The Second Disassociation

What AI takes from people, and what it does not

I spent four hours on a golf course this week.

During those four hours I also cleaned up a large and badly structured dataset, worked through several rounds of research, and made a series of decisions about what the data needed to become.

I did none of that work. I directed it, from my phone, between shots.

Six years ago I wrote that the pandemic was accelerating a separation between activities and the places where they happen. I called it the Great Disassociation and I meant it literally: banking leaving the bank, shopping leaving the store, work leaving the office.

What did not occur to me was that the same separation would run again on a different axis. Not the activity leaving the place. The activity leaving the person.

What I said in 2020

A two-column framework, published twice, still checkable

In August 2020 I published a piece in Propmodo arguing that technology had flattened the enterprise, the school, the shopping center, and therefore the city.

In November of that year the Mann Report ran a longer version of the argument in print, under the title that stuck: the Great Disassociation.

The useful part was not the diagnosis. Everyone could see that things had moved online.

The useful part was a two-column sort, published in the second half of the piece, which claimed that not all of those separations would behave the same way afterward.

One column held what I called the accelerated, expected, and natural disassociations: shopping leaving the store, banking leaving the bank. Pre-existing trends the pandemic pushed forward.

The other held the acute, unexpected, and unnatural. Work leaving the office. Dining leaving the restaurant. Entertainment leaving the theater. The workout leaving the gym. Teaching leaving the classroom. Forced by circumstance rather than chosen.

Accelerated, expected, natural	Acute, unexpected, unnatural
Shopping, from the store	Work, from the office
Banking, from the bank	Dining, from the restaurant
	Entertainment, from the theater
	The workout, from the gym
	Teaching, from the classroom

Then I named the mechanism underneath, in one sentence:

The fundamental difference between the two columns is one of transaction versus experience.

The claim that followed could be checked, and it came with a rough date attached.

Where an activity was a commoditized transaction, the separation would stick, because nobody misses a transaction. Where the presence was the point, people would come back, because the thing they were buying could not be delivered any other way.

The sort was not clever. Retailers had been separating transaction from experience for a decade, which is why malls kept adding restaurants. Banks had been at it since the ATM.

I did not discover the mechanism. I noticed it was about to run through seven categories at once, under duress, in eighteen months instead of thirty years.

Saying so matters, because it sets the terms for what follows. The framework is not a prediction engine. It is an old sorting instrument, and the interesting question is never whether it works but whether you have put a given thing in the right column.

Grading the columns

The transaction column emptied out. The experience column filled back up.

Six years is long enough to check.

Banking left and did not come back. The National Community Reinvestment Coalition found that branch closure rates roughly doubled, from 99 per month before the pandemic to 201 per month after. One analysis of FDIC data counts 2,442 branches lost across 2021 and 2022 alone.

The trend has since decelerated, with 584 net closures between 2024 and 2025 and a handful of banks now expanding again in the Southeast, but the level never recovered.

I had quoted an investment bank's estimate of up to 30% further branch consolidation. That was the right order of magnitude.

Restaurants came back, and the doom forecast I repeated was wrong by a factor of three. In 2020 I cited an estimate that up to one third of America's 660,000 restaurants faced permanent closure. A Washington Post analysis of Bureau of Labor Statistics data later put the actual figure at roughly 72,700 closures in 2020, about 11%. Not nothing, and worse than any normal year. But a third of the industry did not disappear.

What happened next is the stronger evidence. Datassential recorded 886 restaurant closures in April 2025, an 82% decline from January 2018 and the lowest monthly total in their data. First-year failure rates, which doubled during the pandemic, fell to 0.9% in 2025, the lowest since at least 2018.

The gym call is the cleanest one in the set. In 2020 I listed home fitness as an outlying case and named Peloton, Mirror, FightCamp, NordicTrack, and Tonal as the companies capturing it. The framework said the workout would return to the gym, because a workout is an experience rather than a transaction.

The Health and Fitness Association reported that US commercial fitness facilities closed 2025 with nineteen consecutive quarters of visitation growth. Peloton peaked in 2021 at roughly $4 billion in annual revenue and 2.8 million connected fitness subscribers. By 2025 revenue was $2.4 billion, quarterly revenue was down more than half from its December 2021 high, and the stock had lost 97% of its value from the all-time peak. Insider Intelligence had forecast in 2022 that Peloton would shed nearly 900,000 US users in a single year as gyms reopened. (I own one of those bikes. I rode it constantly in 2020 and 2021, and I could not tell you the last time I did. I was in my own data and did not notice for years.)

Bowling shows the shape of the return. Bowlero ran 57% above pre-pandemic revenue on a trailing twelve-month basis by 2023 and reached roughly 30 million guests a year by 2025. Meanwhile the total number of US bowling centers has been falling about 2.1% annually, to roughly 2,484.

So the category grew while the venue count shrank, and food and beverage now runs 35 to 40% of Bowlero's revenue. That is what a healthy return looks like in an experience category: fewer rooms, each leaning harder into the part that cannot be delivered any other way.

Sport is the call that has aged best and drew the least attention at the time. I wrote in 2020 that most professional leagues had already separated the game from the fans decades earlier, through radio and television, and that the empty stadiums were therefore not a new disassociation but an old one running at its limit.

Six years later the leagues are healthy, the stadiums are full, and the broadcast business is larger than ever. Hold onto that one. It comes back.

People went back to the gym. They went back to restaurants. They went back to the bowling alley and the stadium. They did not go back to the bank.

The mechanism held in six of seven categories.

The seventh is where this gets interesting.

The movie theater

The category came back. Cinema did not, and I had them in one box.

Entertainment sat in my acute column as a single line: separate the entertainment from the theater. I supported it by noting that Hamilton, Frozen II, and other films had gone to streaming ahead of their windows, and that Universal had struck a seventeen-day theatrical agreement with AMC. Then I wrote that watching Hamilton on a streaming service pales next to seeing it performed live, and moved on satisfied.

Live performance did come back. Cinema did not.

The 2025 domestic box office came in around $8.6 billion against $11.4 billion in 2019.

Attendance is the harsher number: roughly 769 million tickets sold, down from more than 1.2 billion, a decline near 37%. North America has 5,691 fewer screens than it did before the pandemic, according to Omdia. Regal, Pacific Theatres, and Alamo Drafthouse went through Chapter 11, and Cineworld filed in 2022 listing about $5 billion in debt. Studios and exhibitors estimate that 15 to 20% of moviegoers stopped going and have not come back.

The industry has been flat for two years at roughly 78% of its pre-pandemic output. One analysis put it more plainly than I could: theaters are charging considerably more to considerably fewer people and calling it a recovery.

So I had one box holding two things that move in opposite directions.

Cinema and live performance diverge for a reason my two columns were too coarse to catch. A film watched at home is the same film. The substitute is excellent. Nothing about the story, the performances, or the photography degrades on the way to a living room, and what you give up is a bigger screen and a stranger's popcorn.

A bowling night does not survive that translation, and neither does a workout, a restaurant meal, or a play.

Which means the real axis was never quite transaction versus experience. It is how completely the remote version delivers the thing being bought. Banking substitutes perfectly, so it went and stayed gone. Cinema substitutes well, so it lost a third of its audience permanently. Bowling does not substitute at all, so it came back stronger and consolidated upmarket.

Transaction and experience were a good first approximation of that spectrum. They were not the spectrum.

And here is the part that should have caught my attention in 2020 and did not: I cited the evidence of my own error in the same paragraph where I made it. The seventeen-day window and the early streaming releases were the cinema business eroding in plain sight. I wrote them down as color, and then I sorted the whole category by its most vivid member, which was a Broadway musical.

The one I got wrong

Work was in the right framework and the wrong column

Work sat in my acute column, and I predicted that collaborative needs would pull professionals back into physical space, on the grounds that culture is created more fully when teams are together.

Half of that is true, and the half that is true did not save it. Culture and collaboration are genuinely experiential. But a large share of knowledge work is not collaboration. It is producing a deliverable at a desk, alone, and that is a transaction. I had put the whole category in one column when it needed splitting.

I have written about that failure at length elsewhere, along with what it taught me about the difference between a mechanism being right and a classification being right, so I will not repeat it here.

The short version is that the framework was sound and I misapplied it. That is a more useful kind of error than the framework being wrong, and a more dangerous one, because a sound framework misapplied still feels correct while you are applying it.

Notice that this is the same error as the movie theater, run twice.

In both cases the mechanism was sound. In both cases I took a category that contained two things behaving in opposite directions and treated it as one thing. In both cases I sorted the category by its most visible member: entertainment by the Broadway musical rather than the multiplex, work by the meeting rather than the eight hours of production surrounding it.

Two instances is not a coincidence. It is the failure mode of the instrument, and it is worth stating as a rule.

A sorting framework is only as good as the granularity of the things you feed it, and the natural granularity of business categories is far coarser than the granularity at which the mechanism actually operates.

Nobody sorts "work." They sort a hundred different activities that happen to share a payroll line.

What made it invisible for eighteen months is worth naming, because the same thing will make yours invisible.

A misclassification does not announce itself. Had the mechanism been wrong, the world would have contradicted me loudly: banks would have reopened branches, gyms would have stayed empty. Instead most categories behaved exactly as predicted, and a framework that keeps passing its tests is precisely what stops you from re-examining any single reading.

Meanwhile the one item it was failing on was the one nobody could check yet, because offices in 2021 were empty for reasons unrelated to whether anyone wanted to be in them.

Hold onto that, because it is about to matter again.

The one that ran before mine

Television left the schedule before anything left the building

There is an earlier run of this that I did not write about in 2020, and I should have, because it would have made the office error harder to commit.

Television never had a place problem. It was already at home. It had a time problem. Appointment viewing meant the program happened at eight on Thursday and the household arranged itself around it, and that arrangement was largely gone before the pandemic started.

Run the same sort. Consuming an episode of a drama separated from the clock completely and never came back. The experiential half never separated at all.

Live sport, awards shows, election night. Those are the only appointment television left, and they are appointments precisely because watching them a day later is watching a different thing.

Same mechanism, different axis, mostly finished before 2020 began.

Which makes sport the control case for the whole argument, because it appears in all three rounds and behaves consistently in each. It is the last thing tied to a time. It filled its stadiums back up when the rooms reopened. And it is irreducibly about these specific people doing this specific thing, which is why no one wants a simulated one.

Three axes, then, in order: time, then place, then person.

The second disassociation

The first separated activity from place. This one separates activity from person.

Return to the golf course. What happened there was not remote work.

Remote work in 2020 meant I was somewhere other than the office, sitting at a machine, doing the work with my own hands. The activity had left the place. It had not left me.

On the course, the activity left me.

I supplied judgment about what the data needed to become and what finished looked like. Everything between those two things happened without me, and I checked it the way you check work rather than the way you do it.

Before I extend any of this, the obvious objection deserves an answer, because it is the one I would raise.

You have just watched me be wrong about the largest category in my own framework, at roughly year one of a transition, and I am now about to make a forward call at roughly year three of a different one. That is not a great look and there is no version of it that becomes one.

What I can offer is narrower than a defense. The 2020 sort was published before the outcomes, in print, with the categories named and the mechanism stated, and six of seven categories behaved as it predicted. Two classifications inside those seven were misapplied, and both failed the same way.

The instrument has a record. I have a record of misapplying it in a specific and now well-documented way, which is at least a known failure mode rather than an unknown one.

Read what follows as a sort you can check rather than a prediction you have to trust, and check it the way I have just checked 2020.

The golf course episode is that same separation running again on the third axis. The time round asked whether an activity had to happen at a particular hour. The place round asked whether it had to happen in a particular room. This one asks whether it has to happen in a particular person.

And the mechanism transfers.

In 2020 the sorting question was transaction or experience. Does the presence of the customer in the room constitute part of what is being bought, or is the room just where the transaction is processed? Banking was processing. Dinner was not.

Point that same question at people instead of places and it reads like this. Is the output the point, with the person interchangeable? Or is whose judgment it is part of what is being bought?

Transactional cognitive work will disassociate from the person and it will not come back, for the same reason the bank branch did not come back. Nobody is nostalgic for the person who reconciled the spreadsheet. Judgment work will not disassociate, for the same reason the restaurant filled up again. The whole product is that a particular person with a particular record made the call.

Every AI planning conversation happening right now is an argument about which of those two piles a given task belongs in. Most of them are being conducted without anyone naming the piles.

There is one asymmetry between the two rounds that matters, and it cuts against the comfortable reading of all this.

When banking left the bank, the bank branch closed and the teller found other work. The disassociation was visible, it had a physical footprint, and it produced a number somebody had to report. Two thousand four hundred branches is a countable thing. Every party to it knew it was happening while it happened.

When an activity leaves a person, nothing closes. The person is still at the desk, still on the payroll, still attending the meeting. What changes is the proportion of their day that constitutes the thing they are actually being paid for, and that proportion is not reported anywhere, by anyone, in any system. There is no branch count for it.

Which means the second disassociation will be substantially further along than anyone's numbers say before anyone's numbers say anything. That was not true the first time. It is the single most important structural difference between the two rounds, and it is the reason the sorting question is worth an hour of your time now rather than when the data arrives.

Where the expensive errors will be

The office was misfiled because it looked like collaboration. Yours will be misfiled the same way.

My 2020 error was not sorting badly across the board. Six of seven categories went in the right column. The error was one misfile, in the largest and most consequential category, and it happened because work looked like an experience good from where I was sitting.

Why it looked that way is worth being precise about. I was thinking about the parts of work that are visible when you walk through an office: the meetings, the whiteboard, the conversation by someone's desk, the culture that a leader can feel. All of that is real and all of it is experiential. What I was not counting was the hours between those moments.

Those are most of the hours, and they are transactional.

I sorted the category by its most visible activity rather than by its most common one.

Misfiling by visibility is the error to expect in AI planning, and it will run in both directions. It will also run at the wrong granularity. The movie theater taught me that one twice.

Some work gets called judgment because it requires expertise, and expertise is not judgment. A task can demand years of training and still be a transaction, in the sense that any two qualified people produce the same answer and nobody cares which one did. Lease abstraction is like this, as is much of legal review, medical coding, and financial reconciliation. Difficult and transactional are not opposites.

Other work gets called transaction because the output looks like a document, when the document is only the receipt. A recommendation to a board is a document. What is bought is that a specific person, with a specific record of being right and wrong, put their name on it.

Take a pair from the same job. A regional property manager, eleven years in, responsible for fourteen assets, spends the first Tuesday of every month producing a variance report. Later that quarter she decides which of three capital projects gets funded. Same expertise, same data, same person, both arriving as documents. Sorted by difficulty or by title, both land in the same column.

Sorting them together is the mistake. Any two competent people working the same ledger produce the same variance report, and if it is wrong you fix it. The capital allocation is judgment, and not because it is harder. A year from now somebody will ask why that project and not the other one, and the answer has her name on it.

The tell, in both directions, is to ask what happens when the answer turns out to be wrong. If the answer being wrong is a defect to be corrected, the work is transactional. If the answer being wrong is something a particular person has to account for, it is judgment, and disassociating it moves the accountability somewhere nobody has agreed to put it.

Note that this test says nothing about difficulty, seniority, or how interesting the work is.

Those are the three axes people reach for instinctively, and all three of them will sort your list wrong.

The sort

One hour, your own list, and the split matters more than the sort

Take the list of things you are planning to hand to an AI system this year. Not the categories. The actual tasks.

• Put each one in a column. Transaction, where the output is the point and any qualified party would produce the same result. Or judgment, where whose call it is forms part of what is being bought.

• Split anything that is a category rather than a task. "Customer support" is not a task. "Financial reporting" is not a task. If an item on your list is the name of a function, it contains both columns, and sorting it whole guarantees you get one half wrong. My two errors were both this error. Break it down until each line is a thing somebody actually does on a Tuesday.

• Expect the judgment column to be too long. Everyone believes their own work is judgment, and a framework that lets people declare themselves unautomatable will be used exactly that way in every room where it is run. Mine would be too. The check is not introspection. Ask whether a customer would notice or care if somebody else did it, because the buyer decides this and not the person doing the work. If a competitor disassociates something you called judgment and nobody complains, you were wrong regardless of what you believed while sorting.

• Then apply the office test to the judgment column. For each item you called judgment, ask whether you sorted it that way because of its most visible activity or its most common one. My office call failed on exactly this. If the item is mostly hours of undifferentiated production punctuated by moments of real judgment, you have one item where you need two.

• Then apply the accountability test to the transaction column. For each item you called transaction, ask what happens when it is wrong. If the answer is that someone specific has to answer for it, you have misfiled it, and the misfiling will not surface until the first time it matters.

• Then look at what is in neither column. Some tasks will resist the sort, and the resistance is information rather than a failure of the exercise. In 2020 the categories that resisted were the ones where the choice was not the customer's to make. If a task will not sort, the usual reason is that its classification depends on somebody else's rules rather than on the nature of the work, and that is a different conversation with a different set of people in the room.

The sort takes an hour. What it produces is not a plan. It is a list of the places where your plan is resting on a classification nobody has said out loud, which is where my 2020 error lived for eighteen months before I noticed it.

Run it with the people who will do the work rather than the people who will approve it. The variance report and the capital decision look identical from two levels up, and they never look identical to the person who produces both.

What the framework does not tell you

Sorting correctly and moving correctly are different problems

Someone will point out that the two columns do not tell you what to do, and they are right. Three other objections are worth stating before anyone else states them.

• The framework can be accused of grading itself. Transaction and experience are defined partly by what happened: it came back, so it must have been experience. Then cinema did not cooperate and I adjusted the axis to substitute quality, which is what a man rescuing his own argument would do.

  
  

  The answer is simple and it is the only one I have. The sort was published before any of it happened, in print, with the categories named and the columns filled in. You can go read which box I put things in. An argument that only works looking backward is a horoscope. This one has a date on it.

  
  

• Substitute quality moves, and cinema is the proof. The at-home film experience improved enormously between 2000 and 2020: bigger screens, better sound, everything available at once. Cinema did not lose because a fixed substitute finally won. It lost because the substitute kept improving until it crossed a line. AI substitute quality is improving considerably faster than broadband ever did, which means an item you correctly file as judgment today can become transactional in eighteen months without your classification having been wrong when you made it.

  
  

  The sort is a snapshot with a short shelf life, and that is the sharpest disanalogy between this round and the previous two.

  
  

• Six years contains more than one variable. Stimulus, zero rates, a rate shock, inflation, a labor shortage. Cinema specifically lost most of a release slate to the 2023 strikes. Attributing any of these outcomes purely to substitution overstates the case, and the honest version is that the sort explains direction better than it explains magnitude.

Three objections about the instrument, then one about the reader.

A category returning is not the same as an operator surviving. Restaurants came back, and roughly 72,700 of them did not. Bowling revenue grew while the venue count fell. Being in the right column tells you the category has a future. It tells you nothing about whether you are in it.

A sorting instrument predicts where separations stick. It says nothing about timing, sequencing, cost, or whether an organization is capable of executing the change it implies. Knowing that bank branches would consolidate did not tell any specific bank which branches, in what order, or how fast. Every hard question in a real plan sits downstream of the sort.

There is also a category the columns handle badly, and it was the same one in 2020.

Teaching is experiential, the framework said it would return, and it did. But the return was uneven, contested, and expensive in ways the framework never anticipated, because education is not a market clearing on preference. It is a public obligation, and a mechanism built on what people will choose says nothing about what they are required to provide.

The AI equivalent is any work where the disassociation is legally or professionally constrained rather than economically chosen. The columns will tell you a task is transactional. They will not tell you that a licensing body, a regulator, or a liability regime requires a named human to hold it anyway.

So the sort is a starting instrument, not a plan. It was that in 2020 too. What it does is stop you from arguing about the wrong thing, which is most of what a framework is for.

Back to the course

The activity left. The accountability did not.

The data cleanup that happened while I was on the golf course was transactional work, and I would make the same call again. Nobody was buying the fact that I personally performed the steps.

But notice what did not leave. I decided what the data needed to become. I decided what finished looked like. And if that dataset is wrong, it is wrong under my name, in front of the person I hand it to, and no part of the arrangement moved that anywhere else.

The place round moved activities out of rooms and left us arguing for years about which ones would come back. This round is moving them out of people, and the argument will be harder, because a schedule has no reputation, a room has no reputation, and a person has one.

There is a version of this piece that ends with a warning about what AI will take, and I do not think that version is true.

The bank branch did not take anything from the teller that the teller was being paid for; it took the part of the job that was processing, and the industry spent the next decade discovering that the part it had left behind, the relationship and the credit judgment, was the part that had been carrying the economics all along.

The National Community Reinvestment Coalition found that when branch density fell 8%, small business lending dropped 22%, which is the sound of an industry learning what was actually in the room after the room was gone.

Something similar is available in the second round, and it is not job loss. It is stripping the transactional hours out of a role and then discovering that those hours were where the judgment was being formed.

A person who has never produced the variance report may not be able to make the capital call, not because the report is valuable but because producing it is how you learn what the numbers do. Nobody has a good answer for that yet, including me, and any framework that tells you it does is selling something.

What I got wrong in 2020 was not the mechanism. It was two classifications, made confidently, in the two largest categories, both by sorting a thing whole that needed splitting first.

My error is sitting in somebody's planning document right now.

There are four seats in the name. The fifth one is yours.