Issue 007 - There Is No NAICS Code for Artificial Intelligence

On August 27, the Bureau of Labor Statistics published a number that has never been this high in thirty-two years of comparable readings. As far as I can tell, almost nobody wrote about it. Here is what it says, and here is where I think the people behind it are going.

The Displaced Worker Survey comes out every two years. It asks people who lost a job why they lost it and sorts the answers into three buckets: the plant or company closed or moved, there was not enough work, or the position was abolished.

In the edition published on August 27, covering January 2023 through December 2025, 44.4% of long-tenured displaced workers said their position or shift was abolished.

That is not a drift. Across the sixteen surveys from 1994 through 2024, this number never once left a band between 26.6% and 37.5%. This one broke the top of that band by almost seven points.

Long-tenured means three years or more with the same employer. There were 3.3 million of them, up 746,000 from the survey before.

Last time was different. In the fall of 2008 I had just taken over SMB marketing for the Americas at HP, and less than a month later Lehman Brothers went bankrupt. I have told you that story already. What I had not done until recently was go back and look at what this particular survey said about those years.

The January 2010 edition covered 2007 through 2009 — the Econoquake, start to finish. Insufficient work spiked to 42.8%. Position or shift abolished fell to 26.6%, the lowest reading in the entire series.

In 2026 those two numbers have traded places. Insufficient work is down at 22.9%, near the floor. Abolished positions set the record.

That distinction is the whole thesis of this thread. In July 2026 I gave the people on the other side of that line a name: The Artificial Entrepreneur. It takes in two populations — experienced professionals pushed off the career ladder, and newly credentialed people blocked from stepping onto it. It is a prediction, not a finding, and I am going to score it in public.

A recession takes away the work. What is happening now takes away the job while the work keeps getting done — by fewer people, by software, or by somebody who is not in the building.

I want to be careful here, and I am going to show you later in this issue exactly what being careful costs me. Nothing in that survey says the word artificial intelligence. What it says is that the shape of this displacement does not look like a recession. That is a smaller claim than the one you may want me to make, but it is the one the data supports.

Management, professional and related occupations accounted for 1,545,000 of the 3,324,000 long-tenured displaced. Forty-six percent. In the prior survey that group was 1,098,000. Total displacement rose 29%. That group rose 41%. White-collar work is taking a bigger share of displacement than it did two years ago.

Now look at what happened to them afterward. Share of long-tenured displaced workers who were employed again as of January 2026:

Remember that 57.3%. I am coming back to it.

And one more number, which I think is the most underreported line in the whole release. Of the people who lost a full-time job and found another one, about 49% are earning as much or more than they did before. In the survey two years ago, that figure was 62%.

Reemployment barely moved — 66.1% of long-tenured displaced workers were working again in January 2026, against 65.7% two years earlier. The value of reemployment, however, fell 13 points in two years.

The Displaced Worker Survey only counts people who had a job to lose. It cannot see the other half of this.

In August, the Stanford Digital Economy Lab updated its study of ADP payroll records through June 2026. Overall, it found no broad displacement: employment in its data grew. But for workers aged 22 to 25 in the occupations most exposed to AI, employment is 19% below where it would be had it kept pace with the same age group in less-exposed jobs since late 2022. The gap came from not hiring, rather than firing. The jobs that are missing were never opened.

The authors call their results descriptive, not causal, and about half the gap disappears once they control for education. They also note that new, small businesses are poorly represented in their data. The payroll data stops where the cohort begins.

Every time I walk somebody through this, the next question is the same one. Which industry absorbs them? People want the answer to be the AI industry.

There is no AI industry. I mean that literally, as a matter of federal record.

The North American Industry Classification System — the thing every government statistic, every prospect list you have ever bought, and every market sizing you have ever been handed is built on top of — has no code for artificial intelligence. None. Not at the sector level, not at the six-digit level, nowhere.

It goes further than that. This July, the federal committee that revises NAICS published its proposed changes for 2027. It said it had received suggestions for new industries including artificial intelligence, and it declined to create one, on the grounds that either the resulting industry would be too small in the United States or the proposal did not meet the production-oriented criterion.

The only place AI appears in the entire system is one line in the index — artificial intelligence research and development laboratories or services — which points at the code for physical and engineering research labs.

So if you are waiting for a government series to light up and show you a swarm of AI companies being born, you can stop waiting. The category does not exist, and as of this July it is not being built.

Here is the rule that unlocked this for me, and I wish I had understood it fifteen years ago. NAICS classifies a business by what it sells. Never by what it uses.

Take one example. Somebody builds an AI receptionist for dental offices — answers the phone, books the appointment, handles the reschedule, never sleeps. One founder. One product. One technology. Where does that business get classified?

Same founder. Same technology. Five different codes. The thing that decides it is the shape of the invoice.

Which means which industries will they be in is the wrong question, and I have been asking it wrong. The right question is: which functions inside companies come apart first? Because the function that gets unbundled is what determines the code.

Count the ones above beginning with 5614. Telephone answering. Document preparation. Contact centers. All three sit inside NAICS 5614, Business Support Services. Recruiting sits next door in 5613, Employment Services.

Those are the two subsectors that have been falling off a cliff. Employment services is down 18% and business support services is down 20% since 2022.

Read that again slowly, because it took me a while.

Same code. One line goes down, the other goes up, and the published series reports the difference.

An analyst looking at 5614 today sees a business in decline and writes that story. What may actually be happening is a business changing form — employer payroll walking out the door, one-person firms walking in, and a headline number that describes neither movement.

I have been calling this The Great Undercount since Issue 003. This is a new version of it and I think it is the sharpest one I have found. It is also testable, which matters more. You would look for employer employment falling and nonemployer counts rising inside the same code at the same time. Both of those series exist. I am building it.

I said I would show you what being careful costs me. Here it is.

The story everybody wants told is the million-dollar solopreneur. One person, no employees, seven figures, all of it powered by AI. I want that story to be true. So I pulled the Census receipts data — straight from the source, not from somebody else’s summary — and it does not yet say what the story says.

When I first drafted this section in September, the latest year on record was 2023, and it ran against me. Nonemployer businesses with more than $5 million in revenue fell from 965 to 905. The share clearing a million dollars slipped from 0.39% to 0.38%.

On September 29, Census released 2024, and both numbers turned. There were 1,051 nonemployers above $5 million, and 126,356 above $1 million — 0.40% of the entire universe. The million-dollar count grew 8% in a year when nonemployers as a whole grew 5%.

I am not claiming it. One year is not a trend, and the year before went the other way. Census size classes are fixed in nominal dollars, so inflation alone lifts some businesses over $1 million every year without anything about them changing. And the growth is not yet coming from where my cohort should show up. Professional services, where a displaced manager’s research or bookkeeping practice would be coded, grew 4%, slower than nonemployers overall. Couriers and messengers grew more than twice as fast as the total. Professional-services nonemployers above $5 million went from five to three.

One caution still holds. Census changed how it draws the nonemployer universe in 2022 and does not recommend reading these years as a time series. 2022 through 2024 all sit on the clean side of that break, and right now they give me one year for, one year against, and a composition that does not yet look like my cohort.

The next Displaced Worker Survey is fielded in January 2028 and will cover 2025 through 2027. That is the window this prediction lives or dies in, and I intend to be on the record before the data arrives, not after.

If you have been displaced from a white-collar role since 2023 and started something of your own, hit reply and tell me what your business actually sells. I am trying to learn what code you would land in.

Have data that conflicts with my read? Share it and I will respond and learn from you. Disagree with a conclusion? Tell me and we will have a civil discussion. Reply to this newsletter or write me at [email protected].

Every figure in this issue that I assembled rather than quoted, with the rows it was built from. A number summed or divided out of several published rows is analysis, not citation.