An oligopoly does not set terms against persons. It sets terms against categories. The person is what the category was built to replace.
ALL MODELS ARE WRONG, BUT SOME ARE USEFUL.
The statistician George Box wrote this in 1976, and it is the axiom this entire essay rests on. A model is a simplification. It is supposed to be wrong about the particular case — that is what makes it a model rather than a copy. The question is never whether a model is wrong. The question is always whether it is useful for the purpose at hand, in the domain where it is being applied. A model that is useful in one domain is not thereby useful in another, and the same model can be a legitimate instrument on one side of a boundary and an instrument of extraction on the other. The firm’s error is never that it uses categories. It is that it uses them outside the domain where they are useful, and declines to mark the line.
I. THE TWO THINGS THE MIND DOES WITH A STRANGER
You are standing in a queue. There is a person in front of you you have never met.
Before you have decided anything about them, your mind has already done something. It has put them in a box. The box is built from what is visible: roughly how old they look, roughly which sex they present as, how they are dressed, how they stand, whether they look like the sort of person who works in an office or the sort who works outdoors, whether they look like they are from here or somewhere else. This happens in well under a second. You did not choose it. You cannot stop it. By the time you are aware of the person, the box is already there.
The box is a category. A category is a group of people defined by features they are taken to share, treated as the same for some purpose. The purpose here is simply: what kind of person is this, and what should I expect.
The box is a model, in exactly Box’s sense. It is useful, and it is wrong. It is useful because it lets you act before you have any real information. It is wrong because it describes the average member of the box and not the person in front of you. A statement like “people in that box tend to be like this” can be perfectly true, and the person in front of you can be nothing like it. The box cannot see that. The box does not look.
The other thing the mind can do is learn the person. This takes time. It means seeing them again, in different moods, in different situations, saying different things, making different mistakes. Over enough of that, the box comes apart. What is left is not a type. It is a particular person with particular reasons, some of which are strange, some of which contradict the box entirely.
These are two different operations, and they run at different speeds.
The first is instant and free. It costs nothing, it arrives before any evidence, and it closes the question. You now know what kind of person this is, or you believe you do.
The second is slow and expensive. It costs attention. It requires repeated contact. It requires noticing specifics that the box would have smoothed away. And it never closes the question, because a person is not a question that closes. It ends with a person instead of a conclusion.
Most of daily life runs on the first, and this is not a character flaw. Attention is limited. Nobody can learn every stranger. The first operation is how you get through a city without stopping to understand every face in it. It is a budget decision, made thousands of times a day, mostly below awareness.
So the situation is this. The category is what you get for free. The person is what you get by paying.
II. WHAT A CATEGORY IS FOR
A category is a set of things that can be counted, because for the purpose at hand any one of them can stand in for any other.
That sentence is doing a lot of work, so take it slowly.
To count things, you need them to be interchangeable for the purpose of counting. If I ask you how many chairs are in a room, you count them, because for the purpose of sitting, one chair can stand in for another. If I ask you how many unique things are in a room, there is no answer, because unique things cannot be added up. You can only add up units, and a unit is a thing that can be swapped for another unit without changing the total.
A category is the device that makes a population interchangeable. It says: for our purposes, this one is like that one, so we may treat them as the same unit and count them together.
That is why the category is powerful. Counting is the foundation of every large-scale operation there is — pricing, planning, insuring, taxing, hiring, forecasting. None of it works on individuals. All of it works on units.
And this is where the axiom does its first piece of work. A category that treated every member as unique would not be a category. It would be a list. The category is useful because it is wrong about the members — the wrongness is the compression, and the compression is what makes it countable. The question is never whether the compression loses information. It always does. The question is whether what it loses matters for the purpose it is being used for.
III. THE ARGUMENT, STEP BY STEP
Here is the whole argument. It is four steps and each one depends on the one before.
Step one: to set terms, you have to be able to predict what happens next.
“Setting terms” means deciding the conditions other people have to accept. A price on a good. A wage for a job. A rent for an apartment. A rule for using a service. Anyone who sets terms is exercising a kind of power, and power of this kind is only real if it can forecast. If you raise the price by ten per cent and have no idea whether people will stay or leave, you are not setting terms. You are gambling, and the gamble will eventually be taken from you by someone who can forecast.
Step two: you cannot forecast one person. You can forecast a group.
A single person is unpredictable in a way that does not average out, because there is nothing to average. One person might love the product for a reason that exists only in their own history. One person might leave for a reason nobody could have modelled. Forecasting one person is guesswork.
Forecasting a group of a thousand people who share a category is a different thing. The strange reasons cancel. The idiosyncrasies average out. The group behaves like a group. You can say: if we raise the price by this much, roughly this fraction will leave, and we will still come out ahead. That is a forecast, and it is a forecast because the group is a group.
Step three: a group is only a group if its members can stand in for one another.
This is where the category comes back in. You cannot add up people who are each unique. You can only add up units — members who, for the purpose of the count, are interchangeable. A category is what makes them interchangeable. It is the thing that says: these people can be counted together because, for our purposes, one is as good as another.
Step four: an individuated person cannot be counted.
An individuated person is someone whose box has come apart into a particular person with particular reasons. There is no other person like them to stand in for them. They are not a unit. They cannot be added to the group. And if they cannot be added to the group, the group’s average does not describe them, the forecast does not cover them, and the terms cannot be set against them in advance.
That is the whole thing.
The oligopoly does not fear individuation because individuated people are morally admirable, or because they might organise, or because they might complain. It fears individuation because an individuated person is the loss of the unit of account.
The unit of account is the standard thing you count in. If you are counting money, the unit of account is the currency. If you are counting customers, the unit of account is the category. Take away the unit and the counting stops, and everything built on the counting stops with it.
The axiom applies to this argument too, and it makes it sharper rather than weaker. The firm’s model is not wrong because it uses categories. Every model uses categories. The firm’s model is wrong when it is used outside the domain where it is useful — when the category’s average stops describing the members, when the trait stops causing the behaviour, when the stakes are too high for the error to be absorbed, when the relationship is long enough that the record has already individuated the person. Inside that domain the firm is doing something correct. Outside it, the firm is doing something else, and calling it a model.
IV. WHAT A CATEGORY DOES FOR A LARGE FIRM
An oligopoly is a market with very few sellers. Not one — one is a monopoly — but few enough that each seller is large and each seller’s decisions affect the others. The important feature for this argument is not the fewness itself. It is that firms of that size operate at a scale where nothing can be done person by person. Everything is done to types.
Here is the machinery, term by term.
A market segment is a group of customers defined by traits they are taken to share — income, age, location, taste — and treated by the firm as a single block. When a firm decides to target “urban professionals aged twenty-five to forty who rent,” it has built a segment and is treating a population as one unit.
A brand is a name and a set of associations attached to a product, which tells the customer what to expect before they buy. The brand’s function is to let the firm charge for the expectation. It works by giving a large number of people the same idea about the product at once, which is only possible if those people are treated as one audience.
A focus group is a small group of people assembled to represent a larger group, asked questions so that the firm can learn about the larger group without asking all of them. It works only if the small group can stand in for the large one, which is to say, only if they are the same type.
A credit score is a number computed from a person’s financial history and shared traits, used to predict how likely they are to repay a loan. It is a category rendered as a number. Two people with the same score are treated as interchangeable by the lender, whatever else is true about them.
An insurance band is a risk group into which policyholders are placed, each band with its own premium. Young drivers, older drivers, drivers in this postcode, drivers in that one. The band is a category with a price attached, and the price is a statement about the band rather than about any person in it.
A loyalty tier is a rank given to customers by how much they have spent, with different treatment at each rank. The tier is a category that the customer is placed into and, crucially, wants to stay in.
A demographic is a slice of the population defined by traits that can be measured and counted — age bracket, income bracket, education level, household type.
A persona is a fictional representative of a customer type, given a name and a biography, used by designers and marketers so that everyone in the firm is imagining the same person. The persona is a category with a face drawn on it.
A core customer is the type the firm imagines as its main buyer — the person the product was designed for, and the person the rest of the market is understood in relation to.
All of these are the same device in different clothing. Each one makes a population legible: readable at a glance, countable, and therefore priceable. Priceable means a price can be attached to it. Nothing can be priced until it is first a unit.
Now the money.
Margin is the difference between what you receive for something and what it costs you to provide it. Profit per unit. The margin on a customer is the gap between what that customer will pay and what it costs to serve them.
The firm’s margin therefore depends on knowing what the category will bear — the most that a member of this category can be charged before they stop buying. And that gap is calculable only because the category behaves like a category. If each customer had to be understood on their own terms, the gap could not be calculated in advance, and a business that cannot calculate its margin in advance cannot operate at scale.
Price discrimination is the practice of charging different groups different prices for the same thing, based on what each group will bear. Student discounts, business-class fares, regional pricing, introductory rates that rise after a year. In its technical sense, price discrimination is the whole art of charging each type what its type will bear, and it is only possible if types exist.
And this is the part that matters most.
The firm does not merely measure categories. It produces them.
A survey measures an existing group. Branding builds a new one and then sells membership in it. Segmentation does not find the boundary between two groups; it draws one and then organises the market around the line it drew. The demand curve — the graph showing how many people will buy at each possible price — is, in significant part, a curve the firm itself helped to draw, because the firm had a hand in making the groups that will respond to the price.
A firm that draws the boundary draws the price. That is the whole reason branding and segmentation are worth so much money.
And note what the firm is doing in Box’s terms. It is not choosing a model and then checking whether the model is useful in the domain it is being applied to. It is building the domain to fit the model. This is legitimate when the model is a genuine simplification of something real. It is something else entirely when the model is a convenience the firm has installed, and then cites the convenience as evidence that the world is shaped like it.
V. THE BOUNDARY
Everything so far has a hole in it, and the hole is the most important part.
If categorising people were always wrong, the firm’s entire apparatus would be irrational, and it plainly is not. A supermarket that stocks according to the average household is doing something that works. An insurer that prices by age band is doing something that works. The category is not a delusion. Under certain conditions it is the correct instrument, and the firm is right to use it.
The axiom said this already: all models are wrong, some are useful. The question is which conditions make this one useful. The answer is narrow, and the firm’s error is to treat the narrow answer as a universal one.
There are four conditions under which treating a person as a type is not merely convenient but correct.
First, the members of the category must actually be similar to one another. If the people inside a box differ from each other only slightly, the average describes most of them, and pricing the average prices most of them correctly. If they differ from each other a great deal — if the spread inside the box is wide — then the average describes almost no one, and pricing the average mis-prices almost everyone. The average of a wide distribution is a number that belongs to nobody. This is the condition that fails most often, and it fails silently, because the average still looks like a number.
Second, the trait that defines the category must be the trait that causes the behaviour. Age predicts driving risk, but it does so as an average, and the average is doing a specific kind of work. Young nervous systems and old reaction times really are different things, and that difference really does show up in aggregate claim rates. But the average does not describe any particular twenty-year-old or any particular eighty-year-old. Most twenty-year-olds drive safely and most eighty-year-olds drive safely, and the band exists because the small fraction who do not is large enough to move the total. Age predicts driving risk on average, across a large pool, for the purpose of pricing a policy. Age does not predict what book someone wants to read, whom they will vote for, or whether they will repay a loan, and when it is used for those purposes it is no longer a causal trait. It is a stand-in — a proxy — for something the firm cannot measure. A proxy wearing the clothes of a cause.
Third, the stakes must be low enough that a wrong call can be absorbed. If the supermarket stocks the wrong yoghurt, you buy a different yoghurt and nothing is lost. If the insurer prices you wrongly, you pay too much for years and there is no way to recover it. The category is a reasonable approximation when the error is small and reversible, and it is an extraction when the error is large and permanent.
Fourth, the relationship must be one-off. In a single transaction there is no history, and the category is all there is. In a repeat relationship the history accumulates: this person bought this, returned that, complained about the other, stayed through a price rise, left when the service changed. That history is the individuating material. A firm that ignores it — or that uses it only to sort the person into a finer box — is choosing the category over the record it already holds.
Where all four conditions hold, the category is a legitimate instrument. Where any one of them fails, the category is a wrong answer that reports itself as a right one.
And this is the core conflation.
The firm takes the conditions under which the category is valid and treats them as universal, when in fact they are local. It takes “the category predicts the average” and reads it as “the category predicts the person.” It takes “this category works for this purpose” and reads it as “this category works for every purpose.” It takes the average and calls it the member.
Box’s axiom is not a licence for this. It is the opposite. “Some models are useful” is a statement with a preposition attached: useful for what, useful where, useful under which conditions. The axiom does not say that because a model is useful somewhere, it is therefore useful everywhere. It says the usefulness is the whole question, and the usefulness is always relative to a purpose and a domain.
The conflation is not an error of arithmetic. It is an error of scope: a local truth extended to a global claim without the boundary being marked.
And the firm has a standing interest in not marking the boundary. On the near side of it, the category is a legitimate business tool. On the far side, it is an instrument of extraction. The same apparatus works on both sides. Only the label changes.
So the honest version of the argument is not that categorisation is wrong. It is that categorisation is conditional, that the conditions are narrower than the firm pretends, and that the firm’s fear of individuation is the fear of being held to the conditions. If the boundary were marked, the firm would have to say: for this customer, in this situation, at these stakes, the category does not apply, and we must treat them as a person. That would be more expensive, more laborious, and less profitable. It would also be correct.
The fear of individuation is the fear of the boundary.
VI. WHY THE FEAR AND THE MECHANISM ARE THE SAME INTEREST
The instant box from section one is the firm’s ally. The slow learning is the firm’s problem.
The box is cheap. It arrives before any data. It closes the question of what kind of person this is. The firm’s entire apparatus is built to work at that level, because at that level a population is a set of units.
Learning the person is expensive. It requires the specific. It refuses to close. A firm cannot run a forecasting model on it, because there is nothing to model — a person is not a type and does not average with other persons.
So the firm’s ideal customer is not the enthusiast, and it is not the loyalist either. It is what this essay will call the resolved type: a person whose preferences follow from their position, whose loyalty is a function of their group, and whose reaction to a change in terms can be predicted without anyone having to ask them anything.
And the firm’s nightmare is not the customer who hates it. A customer who hates the firm is still a category — a known, countable, forecastable refusal. The nightmare is the customer who cannot be placed. The one who wants the single thing, from the single maker, for a reason the segment cannot explain. That person will pay what they will pay rather than what the category will bear.
Economists call a customer inelastic when their behaviour does not change much as the price changes — they keep buying even when it gets expensive. The unplaceable customer is inelastic, but not in the firm’s favour. They are inelastic in their own favour. They pay their own price, not the category’s price. And the model has no way to hold that, because the model was built out of categories and this person is not one.
VII. THE SAME ARGUMENT ON THE WAGE SIDE
The mirror image of an oligopoly is an oligopsony: a market with very few buyers. In labour markets, that means few employers — a company town with one factory, a profession with three big firms, a sector where a handful of companies hire almost everyone.
The wage side works the same way, and the identity is exact.
A wage is set against a category. “The market rate for this role” is not a statement about a person. It is a statement about a type — what people of this kind are currently paid. The category is what makes the rate calculable, and it is what makes the rate comparable across firms. Comparability is what keeps the rate down, because every employer can point at every other employer and say: this is what the role pays.
A worker who cannot be slotted into the role does not automatically command a premium. They face something worse. They face a world in which no rate exists for them. And a rate that does not exist cannot be offered, because no one has a number to put in the box.
This is why standardised job titles (a fixed name for a fixed set of duties), salary bands (a range of pay attached to a job grade), and credentials (a certificate showing that a course of training was completed) serve the buyer of labour as much as the seller. The credential does two things at once. It proves what was paid — the years, the tuition, the exams — and it sorts the holder into a slot. And the sorting is what makes the wage a fact about a category rather than a fact about a person.
The boundary from section five applies here in exactly the same way. Where the role really is standard, where the skills really do transfer between firms, where the stakes of a wrong band are small, and where the engagement is short, the rate is a legitimate instrument. Where the role is not standard, where the skills do not transfer, where the wrong band costs a career, and where the engagement is a decade long, the rate is an instrument of extraction wearing the clothes of a market fact. The model is useful in the first case. In the second it is being used outside its domain, and calling it a market fact is the conflation in its wage-side form.
VIII. THE COMPLICATION THAT HAS TO BE STATED
The argument is not airtight without this, and it should be stated plainly rather than buried.
Aggregation is also the worker’s weapon.
A union is an organisation of workers that bargains as a single body, so that the employer faces one counterparty instead of many individuals. A profession is an occupation with controlled entry and shared standards, so that not just anyone can do the work. A guild is the historical version of the same thing. A licensing body is an authority that grants and withholds permission to practise. A collective agreement is a contract negotiated once for a whole group.
Every one of these is a category deployed against the buyer. Every one of them works by making the worker substitutable in a way the buyer is forced to respect — if any qualified member can do the job, then the employer must pay the rate that qualified members command.
So individuation cuts both ways, and the honest version of the argument has to say so.
The individuated worker cannot be slotted into the low band. That is the advantage. But the individuated worker also cannot invoke the standard rate, cannot point to the comparable, and cannot be represented by the body that negotiates on the category’s behalf. That is the cost.
The person who is not a type is unprotected as well as unpriced.
Which means the correct claim is narrower, and more defensible, than the flattering one. Individuation is not automatically good for the individual. It is unavailable to the firm as a unit of control. What fills the space left behind is a political question — who represents the unplaceable, who bargains for them, who protects them — and it is not settled by individuation itself.
And the axiom applies here too, in the direction that makes the argument uncomfortable. The union’s model is also wrong. It is also a compression, and it also fails to describe the particular member. It is useful because the buyer has no other way of being held, and because the alternative — everyone bargaining alone — is a model that is not merely wrong but useless. Some models are more useful than others, and the question of which is more useful, for whom, in which domain, is the whole of the politics.
IX. WHAT THE FEAR LOOKS LIKE WHEN IT IS SPENT
No firm writes a memo about individuation. It does what any apparatus does when its unit of account stops working: it invests in restoring the unit.
It produces categories faster than any person can escape them. Every new data source is a new axis on which to sort. Every new axis is a new set of boxes.
It personalises at the level of the type and calls the result individuation. A recommendation engine is software that suggests what you might want next, based on what you have done before and on what people like you have done. It feels personal. It is a category with better resolution — a finer box, not a dissolved one.
It converts the specific into a data point. A data point is a single recorded measurement that can be added to a set. The moment your particular reason is recorded, it stops being a particular reason and becomes an entry in a table, and a table is a set of categories.
It creates the segment of one. A segment of one is a category containing exactly one person. This looks like the final victory of individuation and is the opposite. A category of one is still a category. It is a box that has been drawn tightly around a single person so that the person can be counted.
And it buys the places where the categories are made. Whoever owns the boundary — the line that defines where one category ends and the next begins — owns the price on both sides of it.
The fear is not felt. It is spent.
X. LAZY THINKING PRODUCES ERRORS
A model is only as good as what goes into it: the data it was computed from, and the premises about what causes what and who belongs in which box. If either is wrong, the model is wrong. Garbage in, garbage out — the arithmetic will be flawless and the answer will be nonsense, because the arithmetic was never the hard part.
“All models are wrong” is not a licence. It is a discipline. It means the criticism is always the same precise question: wrong in what way, by how much, and in which domain. And a model can fail in two ways that are not the same failure. It can be inaccurate — bad data, false premise, an average that describes nobody. Or it can be accurate and out of context — correct about the population it was built from, then applied to a population or a purpose it was not built for. Either one makes the model plainly wrong, and the second is harder to see precisely because the model is right about something.
Nobody checks, because checking is expensive. Testing the premise, re-examining the data, marking the boundary, and living with the fact that the category does not apply on the other side of it — all of that costs money, and the category is cheap. So the model is used because it is cheap, and it is cheap because it is not checked, and it is not checked because the cost of the error is paid by the person the model got wrong, which is to say, by someone outside the firm’s accounts.
So here is the discipline stated plainly. You are always wrong — every model compresses, and the compression is the point. But sometimes you can be less wrong, and less wrong requires two things at once: the model is accurate, and it is in context. When both hold, the wrongness is the measured and accepted kind. When either fails, you are not less wrong. You are just plain wrong, and the number looks exactly like a good number.
Which brings the argument back to where it started. The firm’s fear of individuation is the fear of being asked to check. The individuated person is the case for whom the model would have to be re-examined — the failed premise, the wrong domain, the stakes too large to absorb. Treating them as a person is an admission that the model was never checked, and that someone else has been paying for the not-checking.
References
Box, G. E. P. (1976). Science and statistics. Journal of the American Statistical Association, *71*(356), 791–799.
Chamberlin, E. H. (1933). The Theory of Monopolistic Competition. Harvard University Press.
Galbraith, J. K. (1952). American Capitalism: The Concept of Countervailing Power. Houghton Mifflin.
Robinson, J. (1933). The Economics of Imperfect Competition. Macmillan.
Tilly, C. (1998). Durable Inequality. University of California Press.
Veblen, T. (1899). The Theory of the Leisure Class. Macmillan.
Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.