Slack Is a Forecast

Every buffer in the world is a sentence written in steel. In 1888, Francis Edgeworth asked how much gold a bank must keep in the vault, and noticed the answer had nothing to do with the average day. The average day is already handled. Reserves exist for the scatter around the average — the days the model misses — and so their size is set by the width of that scatter, not by the center of it. In 1951, Kenneth Arrow and two colleagues made it doctrine for everything on a shelf: the buffer a rational system holds is a multiple of σ, the standard deviation of its forecast error. Every warehouse, every hospital bed held open, every reserve generator spinning unloaded on the grid has been sized by that law ever since.

Read the law backward and it says something better. If slack is sized to forecast error, then slack is a forecast — a confession, in square feet and payroll and spare parts, of exactly how much a system trusts its own prediction machine. Walk any warehouse and you are reading a belief. A packed one says: we expect to be badly wrong, often. A lean one says: we expect to be nearly right, nearly always. Neither is a virtue. Each is only honest if the confession matches the sins.

A warehouse is not a building. It is a belief about the future with a roof on it.

The most famous act of slack-cutting in industrial history is remembered wrong. At Toyota, Taiichi Ohno named inventory a form of waste — muda — and built the just-in-time system that deleted it. The world remembers the deletion. It forgets the twin. Ohno paired just-in-time with heijunka — leveling — a relentless smoothing of the production schedule, suppliers pulled close, demand steadied upstream before a single shelf was emptied. He never defied the law of the buffer. He shrank σ first, honestly, and then let the shelf follow it down. The empty shelf was not confidence. It was arithmetic, earned.

His imitators copied the empty shelf and skipped the earning. For decades it worked, because calm years kept the difference invisible — a system sized to a small σ and a system that merely assumed a small σ look identical until the world tests them. Then the tests came. In 2011 an earthquake off Tōhoku damaged, among much else, a single supplier of automotive chips, and Toyota’s own lines — the master’s lines — ran short for months: the quake was variance from outside the leveled zone, a σ the system was never sized for. A decade later the pandemic ran the same lesson through every port on Earth at once. Both times the verdict was announced as “just-in-time failed.” It did not. A forecast failed, and the slack that would have absorbed the miss had been traded away by people who mistook Ohno’s conclusion for his method.

Now run the law forward into this decade. AI is shrinking forecast error honestly, nearly everywhere, all at once — demand, traffic, staffing, maintenance, the whole territory of the merely complicated that I mapped in The Surprise Floor. This is not hype; inside that territory the machines are simply better. And the law does not care why σ fell. It says: shrink the error and the buffer follows. So the optimizers — human ones first, and increasingly the agents we hand the keys — will walk through the economy emptying shelves, thinning rosters, packing calendars, deleting spares. Each deletion will be locally rational. The coming economy will hold less slack than any economy in history, and it will be right to, ninety-nine days in a hundred.

Two things hide from that optimization, and they are both things its own training record cannot show it. The first is the floor. Chaos guarantees a residual σ that no model shrinks, and a quiet decade of data will understate it, because the rare miss is, by definition, rarely in the sample. The model’s confidence is a measurement of calm weather. The second is correlation. When a thousand firms cut their buffers on a thousand different judgments, their misses scattered; one firm’s bad week was its neighbor’s ordinary one, and shared spares worked like insurance. When every forecast drinks from the same few models — the common denominator I wrote about in September — the misses arrive together. A pool of spare capacity sized for independent failures meets a synchronized one, and every system leaning on a neighbor finds the neighbor leaning back.

Surprise, remember, is conserved. Cutting slack does not reduce the surprise the world will deliver; it only changes the packaging. What used to arrive as a daily sprinkle of friction — stockouts, delays, dropped balls — gets swept up and repackaged into the rare cascade. Fewer bad days. Worse bad years.

1. Slack becomes a sales pitch.

   After seventy years as the definition of waste, buffers become features. Firms will advertise held inventory, humans on call, open seats, idle capacity — the way food learned to advertise what was left out of it. Watch for “we keep it in stock” and “a person answers” sold at a premium, as promises about the day the model misses.

2. Failure changes texture: rarer, larger, synchronized.

   Day-to-day friction keeps falling — fewer stockouts, fewer delays, fewer dropped balls — while the failures that do occur grow bigger, move faster, and arrive in many places at once. The record will read as triumph right up to each cascade, and the calm stretches will be cited as proof the spares aren’t needed.

3. Insurance becomes the keeper of last-resort slack, and prices it directly.

   Insurers — the institutions I argued will write the rules of AI — start underwriting the confession itself: premiums that discriminate on inventory held, redundancy kept, humans retained, the way fire policies once priced sprinklers. A firm’s buffers become a quoted, audited number.

4. Your calendar is the last warehouse.

   Personal stacks will pack days to the minute, because a life’s σ looks small on a quiet month of data. White space stops being a habit and becomes an explicit instruction you must give your machines — and keep giving, because an optimizer reads an unused hour the way Ohno read a full shelf.

5. The marquee failure of the agentic era will trace to a forecast that was almost always right.

   Not a bad model — the best model ever built, held with total confidence by a system that had traded away the capacity to be wrong. The postmortem will blame the model. The model will have performed exactly as specified. The missing line item will be the spare.

I run leaner than almost anyone I know. The drafting, the filing, the publishing of these very words — handed to machines. And I publish every day, which readers take for discipline and which is actually heijunka: a fixed cadence takes the variance out of my production, and that leveling — not bravado — is what lets the pipeline run without padding. Ohno would recognize the arrangement. But where I cannot level — the conversation that turns in its third minute, the question no curriculum predicted, the weather — I hold capacity I am not using. Most days it is waste. It is supposed to be. That is what the word spare means.

Already true

The buffer law is in every operations textbook, and it already runs the other way on command: carmakers deleted the spare tire the moment the forecast said they could. Lean’s imitators who skipped the leveling were caught by 2011 and again by 2021. AI forecast error, inside the territory of the complicated, is falling measurably year over year.

What has to happen

Agentic optimization has to reach the capacity decisions — staffing, inventory, grid reserve, schedules — and be adopted at the confidence its calm-weather record will seem to justify. And the convergence has to hold: most forecasts drawn from the same few foundation models, so that the misses stay synchronized rather than scattered.

Where I am probably wrong

The machines may be better accountants of their own error than any human ever was. A model that genuinely knows its σ — and its correlations — would buy slack with a discipline no CFO ever had, and the lean economy would also be the safe one. If that happens, this piece is a story about human overconfidence only, and the agents are its cure rather than its accelerant. The clean test: watch whether AI-run systems hold more reserve than the human-run systems they replace, or less. I am betting less. I would be glad to lose.

Back on the dark road, the forecast in your trunk was right tonight for a hundred thousand drivers, and the nail did not check it. That is the whole argument. The optimizer calls the fifth tire dead weight, and the optimizer is almost always right; the nail prices it at the whole car, and the nail only needs to be right once. Keep the spare.