PDCA / PDSA Cycle
The PDCA cycle is the smallest complete improvement loop: plan a change, run it on a small scale, study the result, then adopt it, abandon it or start again differently. Walter Shewhart published its shape in 1939, as a three-step loop that applies the scientific method to a production line. W. Edwards Deming carried it to Japan in 1950 as a four-step wheel. Japanese executives recast it as Plan-Do-Check-Act from 1951 on, with no author credited, and Deming never accepted it. He rejected the acronym in 1980, in 1990 and in 1991, and he settled his own version in 1993: Plan-Do-Study-Act. The objection is about the work, not the vocabulary. A check verifies that a result conforms. A study compares the observed result against the prediction written at planning time, then corrects the theory when the two disagree.
Goal
The PDCA cycle runs a change as an experiment. The team writes down what it will do, what it expects the change to produce and how it will read the observations, then starts again: the result comes from the number of cycles.
Its deliverable is a cycle register: one row per cycle carries the theory under test, the numeric prediction, the observation, the gap between the two and the decision. That register makes the decision traceable. It also accumulates what the department has learned about its own process, including from the cycles that contradicted the prediction.
The loop settles one question: does our theory of the problem hold up against the measurement? A cycle that answers no has delivered what was asked of it.
Usage
When to use it
- A theory of the problem stated but never verified: the cycle puts it to the test against a dated observation.
- A change whose effect is still uncertain: the trial says what it produces before anyone extends it.
- A trial that is possible on a small scale: one counter and two weeks cost less than the roll-out they prevent.
- A result measurable in days or weeks: the loop is only as good as its feedback delay.
- A process already measured: an existing indicator gives the basis for comparison.
- A team that takes the observation itself: the measurement is taken from the process, not from a monthly report.
- A climate where a failed cycle is reported without repercussions: a contradicted prediction is the most useful information the cycle carries.
- A distributed team with a shared register: the prediction written before the action holds the loop together at a distance.
- Improvement held to a cadence: the cycle is the base unit of Toyota kata and of continuous improvement work.
When not to use it
- A single irreversible decision: no small-scale trial is possible, so use decision analysis, which weighs the options against criteria.
- The cause of the problem still unknown: the Plan step calls for a theory, so build it first with root cause analysis.
- No measurement of how the work runs today: the Study step has nothing to compare, so establish the baseline with process analysis.
Description
Shewhart's three-step loop
Walter Shewhart published the original shape in 1939, in Statistical Method from the Viewpoint of Quality Control. It has three steps: specification, production, inspection. Shewhart gives the equivalence himself, on page 45.
"These three steps must go in a circle instead of a straight line [...] specification, production, and inspection correspond respectively to making a hypothesis, carrying out an experiment, and testing the hypothesis. The three steps constitute a dynamic scientific process of acquiring knowledge."
Deming edited the lectures this book came out of. The argument he would defend for forty years is already in it: inspection does not record whether a part conforms, it tests the hypothesis that produced the specification.
The 1950 wheel and the 1951 acronym
In 1950, JUSE invited Deming to give an eight-day seminar to Japanese executives and engineers. He presented a four-step wheel there: design the product, make it and test it in production and in the laboratory, put it on the market, then test it in service through market research before redesigning it. He insisted that the four steps rotate without stopping. The Japanese named it the Deming wheel.
The PDCA acronym is not Deming's. According to Moen and Norman, Japanese executives recast the wheel as Plan-Do-Check-Act after the 1950 seminar. Masaaki Imai reports this in 1986 in Kaizen without naming anyone, and authorship was never claimed. Kaoru Ishikawa redefined the quadrants again in 1985, lodging goal setting and training inside Plan and Do.
Deming took up his own version again in 1986 in Out of the Crisis, as a six-step procedure, then settled it in 1993 in The New Economics under the name Shewhart Cycle for Learning and Improvement. The step that follows execution is labelled Study, in his words: "Examine the results. What did we learn? What went wrong?"
The four steps
Deming states the four steps in 1993 in The New Economics.
- Plan: plan a change or test aimed at improvement.
- Do: carry out the change or test, preferably on a small scale.
- Study: examine the results, asking what was learned and what went wrong.
- Act: adopt the change, abandon it or run another cycle with different conditions.
The 1986 version loads the Plan step with a second job: decide how the observations will be used. The plan names the measure, the person who takes it, the population and the period, then it writes the decision rule for the Act step: at what result the change is adopted, at what result it is abandoned, at what result the cycle starts again under different conditions. The rule written at Plan time is read at Act time. Written at Act time, it is negotiated.
To that same 1986 version, Deming adds a note that holds for every step: each one may call for the support of statistical method, for economy and for speed. That support also protects against the false conclusions produced by a trial that has neither tested nor measured the effects of interactions between factors.
Abandoning is an outcome in its own right, on the same footing as adopting. A cycle whose only planned exit is deployment is no longer an experiment.
Check or Study
The Japanese PDCA of 1951 fits in four lines, which Moen and Norman reproduce.
- Plan: define the problem and hypothesise its possible causes and solutions.
- Do: implement a solution.
- Check: evaluate the results.
- Act: standardise the solution if the results are satisfactory, otherwise return to the Plan step.
This Check evaluates the results with no second number to compare them against, and this Act has only two outcomes. Abandoning, which the 1993 version sets as an exit in its own right, is absent from it.
ASQ, the American professional body for quality, still teaches the popular form under the name PDCA: spot an opportunity and plan a change, try it on a small scale, examine the trial and analyse its results, then act on what has been learned, either extending the change if it worked or running another cycle with a different plan. The body already puts the words "identify what you've learned" inside its own Check. That is Deming's vocabulary. Common practice mixes the two forms.
Deming rejected that blending. He warned his audiences that the PDCA form is inaccurate because the English word check means "to hold back". Moen and Norman report it. The W. Edwards Deming Institute states the practical consequence: Check is about whether an implementation succeeded or failed, where Deming wanted the result predicted, the result obtained studied and the two compared in order to revise the theory.
| The third step | Check | Study |
|---|---|---|
| The question asked | Did it work? | What did we learn, and what can we predict now? |
| What is compared | The result obtained, against a conformity criterion. | The result observed, against the prediction written at Plan time. |
| What the step can correct | The rest of the roll-out: extend, correct, redo. | The theory of the problem, hence the content of every cycle that follows. |
| A result contrary to expectation | A failure, to be redone. | The cycle's main yield: the gap shows where the theory is wrong. |
The team reads the distance between the two numbers. That is what Deming calls learning.
On 19 August 1980, at a roundtable of the GAO, the United States General Accounting Office, Deming was asked how the plan-do-check-act of the quality circles related to his own design-make-sell-test wheel. He answered: "They bear no relation to each other." He then described his own wheel as a plan for management, and the quality circle as a group of people working on the faults that arise locally. The separation he draws runs between levels of the organisation, not between third steps. In a 1990 letter to Ronald Moen: "call it PDSA, not the corruption PDCA". In a 1991 letter: "What you propose is not the Deming Cycle. I don't know the source of the cycle that you propose. How the PDCA ever came into existence I know not." Ronald Moen and Clifford Norman report these three positions in "Circling Back", published in Quality Progress in 2010.
The prediction written before the action
The Study step rests on one condition: the plan must state a prediction. Moen, Nolan and Provost set that rule in 1991. Without it, the study has only one number, and it falls back into a check whatever the step is called.
A prediction is a number, with a direction, a magnitude, a measure that already exists and a period. "The follow-up rate will fall from 38% to under 15% for this month's requests" is a prediction: an observation can contradict it. "The new form will improve the quality of requests" is a wish, and nothing can contradict it.
The gap between the predicted number and the recorded number measures the distance between what the team believes about the process and how the process actually runs. That is what the cycle produces. Moen, Nolan and Provost say why: finding that a change produced a gain during the trial is not enough. The team will have to say whether it holds under the conditions it meets next, and that extrapolation calls for a theory that only a prediction puts to the test.
Langley, Nolan and Nolan attach three questions upstream of the loop in 1994, under the name Model for Improvement: What are we trying to accomplish? How will we know that a change is an improvement? What change can we make that will result in improvement? The second forces the team to name the measure before touching the process. A prediction is written on that measure. Moen and Norman report both contributions.
The cycle's successors
Toyota adopted PDCA from 1961 on, with company-wide quality control. Art Smalley says the loop became the reasoning discipline expected there of every manager and every team leader, rather than a tool reserved for the quality function.
Toyota kata makes the cycle its unit of progress: each step toward the target condition is a cycle, and the coach's card calls for the written expectation before the action. DMAIC and QC Story take up the same loop, adding named stages and deliverables. The loop remains the smallest complete form, and one cycle runs in an afternoon when the subject allows it.
The mistakes that hollow out the cycle
A prediction written after the fact is the commonest mistake and the hardest to spot once the register is filled. The cycle runs, the number comes in, someone fills the "expected" column with a value close to it. The loop then returns a perfect success rate that has tested nothing. The guard against it is a date: the prediction carries the date of the day the cycle began.
A check kept under the name Study changes nothing as long as the question asked in the session stays "did it work?". The guard is the reading order: in the session, the team reads the prediction aloud before the observation, and the comparison of prediction with observation is forced by the order rather than by the name of the step.
A single cycle is a trial, not the technique. The first one mostly serves to find out which theory to test next. The value arrives on the second cycle or the third, when successive gaps map where the process resists. Deming writes this into his own procedure. His 1986 version has six steps and carries no Act step: it provides no exit, only the next cycle. Its last two steps call for repeating the first with the knowledge accumulated, then the second and so on.
Full-scale execution costs twice over. A change that turns out to be bad has been paid for everywhere, and the conditions of a full roll-out move too many variables at once for the gap to be attributed to the change any longer. That is why Deming writes "preferably on a small scale".
The cycle filled in for the steering committee loses the loop: four quadrants on a slide, filled in once the work is finished, describe a completed project in the vocabulary of continuous improvement.
A missing register ruins the series. The second cycle starts the first one over and nobody can read ten cycles in a row. What the team has learned then leaves with the person who remembered it.
AI considerations
An assistant keeps the register. After a session, the team dictates the cycle to it, and it files each element in its column. It then flags the rows where the prediction carries the same date as the observation. That is the mechanical trace of a prediction written after the fact.
It also contests the team's reading of the gap. Faced with a result that contradicts the prediction, a model lists rival explanations the team no longer sees, and it says whether the size of the trial supports the conclusion the team wants to draw from it. With 74 requests, a two-point gap stays inside the noise.
The prediction has to stay the team's own. A model asked about the expected effect of a change returns a reasoned range, which the team is tempted to copy. The cycle then no longer tests what the team believed, so it no longer corrects its theory. The useful practice runs the other way: the team writes its prediction first, then asks the model for its own, and the gap between the two is discussed before execution.
The observation is out of its reach. Asked about a process it has not seen, a model returns plausible values. An invented observation corrupts the gap, so the revised theory, so every cycle that follows. The decision at the Act step is out of reach too, because adopting or abandoning commits people and a budget.
Examples
The contrôle des habitants of a Vaud commune, the office that keeps the residents' register, issues certificates of residence. Some requests arrive incomplete and trigger a follow-up with the applicant. The register carries the first two cycles.
| Cycle 1, 4 to 17 March | Cycle 2, 18 to 31 March | |
|---|---|---|
| Theory | The follow-ups come from fields left empty on the paper form. | The follow-ups come from the supporting document, which the form does not require. |
| Measure | Share of certificate requests that give rise to at least one follow-up, counted on the counter's register of requests. Baseline: 38% on the 1'640 requests of the previous year. | The same measure, on the same register, at the same counter. |
| Prediction | An online form with mandatory fields brings the follow-up rate down from 38% to under 15%. | Blocking submission until the document is attached brings the rate under 15%. |
| Execution | One counter, two weeks, 74 requests. | The same counter, two weeks, 81 requests. |
| Observation | 31% follow-ups, or 23 requests out of 74. | 11% follow-ups, or 9 requests out of 81. |
| Gap | 16 points above the predicted ceiling. All 23 follow-ups are about a missing document, none about an empty field. | 4 points under the predicted ceiling. The 9 follow-ups are about illegible documents. |
| Decision | Theory ruled out, new theory written, second cycle at the same counter. | Adopt at all four counters, re-measure after a month. |
The cycle that yielded most is the first, the one whose prediction was wrong. It cost one counter and two weeks, and it ruled out an explanation the office had taken for granted for years. A check would have recorded it as a failed trial. The "gap" row carries the next theory instead, already written: the cause is not the empty field but the missing document.
Visualisations
The cycle is a loop, and the dispute is about one of its four steps only. A drawing shows that at a glance: the four arcs, the direction of rotation and the third arc with its double label. A grid of rows and columns would not say that the loop rotates. The lineage calls for a drawing too, because what it shows is a branching: the acronym's branch leaves the line in 1951 and Deming's line runs on elsewhere until 1993. The comparison of Check and Study is read criterion by criterion and stays a table. The cycle register is a table as well, and it is shown filled in rather than described.
Cost
| Dimension | Level | Justification |
|---|---|---|
| Preparation | Low | No template to produce and no workshop to convene. It takes a written theory, a numeric prediction and a measure the process already produces. |
| Execution | Medium | One cycle runs on a small scale in one to two weeks. The technique calls for several in a row, and that is where the cost sits. |
| Documentation | Low | One row per cycle. The prediction is written there before execution. |
Tooling
One column on an improvement board posted next to the process is enough to keep the register, and the prediction becomes public there before execution.
A spreadsheet holds the series well, with the date of each cycle. It becomes necessary as soon as anyone wants to read ten cycles to see where the process gave way and where it still resists.
The learner's storyboard of Toyota kata carries a PDCA cycle register already laid out. The A3 problem solving sheet records the outcome of a series on a single page, when it has to be taken to a management body.
A control chart separates a real shift from the ordinary variation of the process. Deming's note calls for that support. The chart becomes necessary as soon as the predicted gap is of the same order of magnitude as the noise of the measurement.
A task-tracking tool is a poor fit for cycles. The ticket has no field for the prediction, and the cycle reverts to a task to be executed. Anyone who keeps their cycles in Jira or Azure DevOps anyway adds that field before opening the first ticket.
Sources
- Shewhart, Walter A. (1939), Statistical Method from the Viewpoint of Quality Control: the three-step loop and the equivalence between specification, production, inspection and hypothesis, experiment, test of the hypothesis, page 45.
- Deming, W. Edwards (1986), Out of the Crisis, MIT Press: the six-step procedure, with no Act step, as well as the note on the support of statistical method, page 88, as Moen and Norman reproduce them.
- Deming, W. Edwards (1993), The New Economics for Industry, Government, Education, MIT Press: the Shewhart Cycle for Learning and Improvement and the wording of the four steps, as Moen and Norman reproduce them. Moen (2009) gives page 135.
- Moen, Ronald and Norman, Clifford (2010), "Circling Back", Quality Progress, ASQ: the full lineage from 1939 to 1993, the Japanese PDCA of 1951 after Imai, Ishikawa's redefinition of 1985, Deming's warning about the word check and the three positions taken against the acronym, with their dates.
- Moen, Ronald (2009), "Foundation and History of the PDSA Cycle", Asian Network for Quality, Tokyo: the long version of the same lineage, with the series of drawings of every version of the cycle.
- The W. Edwards Deming Institute, "PDSA Cycle": the practical consequence of the difference between Check and Study, as well as the four steps in plain language.
- American Society for Quality, "PDCA Cycle": the form the professional body teaches, with the detail of its four stages.
- Moen, Ronald, Nolan, Thomas and Provost, Lloyd (1991), Improving Quality Through Planned Experimentation, McGraw-Hill, p. 11: the obligation to state a prediction at the Plan step and the reason for that obligation, as Moen and Norman report them.
- Langley, Gerald, Nolan, Kevin and Nolan, Thomas (1994), "The Foundation of Improvement", Quality Progress, June 1994, p. 81: the Model for Improvement and its three questions, as Moen and Norman report them.
- Smalley, Art, "PDCA", Art of Lean: the adoption by Toyota from 1961 on and PDCA as the reasoning discipline expected of every manager.