PDCA is the scientific method behind kaizen. It turns improvement ideas into experiments and keeps teams from rolling out untested fixes.
Where it comes from
PDCA builds on Walter Shewhart's cycle and was popularized in Japan by W. Edwards Deming, who later preferred the term PDSA for plan, do, study, act. Toyota made it central to its problem solving.
At a glance
The PDCA wheel
Swipe sideways to see the whole diagram.
How to apply PDCA
- 1
Plan
Define the problem, the current condition and the target. Form a hypothesis about what change will close the gap and how you will measure it.
- 2
Do
Run the change on a small scale, such as one shift or one cell. Record what actually happens.
- 3
Check
Compare results to the prediction and the baseline. Look for side effects as well as the intended gain.
- 4
Act
If the change worked, standardize it and spread it. If not, adjust the hypothesis and start another cycle.
Worked example (illustrative)
A warehouse team predicts that moving the 50 fastest-moving items closer to packing will cut walking time per pick route by at least 15%. They try it in one aisle for a week and time 40 pick routes. Average walking time falls from 6.0 to 4.8 minutes, a 20% drop, which beats the prediction, so they re-slot the remaining aisles and update the slotting rule.
Common mistakes
- Skipping the check step and treating every change as a success.
- Planning large rollouts instead of small, fast experiments.
- Not writing down a prediction, so there is nothing to learn from the results.
When PDCA is not the right tool
PDCA is for change with an uncertain outcome. Routine work that already meets its standard runs on SDCA instead: standardize, do, check, act.
It also needs a prediction. Without a stated expected result, the check step has nothing to compare against.
PDCA meaning
PDCA stands for plan, do, check, act. It is a four-step cycle for testing a change on a small scale before adopting it.
Plan the change from the cause of a problem. Do it on a small scale. Check the result against the plan. Act by adopting, adjusting or dropping the change.
PDCA cycle example
This example is illustrative. About 12% of a team's orders ship late because of missing paperwork. Plan: the team adds a checklist at order entry and predicts late orders will fall below 5%.
Do: one shift uses the checklist for two weeks. Check: late orders on that shift fall to 4%, while the other shifts stay near 12%. Act: the checklist becomes the standard for every shift.
The Deming cycle and the Shewhart cycle
The cycle grew from Walter Shewhart's work on quality in the 1920s and 1930s. W. Edwards Deming taught it in Japan from 1950, so it is often called the Deming cycle.
Deming later preferred plan, do, study, act, or PDSA. Study stresses learning from the result rather than only checking it. Toyota made the cycle central to its problem solving, and the A3 report puts one PDCA cycle on a single page.
PDCA and SDCA
SDCA keeps a process at its current standard. PDCA raises the standard. Teams move between the two: improve with PDCA, then hold the gain with SDCA.
Standard work is the S in SDCA.
PDCA in Toyota Kata
Mike Rother's book Toyota Kata describes short, repeated PDCA experiments with a coach. Each one moves a team a step toward a target condition.
A one-page PDCA log
Keep one line per cycle with five fields: the change, the prediction with a number, how you measured, the result, and the decision to adopt, adjust or drop.
For the illustrative warehouse test: re-slot the 50 fastest items in aisle 3; predict 15% less walking per route; time 40 routes over a week; walking fell 20%, from 6.0 to 4.8 minutes; adopt in all aisles.
A log like this shows which predictions were wrong, which is where the team learns most.
What each stage should produce
Give every stage a concrete output, so moving round the loop means more than moving a card:
1. Plan
What will we test, and why?
OutputA defined problem, a prediction and a test plan
2. Do
What happened during the test?
OutputImplementation notes and the evidence collected
4. Act
What happens next?
OutputA decision to adopt, adapt or abandon, with an owner
3. Check
What did the evidence teach us?
OutputA comparison with the prediction, and its limits
Before approving the next cycle, point to the finding that justifies it.
Planning a test in seven steps
Start from the gap, not the fix. We need automation names a solution; approved requests wait in a queue before assignment names a problem. Then:
- 1
Describe the problem. What happens, where, and who feels the consequences.
- 2
Watch the real work. Follow actual cases, including exceptions and handoffs, not only the written procedure.
- 3
Set a baseline. Record current performance with definitions you can reuse during the test, including where the process starts and ends.
- 4
Name a likely cause. Separate observations from assumptions, and note what evidence would prove you wrong.
- 5
Bound the change. Say which people, cases, place or workflow the test covers.
- 6
Write the prediction. What you expect to change, by how much and why.
- 7
Set review and stop rules. When you will look at results, and what would pause the test.
Replace improve communication with test a required owner field on incoming requests, because unassigned requests seem to wait longer. That names both a change and an explanation to check.
What to log during Do
Run the change small, brief the people involved and keep a short log:
- Coverage: Which cases got the change and which did not.
- Fidelity: Whether people followed the intended method.
- Conditions: Interruptions, staffing changes or unusual workload.
- Extra effort: What it took to make the test work.
- Concerns: What the people doing or receiving the work said.
If someone bypasses the new step because it does not fit a real case, record the case instead of quietly dropping it. And do not add more fixes halfway through without noting when and why, or you cannot tell which version produced the result.
Four questions for Check
Deming preferred the word study for this stage, because the point is learning from the gap between prediction and result, not pass or fail. Review the outcome and how the test ran:
Result
What changed in the agreed measures?
Execution
Was the planned change actually delivered?
Context
What else changed during the test?
Learning
Which assumptions gained support, and which are still open?
Keep observations apart from explanations. Three requests lacked an owner is an observation; the form instructions were unclear is an explanation to check. When evidence is thin, say so: promising, but not yet tested across shifts is more useful than a verdict the data cannot carry.
Act: adopt, adapt or abandon
Act turns learning into a decision that someone else could carry out:
What does the evidence support?
Adopt
Update instructions, remove old versions, name the process owner and agree how it will be monitored. This is the hand-off to SDCA.
Adapt
Change the idea or the test conditions and run another cycle, for example with a different shift.
Abandon
Drop the idea, and record what the test ruled out and what is still unknown.
Continue improving intake is too vague. Revise the owner-field instructions, test with the evening team and review missing assignments at next Tuesday's meeting is a next step. Close every cycle with an owner and a review date.
A worked example: request assignment
An illustrative internal support team receives requests through a shared form. Some wait because nobody can tell who should own them:
- Plan
Assignment time is defined as business time from a complete request to a named owner. Of 50 recent requests, 15 lacked enough category information to route. The test: a clearer category field with short descriptions, for one request type over two weeks. Prediction: fewer clarification messages and faster assignment. Countercheck: requesters must not need more help to submit.
- Do
The pilot group uses the revised form while everything else stays the same. The log shows several requesters picking an overly broad category, and the descriptions display badly on small screens. Both are recorded, not fixed quietly.
- Check
4 of 30 pilot requests need clarification, against 15 of 50 before. The earlier sample covered more request types, so the team compares like with like and reads the assignment times too. The broad category suggests the explanation was only partly right.
- Act
Adapt. Split the broad category, fix the small-screen layout and run the next cycle with a different group of requesters before making the form standard.
Requests needing clarification
Keep denominators beside every share. Four clarifications means something different in 10 requests than in 100.
Measuring result, delivery and side effects
Choose measures from the decision you need to make, and include what people say as well as numbers:
| Evidence | Purpose | Request example |
|---|---|---|
| Desired result | Whether the original problem improved | Time until a named owner is assigned |
| Delivery of the change | Whether the test ran as planned | Share of eligible requests using the revised form |
| Unwanted effects | Whether the difficulty moved elsewhere | Requester confusion and help needed to submit |
Desired result
Whether the original problem improved
Request example: Time until a named owner is assigned
Delivery of the change
Whether the test ran as planned
Request example: Share of eligible requests using the revised form
Unwanted effects
Whether the difficulty moved elsewhere
Request example: Requester confusion and help needed to submit
Define each measure before collecting it: start and end events, cases included and excluded, source and owner. Look at unusual cases as well as averages, and ask participants what got easier, what got harder and what needed a workaround.
PDCA or DMAIC?
Choose by the investigation and control the problem needs:
| Point | PDCA | DMAIC |
|---|---|---|
| Stages | Plan, do, check, act | Define, measure, analyze, improve, control |
| Shape | A short loop that repeats | Distinct phases for diagnosis before the fix |
| Holding the gain | Part of Act | A separate Control phase with a control plan |
| Ask first | Can we frame a useful, bounded test? | Do we need a full roadmap for diagnosis and control? |
Stages
Plan, do, check, act
DMAIC: Define, measure, analyze, improve, control
Shape
A short loop that repeats
DMAIC: Distinct phases for diagnosis before the fix
Holding the gain
Part of Act
DMAIC: A separate Control phase with a control plan
Ask first
Can we frame a useful, bounded test?
DMAIC: Do we need a full roadmap for diagnosis and control?
PDCA does not forbid analysis, and DMAIC includes pilots. The DMAIC page covers when a full project is worth it.
Good first PDCA tests
Pick a process you can watch directly, with an owner and accessible evidence. Illustrative starting points:
Administrative handoffs
Test a standard handoff note, and watch missing information, reassignment and waiting.
Production preparation
Test a revised setup checklist, and track omissions, preparation time and operator feedback.
Customer support
Test a clearer reply template, and look at repeat contacts as well as speed.
Personal planning
Test a protected work period, and record finished priorities, interruptions and what got pushed aside.
A small pilot is not permission to skip safety or approval rules. Get the usual sign-off for changes with real consequences.
PDCA mistakes and fixes
- 1
Stopping at Check. Plan, do, check without a recorded decision leaves nothing changed. Add the decision, an owner and a review date.
- 2
Solution before problem. Rewrite we need a dashboard as the missing information, who needs it and which decision it affects.
- 3
Several changes at once. Narrow the test, or label it a bundle and do not claim to know which part worked.
- 4
One good result as proof. Retest under normal load, other shifts and hard cases before making it standard.
- 5
A change that depends on one person. Have someone outside the pilot follow the new instructions without coaching before closing.
A PDCA working template
For tests bigger than one log line, keep one sheet per cycle and keep the sheets in order. Fields:
PDCA test record · one sheet per cycle
Cycle
- Process
- Cycle number
- Owner
- Participants
- Review date
- Link to the previous cycle
Plan
- Problem
- Scope
- Baseline
- Likely cause
- Prediction
- Test plan
- Evidence to collect
- Approvals and stop rules
Do
- Cases included
- Real dates
- Observations
- Exceptions and workarounds
- Missing data
Check
- Result against prediction and baseline
- Evidence the change was delivered
- Other explanations
- What we learned
Act
- Adopt, adapt or abandon, and why
- Updates to instructions or training
- Next test
- Owner and next review
Before filing it, ask another participant to read the learning statement and say whether the next step follows from the evidence. The A3 report puts the same loop on one page for bigger problems.
Guides on this topic
- Kaikaku: when a process needs redesign, not another small fixKaikaku explained: what radical change means in lean, how it differs from kaizen, when it fits and when to pause, a six-step way to run it, examples from factories to purchasing, measures and how to protect the change.Read the guide
- Leader standard work: routines that keep the management system runningLeader standard work explained: what to standardize, linking routines to a purpose, observation and coaching, a template, cadence by level, a five-step pilot, coaching questions, choosing an app, measures and common traps.Read the guide
- Kaizen event: meaning, steps and a 3-5 day agendaA kaizen event is a focused workshop that improves one process in 3-5 days. Learn the roles, agenda, steps and measures, with a free kaizen charter template.Read the guide
- Continuous improvement in lean: examples and strategiesContinuous improvement in lean explained: the PDCA cycle, daily kaizen and kaizen events, with practical examples and strategies that keep improvement going.Read the guide
Sources
Books and papers show their publication year. Web pages show the month we last checked them.
- 1.W. Edwards Deming, Out of the Crisis (MIT Center for Advanced Engineering Study, 1986)
- 2.W. Edwards Deming, The New Economics for Industry, Government, Education (MIT Center for Advanced Engineering Study, 1993)
- 3.Walter A. Shewhart, Economic Control of Quality of Manufactured Product (D. Van Nostrand, 1931)
- 4.Ronald D. Moen and Clifford L. Norman, Circling Back: Clearing Up Myths About the Deming Cycle and Seeing How It Keeps Evolving (Quality Progress, November 2010)
- 5.Gerald J. Langley, Ronald D. Moen, Kevin M. Nolan, Thomas W. Nolan, Clifford L. Norman and Lloyd P. Provost, The Improvement Guide: A Practical Approach to Enhancing Organizational Performance, 2nd edition (Jossey-Bass, 2009)
- 6.Mike Rother, Toyota Kata: Managing People for Improvement, Adaptiveness and Superior Results (McGraw-Hill, 2009)