A probability is not a price.

Two delivery slips are clipped to a board on a receiving desk, and each carries the same stamp: a 20 percent chance of arriving late. The first is for a display sample, a demonstration unit for a showroom shelf that could be swapped for another without anyone outside the building noticing. The second is for a gearbox, the one part a stopped production line needs before it can run again. The company, the parts and every number that follows are invented. The scheduling model that stamped the slips did its job well. It looked at the carrier, the route and the weather, and it concluded, for both shipments, that one in five shipments like this one arrives after the promised time.

Should the company pay to rush either of them? The stamp cannot say, and that is the point of this essay. Two things travel together on every scored decision, and they are easy to mistake for each other. One is the predicted probability, how likely an outcome is. The other is the economic consequence, what the outcome would cost if it happened, and what acting would cost to prevent it. A prediction supplies the first. It does not, by itself, supply the second, and a business decision needs both, joined by a rule that someone has written down.

This is not an argument against better predictions, which matter whenever they change what a business does. Nor is it a claim that a person must weigh every shipment by hand. Consequences can often be estimated, and decisions can often be automated. The claim is narrower: somebody has to state what an error costs, and a system that acts on scores is optimizing whatever it was told, whether or not anyone meant to tell it.

What the 20 percent says

Start with what the stamp means, because it means something precise. The model was asked about a specified event, that the shipment arrives after five o'clock on Thursday, and it answered with a frequency. If the model is well calibrated, then of all the shipments it stamps at 20 percent, about one in five will in fact arrive late. That is a statement about the world, and a useful one.

It is also a statement that has to be checked rather than assumed. In 2017 Chuan Guo, Geoff Pleiss, Yu Sun and Kilian Weinberger showed that modern neural networks, unlike those of a decade earlier, are poorly calibrated: the confidence they report does not match how often they turn out to be right.⁴ They also showed that a simple correction applied after training, which they called temperature scaling, repairs much of the problem. The lesson for a buyer is modest but real. A score printed to two decimal places is a claim about frequencies, and it can be tested against what actually happens.

Suppose the test passes. The number is honest. It still says nothing about the showroom or the production line. It is not a recommendation, because it does not mention any action. And it is not a sum of money, because it does not mention any cost. Those are different questions, and the stamp was never asked them.

Cheap prediction, and the part that stays

The clearest account of why this distinction matters now comes from three economists at the University of Toronto. In Prediction Machines, published by Harvard Business Review Press in April 2018, Ajay Agrawal, Joshua Gans and Avi Goldfarb recast the rise of AI as a drop in the cost of prediction.¹ The economics of that framing are familiar from any other input. When something becomes cheap, people use more of it, and the things that work alongside it become more valuable.

The thing that works alongside prediction, in their account, is judgment. In a working paper released in January 2018, the same authors adapted standard models of decision-making under uncertainty and labeled the process of determining the payoffs "judgment."² Their example is a bank deciding whether to approve a card transaction. A model can estimate the chance that the transaction is fraudulent, but the bank must also know what it gains by approving a legitimate purchase, what it loses by refusing one, and what it loses by approving a fraud.² Prediction says which state of the world is likely. Judgment says what each action is worth in each state. In their model the two are complements, as long as judgment is not too difficult, which is to say that cheaper prediction makes clear payoffs more valuable rather than less.

A companion paper, released that May, sharpens the point and complicates it.³ It argues that better prediction allows riskier decisions to be taken, which raises productivity but can also widen the spread of outcomes. It treats judgment as what a decision needs when its objective cannot be described, or coded, in advance. And it shows that better prediction does not raise the value of every kind of judgment: it complements judgment about hidden costs and substitutes for judgment about hidden opportunities.³ Both papers are theoretical models issued as NBER working papers. They show how the pieces fit, not how much any firm gained.

The framework has limits worth stating. A language model today does much more than a narrow statistical forecast. It can draft the contract clause, summarize the downtime log, and estimate what a stopped line costs per hour. Those estimates are themselves predictions about consequences, and they can be useful. But they do not remove the distinction; they move it. Someone still decides which costs count, and the book's lens is a way to analyze a single decision, not a complete theory of what such systems can do.

Give the decision a price

Back to the receiving desk, with numbers that are deliberately simple. Suppose a rush shipment costs $100 and, for now, removes the risk of lateness entirely. Suppose a late display sample costs $200, a day of an empty showroom shelf, and a late gearbox costs $2,000, a day of lost output on the line.

For the display sample, the expected loss from doing nothing is a 20 percent chance of losing $200, or $40. Paying $100 to avoid an expected $40 is a bad trade, so the sample ships standard. For the gearbox, the expected loss is a 20 percent chance of losing $2,000, or $400. Paying $100 to avoid an expected $400 is a good one, so the gearbox ships fast. The probability was identical. The consequences were not, and so the actions differ.

The arithmetic generalizes into a rule. Act when the probability times the loss exceeds the cost of acting, or, turned around, when the probability exceeds the cost divided by the loss. For the sample, the threshold is $100 over $200, or 50 percent. For the gearbox, it is $100 over $2,000, or 5 percent. The same scoring model can serve both decisions. What it cannot do is set the threshold, because the threshold is made entirely of costs. An earlier essay here argued that a forecast is not an order; this is the same idea at the scale of one shipment.

What the rush shipment actually buys

The strongest assumption so far is that paying for speed removes all the risk. It rarely does. A faster carrier can still meet a closed road. So relax it: suppose the rush shipment cuts the chance of lateness from 20 percent to 5 percent.

Now the right number to multiply is not the original 20 percent. It is the 15 points the action removes. For the gearbox, 15 percent of $2,000 is $300, still well above the $100 fee, so the decision holds. But take a part that costs $600 if late. With full prevention, 20 percent of $600 is $120, and rushing it pays. With the more realistic three-quarters reduction, 15 percent of $600 is $90, and it does not. The decision flipped, and the late-arrival model did not change at all.

That reveals a second prediction hiding inside the decision. The scoring model estimates what happens if the company does nothing. The decision also needs an estimate of what happens if it acts: how much the rush service actually reduces the risk. That is a question about a counterfactual, and the model that stamped the slips was never trained to answer it. Figure 2 lets you vary the loss and switch between the two assumptions.

The argument hiding in the spreadsheet

Every number in that rule has an author, and the authors often disagree. Who said a late gearbox costs $2,000? The plant manager might count lost output per hour. The sales director might add the customer whose order slips, and the chance that the customer does not come back. Finance might count only the contractual penalty, because it is the only cost that appears on an invoice. Each figure is defensible, and each implies a different threshold: at a loss of $800 the gearbox threshold is 12.5 percent, and at $3,000 it is about 3.3 percent.

The same is true of the other terms. The cost of intervention includes the fee, but also the planner's time and the habit it creates, since a plant that rushes everything soon has no standard shipping at all. The tolerated downside, how much variation the business is willing to live with, is a choice about risk, not a fact about freight. These are commercial and organizational questions. A model cannot settle them by producing more decimal places, because the dispute is not about the probability.

Disagreement is not a reason to give up on numbers. It is a reason to use ranges. If everyone agrees that a late gearbox costs somewhere between $800 and $3,000, the threshold falls somewhere between 3.3 and 12.5 percent, and a 20 percent score clears the whole range. The decision is robust even though the cost is contested. The arguments worth having are the ones where the range straddles the score, and a written rule makes it obvious which ones those are.

A card for each decision

The practical form of all this is short enough to fit on an index card, one for each kind of decision the model informs. It names the predicted event, in terms specific enough to check: arrives after five on Thursday, not "delayed." It names the available action and its cost. It states the residual risk after acting, which is the estimate most often left blank. It states the consequence, as a range if the range is honest, and who supplied it. And it states a review condition: the date or the event that sends the card back for revision.

Then the card has to meet the world. Track what happened to the shipments that were rushed and the ones that were not. That record tests the calibration of the score, and it also tests the costs, which drift as quietly as models do. A penalty clause gets renegotiated, a second supplier appears, a line acquires a spare gearbox. A decision rule built on last year's costs can be wrong while the model is perfectly accurate.

Some things should never go on the card as a price. If a shipment carries a medicine a patient is waiting for, or a part that a safety inspection requires, the right form is a rule, not an expected value. Legal obligations, risk limits and harm to people are constraints the calculation must respect. Converting them into convenient dollar figures, so that a large enough fee can outweigh them, is exactly the mistake a written decision rule should prevent.

Objections, considered

Expected values hide catastrophes. They do. A 1 percent chance of losing the business and a 50 percent chance of losing a fiftieth of it have the same expected value, and nobody sensible treats them alike. Averages also hide who bears the loss: a rule that minimizes average cost can still put all of it on one customer. Where the downside is ruinous or unevenly spread, the card needs a constraint or a worst-case test alongside the expected value. The arithmetic here is the simplest case, not the only one.

The costs are too uncertain to write down. Often they are uncertain, and sometimes they are contested for political reasons. But a decision made without stating the cost still uses one; it is simply implied by the threshold someone picked, and nobody can inspect it. Writing a range, with the name of whoever supplied it, is more honest than false precision and more honest than silence.

A good enough system can do all of this itself. It may. A sufficiently specified system can estimate the probability, the effect of the action and the consequence, and choose the action without a person in the loop, and for many routine decisions it should. But automation strengthens the case for writing the objective down. A person deciding by feel can be asked what they were thinking. A system acting on thousands of shipments a day is optimizing exactly the costs it was given, and the only way to check them is to read them.

Accuracy is what you are paying for. Better predictions are worth paying for, and nothing here says otherwise. A sharper score moves more decisions to the right side of the threshold. But the value of an extra point of accuracy depends on how many decisions sit near the threshold and what an error costs there. Without the costs, a buyer cannot tell whether a more accurate model is worth its price.

A more accurate probability cannot tell you what you are willing to lose.

Before you act on a score

• What exactly is predicted? The event, the deadline and the population the score was calibrated on, and when the calibration was last checked against outcomes.
• What action changes it, and by how much? The cost of acting and the risk that remains afterward, estimated separately from the score.
• Whose consequences enter the calculation? Who supplied the cost of an error, as what range, and whose losses were left out.
• What is not for sale? The obligations and limits that apply as rules, whatever the expected value says.

Back to the clipboard on the receiving desk. Both slips still carry the same stamp, and both stamps are right. The display sample goes by ground and the gearbox goes by air, and nothing about that is inconsistent. The two scores matched because the shipments were equally likely to be late. The actions differed because being late meant different things, and that was never the model's to know until someone wrote it down.