Out of the Crisis

The argument

The book was written in anger, and the anger was earned. Deming had spent the 1950s teaching Japanese engineers and executives the statistical control of quality, watched them take it seriously for thirty years, and then watched American management discover the results in its own falling market share and call them unfair. Out of the Crisis is his answer to the question American industry was asking in the early 1980s, and the answer is the diagnosis nobody wanted: the crisis was not made by Japan, by unions, or by workers. It was made in the boardroom, by habits of management so normal that nobody thought to call them causes.

The foundation is a distinction Deming took from Walter Shewhart at Bell Laboratories and carried to its conclusions. Every process varies. Some of that variation comes from the system itself: the machines, the materials, the training, the procedures, the thousand small conditions management has arranged. This is common-cause variation, and it is as permanent as the system that produces it. Some variation comes from outside the system: a broken tool, an untrained substitute, a bad batch. These are special causes, findable and fixable one at a time. A process with only common causes is stable: it fluctuates within predictable limits, and nothing a worker does by trying harder will move those limits, because the limits are properties of the system, not of the effort inside it.

Two errors follow, and Deming insists both are expensive. Treat common-cause variation as if it were special, and you tamper: you adjust a stable process in response to its own noise, make the output worse, and manufacture explanations for every wiggle. Treat a special cause as if it were common, and you miss the signal: something has actually changed and you average it away. Management without the distinction commits both errors daily, punishing people for weather and sleeping through storms.

From this statistical fact the book draws its moral one, and the transfer is the whole argument. The system belongs to management. The worker works within it. Deming’s estimate is that something like ninety-four percent of troubles and possibilities for improvement belong to the system and six percent to special causes. If that is even roughly right, almost everything management does about performance is a category error: it locates in individuals causes that live in the structure, and no amount of appraising, ranking, exhorting, or firing will improve a defect rate the system has already fixed.

He demonstrates it with beads. Willing workers draw paddles of beads from a box in which red beads, the defects, are mixed with white in a fixed proportion. Procedures are strict; motivation is applied; the counts are posted; the best worker is praised and the worst put on probation. The proportion of red is decided entirely by the mixture in the box, which is to say by management. The rankings are noise. The experiment plays as comedy until you recognize the annual review in it.

The fourteen points, the part of the book everyone cites and few derive, are consequences of this, not commandments beside it. Create constancy of purpose, because a company that changes direction every quarter is a special cause inflicted on its own system. Cease dependence on mass inspection, because quality is decided upstream where the work is done, and inspecting defects out at the end is paying twice for the same failure. End the practice of awarding business on price tag alone, because a supplier’s variation flows into your process and total cost, not sticker price, is the real number. Drive out fear, because a frightened organization reports fiction, and every figure that reaches the top has been laundered on the way. Eliminate slogans, quotas, and numerical targets for the workforce, because they address the six percent as if it were the ninety-four. Abolish the annual merit rating, which rewards the noise, sets colleagues against each other where the work needs cooperation, and teaches the whole company to manage the visible number.

The deadly diseases complete the picture: emphasis on short-term profits, mobility of managers who reorganize and depart before the consequences arrive, and running a company on visible figures alone. He repeats a remark he credits to Lloyd Nelson like a refrain: the figures that matter most to management are unknown, and some are unknowable. The value of a loyal customer, the cost of a demoralized engineer, the price of fear: none of it appears on any report, and management that steers only by what appears will optimize the measurable and lose the business.

The quietest chapter may be the most durable: operational definitions. A specification without an agreed procedure for measuring it means nothing; clean, defective, uniform, on time are empty words until two people applying the same test reach the same verdict. Deming treats meaning itself as something that must be engineered, and half of every quality dispute as two parties measuring different things with the same word.

Working notes

The book is a sermon wearing the clothes of a textbook, and it converts the way sermons do: by repetition, by indignation, by the one demonstration you cannot unsee. I retain the red beads the way I retain nothing else in the quality literature. Once you have watched a ranking manufactured from pure noise, every leaderboard and every forced distribution asks the same question: what is the mixture in the box, and who mixed it?

Deming and Goldratt work the same seam from opposite ends. The Goal narrows attention to one constraint and says throughput is decided there; Deming widens attention to variation everywhere and says quality is decided by the structure. Both relocate cause from the worker to the design, and both name the same enemy: the local measure, efficiently optimized, globally ruinous. Meadows completes the set. Thinking in Systems says behavior arises from structure; Deming is that sentence with a control chart attached and a temper.

Sloan is the tension worth keeping. My Years with General Motors is the constitution of management by visible figures: decentralized divisions, centralized financial control, return on investment as the sovereign number. It built the largest company on earth, and Deming’s deadly diseases read like an autopsy of its heirs. I hold both without flinching. Financial control is how a giant is kept honest, and the moment the visible figures become the only reality, the giant begins manufacturing them. Sloan built the instrument panel; Deming states what happens to the pilot who flies by it alone.

Grove sits closer to Deming than either would admit. High Output Management pairs every indicator with its counter so that no number can be gamed in isolation; Deming goes further and removes numerical targets from the workforce entirely. The difference is who each man trusts. Grove trusts instrumentation with tension built in; Deming trusts nobody’s numbers once fear or reward has touched them, because a measure made into a target stops measuring. My resolution is narrower: figures used to understand a system are medicine, and the same figures handed down as quotas are poison; the substance is identical, and the delivery decides which.

The chapter that grows on rereading is fear. Drive out fear looks like the soft point among the fourteen, the one destined for the human resources slide. It is the hardest engineering claim in the book: fear corrupts data at the source. A frightened organization does not merely underperform; it reports fiction, and every chart drawn on fiction is decoration. Regression sharpens the cruelty. Punish the bottom of a stable system and it improves by pure reversion to the mean, teaching the manager that punishment works. The superstition confirms itself, which is why it has outlived every management fashion since.

Where I push back

The abolitions are cleaner than the replacements. Deming is right that the annual merit rating ranks noise, and right about what the ranking does to cooperation. But organizations still must decide pay, promotion, and dismissal, and his answer, institute leadership, is a job description rather than a mechanism. It presumes managers statistically literate enough to tell system from noise and honest enough to act on the difference: the rare case, and the book offers no design for the common one. Evaluation abolished without that leadership shelters the mediocre and starves the excellent, and the excellent leave first.

The single-supplier doctrine prices variation and ignores correlation. One supplier, held close in a long relationship of trust, does reduce incoming variation, exactly as he argues. It also concentrates failure: one fire, one insolvency, one port, and the whole line stops. A statistician of Deming’s caliber knew that reducing variance can raise fragility; the book never weighs the trade, and trust is not a hedge.

The doctrine also invites its own superstition. Deming leaves room for special causes; his disciples often do not, and everything becomes the system, which means nothing is ever anyone. Some failures are in fact people: the negligent, the dishonest, the wrongly hired. A theory that forbids saying so is as blind as the one it replaced, and considerably more comfortable.

And the frame has aged badly in one respect: Japan as the verdict of history. The lesson was real, but the country he held up as proof spent the following decades demonstrating that quality religion does not repeal strategy, demography, or finance. The theorems survive; the triumphalism does not. Take the book for Shewhart’s distinction carried to its consequences, and hold the prophet’s certainty at arm’s length. Certainty was his method. It is the one thing in the book that should not be adopted.

How it enters the work

Half of what Intelliblitz is hired to build is measurement, which means half of what I am hired to prevent is tampering. The monthly review that interrogates every downtick of a stable metric is the client’s most expensive meeting: the process answers with its own noise, explanations are manufactured, and the corrections widen the variation they were meant to remove. So the reporting I architect carries Deming’s distinction in its structure: control limits alongside the trend line, so a number can be read as weather or news before anyone reacts to it. A dashboard without that distinction is an anxiety machine.

Operational definitions turned out to be the most billable chapter. When two departments report different churn from the same warehouse, the disagreement is almost never in the data; it is in the word. The fix is Deming’s: one definition, one procedure, one verdict, written down. In enterprise architecture the same law returns as data quality, and it obeys his rule about inspection: quality cannot be added downstream by a cleansing layer any more than it can be inspected into a product at the end of the line. It is built at the source system or it is absent, and the architecture either enforces that or apologizes for it.

Postmortems run on his estimate. When a deployment fails or a report ships wrong, the first question is not who but what made this easy: the missing gate, the ambiguous definition, the deadline that priced in the shortcut. Blame the system first; it is usually guilty, and interrogating it costs nothing but pride.

The trading application is the strictest, because there the tamperer and the victim are the same man. A drawdown inside the strategy’s historical variation is common cause; revising the process in response to it is tampering with a stable system, and the revision usually buys the next loss. A drawdown that breaks the envelope is a special cause: a regime has moved, and averaging it away is the opposite error, the missed signal. The whole discipline of running a thesis through a cycle is the discipline of that one distinction, held under the pressure of an open position. Deming drew the chart for factories. The equity curve reads by the same rules, and its reader is worse behaved.