Jump to text
On this page
Copy and source
Published source confirmed at last check

Source changed 2026-11-16 21:18:08 UTC · snapshot created 2026-11-16 21:19:40 UTC · last check 2026-11-17 20:05:03 UTC

E.23:4.4.1 - Improve a work process through target conditions and experiments

Use this branch when a team must improve an actual process but cannot responsibly prescribe the whole path to the desired state. The working question is: What should the process do next, what currently prevents that, and which observation or experiment can justify the next move? Rother’s Improvement Kata provides a way to answer it repeatedly. A useful local result can be a bounded target condition and an informative next step; it need not be a complete solution to the longer-term challenge.

Choose an observable process and an attainable period of work. When people are learning the way, a small process with frequent, visible repetitions makes the connection between action and consequence easier to examine. Later applications can address harder or slower work. In a long-cycle process, an offline trial or simulation can provide earlier information, while the conditions that only real operation can establish remain to be checked there. Use an already adequate way directly when the receiving task needs no inquiry or practice.

Establish direction and study the current process. State the challenge that makes this improvement worthwhile, the desired outcome and the qualities that must be protected. A direction may come from a larger organizational target, but the team still needs to understand its own process. A long-term direction is not a predetermined list of intermediate solutions.

Observe where the work happens. Follow the material, information or customer request through the relevant steps; identify how work arrives, how an output is completed and where waiting, rework or variation occurs. Compare recent outcome observations with the required performance. Then examine the operating pattern that could explain that difference. For repeated output, observations at a consistent completion point distinguish the interval between outputs from the time spent on an individual task. Retain unusual observations and the circumstances that may explain them. An average can conceal the interruption or variant that should change the next action.

Use the analysis that the process requires. Where demand and available time govern output, derive the needed rate from those values. Where equipment is involved, distinguish unattended running, handling that delays the next cycle, and work that can occur in parallel. Where staffing capacity is the question, examine repeatable work content and variation before treating a simplified capacity estimate as available capacity. For office or project work, trace actual cases, their information dependencies and delays. These distinctions select a useful next question; they do not impose one manufacturing worksheet on every process. Initial observations can be sufficient to choose a step even while other facts remain unknown. Refresh them after intervening changes.

Construct the next target condition. Combine a near achieve-by date, the outcome sought within that scope, and the operating attributes expected to produce it. For example, a target might concern how complete inputs become available, how often an output is completed, or how work passes between two operations. Keep safety, correctness and other protected qualities explicit. The target describes a desired operating state; buying software or instructing operators describes a possible means of reaching it.

Derive this target from the current observations and the direction. Choose enough stretch to require useful learning, with a horizon and scope in which the team can act and see consequences. A wider ambition may first require a smaller target. Keep the operating attributes already adequate and leave genuinely unknown details unresolved rather than filling a form with invented knowledge. When a missing fact prevents even a meaningful target, obtain that fact first. The next target after this one will depend on what is learned; it need not be fixed now.

Select an obstacle and the boundary of present knowledge. Ask what observed condition prevents this target. A current queue, an inconsistent input or an unexplained variation can be an obstacle. “We have not installed our preferred solution” has not yet identified one. Describe what that solution would change and check whether the corresponding problem is present. An acknowledged lack of knowledge can itself determine the first step.

Keep a provisional list when several possible impediments would otherwise be lost, and select one relevant, manageable obstacle for the next investigation. The list is not a commitment to remove every item. Later observations can show that an item was mistaken, has disappeared or no longer matters to the target. A beginner may learn from a tractable obstacle; selecting the largest one is not a universal prerequisite.

Choose the next step from what is not yet known. Three useful choices have different results:

  • Observe the process without changing it when the current condition or the supposed obstacle is insufficiently known. State what the observation should distinguish, where to look and what facts to bring back. An expectation about what will be seen helps expose assumptions; it need not be a fabricated numerical effect prediction.
  • Try a bounded exploratory change when a small, permissible perturbation can reveal how the process responds. Predict the response as far as the present knowledge permits and identify what remains uncertain. Protect the people, product and continuing work if the prediction fails.
  • Test a particular explanation when the proposed causal relationship is clear enough to discriminate. State the predicted effect, the observations and comparison that could support or challenge it, and the conditions needed for that inference.

Match the comparison to the claim. To learn how a component affects performance, arrange variation that distinguishes its contribution from other changes. When components may interact, compare the relevant combinations: the effect of one change may depend on whether another is present. A qualified factorial design can support those distinctions.

When the decision concerns adopting a complete new way of working, compare that way with the available alternative under comparable receiving conditions; the warranted conclusion concerns their overall difference. A comparison of whole ways does not by itself identify their components’ separate effects. The subject’s experimental method determines allocation, replication, measurement and analysis. Carry changes in inputs, assistance or operating conditions into that design or restrict the resulting claim.

Choose the size and timing of the next contribution. Select the least burdensome sufficient action or coordinated set of actions for the decision, including observation, analysis or preparation where needed. One obstacle may require several factors or several experiments to be considered together. When the experimental arrangement remains unsettled, compare feasible alternatives, such as a single trial, a planned batch or an adaptive sequence, by the information they can supply, their full preparation and execution burden, delay and risk.

With long feedback, a batch can examine useful combinations in one waiting period; an adaptive sequence can justify additional waiting when an earlier result can usefully change a later experiment. A planned sequence may also need preparation whose value appears only in the later comparison. C.40:4.7 supplies the related choice of how to search and compare candidate ways; the relevant domain method supplies the experiment itself.

Before acting, agree the intended action or set, the expected result or learning, how it will be observed, the limits on harm and the point at which useful information should be available. A short trial is useful only if the process has had time and opportunity to produce the relevant observation. Involve the people who perform and understand the work. A coach can help the learner design this contribution while leaving the learner’s needed judgement observable.

Compare and decide what follows. After the step, distinguish the observed facts from the explanation drawn from them. Compare them with the expectation stated before acting. The result may support the explanation, contradict it, partly fit it or leave the question unresolved. Use the difference to update the current condition and choose what now needs to be learned or changed. A successful prediction can also provide useful knowledge. Rother uses experimentation broadly for this learning sequence; an observation or exploratory trial does not acquire the warrant of a controlled causal test by that name.

Retain a useful change only within the conditions its evidence supports. Before sustained use, check the protected qualities, the affected process as a whole and the operating conditions under which the change must continue to work. A successful trial and a stable working arrangement are different results. Return to §4.3 for the justified improvement continuation or stop; observation undertaken to fill a missing basis retains its preparatory status. Keep an adequate current arrangement when another change has no worthwhile contribution.

When the target is achieved or its agreed date arrives, compare the result with the original target and review what the process and the participants have learned. A missed target remains a missed target for that interval. Decide the next target from the resulting current condition and direction. If demand, safety, scope or the larger purpose has materially changed, reconsider the target and its grounds; do not preserve an obsolete target merely to keep a routine going or rewrite the old result as success. HCD.9:4.10 supplies the separate review of the learner’s reasoning and the coach’s help.

Worked example — a contract-entry bottleneck. This constructed example adapts the reasoning difficulty in the Practice Guide’s sales case; its times and subsequent trials are illustrative, not measurements reported by Rother.

A team wants an ordinary contract completed within 20 minutes. Older observations suggest 31–39 minutes overall, including 10–14 minutes entering data. The learner proposes automating entry. The coach asks when those observations were made and what now consumes the entry time. Because the available observations predate a change to the input form, the next useful step is to observe current contracts, separate typing from lookup and correction, and bring back the sequence and times. The team does not yet have a justified automation proposal.

Current observations of the selected ordinary variant show 11–13 minutes of entry, including three to five minutes repeatedly searching for approved item codes. The team agrees a near target: by Friday, entry for this variant should take at most eight minutes, with verified inputs available when entry starts, no invalid codes and no added delay elsewhere. The 20-minute overall challenge remains open. The desired input condition does not prescribe whether its means should be software, a changed handover or another arrangement.

The learner now predicts that a view of the existing approved code source, available at entry, will remove repeated lookup without adding a separate preparation task. A safe offline comparison uses matched saved contracts, keeps their content and correctness checks comparable, and alternates which procedure is tried first. This helps distinguish the changed lookup procedure from simple familiarity with a particular contract. Both code correctness and time are observed. In the constructed result, entry takes seven to eight minutes and the codes remain correct for the tested variant. That is a reason to examine a bounded supervised application and the whole-process consequences. It is not proof that every contract can now be completed within 20 minutes. Representative live work and maintained access still determine whether the change can become the ordinary arrangement.

The same contract work can pose a different experimental question. Suppose the proposed route changes both the input handover and the code view. If the team needs to understand whether the changes work separately or together, it needs observations that distinguish the four combinations, with the assignment of contracts, repeated observations and analysis chosen to support that inference. If it only needs to choose between the complete new route and the current route, compare those two arrangements for correctness, completion time and total work. That second comparison can support the route choice without identifying which component caused the difference.

A condition that changes the next move. A special contract needs a customer choice before an item code can be approved. A faster lookup cannot supply that missing decision. If this contract type is now required, the team first establishes when and by whom the choice can be obtained, and revisits the input handover or target scope. It preserves the demonstrated benefit for the ordinary variant without claiming the broader target. If the customer response time makes Friday’s broader target infeasible, that fact changes the target decision; it is not a reason to keep running identical lookup experiments.

The process and coaching sequence is developed from Mike Rother’s The Toyota Kata Practice Guide, the 2018 electronic edition, Chapters 5–9 and 11–12. It supplies current-condition analysis, target-condition construction and learning at the boundary of knowledge. The experimental distinctions use the NIST/SEMATECH handbook’s interaction, factorial-design and randomized-comparison explanations. Treating a complete way as the compared alternative is the application made here. Huan, Jagalur and Marzouk, revised 2026, §§5 and 6.3 distinguish batch and adaptive experimental design and its modeling and computational demands; model-based experimental-design optimisation requires a suitable model and enough expected benefit to justify its computation. The contract examples are constructed applications. Rother supplies a practice prescription and teaching cases; the experimental-design sources justify comparisons under their stated model conditions.