On the Spot: The Knowledge Problem and the Planner That Lives Where You Are

Prologue

The Clipboard

She knows things. The shrimp came in short this morning, two cases instead of three, and the driver said the boat had a bad week. Her best grill cook texted at noon that his daughter has a fever, so he can work until eight and not a minute after. There is a graduation across the river tomorrow, which means tonight the families arrive early and loud and want the big tables. The couple at table nine have been coming every other Friday for six years, and the husband has stopped drinking, and nobody is supposed to mention it. The walk-in is running a degree warm. The parking lot next door is closed for repaving, so the regulars will circle the block twice and walk in annoyed.

None of that is in any system. Not the point-of-sale, not the reservation software, not the supplier’s portal, not the dashboard the franchise consultant once tried to sell her. Some of it could be typed in, if she had the time. Some of it she could not type if she tried. She could not tell you how she knows the walk-in is warm. She just opened the door and knew.

In the next four hours she will make perhaps two hundred decisions. Which tables to hold. When to stop selling the shrimp special and what to steer people toward instead. Who goes home at eight and who covers. Whether to comp the dessert at table nine and how to do it without saying why. Every one of those decisions is made with knowledge that exists in exactly one place, which is her, and is good for exactly one evening, which is this one.

That clipboard is the subject of this book.

In 1945 an Austrian economist named Friedrich Hayek wrote a short essay about the woman with the clipboard. He did not use her, he used a shipper and a real estate agent and a man who deals in tin, but she is who he meant. His claim was that the most important knowledge in any economy is not scientific and not statistical. It is knowledge of the particular circumstances of time and place. It is scattered across millions of people. Most of it is fleeting, and much of it cannot be written down. And so, he argued, no central planner can ever gather it, no matter how clever the planner or how large the office. The decisions have to be left to the people on the spot.

For eighty years that argument has held. Every attempt to run things from the center has run into it, from Moscow to Santiago to the corporate headquarters that rolls out one scheduling system to four hundred stores and wonders why the good managers quit.

Now the argument is about to be tested again, harder than ever, because we have built something that looks very much like the planner Hayek said could not exist. A machine that reads everything, remembers everything, never sleeps, and reasons. The temptation is obvious. Put it at the center. Feed it the data. Let it decide.

This book is about why that temptation will fail in the same way it always has, and about the other thing AI makes possible, which is far more interesting. For the first time, we do not have to move the knowledge to the intelligence. We can move the intelligence to the knowledge. We can put a planner on the spot.

The owner with the clipboard has never had a staff that knew what she knows. She is about to get one.

Chapter 1

The Tin Example

The essay is called “The Use of Knowledge in Society.” It appeared in the American Economic Review in September 1945, eleven pages long, and it is one of the few economics papers that people outside economics still read. It is worth reading for the same reason a good proof is worth reading: it starts with a question everyone assumes is settled and shows that it is not.

The question is this. What problem is an economy trying to solve?

The usual answer is allocation. We have so much steel, so much labor, so much land, and people want various things. Find the best way to turn the one into the other. Put that way, it sounds like a math problem, and if it is a math problem, then the obvious thing is to gather all the numbers in one place and solve it.

Hayek’s move was to say that this is not the problem at all. The math is easy compared with what comes before it. The real problem is that the numbers do not exist in one place and never can. The knowledge that would go into the calculation is not given to anyone as a whole. It is dispersed, in bits and pieces, among all the separate people who hold it, and much of what each person holds is known to that person alone.

He was careful about what kind of knowledge he meant. Not the knowledge of experts, not science. Science is the one kind of knowledge that does gather well. You can write it down, put it in a book, teach it to a committee. He meant something humbler and much larger: knowing that a machine is idle this week, that a certain worker is better than his record shows, that a stock of material could be used in a way nobody has noticed, that a cargo ship is about to make its return trip half empty.

He gave examples. The shipper who earns his living by filling the empty or half-filled journeys of tramp steamers. The real estate agent whose whole value is knowing temporary opportunities. The arbitrageur who profits from local differences in price. People tend to look down on this kind of knowledge, he said, as if it were not really knowledge at all. But a society that wasted it would be poor, and a society that used it well would be rich, and the difference between those two is most of what we mean by an economy working.

Then he asked how such knowledge could ever be put to use, if it cannot be gathered. And his answer was the price.

The tin

Suppose, he wrote, that somewhere in the world a new use for tin has been found. Or a tin mine has closed. It does not matter which. And it is significant, he said, that it does not matter which.

Here is what happens. The people who want the tin for the new use bid a little more for it. The price rises. Every user of tin, everywhere, now faces a slightly higher price. Each one economizes a little. The can maker uses a thinner coat. The solder maker substitutes where she can. The people who use tin for things that matter least give it up first, because to them it is no longer worth the price. The people who use substitutes for tin find their substitutes in slightly more demand and economize on those too. The adjustment ripples outward through people who will never know why it happened.

Not one of them needs to know about the new use, or the closed mine. Not one of them needs to know anything except the price and their own circumstances. And yet the whole system behaves as if a single mind had seen the change and reorganized everything around it.

He called this a marvel, and said he chose the word on purpose, to shock people out of taking it for granted. The price system, he said, is best understood as a system of telecommunications. It lets each person watch a few pointers, the way an engineer watches a few dials, and adjust to changes far away that they will never see.

The price is not the knowledge. The price is the shortest possible message about the knowledge, sent to people who need to act but do not need to understand.

The man on the spot

The conclusion follows. If the knowledge that matters is knowledge of particular circumstances, and it lives with the people who are in those circumstances, then the decisions have to be made by those people. The man on the spot, Hayek called him. He cannot be told what to do by someone far away, because the person far away does not know what he knows. What he needs from the rest of the world is not orders. It is a signal small enough to act on, like a price, that tells him how his little piece of knowledge fits the whole.

Everything in this book comes back to that phrase. The person on the spot. The owner with the clipboard is on the spot. So is the nurse who notices the patient is quieter than yesterday, the contractor who sees that the lumber is wet, the teacher who knows which kid did not eat breakfast. Hayek’s point was not that these people are wise. It was that they are there, and being there is a kind of knowledge nothing else can replace.

Chapter 2

The Calculation Debate

Hayek was not starting a fight. He was finishing one.

In 1920 his teacher, Ludwig von Mises, published an essay arguing that a socialist economy could not calculate. If the state owns all the factories and all the materials, then nobody buys or sells them, and if nobody buys or sells them, there are no prices for them. And without prices, Mises said, the planner has no way to know whether it is cheaper to build a railway or a canal, to make shoes from leather or from rubber. He can count things. He cannot compare them. He is flying blind.

The socialists answered, and their best answer came from a Polish economist named Oskar Lange in the 1930s. Lange said: fine, prices are needed, so the planners will set prices. They will start with a guess. If a product piles up in warehouses, they lower its price. If shelves go empty, they raise it. By trial and error, the planners can find the same prices a market would find, without private owners. Lange’s paper was elegant, and for a generation many economists thought it had won.

Hayek’s 1945 essay was, in part, his reply. Lange had treated the problem as one of computation, a matter of finding the right numbers. Hayek said the numbers were never the problem. The planners could not set the right price for tin because they could not know, in time, all the particular facts that should move it. By the time the warehouse reports reached the planning office, the circumstances that produced them had changed. The knowledge was perishable, and it spoiled in transit.

So what’s the trouble?

Then came the computers.

In an essay published in 1967, two years after his death, Lange returned to the argument with obvious satisfaction. If he were writing his original paper today, he said, his task would be much simpler. His answer to Hayek would be: so what’s the trouble? Put the equations on an electronic computer and get the solution in less than a second. The market, with its slow groping toward the right price, could be seen as a calculating machine of the pre-electronic age.

It is one of the great confident sentences in the history of economics, and it has aged in an interesting way. In one sense Lange was right. The arithmetic of planning became trivial. A Soviet mathematician, Leonid Kantorovich, had already invented the core method for solving allocation problems with many constraints, and later shared a Nobel Prize for it. Today a laptop can solve problems that would have filled a Soviet planning ministry.

In the sense that mattered, he was wrong. The Soviet economy had the mathematicians and, later, the computers. What it never had was the knowledge. The factory manager who knew his machines were worn out had every reason to hide it, because the plan rewarded output, not truth. The data that reached Moscow was late, coarse, and shaded. The computers solved the equations perfectly. The equations were about a world that did not exist.

The room in Santiago

The most beautiful attempt came in Chile. In 1971 the government of Salvador Allende, nationalizing large parts of the economy, invited a British cybernetician named Stafford Beer to design a system to help manage it. Beer’s team built a network of telex machines linking factories to a computer in Santiago, software to watch the daily figures for trouble, and an operations room that still looks like the future: a hexagonal space with swivel chairs in a circle, buttons in the armrests, screens on the walls.

It was called Cybersyn, and the room is what everyone remembers. But the room never really ran the economy. It mostly showed prepared slides.

What worked was the network. In October 1972 a national truckers’ strike threatened to stop the country. The government used the telex lines to find out, hour by hour, which roads were open, where fuel was, which trucks were still driving, and to move supplies around the blockade. People on the spot reported what they saw, and other people on the spot used it. The center mostly relayed. Beer himself had insisted that the system should protect the autonomy of the factories, not override it.

The project ended with the coup in September 1973, and the operations room was destroyed. It is usually told as a story about a central planning machine. I think it is better told as a story about which half of it worked. The half that carried messages between people who knew things worked. The half that tried to know everything from a chair did not.

Chapter 3

Four Reasons the Knowledge Does Not Travel

It is worth being precise about why the knowledge on the spot cannot simply be sent to the center. “It’s too much data” is the lazy answer, and it is the answer AI seems to defeat. The real answer has four parts, and AI defeats only one of them.

One: it is dispersed

The knowledge is in millions of places at once. Each piece is small and each is held by someone different. Gathering it means asking everyone, all the time.

This is the part AI does help with. Collection is cheap now. Sensors, logs, transactions and messages produce more data than any planner of 1945 could imagine, and machines can read all of it. If dispersal were the only problem, it would be solved.

Two: it is fleeting

Hayek’s phrase was “time and place,” and the time part matters as much as the place. The shrimp came in short this morning. The grill cook leaves at eight. The parking lot is closed today. Knowledge of this kind has a half-life measured in hours. It is true now and false by tomorrow, and its whole value is in being acted on while it is still true.

Every step between the spot and the decision costs time, and time is exactly what this knowledge cannot afford. A report that is perfect but arrives tomorrow is worth nothing. This is not a problem of bandwidth. It is a problem of distance, and distance in this sense means anything between knowing and acting: forms, queues, approvals, meetings, a person who has to read it. Faster networks shrink some of that distance. They do not shrink the approval.

Three: it is tacit

In 1966 the chemist and philosopher Michael Polanyi put the third reason in one short sentence: we can know more than we can tell. You can recognize a face among thousands and not say how. A good mechanic hears a problem before he can name it. The owner opens the walk-in and knows it is warm.

Tacit knowledge is not secret. The person holding it would happily share it. They cannot, because it was never in the form of words. It lives in trained perception, in habits of attention, in what feels wrong. You can learn it, but only the way apprentices always have, by being near the work for a long time and being corrected.

No reporting system captures this, however good. The form asks for what can be told, and the tacit part is precisely what cannot.

Four: it does not exist yet

The last reason is the deepest and the least appreciated. Hayek came to it later, in a 1968 lecture called “Competition as a Discovery Procedure.” Much of the knowledge an economy needs is not sitting anywhere waiting to be collected. It is created by people acting.

Nobody knows what customers want from a new kind of shop until someone opens one. Nobody knows the cheapest way to make a thing until someone tries a new way. The owner does not know whether table nine wants the dessert comped until she reads the moment. Preferences, methods and opportunities are discovered in the doing. A central planner cannot collect them, because before someone acts, there is nothing to collect.

This is why competition matters even when it looks wasteful. Ten shops trying ten approaches are not duplicating effort. They are running ten experiments, and the results are knowledge that did not exist before.

Which one AI fixes

Put the four side by side and the honest picture appears. AI fixes the first one, dispersal, almost completely. It can read everything. It does very little for the second, because the bottleneck is the distance from knowing to acting, not the speed of reading. It struggles with the third, because tacit knowledge was never written down to be read. And it cannot touch the fourth at all, because no amount of intelligence can collect what has not yet happened.

A central AI, however brilliant, is a better collector of the kind of knowledge that was already the easiest to collect.

Chapter 4

Lange’s Computer, Finally

If you want to see the temptation clearly, imagine Oskar Lange given a modern AI system.

He would not need to set prices by trial and error. He would not need to solve equations. He could ask it: here is everything we know about our factories, our stores, our people, our weather, our trade. What should we make, and where should it go? And it would answer, fluently, with reasons, in seconds. It would write the memo, the plan, and the directive to each factory manager, personalized and polite.

This is not a fantasy about governments. The same temptation is arriving in every large organization right now, under friendlier names. A company buys an enterprise AI platform. It connects the platform to every system it owns. It builds a single brain for the business, a dashboard that sees all, an assistant that answers every employee’s question from the same central store. Scheduling, pricing, staffing, purchasing, all optimized from the middle.

Some of this will work well, and I want to be fair about what. Anything that is genuinely legible, meaning already in the form of records and rules, centralizes beautifully. Inventory counts. Contract terms. Payroll. Compliance. The things that were always easiest to see from headquarters become even easier. The central AI is a superb clerk.

The trouble starts at the boundary, where the legible meets the particular.

The memo arrives

Picture the restaurant again, but now it belongs to a group of forty. Headquarters has an AI planner. At 3:00 p.m. it sends the owner, now a general manager, a staffing plan for the evening, optimized against last year’s Fridays, the weather forecast, and the reservation book. It is a good plan. It is better than most human planners would produce from the same data.

It does not know the grill cook leaves at eight. It does not know about the graduation. It does not know the shrimp came in short, because the receiving log has not synced. It does not know table nine.

So the manager does what managers have always done with plans from headquarters. She works around it. She makes her real decisions on the clipboard and enters whatever the system needs to stop complaining. The plan and the reality drift apart, and the system, reading its own records, believes the plan was followed.

That last part is the danger. A central planner that is ignored is merely wasteful. A central planner that is ignored without knowing it becomes confident about a world that is not there.

What is new, and what is not

What is new about AI is that the plan can be so good. It can be personalized, explained, adjusted, even humble in its wording. It removes every excuse for the center not to try.

What is not new is the gap. The plan is still built from what reached the center, which is the dispersed part of the knowledge, collected late, with the fleeting part already spoiled, the tacit part missing, and the not-yet-made part impossible. Lange’s computer has arrived. Hayek’s problem has not left.

Chapter 5

Seeing Like a Model

In 1998 a political scientist named James C. Scott published a book called Seeing Like a State, about why grand schemes to improve the human condition so often fail. His central idea was legibility. A state cannot manage what it cannot see, so it reshapes the world to be seen. It gives people permanent surnames so they can be taxed. It surveys land into neat parcels. It standardizes weights and measures. Each step makes society easier to read from the capital, and each step throws something away.

His most famous example is a forest.

In the late eighteenth century, foresters in Prussia and Saxony wanted to manage their woods for revenue. A natural forest is a mess from that point of view: many species, many ages, undergrowth, deadfall, animals, the villagers gathering kindling and herbs and fodder. So the foresters simplified. They measured trees by the volume of timber they would yield. Then they planted new forests to match the measurement: one species, usually Norway spruce, in straight rows, all the same age, the undergrowth cleared away.

The first rotation was a triumph. The timber came in uniform and predictable, easy to count and easy to sell. The second rotation was a disaster. The trees grew poorly. The soil had been drained of the life the messy forest had kept in it. Pests found endless identical food. The Germans gave the decline a name, Waldsterben, forest death.

The foresters had not been stupid. They had been seeing through an instrument that recorded one thing, and they had rebuilt the world to suit the instrument. Everything the instrument could not see, they treated as noise, and the noise turned out to be what kept the forest alive.

Metis

Scott borrowed a Greek word for what the scientific foresters threw away: metis. It means practical, local, experienced knowledge, the cunning of someone who has done a thing for a long time in a particular place. The sailor’s feel for weather. The farmer’s sense of when the ground is ready. Metis is Hayek’s knowledge of time and place plus Polanyi’s tacit knowledge, held by people with their hands in the work.

Metis is not anti-intellectual. It is simply knowledge in a form that does not survive being written down in a standard format. Ask a farmer to fill in a form about when to plant and you will get a date. The date is not the knowledge. The knowledge was the looking at the sky and the feel of the soil and the memory of the bad year.

The model as forester

A central AI is a new kind of instrument, and it carries the forester’s temptation in a new form. It sees what is in the data. It is very, very good at what is in the data. And an organization that runs on it will slowly rebuild itself to be legible to it.

You can watch this happen already. The customer service call gets reduced to the categories the system can route. The nurse’s judgment gets reduced to the fields the record can hold. The teacher’s sense of a child gets reduced to scores. Each simplification makes the work easier to optimize from the center. Each one throws away some metis. The first rotation looks great. The numbers go up. It is the second rotation you have to worry about, when the people who held the metis have left or stopped bothering, and the system is running on a forest of its own making.

Chapter 6

The Pretence of Knowledge

In 1974 Hayek shared the Nobel Prize in economics, and he used his lecture to scold his own profession. He called it “The Pretence of Knowledge.” Economists, he said, had fallen in love with the methods of physics. They measured what could be measured and built models around it, and then they acted as if the model were the economy. The things that could not be measured were left out, and leaving them out made the models look more scientific than they were.

His warning was not that knowledge is impossible. It was that partial knowledge, dressed as complete knowledge, is more dangerous than honest ignorance. An ignorant planner hesitates. A planner with a confident model acts, at scale, on everything except the part that mattered.

The fluent center

AI raises the stakes on this warning in a specific way. The old central plan at least looked like a plan. It was a spreadsheet, a memo, a set of quotas, and everyone could see its edges. You knew it had been made far away by people who did not know your store.

The AI plan does not look like that. It speaks. It explains itself. It addresses you by name, refers to your history, anticipates your objection. It sounds like it was written by someone who knows your situation, because it has absorbed the part of your situation that reached it, and it is extraordinarily good at sounding complete.

This is the pretence of knowledge with a voice. The danger is not that the central AI is dumb. It is that the central AI is articulate about the legible part of the world in a way that makes the illegible part feel like it does not exist.

The Bundlers

I have argued elsewhere that we are heading into what I call the Bundled Economy. A handful of very large companies will compete to supply more and more of each person’s life on subscription: transport, food, housing, health, and the rest. In that world, the Bundlers become something like planners. They forecast demand across millions of members, stock warehouses, route vehicles, buy energy and materials years ahead.

I still believe that is coming, and I think much of it will be good. The Bundlers will be excellent at the slow, legible layers of life: tires, plywood, power, the commodities whose demand moves predictably and whose supply can be planned. Scale and AI together make that layer cheaper every year, the way I described in Cheaper Every Christmas.

But a Bundler that tries to plan the particular layer, your Tuesday, your knee, your daughter’s game moved to Thursday, your mood about a restaurant, will commit the pretence of knowledge at enormous scale. It will build a fluent model of you from what reached it, and it will be confidently wrong about you in all the ways a forester was wrong about a forest. The question for the Bundled Economy is not whether the Bundlers will plan. It is where their plans stop and something else begins.

That something else is the subject of the rest of this book.

Interlude

One Reality per Actualizer

In my own framework, Reality is a ratio. R equals A over E: what arrives, over what was expected. The Actual is what the world hands you. The Expectation is what you bring to meet it, and you do not bring it on purpose. It is already there when the moment arrives, unconscious, built out of everything that has happened to you before.

The first question the equation asks is not what happened. It is whose Reality we are talking about. There is one Reality per Actualizer at each instant. A person has one. So does a restaurant, a city, a company, a nation. Each is having its own, against its own denominator.

Read Hayek with that in mind and his problem gets sharper. A planner at the center can, in principle, gather what arrived: the shipments, the sales, the weather. It is trying to compute a world of ratios from the numerators alone. Every one of those Realities has a denominator that lives inside the Actualizer having it.

And no one can reach in. That is the hardest rule in the framework, and the one I hold most firmly: no entity can touch the numerator or the denominator of anyone’s Reality, by any means. An entity can only act. Its actions leave artifacts. The artifacts become history. History is Actual, and only Actuals retrain expectation. The road to anyone’s denominator runs through the numerator, one way, and slowly.

That is why a central plan lands like weather on the people inside it. It can change what arrives. It cannot reach what they expected. The only planner that can sit beside a denominator is one that lives with the Actualizer, learns from the same history as it arrives, and is corrected by the same person, day after day. For a person, that is their own AI.

Chapter 7

Move the Computer, Not the Knowledge

For eighty years, the whole calculation debate rested on an assumption so basic that neither side stated it. Intelligence was scarce.

There were only so many planners, so many economists, so many clerks, so many computers. Intelligence was expensive, so it had to be concentrated. The planners thought it should be concentrated at the center, in a ministry or a headquarters. Hayek thought it was already concentrated in millions of individual heads, each attending to its own little corner, and that prices were how those heads coordinated. But both sides took it for granted that you could not give every corner its own planner. That would be absurd. Who would pay for it?

That assumption is now false.

Intelligence is becoming cheap enough to station anywhere. Not intelligence in the grand sense, not wisdom, but the working kind: reading, remembering, comparing, drafting, calculating, following up, noticing that two facts conflict. The kind a good chief of staff provides. For the first time, every person, every household, every small business can have that kind of intelligence sitting beside them, working only for them, all day.

Once you see that, the old debate flips. The planners wanted to move the knowledge to the intelligence, because intelligence was scarce and centralized. Hayek showed that the knowledge cannot move: it is dispersed, fleeting, tacit, and not yet made. So the only remaining option, the one nobody could afford before, is to move the intelligence to the knowledge.

The calculation debate assumed intelligence was scarce. When intelligence is cheap, you stop shipping knowledge to the planner and start stationing a planner beside the knowledge.

The flip

Two ways to bring knowledge and intelligence together

1945 to now: move the knowledge

• Intelligence is scarce, so it sits at a center
• Particulars must travel to it, and spoil on the way
• The tacit part never arrives
• The decision moves away from the spot

Gosplan, Cybersyn’s room, the enterprise brain

cheap intelligence

From now: move the intelligence

• Intelligence is cheap, so it can sit on every spot
• Particulars stay where they are true
• Tacit knowledge passes by correction
• The decision stays with the person on the spot

A personal delegate, and prices between delegates

A planner for one

What does that look like? It looks like personal AI, which I have been arguing for since before it had a name. Not an app you open. A delegate that works for you, holds your context, acts within limits you set, and reports back. I wrote about why I prefer the word delegate to agent, and about why the right word for the human is manager, not user. This book gives the economic reason those words matter.

A personal delegate is a planner whose entire jurisdiction is one life, or one small business. It is not trying to plan the economy. It is trying to plan Friday. And for Friday, it has advantages no central planner can ever have, because it is on the spot.

It hears about the grill cook’s daughter at noon, because the text came to the owner and the owner’s delegate reads the owner’s messages. It knows the shrimp came in short, because the owner photographed the invoice. It knows about the graduation because it read the city calendar last week and connected it to last year’s sales on the same weekend. It knows table nine because the owner told it, once, two years ago, and asked it never to bring it up.

None of that knowledge traveled to a center. It stayed where it was. The intelligence came to it.

Hayek’s answer, completed

I want to be careful here, because this is not a refutation of Hayek. It is the completion of his argument. He said the decisions must be left to the person on the spot because only that person has the knowledge. The weakness of that answer, the thing that kept central planning tempting for a century, was that the person on the spot is limited. She is tired. She has two hundred decisions and four hours. She cannot read the city calendar, compare three suppliers, redo the staffing plan, and greet table nine all at once.

So the center kept saying: let us help. And the help always came with a price, which was moving the decision away from the spot.

A personal delegate is help that does not move the decision. It makes the person on the spot more capable without making her less present. That is the thing the twentieth century could not build, and it is why I think the planning debate is about to end, not with the victory of either side, but with the disappearance of the dilemma.

Chapter 8

What a Delegate Knows That No Center Can

Take the four reasons the knowledge does not travel, from Chapter 3, and ask how each one looks from the position of a personal delegate.

Dispersed: it only needs one piece

A central planner needs everyone’s knowledge. A delegate needs one person’s. The dispersal problem simply disappears, because the delegate is not trying to gather the world. It is trying to know one corner of it very well. The corner is small enough to hold.

Fleeting: there is no distance

The fleeting part of knowledge spoils on its way to the decision. But a delegate sits at the decision. When the grill cook texts at noon, the delegate knows at noon. It does not file a report for someone to read on Monday. It can redo the staffing plan in the next minute and put it in front of the owner, who is right there, who can approve it with a word. The distance between knowing and acting, which ruined every central plan, is close to zero.

Tacit: it learns the way apprentices learn

This is the most interesting one. Tacit knowledge cannot be told. But it can be learned, by watching and by being corrected. That is how every apprentice in history has learned the part of the trade that is not in the book: the master does it, the apprentice tries, the master says no, not like that, like this.

A delegate that lives with one person gets exactly that kind of education. It drafts the email, and the owner rewrites the second sentence. It proposes the table plan, and the owner moves two parties without explaining why. It suggests comping the dessert with a note, and the owner says just bring it, say nothing. Each correction carries a little of the tacit knowledge across. The owner still cannot tell you how she knows. But over hundreds of corrections, the delegate starts to make the moves she would make.

I have watched this happen in my own work. The instructions my AI follows for me were not written in one sitting. They grew out of corrections, one at a time, each one something I could not have specified in advance but recognized instantly when it was wrong. The written instructions are only the residue. The real knowledge is the history of being corrected.

No central system gets this education. The center receives reports, not corrections. It never stands at the elbow of the person doing the work.

Not yet made: it is there when it is made

Knowledge that is created by action cannot be collected in advance. But a delegate does not need to collect it in advance. It is present when the action happens. The owner tries a new special, and the delegate sees how it sold, what people said, what came back on the plates. The owner opens on a Monday for the first time, and the delegate learns what Mondays are like. It is not forecasting from a distance. It is keeping the notebook of someone who is discovering things.

The asymmetry

Put it together. On every one of Hayek’s four counts, a personal delegate has the advantage over a central one, not because it is smarter, but because of where it sits. The central AI may be a far more powerful model. It does not matter. Position beats power when the knowledge is local.

Position beats power

Hayek’s four reasons, seen from the center and from the spot

Dispersed

From the centerAI can now read it all. Solved.

From the spotNeeds only one corner, small enough to hold.

Fleeting

From the centerSpoils between knowing and acting. Barely helped.

From the spotNo distance. Known at noon, acted on at noon.

Tacit

From the centerNever written, so never read.

From the spotLearned the apprentice’s way, by correction.

Not yet made

From the centerCannot collect what has not happened.

From the spotPresent when it is made, and keeps the notebook.

Chapter 9

Prices Between Delegates

If every person and every small business gets its own planner, the obvious worry is chaos. Millions of delegates, each optimizing for its own person, colliding. Who coordinates them?

Hayek already answered this. The same thing that has always coordinated millions of people on the spot: signals small enough to act on. Prices. Offers. Terms. Refusals. The delegates do not need a central planner above them any more than the tin users did. They need a way to send each other short messages about what they want, what they have, and what they will accept.

The tin, again

Go back to the tin. A new use is found, the price rises, and every user economizes a little without knowing why. Now give every user of tin a delegate. The price still rises. But the delegates respond to it faster and more precisely than their people ever could. The can maker’s delegate sees the price move the same day, checks the coating specs, calculates the savings of a thinner coat, and puts the decision in front of the can maker before lunch. The solder maker’s delegate starts quietly asking suppliers about substitutes.

The price system works exactly as Hayek described. It just runs on better attention at every point. The pointers on the dials are watched by something that never looks away.

Richer messages

Something else happens, too. The price was always the shortest possible message, and it was short partly because people are slow. A human cannot negotiate every purchase; the cost of haggling is too high. So we invented the fixed price tag, which, as I argued in The Price Tag Was a Machine, was a nineteenth-century device for a world of hired clerks who could not bargain at scale.

Delegates can bargain at scale. So the messages between spots can get a little richer without getting slower. Not just a price, but a price and a delivery window. A price that drops if you can wait until Tuesday. A price that includes the return. An offer that says: we have two cases of shrimp, not three, at this price, available by four. The owner’s delegate can take that message from three suppliers, compare them against what it knows about Friday, and present one recommendation. The owner says yes. Nobody at any center was involved.

This is still Hayek’s world. The messages are still small. Each supplier’s delegate still knows nothing about the graduation or the grill cook. It does not need to. It needs to know what is wanted, by when, and at what price, and it can answer in seconds.

Protocols, not platforms

There is a design lesson here for anyone building AI systems. The valuable thing is not a central platform that knows everything. It is a set of agreed messages that let delegates on different spots talk to each other: here is what I need, here is what I offer, here are my terms, here is my refusal and why. The internet itself won this way, through protocols between machines, not through one central computer. Cybersyn’s one success was its network of messages, not its room.

The future of coordination is many small planners, each on its own spot, speaking a common language of offers and prices. The center does not disappear. It shrinks to the job of keeping the language honest.

Chapter 10

The Enterprise Mistake

Most large companies are about to make the planning mistake in miniature, and many already have.

The pattern goes like this. The company decides AI is strategic. It buys an enterprise platform. It bans the AI tools employees were already using, on grounds of security, which are real. It builds one central system, connected to company data, with one set of instructions, and gives every employee access to it. The employee’s own AI, the one that knows how they work, stays at home.

From headquarters this looks responsible. From the spot it looks like this: a skilled person arrives at work and is told to leave their trained assistant at the door and use the company’s generic one, which knows the company’s files but nothing about the way this person does the job. The tacit knowledge that person had built up with their own delegate, months of corrections, is thrown away every morning at the front desk.

It is Gosplan with a friendly interface. The company has moved the intelligence to the center and asked the knowledge to come to it. The knowledge, being tacit and personal, does not come.

The charge belongs to the person

I wrote about the underlying principle in The Charge Is Yours. A hard decision is counted against a person, not against a job. The model that matters most, for doing good work, is the model of how one particular person works: their standards, their shortcuts, their refusals, what they always check. That model is built mostly outside any one job, and it is worth something only if it travels with the person.

This is the man on the spot again, inside a firm. The employee is the one with knowledge of the particular circumstances. The company has the files. The company’s mistake is thinking the files are the knowledge.

Bring your own

That is why I argue every business needs a bring-your-own-AI policy, and why I wrote a short book, Bring Your Own AI, on how to do it safely. The principle is simple. The company owns its facts. The person owns their method. The policy draws the line between them and builds a gate that makes the line real: company data comes into the person’s delegate under rules, for the job, and does not leave with them.

This is the Hayekian design. Decision rights sit with the person on the spot, augmented by their own planner. The center provides the signals, the facts, the rules and the boundaries, the way a market provides prices and law. It does not try to be the brain.

The companies that get this right will find something surprising. Their people will make better decisions than any central system could, because their people are on the spot and are no longer limited by attention. The companies that get it wrong will have the best central AI money can buy and wonder why the good people keep leaving.

Chapter 11

The Hyper-Local Business

For most of the last century, local knowledge was a handicap at scale. The corner shop knew its customers by name, but it could not match the chain on price, selection or systems. The chain could not know its customers by name, but it did not need to: it had the warehouse, the buying power and the software. Local knowledge was charming and, economically, a little sad.

The reason was the same scarcity of intelligence. The chain could afford central planners, analysts, systems people. The corner shop had the owner, and the owner was already working eighty hours. Her local knowledge was real, but she had no staff to multiply it. So the chain’s thin, central, legible knowledge beat her rich, local, tacit knowledge, because the chain’s knowledge had intelligence attached and hers did not.

Now give the corner shop a staff. Not people, which she cannot afford, but delegates: one that answers the phone and knows the regulars, one that handles ordering and talks to suppliers’ delegates, one that keeps the books, one that watches the reviews and the city calendar. Each one built around her particulars, corrected by her, working on her spot.

Suddenly the balance changes. The chain still has scale, and scale still matters for the legible layer: the commodities, the logistics, the buying. But at the particular layer, the corner shop’s knowledge now has intelligence attached, and the chain’s does not. The chain’s central AI knows a million customers thinly. Her delegates know four hundred customers deeply, and they are standing in the room.

Why I built Charleston AI the way I did

This is the bet behind Charleston AI. It is deliberately hyper-local. In January I opened a three-thousand-square-foot space in Charleston where local owners and people can walk in, sit down, and get help understanding and using AI. We do not sell a platform. We deliver finished work for local businesses, much of it in hospitality: the phone that gets answered, the system that knows the reservations, the record that actually holds what the owner knows.

The reason to be local is Hayek’s reason. The knowledge that makes an AI useful to a restaurant in Charleston is knowledge of particular circumstances of time and place: this street, this parking lot, this season, these regulars, this owner’s way of doing things. A national platform cannot hold it. Someone on the spot can help the owner put it into her own delegate, where it belongs. I even wrote up the plain particulars of one such spot, parking at Miller’s All Day, because those details are exactly the kind of knowledge no central system bothers to keep.

The new advantage

I think we are about to see a quiet reversal. For decades, the advice to small businesses was: standardize, systematize, look more like the chain. The advice for the next decade is the opposite. Your particular knowledge is now your advantage, because it can finally be multiplied. Do not throw it away to fit someone else’s system. Write it into your own.

Chapter 12

The Bundle Must Ask

I have spent the last month writing about a future in which most people own very little and subscribe to almost everything, supplied by a few large Bundlers. A reader might fairly ask how that squares with this book. Isn’t the Bundled Economy exactly the central planner I have been warning about, a company that decides what your life needs?

It would be, if the Bundler held your knowledge. It will not, if you do.

Here is the distinction. A Bundler can plan supply superbly. It can forecast how many cars a city needs, how much power, how many meals, and it can buy and build at a scale no household could. That is the legible layer, and the Bundlers will be very good at it.

What a Bundler cannot do, for all the reasons in this book, is know your particulars. That your knee hurts on stairs this month. That you would rather walk on Tuesdays. That your mother is visiting and needs the car with the higher seat. That you have stopped drinking and nobody should mention it. That knowledge is on your spot. It belongs with your delegate.

Asking is the whole design

So the healthy shape of the Bundled Economy is this. The Bundler plans the supply. Your delegate plans your life. And the Bundler must ask your delegate.

It asks in Hayek’s way, with small messages. Here is what we can offer this week. Here is the price of the higher-seat car on Saturday. Here is the meal plan, accept or change. Your delegate answers from what it knows about you, which the Bundler never sees in full. The Bundler gets a signal, an order, a refusal, a preference, just enough to plan its supply. You keep the knowledge.

I called this the monthly election: every bundle must win you again each month, and your own AI casts the vote. This book adds the reason that arrangement is not just fair but efficient. The vote carries the knowledge on the spot back to the planner in the only form it can travel, a small signal. Without it, the Bundler is Gosplan.

The right that keeps the knowledge home

All of this depends on one right, and I will keep repeating it until it is law: the right to leave with your delegate’s memory and skills intact. If your delegate lives inside the Bundler, then your particulars have moved to the center, and the Bundler is a central planner with a nicer app. If your delegate is yours, portable, able to walk to a competitor with everything it knows, then the knowledge stays on your spot, and the Bundler has to keep asking.

Portability sounds like a consumer-protection detail. It is actually the boundary between a market and a plan.

Chapter 13

What Stays With You

If the delegate is on the spot, what is left for the person? Is the owner with the clipboard just a supervisor now, approving what her delegates propose?

No. And the reason goes to the heart of what being on the spot means.

The body is the spot

Every argument in this book has treated “the spot” as a place where knowledge lives. But the deepest knowledge on any spot is held by a body that is there. The owner opens the walk-in and feels that it is warm. She looks at table nine and reads the moment. She hears the tone of the grill cook’s text and knows he is more worried than he wrote. A delegate can learn a great deal of this from her corrections. It cannot stand in the room.

This is why I tell my students, be the face that your customers trust. As every channel, the email, the phone call, the video, becomes easy to fake, the person physically present carries a kind of trust no message can. The face on the spot is not a leftover. It is the part of the system that cannot be copied or moved.

Rank and order

There is a second thing that stays with you, and it is a matter of honor more than economics. I teach a simple rule about delegation: delegate the small matters, keep what is worthy of your attention. And the most important things worthy of your attention are people. Never send your AI to deal with a person who deserves you. The customer you have served for six years, the friend, the family member. Delegating them down is a kind of dishonor, worse than being used.

The couple at table nine are not a task for a delegate. The delegate’s job is to make sure the owner has the time to be at table nine. That is what all those other decisions, the shrimp and the staffing and the parking, were taking away from her. The point of putting a planner on the spot is not to replace the person on the spot. It is to free her for the part of the spot only she can hold.

Keep your notebook

One last thing stays with you, and it is practical. Your corrections. The tacit knowledge that passes from you to your delegate passes through the corrections you make. If those corrections live somewhere you control, your knowledge stays yours, and any future delegate can learn it. If they live inside someone else’s system, you have given away the most valuable record of your own judgment. Treat that record the way the old craftsmen treated their notebooks. It is the trade.

Chapter 14

A Friday in 2032

At 6:10 this morning the shrimp came in short. Her delegate saw the invoice photo the driver uploaded, saw two cases instead of three, and before she was awake it had sent the same small message to the delegates of four suppliers: one more case of this grade by three o’clock, here is the price we usually pay, what can you do. Two answered. One could deliver by two at a little more. It put the choice on her phone with one line of reasoning. She tapped yes over coffee.

At noon the grill cook texted that his daughter had a fever. Her delegate redid the evening plan in a minute: shift the second cook to the grill at eight, bring in the part-timer who had asked for more hours, move one big party from eight to seven-thirty with a note offering the earlier table and a reason. It asked her once. She said do it.

The graduation across the river was on its list since last month. It had already held the large tables, warned the kitchen, and ordered extra bread. The parking closure next door had gone out in the reservation confirmations that morning, with a map to the lot down the street. The walk-in had been flagged at two, a degree warm, and a technician was coming Monday. She still opened the door to check. It was warm. She smiled.

The group she belongs to, a buying cooperative of independent restaurants, sent its weekly offers through the same channel. Her delegate accepted the oil and declined the new dessert line, because it knew her regulars and she had said no to that kind of thing twice. The co-op’s system got two small messages. It did not get her reasons, and it did not need them.

At seven-forty the couple from table nine come in. They have been coming every other Friday for twelve years now. The owner walks them to their table herself. She has the time. That is the whole point.

Later, she will look at the evening the way she always has, and she will notice that she made perhaps twenty decisions instead of two hundred. They were the twenty that mattered. Everything else had been handled by something that was on the spot with her, knew what she knew, and asked when it should.

Chapter 15

What to Do Now

The future in the last chapter is not far away, and most of it is buildable today. Here is what I would do now, depending on who you are.

If you are a person

Get your own delegate, and keep it yours. Not an app that belongs to a platform, but an AI you manage, with memory you can see and take with you. Start giving it your particulars: how you like things done, the people who matter to you and why, the things you never want to be asked twice.

Correct it, and keep the corrections. Every time it gets something wrong and you fix it, you are handing over a piece of tacit knowledge. Make sure that record lives somewhere you control.

Be careful what you pour into central systems for free. Every platform would like your particulars. Give them the small signals they need, an order, a preference, a refusal, and keep the rest on your spot.

If you run a small business

Write your particular knowledge into your own delegate, not into someone else’s software. The regulars, the quirks, the suppliers, the season, the street. That knowledge is your advantage now, because it can finally be multiplied.

Give your delegate the dull decisions and keep the human ones. Shrimp, staffing and parking are small matters. Table nine is not.

Join with others for the legible layer, and stay independent at the particular layer. Buy oil together. Plan Friday alone.

If you run a large organization

Put decision rights where the knowledge is. Every time you are tempted to centralize a judgment because the central AI is good enough, ask which of the four kinds of local knowledge the decision depends on. If it depends on knowledge that is fleeting, tacit or not yet made, the center will be confidently wrong.

Adopt a bring-your-own-AI policy. Own your facts. Let your people own their method. Build the gate between them, and stop throwing away your employees’ trained delegates at the door.

Build protocols, not brains. The job of the center is to keep the signals honest and the rules clear, the way law and money do, not to know everything.

If you build AI

Build for the spot. The most valuable systems of the next decade will not be the ones that know everything centrally. They will be the ones that let a person’s own delegate act on their behalf, talk to other delegates in small, honest messages, and leave whenever its person wants.

Epilogue

The Clipboard, Again

She knows the shrimp came in short and the cook has to leave at eight and the lot next door is closed. She knows table nine. She cannot tell you how she knows the walk-in is warm.

For eighty years, economists have argued about her without meeting her. One side said a clever enough center could plan her evening. The other said only she could, and that the best the rest of us could do was send her small signals and get out of the way. The second side was right. But it left her alone with the clipboard and two hundred decisions and four hours.

That is the part that is changing. Not where the knowledge lives. It will always live on the spot, dispersed, fleeting, tacit, and made fresh each night. What is changing is that the spot can finally have intelligence of its own.

Do not send her knowledge to the center. Send the help to her.

She is on the spot. She always was.