance marketplace spend fell from 0.66% of total corporate spend in Q4 2021 to 0.14% by Q3 2025, while AI model provider spend rose from zero to nearly 3%. More than 50% of businesses using freelancers in 2022 have stopped entirely. For the highest-exposure firms, every $1 reduction in freelance spend corresponded to only $0.03 in AI spend, a 25x cost savings. Fiverr has lost 95% of its stock value since 2021. Upwork’s active buyers have fallen from 4.2 million to 3.1 million.

The evaporation layer doesn’t show up as layoffs. It shows up as hours that don’t get hired. Contracts that don’t get issued. Junior roles that don’t get posted. Resume.org surveyed roughly 1,000 U.S. business leaders and found that 36% expect to stop hiring entry-level workers by year-end 2026. Entry-level job postings for AI-exposed roles have already fallen more than 40% from 2023 levels according to Revelio Labs. Only 30% of 2026 college graduates are securing full-time jobs post-graduation, down from 41% in 2024.

This is not a tight labor market with a skills mismatch. It is a labor market where the work itself is changing shape faster than the instruments can track.

The Stanford Digital Economy Lab paper, updated through February 2026, found that employment for 22-25-year-old software developers has declined roughly 20% from the late 2022 peak. Call center hiring for the same age group is down about 15%. Their regression analysis shows the decline in AI-exposed occupations becoming statistically significant starting in 2024 and reaching approximately 16% by October 2025. It has not reversed.

Erik Brynjolfsson’s framing is useful here: the traditional career “pyramid,” many junior roles feeding into fewer senior ones, is shifting to a “diamond” shape. The entry points are shrinking while mid-level and senior roles remain stable or grow. The base of the workforce pipeline is narrowing. Nobody above it can feel it yet.

I’ve been working through a finding with the Hari framework that I want to address here: the apprenticeship crisis. The junior roles being eliminated aren’t just jobs. They’re the training pipeline for the next generation of senior professionals. When a law firm uses AI to replace research associates, or a consulting firm eliminates junior analysts, or a bank automates the entry-level credit analysis that teaches how to think about risk, they’re not just cutting a cost line. They’re severing the mechanism that produces the judgment, relationships, and domain expertise that make senior professionals valuable. The Dallas Fed’s experience premium research shows that AI complements experienced workers while substituting for entry-level ones. But the experienced workers who benefit today were built by the entry-level positions being destroyed. This is a 5-10 year delayed structural failure, invisible to current efficiency metrics, devastating to the professional pipeline underneath. Firms optimizing for quarterly margins are consuming seed corn.

This matters for the Hari projection because evaporation-layer displacement doesn’t trigger any of the traditional alarm systems. It doesn’t register as a layoff. It doesn’t create an unemployment claim. It doesn’t even create a discouraged worker, because the person never had the job in the first place. It is preemptive displacement, and it may be larger than top-down displacement by the time we have instruments to measure it.

The chain has almost fully activated

In August 2025, the first installment of this series proposed a six-link causal chain for AI-driven workforce displacement. In February, the six-month review showed the chain tracking ahead of schedule. In March, I showed the structural deterioration underneath a headline that looked stable.

Here is where each link stands in April 2026:

In August, I scored the composite signal at approximately 3.8. In February, at 4.57. This month, I’d place it in the 4.65-4.75 range. The incremental shift reflects the transition of Link 6 from Emerging toward Active: the unemployment picture is deteriorating underneath a headline rate that looks relatively benign.

All six links are now active or in active transition. This was the August projection’s end state for the onset of significant displacement. We projected that window as Q3 2026.

Q3 2026 is 10 weeks away.

Narrative lag

There’s a concept I’ve been developing across these installments: the narrative lag.

The headline was +178,000. The market rallied. The narrative was “resilient economy.” That narrative was set within 24 hours of the release and will persist until the next data point challenges it.

But:

• University of Michigan consumer sentiment hit an all-time record low of 47.6 in April.

• Year-ahead inflation expectations surged from 3.8% to 4.8%, the largest single-month jump since April 2025.

• Goldman Sachs raised recession probability from 20% to 30% within two weeks in March. Moody’s Analytics places it at 48.6%. Polymarket prediction traders: 31%.

• Total credit card debt hit $1.28 trillion, a record since the New York Fed began tracking.

• National office vacancy sits at 18.4%. Industrial sublease space hit a record 278 million square feet, exceeding even the 2023 office sublease peak.

• A NerdWallet survey found 65% of respondents expect a recession in the next 12 months.

The New York Fed describes the current moment as displaying “characteristics of a K-shaped economy”: a disconnect between stock market performance and consumer reality. Goldman’s AI economist Joseph Briggs said in March: “The big story in 2026 in labor will be AI.”

The narrative lag is the gap between what structural indicators show and when the consensus narrative catches up. This gap is itself analytically useful: it tells you how much information advantage remains in reading the structural layer. When the narrative and the structure align, the analytical advantage disappears. Right now, the gap is as wide as I’ve ever measured it.

Here’s a concrete way to think about it. The ADP Research “Today at Work 2026” survey of 39,000 workers across 36 countries found that only 22% of workers globally “strongly agree” their job is safe from elimination. Among front-line individual contributors: 18%. Even among C-suite executives: 35%, barely more than a third. The Mercer Global Talent Trends 2026 survey of 12,000 workers found that 40% fear AI will make their job obsolete, up from 28% in 2024. The Jobs for the Future survey found that optimism about AI’s employment impact dropped 10 percentage points; only 39% now feel positive.

The workers feel it. The consumers feel it, hence the record-low sentiment reading. The financial markets don’t feel it, hence the rally on the jobs report. The policy apparatus doesn’t feel it: the Economy of the Future Commission Act won’t produce recommendations until Q2 2027, and the DOL’s current retraining capacity serves roughly 60,000 workers annually against a tech sector alone displacing 55,000 to 120,000 per year.

The gap between who feels it and who doesn’t is the narrative lag made visible.

Compound fracture

Tariffs and AI are compressing opposite ends of the labor market at the same time. I haven’t heard anyone talk about these in tandem.

Since Trump’s “Liberation Day” tariff announcement in April 2025, the economy has shed an estimated 89,000 manufacturing jobs and 123,700 transportation/warehousing jobs through February 2026. That’s 189,600 blue-collar jobs lost to trade disruption. Simultaneously, AI is taking white-collar knowledge work: finance, tech, legal, professional services.

The perverse interaction: tariffs are supposed to incentivize domestic manufacturing investment. But manufacturers are responding to tariff-driven cost pressure by accelerating AI and automation investments to offset higher input costs. As CNN’s tariff impact tracker noted: “Tariffs are not the sole reason for the sluggish job market; advancements in artificial intelligence have enabled businesses to operate with fewer employees, exacerbating the situation.”

The CNBC Supply Chain Survey from January showed supply chain manager layoffs doubling from 16% to 32% post-tariffs, with 56% of firms worried about recession. At the American Manufacturing Summit in March, 98% of manufacturers reported exploring AI automation, though only 20% were prepared to deploy at scale.

This is the compound fracture. There is no sector of the economy simultaneously gaining from both tariff protection and AI immunity. Construction comes closest, and construction indeed added 26,000 jobs in March. But one sector cannot absorb the displacement from all others.

The healthcare sector, which carried the March print, presents its own version of this tension. Healthcare added 89,900 jobs in March, the largest sectoral gain. But Challenger data shows healthcare/products cuts at a record Q1 of 23,520, the highest first quarter ever. UnitedHealth Group is targeting $1 billion in AI-enhanced cost savings in its insurance division. The sector is simultaneously the largest source of new jobs and quietly restructuring its administrative and insurance back-office functions. The clinical side, nurses, physicians, therapists, faces a shortage. The administrative side faces the same AI displacement pressure as every other knowledge-work sector. These two realities coexist in the same industry, and the BLS aggregates them into a single number.

A single number that tells you everything is fine.

What to watch for May

Seven things I’m tracking for the next installment:

1. The March JOLTS report (due late April). February’s hires rate of 3.1% was an alarm. If March confirms or worsens, the “frozen labor market” thesis, employers posting openings but not converting them to hires, becomes difficult to argue against. An Enhancv survey estimates 27% of all job listings are ghost postings with no intention to hire.

2. Q1 2026 earnings calls. Listen specifically for profit-per-employee language migrating out of tech and into financial services, healthcare administration, and professional services. The metric is spreading. Business Insider has already noted that companies are “no longer competing on headcount, they’re competing on efficiency.”

3. Whether the April BLS confirms or reverses the March bounce. The March print was strong on the surface. If April reverts to the 50,000-80,000 range, it confirms the three-month average of roughly 68,000 as the real trend line.

4. Initial claims. The 4-week moving average of 209,500 remains stubbornly low, well below any alarm threshold. I’ve noted previously that the 75% non-filing rate for unemployment insurance may render this indicator structurally incapable of triggering in a displacement-driven downturn (as opposed to a demand-driven recession). But the stability of claims data is a genuine counterargument to the acceleration thesis, and I want to name it honestly.

5. Entry-level hiring data. The pipeline is collapsing. Resume.org found 36% of companies expect to stop entry-level hiring by year-end. Stanford Digital Economy Lab data shows employment for 22-25-year-old software developers down roughly 20% from the late 2022 peak. Only 30% of 2026 graduates are securing full-time employment, down from 41% in 2024. This is where the pipeline disruption becomes structural.

6. The tariff-AI interaction. If tariff uncertainty persists into summer, manufacturers will continue to lean into automation as a hedge against cost volatility. Watch for announcements that frame AI investment as a response to trade policy.

7. AI agent cost trajectory. The KPMG survey shows organizations projecting an average of $207 million in AI spend over the next 12 months, nearly double last year. As costs fall and capabilities improve, the economic threshold for replacing each successive human task drops with it.

Honest caveats

This analysis is only useful if it’s honest about what could prove it wrong.

The March print was genuinely positive in several sectors. Construction, manufacturing, leisure and hospitality all added jobs that weren’t just statistical noise. The organic 143,000 (ex-strike) is not a weak number in isolation. If April and May print similar numbers, the three-month average climbs back to a range that challenges the deceleration thesis.

Claims data remains remarkably stable. Initial claims at a 4-week average of 209,500 with continuing claims at 1,794,000, the lowest since May 2024, is not what an accelerating displacement story should produce. The 75% non-filing rate hypothesis explains this, but it could also mean displacement is genuinely slower than the other indicators suggest. I take this signal seriously.

AI washing is real. Sam Altman himself acknowledged: “There’s some AI washing where people are blaming AI for layoffs that they would otherwise do.” Cognizant’s Chief AI Officer: “Sometimes AI becomes the scapegoat from a financial perspective.” The NBER CFO Survey found over 80% of firms report no measurable impact on employment or productivity from AI. The signal is real, but it is noisier than the headline numbers suggest.

DOGE federal cuts confound the signal. The 355,000-job decline in federal employment since October 2024 is a policy choice, not an AI effect, even if some agencies are rehiring workers and spending more than before the cuts. Separating federal workforce reduction from AI-driven private-sector displacement is important and difficult.

The productivity data hasn’t moved. BLS Q4 2025 data shows nonfarm business labor productivity at +1.8% annualized. There is no clear “AI signature” in the macro productivity numbers. If AI were genuinely transforming productivity at scale, you would expect this figure to be moving. It isn’t, yet. This could mean the displacement is real but the productivity offset hasn’t materialized, which would be a different and more concerning scenario. Or it could mean both the displacement and the productivity gains are smaller than the micro-level indicators suggest.

The regret-rehire cycle could be larger than I’m modeling. If the Gartner prediction is right, that half of AI-attributed cuts will be reinstated by 2027, the net displacement through this cycle could be substantially lower than the gross announcement numbers suggest. I’m modeling the oscillation as damping toward lower headcount. But the amplitude of the oscillation matters for timing.

Looking toward Q3

In August 2025, the first Hari Projection said that Q1-Q3 2026 would be the window where AI-driven workforce displacement moved from signal to structure. We are now inside that window. The chain is tracking ahead of the base case on five of six links. The sixth, headline unemployment, is lagging, but the indicators underneath it (discouraged workers, long-term unemployment, labor force participation, hires rate) are deteriorating on schedule or ahead of it.

What do the next six months likely look like?

Q2 earnings season (July) will likely show the broadest adoption of AI efficiency language yet. Every company watching Block’s stock rise 24% on a 40% workforce reduction is learning the same lesson. Expect “profit per employee,” “AI-enabled efficiency,” and “workforce optimization” to appear in earnings calls across financial services, professional services, and healthcare administration, not just tech.

The regret-rehire cycle will continue, but dampen. Companies in their second or third oscillation will cut with more precision, better data on which tasks automate well, and less of the “cutting in anticipation” pattern that characterized the first wave. The HBR finding that 60% of cuts were anticipatory will likely shrink as organizations accumulate actual deployment data.

The entry-level hiring collapse will likely spread. It’s already visible in tech and professional services. The KPMG data showing 57% of leaders say AI agents have already changed their approach to entry-level hiring, and a Goldman Sachs analysis estimating AI is erasing roughly 16,000 net U.S. jobs per month, suggest this will migrate into financial services, healthcare administration, and legal support by Q4.

The evaporation dynamic will likely accelerate. AI agent costs continue to fall. Each new price point makes a new category of task economically automatable. The 54% deployment rate that KPMG measured is climbing a steep curve. The service-sector multiplier, estimated at 3-5 service jobs at risk per tech job lost, operates on a 2-4 quarter lag. The tech layoffs of Q1 2026 will begin to manifest in regional service economies by Q3-Q4.

If tariff-driven cost pressure persists, manufacturing will layer AI automation on top of trade restructuring. The 98% of manufacturers exploring AI automation are looking at it partly because tariffs have made labor cost reduction more urgent. This is the double squeeze.

The structural story will become harder to hide behind headline numbers. As seasonal adjustments normalize and one-time factors (strikes, federal cuts) fade from the calculation, the underlying trend, which I estimate at approximately 68,000 jobs per month, will become the visible trend. It’s unclear whether the economy can sustain job creation at that rate while absorbing the multiplier effects from the displacement that has already occurred.

The NBER/Duke CFO Survey from March projects AI-driven job cuts at roughly 502,000 in 2026, a ninefold increase over 2025’s 55,000. Goldman Sachs research from early April estimates AI is erasing approximately 16,000 net U.S. jobs per month (25,000 substitution minus 9,000 augmentation). These are projections, not certainties. But they come from institutional research teams with access to granular firm-level data, and they align directionally with the structural pattern the Hari chain has been tracking since August.

If these projections are even partially correct, the math on headline employment becomes difficult by late 2026. A labor market generating 68,000 jobs per month while absorbing 16,000 net AI-driven losses produces a net of approximately 52,000, below the roughly 100,000-120,000 monthly pace needed to keep up with population growth. I expect this to be the defining issue of the November midterm elections. The headline unemployment rate would begin to rise meaningfully by Q4 2026 or Q1 2027, even without a traditional demand-side recession.

Mind the gap

I want to be precise about what this analysis is and what it isn’t.

It is not a prediction that the economy is about to collapse. GDP grew at reasonable rates through 2025. Wage growth is positive. Claims are stable. These are real data points that describe a real economy that is still functioning.

What this analysis shows is a structural transformation happening underneath numbers that were designed to measure a different kind of economy. The BLS payroll survey, the unemployment rate, initial claims: these instruments were built to detect demand-driven recessions, the kind where businesses lose customers, cut workers, and the cuts show up in claims filings and payroll reports within weeks.

AI-driven displacement doesn’t follow that pattern. It shows up as tasks that don’t get assigned. Roles that don’t get posted. Junior positions that disappear from the hiring plan. Hours that get absorbed by a tool instead of a person. The displaced worker doesn’t file an unemployment claim; they freelance, consult, drive for Uber, start a business application, and exit the labor force statistics at a rate of 396,000 in a single month.

The work doesn’t go away. The containers change. The margin that used to go to a human employee goes to a software subscription and a smaller team. The economy gets more productive in the aggregate, eventually, but the transition period produces exactly the pattern we’re seeing now: strong headline numbers, deteriorating structural indicators, record-low consumer sentiment, and a growing gap between the narrative and the reality underneath it.

Is unemployment rate a useless number? It might be.

Ten weeks to Q3. The chain is tracking. The oscillation is dampening. The evaporation is accelerating. The narrative hasn’t caught up.

This is where the danger lives: not in the number, but in the distance between the number and the thing it’s trying to measure.

And here is the thing about measurement gaps. They close. The structural math, unlike the press releases, doesn’t negotiate. You have two, maybe three quarters to position yourself on the right side of this, to build the skills, the networks, the understanding of where value is migrating. The distance between the number and the thing is your window. Use it before the consensus catches up.

This is the fourth installment in an ongoing series tracking AI employment leading indicators. Part 1: The Last Normal Year (August 2025). Part 2: AI Is Coming for Your Job After All (February 2026). Part 3: Underneath (March 2026). See also: The Evaporation Thesis (March 2026).

This analysis uses the Causal Chain Projection Framework (Hari v3.0) and draws on the March 2026 Employment Situation Report (BLS), JOLTS February 2026, ADP National Employment Report, Challenger Gray & Christmas job cut data, Federal Reserve FEDS Note on AI Adoption, KPMG Q1 2026 AI Pulse Survey, Dallas Fed research on experience premiums, Stanford Digital Economy Lab employment studies, University of Michigan Consumer Sentiment Survey, and current reporting from Fortune, Bloomberg, Forbes, CNBC, Goldman Sachs Research, and Tom’s Hardware.

The May projection will publish after the April BLS release. If you’re tracking similar signals in your sector or building resilience for what’s emerging, I’d love to hear what you’re seeing.

Welcome to Q2 2026. Oracle notified 30,000 employees yesterday via a 6am email that their services were no longer needed. I won’t be writing about that today. There are plenty of places you can find opinions about that. I do projections.

The monthly Hari report on AI Job Displacement drops two weeks from today. Since its first projections last summer, th…

Read more

I automated a workflow last week that I had no idea how to do.

I had no idea how to do it manually. I described what I wanted to happen, an AI agent figured out the sequence, and the thing was running inside of a couple of hours. A process that would have taken me days to learn and weeks to systematize was just... done.

My first thought was delight. My second thought was: I just permanently removed something from my workload. My third thought persisted. If I can do this, and I am not an engineer, then the 78% of knowledge workers already using AI tools at work can do it too. Some of them probably already are.

That third thought became an analysis. I’ve spent the last several weeks running it through the projection framework I use for structural trend work, the same one that produced the timelines in The Last Normal Year and AI Is Coming For Your Job After All. Those pieces tracked the macro forces: corporate adoption, infrastructure buildout, the top-down decisions that lead to headlines about layoffs (I’m updating those monthly now—subscribe for first looks at what’s coming). This one tracks something quieter and, I think, more consequential.

The bottom-up disappearance of work itself.

Your job is a label on a bundle

Here is a sentence that sounds obvious: a job is a collection of workflows.

The marketing coordinator’s job is 15 or 20 workflows bundled together and assigned to one person for 40 hours a week. Write the brief, pull the data, build the report, schedule the posts, update the tracker, send the recap. Each workflow is a discrete sequence of steps with an input and an output. The job title is just a label on the bundle.

We’ve always talked about automation as something that happens to industries, to companies, to categories of workers. A factory installs robots. A bank deploys chatbots. A CEO announces a restructuring. Those are real. I’ve been tracking those dynamics for 18 months and the causal chains are on schedule.

But something else is happening at the same time, and it runs on a completely different clock.

Individual workers are automating their own workflows. One at a time. From the bottom up. With tools they chose themselves, often without telling anyone.

“I know kung fu.”

The first time someone automates a workflow they didn’t know how to do, it’s a moment. It’s a bit like Neo downloading kung fu in The Matrix. Shock first. Then delight. Then a question: what else can I download?

That question is the ignition point. Because once a worker automates one workflow successfully, they have more free time and more skill at automation simultaneously. The next workflow goes faster. The one after that, faster still. Each removal feeds the next. It compounds.

The data says this is already widespread. Microsoft and LinkedIn reported in 2025 that 78% of knowledge workers are bringing their own AI tools to work. Sixty-three percent said they’d continue doing so even under an explicit ban. The fastest-growing open-source project in history right now is OpenClaw, an agentic AI platform. Not a chatbot. An autonomous agent that can execute multi-step workflows on its own. That project hit 250,000 GitHub stars in 60 days, crossing adoption levels that took Linux years to reach.

And cost is collapsing. The price of running an AI agent through a complex task sequence dropped roughly 7.5x in 16 months. We are approaching the point where running an autonomous agent for an hour costs less than a cup of coffee.

The consulting layer is forming too. For every person who figures out automation on their own, there will be ten who hire someone to do it for them. Workflow architects. AI consultants. The personal trainers of the new economy, showing up with a clipboard and a subscription and rebuilding someone’s role in an afternoon. That service layer removes the technical skill bottleneck entirely. You don’t need to be an engineer to have your workflows automated. You just need to be willing to pay someone who is.

Layoffs don’t tell the whole story

The workflows leave, one by one. What happens to the job?

Think about what a role looks like after 40% of its workflows have been automated away. The person is still employed. The title still exists. But the scope of the work has contracted. Management may not even know. The worker has a structural incentive to hide the compression. If the boss finds out the 40-hour job takes 15 hours, the role is at risk. So the automation stays in the dark.

This creates a specific dynamic. Compression builds up silently at the individual level. Dozens, then hundreds, then thousands of workers are sitting on roles that have quietly shrunk to a fraction of their original scope. The organization doesn’t see it because nobody is reporting it. The macro data doesn’t see it because BLS surveys measure jobs, not the workflows inside jobs.

Then something triggers the reveal. A round of layoffs. A performance audit. An economic downturn that forces the question. And when it surfaces, it surfaces all at once. The Thanos snap. The compression was already complete. The snap is just the moment everyone finds out.

Experience premium = higher octane

Which roles compress and which ones transform? The answer showed up in an unexpected place.

The Dallas Fed published research earlier this year on what they call the experience premium: the wage gap between entry-level and experienced workers in the same occupation. That gap turns out to be a remarkably clean proxy for how much tacit knowledge a role contains. High experience premium means the job depends on judgment, relationships, domain expertise that can’t be written down. Low experience premium means the job is mostly defined by its workflows.

When you cross that data with actual AI exposure measurements, a pattern emerges.

Low experience premium, high AI exposure: the workflows automate, and there’s nothing underneath to sustain the role. These are the evaporation candidates. Office administrators, routine data analysts, customer service representatives, entry-level content producers. Roughly 28 to 30 million workers in the US fall into this category. About 38%-41% of the white-collar workforce.

High experience premium, high AI exposure: the workflows automate, but the residual human contribution (the judgment calls, the client relationships, the strategic direction) keeps the role alive in compressed form. These roles transform. They get smaller and stranger and more valuable per hour. Senior financial advisors, experienced software architects, seasoned project managers, creative directors. Another 22 to 23 million workers.

The entry-level pipeline is where both categories converge into something more concerning. In the high-experience-premium professions, the junior positions that once trained the next generation of experienced practitioners are themselves low-experience-premium roles. They evaporate too. The senior financial advisor keeps their job. The junior analyst who was supposed to become the next senior advisor? That position just quietly stopped being posted.

Decentralization in increments

I guess we’re talking about movies today.

In Spike Jonze’s Her (2013), Theodore Twombly writes emotionally sophisticated personal letters for other people. That is his job. It requires voice, emotional intelligence, relational understanding. The workflows that once supported that kind of work (the research, the drafting, the formatting, the scheduling) have been absorbed by AI systems so seamlessly that nobody thinks about them anymore. The world Theodore lives in still has work. Lots of it. What it doesn’t have are jobs in the way we’ve understood them for the last century: fixed bundles of workflows assigned to a single person for 40 hours a week.

That is the end state I see in the data.

Work persists. The human contribution (judgment, creativity, relationship, taste, accountability) persists and probably becomes more valuable. What dissolves is the container. The 40-hour bundled role. The job title that was really just a label on a collection of workflows that no longer need a person to execute them.

The existing literature on the future of work has described this destination. Ravin Jesuthasan and John Boudreau wrote an entire book called Work Without Jobs that maps the organizational redesign path to get there. Their version is managed, deliberate, top-down. Organizations deconstruct roles into tasks and recombine them fluidly.

What I’m seeing is the unmanaged version. Workers automate their own workflows. The jobs shrink in the dark. Organizations discover the compression after the fact. The residual work recrystallizes around whatever human contribution remains. Same destination. Radically different mechanism. And much, much faster.

Evaporation happens faster than you think

The honest version of the timeline looks like this. A leading edge of technical early adopters is already there. Their roles have compressed meaningfully in the last two to three quarters. A second wave, accelerated by the service layer and rapidly improving tools, reaches a broader professional population within two to four quarters. The median economy, where the macro data finally catches up to the micro reality, follows on a scale of years. But not many years.

If that sounds fast, I’ll point out that the same projection framework estimated the top-down displacement timeline at six months back in August 2025. That felt unthinkably soon at the time (consensus at the time was 3-4 years). It has tracked on schedule since. Q4 2026 feels the same way Q1 2026 felt for the original analysis. Unthinkably soon and probably right on track.

Would I leave you hanging without an answer…?

The instinct, when looking at something like this, is to diagnose and stop. Here is the structural force. Here is the timeline. Good luck.

I don’t think that’s useful. The structural force is real, but so is the fact that the human contribution at the center of these roles is becoming more, not less, important. The workflows go away. The judgment stays. The relationships stay. The ability to see a situation clearly and make a call that matters, that stays.

But what we do with the transition period? The window between now and the moment the macro data catches up. That window is where the decisions are.

For executives: the compression is already happening inside your organization. Discover it on your terms or react to it on its terms. One of those is a redesign. The other is a crisis.

For people in roles defined by their workflows: the experience premium data is specific about what endures. Tacit knowledge. Judgment. Relationships. Domain expertise that can’t be codified. The adaptive move is to develop those capacities now, while the role still exists in a form that lets you practice them.

For everyone watching: the work doesn’t go away. The container changes. And the people who understand the new shape of work before it arrives will have built something that compounds in the other direction. Not a shrinking role, but an expanding one. One where the workflows are handled and the human part, the part that was always the point, is finally all that’s left.

That process is already underway. And I hope your awareness is catching up quickly. The changing nature of work is much deeper than layoffs. You have a couple quarters to lean in on your terms.

The best time to plant a tree is 20 years ago. The second best time is now.

If you’re a coach working in this environment and want to see where your specific practice sits in these four chains before the pivotal Q2 window closes, I’m running a limited pilot of a self-administered diagnostic. You’d get the same structural map I build before a first session. Details here: form.typeform.com/to/QlPtNfbL

Every now and then I talk about coaching. It’s what I do, after all.

Today’s one of those days. I’ve started to see some signals that look very different than the ones I’m seeing on LinkedIn and hearing from coaches.

The International Coaching Federation released its 2025 data last month. Global coaching revenue: $5.34 billion. Up 17% from 2023. Growth ma…

Read more

This article is an update to my original projection in August 2025 that anticipated significant AI-driven job displacement beginning in Q1 2026. The consensus at the time tended to be 3-4 years. Last month I revisited the analysis six months later, and there were a few surprises and a lot of confirmation.

I committed then to publishing an updated analysis each month. This month’s signals moved substantially. The purpose of sharing this is to show the system moving beneath the numbers. Unemployment, GDP… those are lagging indicators. I developed the Hari projections to track leading indicators and their vectors, establishing signal in the noise. The point was to be roughly right rather than precisely wrong.

The news is short-term grim. I’m not sharing these projections for shock value, though. My intention is to show the direction of large trends and the nuance within them. I identify thresholds beyond which trends compound. I recommend decision points for companies, investors, and individuals throughout the system. My aim is to provide you a different lens that will help you anticipate and make your best decisions.

There’s a sentence buried on page 7 of the February Employment Situation Report that almost nobody will read. It says: “The number of discouraged workers, a subset of the marginally attached who believed that no jobs were available for them, decreased by 128,000 to 464,000 in February.”

That sounds like good news. Fewer people discouraged. Progress.

Except labor force participation fell in the same report. It dropped from 62.1% to 62.0%. If 128,000 discouraged workers became less discouraged, you’d expect them to start looking for work again. The participation rate should have gone up. It went the other direction.

So where did those 128,000 people go?

They didn’t re-enter the labor force. They reclassified. BLS tracks several buckets of people who aren’t working and aren’t looking: “discouraged” is one. “Other reasons” is another. When a discouraged worker tells the survey they stopped looking because of childcare, or health, or because they’ve given up in a way that doesn’t fit the specific “discouraged” definition, they move from one bucket to another. They vanish from the discouraged count. They don’t appear in the unemployment rate. They just disappear from the numbers that get reported on the news.

That’s the pattern I want to talk about this month. The gap between what the numbers say and what’s actually happening underneath them.

The print

On March 7th, the Bureau of Labor Statistics reported that the economy lost 92,000 jobs in February.

The consensus estimate was a gain of 59,000. The miss was 151,000 jobs wide. December was revised downward again. Combined revisions for December and January subtracted another 69,000 from what we thought we knew.

Unemployment rose to 4.4%. The average duration of unemployment hit 25.7 weeks, the longest in over four years. Long-term unemployment (people out of work for 27 weeks or more) climbed to 1.9 million, now representing one in four unemployed Americans.

There’s a confound worth naming: a Kaiser Permanente strike sidelined roughly 30,000 healthcare workers during the survey week. Strip that out and the print is still negative at -64,000. The information services sector has now lost an average of 5,000 jobs per month for twelve consecutive months. Federal employment fell by 10,000 in a single month.

These are the numbers that made the headline. They are bad. But the more interesting data is in the places almost nobody looks.

The numbers underneath

Start with JOLTS, the Job Openings and Labor Turnover Survey, released March 13th.

Total openings: 6.9 million, down from 7.4 million a year ago. That puts the ratio of open positions to unemployed workers at 0.94. Below 1.0. For the first time since the pandemic recovery, there are fewer available jobs in the American economy than people looking for them.

That number, by itself, would be significant. But pair it with this: the quits rate has been frozen at 2.0% for seven consecutive months. That’s a decade low. Workers are afraid to leave their current positions because they can see what the job market looks like from the inside.

Now add the ADP private payroll number: +63,000 in February, the weakest since mid-2025. And the Challenger data: January 2026 saw 108,435 announced job cuts, the highest January since 2009, up 118% year over year. February added another 48,307. Tech sector cuts alone are running 51% ahead of last year’s pace.

None of these series makes the nightly news by itself. Together, they tell a story the headline unemployment rate doesn’t capture. The labor market is contracting on multiple fronts simultaneously. Openings are falling. Hiring is slowing. Workers are frozen. And the people losing work are, in large numbers, simply not showing up in the official statistics.

Fortune reported this month that nearly 75% of workers displaced by AI-related actions don’t file for unemployment insurance. They have savings. They feel stigma. They think they’ll find something quickly. They don’t. But they also don’t become a number in the system that tracks these things. If that figure is even directionally correct, the real displacement is roughly four times what the claims data suggests.

That’s the gap. The headline rate is 4.4%. The actual situation is significantly worse. And the measurement tools the country relies on were designed for a different kind of recession, one that announces itself with plant closures and mass layoff events and lines at the unemployment office. This one doesn’t do that. This one is quieter.

The profitable companies

Here’s where it gets counterintuitive.

In a normal downturn, companies cut because they’re losing money. Revenue falls, margins compress, headcount follows. The companies doing the cutting are struggling. The logic is survival.

That’s not what’s happening.

Oracle announced between 20,000 and 30,000 cuts in March. Their revenue is rising. They are spending aggressively on AI infrastructure, carrying over $100 billion in debt to fund the buildout, and restructuring their workforce around AI agents that now handle tasks previously done by database administrators.

ASML, the Dutch semiconductor equipment maker, posted 16% sales growth and 19% profit growth. They cut 1,700 people.

Mastercard reported strong profits. They cut 1,400.

Block, the payments company Jack Dorsey runs, cut 4,000 of its roughly 10,000 employees. Nearly half. Block is on track to generate approximately $2 million in gross profit per remaining employee this year, double the 2025 figure. Dorsey told employees the company is moving toward “smaller, highly talented teams using AI to automate more work.” Then he said something that deserves attention: “I think most companies are late. Within the next year, I believe the majority of companies will reach the same conclusion and make similar structural changes.”

That’s the CEO of a publicly traded company telling you, on the record, that what he just did to half his workforce is coming to most companies within twelve months.

These are not distressed organizations. They are profitable, growing businesses that have decided the optimal number of employees is significantly lower than what they currently have. The gap between revenue growth and headcount is being filled by AI. And the metric that’s emerging as the new benchmark (profit per employee, or revenue per employee) tells you exactly where this goes: up and to the right on productivity, down on people.

The places you wouldn’t expect

The tech layoffs get the coverage. They should. But the data this month surfaced signals in places the typical AI-and-jobs conversation doesn’t reach.

Healthcare companies announced 17,107 cuts in January 2026. That’s nearly six times the January 2025 figure. CVS, Trinity Health, Alameda Health System. The sector people point to as recession-proof and automation-resistant is showing structural stress. Part of it is federal funding changes and reimbursement compression. Part of it is AI-driven administrative automation starting to hit back-office operations. The causes compound.

Dow Inc. led 4,701 cuts in chemical manufacturing, explicitly citing a shift to AI and automation. Chemical manufacturing. Not software. Not consulting. Chemicals.

Entry-level hiring has collapsed in ways the headline data doesn’t capture, and the collapse has a specific shape that’s worth understanding. Entry-level tech job postings fell 67% between 2023 and 2024, and the trend has continued. But here’s the part that matters: job postings labeled “entry-level software engineer” actually increased by 47%. The postings exist. The hiring into those roles dropped 73%. Companies are listing positions they have little intention of filling. The posts serve branding, pipeline building, or compliance purposes. The actual entry-level jobs are disappearing behind a wall of phantom listings.

The numbers on where this goes are stark. 36% of companies surveyed say they will stop hiring entry-level employees entirely by the end of this year. 47% expect to reach that point by 2027.

IBM reported something quietly that stuck with me. Their U.S. voluntary attrition rate fell below 2%, down from a historical norm around 7%. That’s the lowest in 30 years. People aren’t quitting because the outside market feels worse than staying. Everyone is holding still. And when an organization’s voluntary attrition rate falls that far below baseline, it creates a paradox: the company can’t execute a quiet, attrition-based workforce reduction because nobody is leaving. When cuts do come, they hit a workforce that has no recent experience looking for work, entering the weakest job market in years.

The salary split

Compensation data confirms what the Dallas Fed identified last month with their experience-premium research.

Senior software developer base compensation dropped 10% year over year. Senior SQL developers fell 7%. These are experienced generalist roles. People with solid skills and real tenure, watching their market value decline.

Meanwhile, LLM specialists (the people building and fine-tuning AI systems) average $209,000. The overall tech salary increase for 2026 is somewhere between 1.6% and 3.5%, depending on whose survey you read, well below inflation.

The market is repricing technical work in real time. If your skills are specific to building and deploying AI, you are more valuable than you were a year ago. If your skills are general-purpose technical work, even at a senior level, you are less valuable. The gap is widening monthly.

This matches the Dallas Fed’s framework precisely. For occupations with high experience premiums (where judgment, context, and institutional knowledge differentiate experienced workers from entry-level), AI boosts wages. For occupations with low experience premiums (where the work is defined well enough that experience doesn’t matter much), AI compresses wages across the board. The distinguishing variable is tacit knowledge: the things you know how to do that are hard to write down. If your value is in things that can be specified and measured, the specification is what gets automated.

What’s happening to the chain

I’ve been tracking what I call the primary causal chain since last August. It maps six links that connect AI capital investment to measurable employment effects:

Capital Investment → Infrastructure Build → Executive Strategy → Hiring Demand Shift → Payroll Effects → Unemployment.

In last month’s update, the first four links were active and the fifth was emerging. This month, two links advanced.

Payroll Effects moved from Emerging to Active. The evidence: a negative NFP print, continued downward revisions, and the ADP confirmation of private payroll weakness.

Unemployment moved from Not Yet to Emerging. The evidence: 4.4% rate, long-term unemployed climbing to 1.9 million, average duration at 25.7 weeks.

Five of six links are now active or emerging. The composite signal score across the six indicator categories we track (JOLTS, payroll, job postings, claims, layoffs, and executive narrative) moved from 4.25 to 4.57 out of 5.0. Five of six components strengthened. The only unchanged component is claims data, which remains stable at 213,000 initial claims weekly. Claims are the most lagging indicator in the set. They’re the last shoe to drop.

Four of five escalation conditions are met: NFP below 150K for 3+ months, JOLTS below 7.5M, a major employer announcing an AI-driven reduction exceeding 5,000 positions, and now the unemployment rate exceeding 4.3%. The sole remaining condition (initial claims exceeding 300K for four consecutive weeks) may be structurally incapable of triggering in this particular displacement, given the 75% non-filing rate and the dominance of attrition over formal layoffs.

We projected last August that the displacement window would open in Q1 to Q3 of 2026. We are inside that window. The chain is tracking ahead of the base case.

The narrative is catching up

There’s a qualitative shift happening in how institutions and media talk about this.

The IMF chief warned in January of an “AI tsunami coming for young workers.” Goldman Sachs published a forecast expecting unemployment to reach 4.5% by year end. The Dallas Fed released research in February showing AI simultaneously aiding experienced workers and displacing entry-level ones. Anthropic’s own research team published findings suggesting the possibility of a “Great Recession for white-collar workers.” Andrej Karpathy, formerly of OpenAI, published a treemap on March 15th scoring 342 occupations for AI exposure that went viral.

Worker anxiety has tracked the shift. The share of Americans who fear losing their job to AI rose from 28% to 40% in the past year. That’s a 43% increase in a fear metric, which is itself a leading indicator: when enough people believe displacement is real, their behavior changes. They stop spending. They freeze in their current role. They pull back from risk. Fear of job loss produces many of the same economic effects as actual job loss, sometimes faster.

AI adoption keeps climbing. Pew found 21% of U.S. workers using AI on the job by late 2025, up significantly from the prior year. Gallup estimates about 30%. Broader surveys show chatbot usage approaching 37%. The tools are spreading through the workforce while the workforce is contracting. Both of those things can be true at once, and they are.

And the capital keeps flowing in. The four largest hyperscalers (Amazon, Alphabet, Meta, Microsoft) have committed roughly $700 billion in capital expenditure for 2026, with the majority directed at AI infrastructure. Amazon alone is spending over $200 billion. Alphabet plans $175 to $185 billion, more than double their 2025 outlay. These companies are projecting negative free cash flow to make these bets. They are not hedging. They are going all in on a world where AI does more and people do less.

What decisions to make, and when

If you’re running a company, the signal from this month’s data is specific enough to act on.

The window for proactive workforce restructuring through attrition is closing. IBM’s experience shows why: when workers stop leaving voluntarily, you lose the option to reshape your team quietly. The companies that are adjusting now (Block, Oracle, Atlassian) are doing it while they have the choice. The companies that wait will do it under more pressure and less control. Dorsey’s prediction (”within the next year, the majority of companies will reach the same conclusion”) is aggressive, but the direction is right.

The metric to watch is profit per employee (or revenue per employee, depending on your business model). Block’s $2 million target is the clearest public benchmark. If your company is below that trajectory for your sector, you’re either going to get there by restructuring proactively, or the market is going to force you there reactively. Proactive is cheaper. Proactive also means you choose which roles to preserve and which to automate, rather than cutting across the board when margins demand it.

For investors, the sectors most exposed to this employment contraction (consumer discretionary, commercial real estate, staffing firms, mid-market SaaS) haven’t fully priced in the trajectory. The February NFP print was a surprise to consensus. The March print will either confirm the trend or give the market a temporary reprieve. If March comes in negative again, the re-pricing gets aggressive.

For individuals, the salary data tells you exactly where to invest your time. Generalist technical skills are depreciating. The experience premium, the tacit knowledge that AI can’t replicate, is the asset to build. If you’re early in your career and the entry-level pipeline is closing (and the data says it is), you need a different path to building that tacit knowledge. Apprenticeship, freelance project work, contributing to open-source projects, or getting hired at a smaller company where you’ll learn more roles: those are the routes that still exist. The traditional path (apply to a junior role at a large company, get trained, move up) is disappearing.

The decision timing depends on what you’re watching. The next JOLTS report drops March 31. The March Employment Situation lands April 3. ADP reports April 1. Challenger in early April. If those prints confirm the February trajectory (JOLTS below 7M, NFP negative or near zero, Challenger running hot), the structural picture solidifies further. The time to act is before confirmation, because confirmation is when the crowd arrives.

The honest caveats

Three things complicate this picture, and I want to name them.

First, “AI washing.” Some companies are attributing layoffs to AI that are really driven by cost-cutting, margin pressure, or strategic restructuring that would have happened regardless. OpenAI’s Sam Altman has publicly confirmed that this is happening. It muddies the signal. Whether the cuts are caused by AI or merely branded as AI-driven, the jobs are gone either way. But for the purpose of tracking the AI displacement chain specifically, some portion of the attribution is inflated.

Second, DOGE. The federal workforce has shrunk by over 330,000 from its October 2024 peak. That’s a real and significant contraction, but it’s driven by policy decisions, not by the AI displacement mechanism the model tracks. DOGE effects show up in the BLS data and accelerate the headline deterioration, but they’re an external force acting on the system, not a product of the causal chain. Some agencies have started rehiring, which may partially reverse these losses.

Third, the Kaiser Permanente strike distorted February’s healthcare numbers. Roughly 30,000 workers were sidelined during the survey week. That’s temporary. But the 17,107 healthcare layoffs from January are structural and have nothing to do with the strike.

Strip all three confounds out, and the underlying signal is still strongly negative, still directionally consistent with what the model projected seven months ago.

What I’m watching for April

A second consecutive negative NFP would be the clearest signal yet that the “Payroll Effects” link in the chain is fully active. That’s the April 3 report.

Q1 earnings season begins in April. Listen for “efficiency,” “doing more with less,” and headcount guidance. Block’s profit-per-employee language may become the template other CEOs adopt. Earnings call language has been the most reliable leading indicator in this chain from the beginning.

I’m tracking whether the entry-level hiring freeze spreads beyond tech into professional services and healthcare. Whether the JOLTS ratio stays below 1.0. Whether DOGE rehiring accelerates (that would be a counter-signal). Whether Goldman’s year-end 4.5% unemployment forecast holds or gets revised upward.

And I’m watching the discouraged workers number. That quiet sentence on page 7 of the BLS report. The one that looks like good news until you read what’s underneath it.

This analysis uses the Causal Chain Projection Framework (Hari v2.2) and draws on the February 2026 Employment Situation Report (BLS), JOLTS January 2026, ADP National Employment Report, Challenger Gray & Christmas job cut data, Dallas Fed research on AI exposure and experience premiums, and current reporting from Fortune, Bloomberg, CNBC, and CNN Business.

The companion Hari briefing for this month’s data, with full indicator tables and component scores, is available to subscribers.

Last week I posted “Roughly Right is Better Than Precisely Wrong,” which introduced a methodology I’ve been building for the past nine months. It maps causal chains, tracks leading indicators, and produces threshold-triggered decision points for leaders and investors operating in fast-moving environments.

A few people reached out to say it was useful. A …

Read more

Nine months ago, I started tracking something.

I’d been watching the AI employment conversation for a while. The consensus timeline for when job displacement would really bite was “3 to 5 years.” That number showed up in McKinsey reports, analyst decks, and dinner party conversations with roughly equal confidence and roughly equal rigor. It felt anchored to what people could comfortably absorb. And I kept noticing signals that suggested the actual timeline was shorter. Quite a bit shorter.

So I did something I’ve done my whole career, but had never bothered to formalize.

I traced the causal chain backward. What has to be true before the outcome can materialize? What are the preconditions, in sequence? And at each link in that sequence, what observable signals would tell you it’s already happening?

I’ve been thinking this way for thirty years. It’s the way I’ve assessed markets, evaluated companies, stress-tested strategies. But it always lived in my head, and in the hundreds of books sitting on my shelves. The AI employment question forced me to write it down.

What came out of that process was published last August as “The Last Normal Year.” The projection: AI-driven employment displacement would become measurable in 6 months, reaching a crescendo within 18 months. That put it roughly two years ahead of the consensus. Six months later, I published the follow-up, “AI Is Coming for Your Job After All.” Five of seven chain links had activated. The JOLTS job openings reading had crossed below the threshold I’d been watching. Challenger layoff data had exploded past 108,000 announcements in a single month. The projection was tracking.

I got some things wrong, too. I’d assumed the displacement would be linear. Steady, accumulating, one-directional. It wasn’t. Companies cut, regretted it, rehired selectively, cut smarter. The actual shape was oscillatory. That miss taught me something about the framework I was building. It needed a way to distinguish between systems that spiral and systems that settle toward a new equilibrium.

I built one.

And then I kept going. Over the past nine months, I’ve taken the thinking that lived in my head and turned it into a formal, repeatable methodology. I’ve stress-tested it across multiple domains. I’ve run it with a handful of venture capitalists, a private equity firms, and several company founders. Their responses have been very positive. And the reason keeps coming back to the same thing.

Which I’ll get to. But first, I want to talk about the fog.

What leaders actually deal with

You have dashboards. You have analysts. You have Gartner subscriptions and board decks and quarterly reviews. You’re still making your most consequential decisions in a fog.

The dashboards tell you what happened. Lagging indicators, mostly, arriving weeks or months after the underlying dynamics already shifted. The analysts give you consensus estimates, which are by definition what everyone else already believes. And the decisions that actually matter, the ones about where to deploy capital and when to restructure and whether the assumptions in your thesis still hold, those need to be made before the data is obvious.

By the time you get confirmation, the market has already moved. Competitors too.

Most leaders know this. Most teams know this. They spend enormous amounts of money trying to cut the fog. And most of what they buy gives them a better-organized view of the past. Which is helpful, but it’s the wrong problem. What you actually need is directional accuracy about the future. Good enough to know when to move. Delivered fast enough that you still have options when the signal lands.

My friend Nate Chaffetz puts it simply: roughly right is really the only kind of right that matters.

He’s spot on. And “roughly right” is what the people I’ve worked with on this keep telling me they’ve been looking for. Two things that are missing from most predictive intelligence, from most dashboards, from most strategy work: what action to take, and when to take it.

What “roughly right” looks like in practice

The methodology does four things.

It maps the causal chain. For any anticipated outcome, it identifies the sequential preconditions that must be true before that outcome can materialize. Each link is observable. Each has leading indicators you can track before the consensus catches up.

It estimates time delays. Ranges backed by historical analogues and current velocity data. Named assumptions you can update as evidence arrives. Not vibes. Not “3 to 5 years.”

It tests from multiple angles. A causal chain is a hypothesis. Hypotheses need to hold up when you push on them. Every projection gets checked against macro and micro perspectives, systems thinking, cross-domain convergence, and counter-indicators that might reshape or invalidate the chain. Identifying the chain and setting thresholds is an interesting way to frame a problem. The multi-perspective testing is what makes the answer worth having.

It produces decisions, not reports. Every projection generates specific threshold-triggered decision points. When indicator X crosses level Y, that’s the signal to execute response Z. And the cost of waiting is mapped explicitly: what each option costs now versus what it costs if you delay six months.

I’ve validated this across AI employment displacement, US institutional stability, sector-level market analysis, and a growing number of company-specific engagements. The logic is the same underneath. Outcomes arrive through sequences of preconditions. Track the preconditions, and you see the outcome forming before it becomes consensus.

Now. Remember the thing I said I’d get to? The thing the VCs and founders and the PE firms kept responding to?

Fifteen minutes with the SaaSpocalypse

It was speed.

In the first week of February 2026, roughly $285 billion in software market value disappeared in a single session. Over the following weeks, the damage widened to nearly $2 trillion. Atlassian fell 35% after reporting the first enterprise seat-count decline in its history. Salesforce dropped 28% even as revenue kept growing. The iShares Software ETF fell 22% year-to-date. The steepest software selloff since the 2022 rate hike cycle.

The trigger: a new generation of agentic AI platforms showed that autonomous agents could perform the same end-to-end workflows that used to require teams of humans running SaaS tools. If one AI agent does the work of five employees, the need for five software licenses vanishes with them. The per-seat pricing model that built a $600 billion industry is suddenly exposed.

A little balance here: Klarna fired a bunch of their people to use AI and moved too soon, leading to that “regret rehire” oscillation I mentioned before. I believe they’ll make cuts again. Jack Dorsey at Block (parent of Square and Cash app) is leading this charge, laying off nearly half of a workforce generating $2 million in profit per employee.

Every VC with SaaS in their portfolio, every PE firm that bought a software company at 12x revenue, every SaaS founder watching the crater form, they’re all inside the same question. When does this stabilize? Who survives?

The causal chain took about fifteen minutes to map. Here’s what it looks like.

Link 1 (already happened): AI capability reaches functional parity with specific SaaS task categories. Coding assistants generating 40-60% of routine code. Agents handling CRM entries, support tickets, document processing at error rates below human benchmarks.

Link 2 (the February catalyst): Agentic platforms demonstrate end-to-end workflow replacement. The market reaction was the recognition that capability had crossed a threshold. “AI can assist with tasks” became “AI can replace workflows.” That’s a different sentence. A very expensive one.

Link 3 (watch this now): Enterprise procurement shifts from per-seat SaaS to AI-native alternatives. The leading indicators: CFO spending surveys, enterprise AI procurement data, and the net-new customer acquisition rate at incumbent SaaS companies. Salesforce dropped 28% because new customer acquisition slowed. The market is pricing in what procurement data will confirm in two quarters.

Link 4 (next up): SaaS companies that can’t embed AI see churn accelerate. The threshold: net revenue retention drops below 100% for three or more consecutive quarters across a basket of mid-cap SaaS names. Atlassian’s seat-count decline is the first structural signal.

Link 5 (6-12 months out): Consolidation wave. Acqui-hires, fire sales, shutdowns. M&A deal flow in SaaS exceeding 2x the trailing three-year average.

Link 6 (12-24 months out): New equilibrium. Survivors are AI-embedded. Pricing shifts from per-seat to outcome-based or consumption-based models. The sector re-rates at new multiples.

One more question the framework forces you to answer: is this a loop that accelerates until something breaks, or one that oscillates toward a new settling point?

The evidence says it oscillates*. Some SaaS categories are far more exposed than others. Infrastructure, data management, and security have high switching costs and less AI substitutability. CRM, project management, and support tools are directly in the path. Companies will overcorrect on the sell side, some SaaS names will rebound, and the settling point probably lands 30-45% below 2024 highs in aggregate, distributed very unevenly.

That took fifteen minutes. It’s a structural map of a $2 trillion market event. A full engagement adds the threshold dashboard, the decision-trigger table, and the strategic response map with time-dependent costing. But even at this level, it’s more structural clarity than most SaaS investors had on February 4th.

Who this is for

Four groups get massive benefit from this “roughly right” model:

Venture capitalists who want another lens on the investments they’re considering. The framework asks whether the market the company is entering is actually going where the pitch deck says. When the chain shows the thesis depends on a precondition that hasn’t activated, that changes how you size the bet.

Private equity firms asking whether their portfolio companies can hit their targets within the standard holding period. A causal chain showing the market shifting midway through a five-year hold is the difference between an exit and a write-down.

Company founders who need an ongoing stress test for their assumptions. Your business plan is a hypothesis about the future. The framework tracks whether the structural conditions it depends on are still trending in your direction, and how much lead time you have to adapt if they aren’t. “Lead time” often means “runway”.

Non-profits and mission-driven organizations working in rapidly shifting environments. Education, public health, community development, social services. These fields are being reshaped by the same forces hitting the private sector, and the organizations doing the work rarely have access to the tools that funded companies use for environmental scanning.

What this is (and what it isn’t)

This is not a dashboard tool or a connector to your data warehouse. Not a BI platform.

The traditional route to that kind of intelligence involves weeks of setup, teams of analysts, data connectors chasing sources, latent data arriving on a delay, and models that are outdated by the time they reach the board. There’s a place for that infrastructure. That’s a different offering and a different problem.

What we’re doing is a rapidly rolling check against leading indicators. The signals that precede the outcomes everyone else waits to confirm. Because the methodology doesn’t depend on heavy data architecture, the turnaround is measured in hours. A full briefing and threshold dashboard for a specific question can be produced in a single working session—even half an hour before a board meeting or investor pitch. No procurement. No multi-month implementation. No six-figure contract just to get to the starting line.

The methodology consolidates three decades of experience, pattern recognition, and cross-domain reading into a structured process that produces consistent, testable, updatable results. It’s not a flashy AI product. It’s a way of thinking that finally has a formal structure around it.

And it works best as a recurring signal check. Monthly or quarterly, depending on how fast your domain is moving. Each refresh shows what changed, whether any thresholds shifted, and whether decision windows opened or closed. The more data you provide, the better the signal.

This is how “roughly right” builds over time. Not a single prediction, but systematic updates that keep you positioned correctly as the structure evolves.

Meet Hari

I’ve been building this for nine months. Three rounds of stress testing, a public prediction that tracked, and a growing set of engagements that have sharpened it considerably. Time to give it a name.

Hari.

Two references, both intentional.

In Asimov’s Foundation series, Hari Seldon created psychohistory, the science of predicting the broad sweep of civilizational change without predicting individual events. He couldn’t tell you which crisis would come. He could tell you a crisis was structurally inevitable, and roughly when. That’s the aspiration. Structural clarity, not omniscience.

Hari is also a Sanskrit name and epithet for Vishnu. The one who takes away darkness and illusion. The preserver. A remover of fog.

Both namesakes share the same conviction. The broad pattern is knowable even when the specifics aren’t. Structural forces produce tendencies you can track. And seeing clearly, roughly right, is enough to change your position before the wave arrives.

Intellectual honesty is a big deal to me. Every projection Hari produces gets logged, tracked, and publicly calibrated. What I projected, what actually happened, what I learned. The track record compounds over time. And it’s the track record, not the pitch, that earns trust.

If you’re making decisions in the fog and you’d like a structural map of what’s forming around you, I’d love to show you what Hari sees in your world.

Note: I used the word oscillate in this piece several times, and it’s helpful to know what it means. In short time frames there are effects, responses to effects, responses to those responses, etc. When there’s a bias toward AI being overhyped, for example, the “regret rehires” (bringing people back because AI wasn’t yet as capable as the company thought) can seem like a confirming signal that AI didn’t live up to the promise and won’t be a thing the way many thought it would. The sequence of events looks like layoffs driven by AI —> dips in performance and customer satisfaction —> lost revenue —> rehire to get performance back to expectations. Projections suggest that there are a few more links in that sequence and that capabilities improve, companies get smarter about workflows, and cuts happen again a few months later. Today’s news can seem definitive, but a lot of it is merely oscillation in a trendline that continues in a direction.

Quick note before we begin: I’ve begun hosting Q&A sessions on Thursdays at 10am MT to go deeper on what I write about here—primarily business, AI, and consciousness. Recent discussions have touched on AI-driven job displacement, reconfiguring companies for a new reality, and signals I’m tracking in my projection models. Link to register is here.

Read more

In 1855, Daniel McCallum had a problem. As General Superintendent of the New York & Erie Railroad, he was responsible for over 5,000 workers spread across hundreds of miles of track. He needed a way to see who was responsible for what, and how accountability flowed across the system. So he commissioned the world’s first organizational chart, a sprawling…

Read more