Process Learning: Turning Analysis into Controlled Improvement

Process Learning Begins with a Verifiable Observation

For a learning cycle, an organization must record which observation raised a question, which change it approved, and which effect it subsequently identified. This keeps the path from the initial question to the operational decision open to later review.

Reliable process events provide the foundation. They document which activity was performed, when, on which object, under which conditions, and with what result. Different process paths can be meaningfully compared only when these observations are sufficiently complete and contextualized. The Process Mining Manifesto accordingly describes process mining as a connection between event data and the discovery, monitoring, and improvement of real processes.^[1]

In Weske’s business process lifecycle, execution logs provide the basis for evaluation: they help assess and improve both the process model and its implementation, including the adequacy of the execution environment.^[5] For the waiting-time example, this widens the question: does the delay arise from the defined sequence, or are the resources needed for preparation unavailable? Changing the model alone would not resolve a resource shortage.

Example: Waiting Time Before an In-Process Test

An illustrative example: Several batches that initially appear comparable show different waiting times before an in-process test. Analysis reveals a recurring pattern: when preparation of the test equipment starts only after the preceding work step is complete, waiting time increases. When preparation starts earlier, the transition is often shorter.

The responsible roles check whether the product group, test equipment, qualifications, room state, and SOP version are comparable. They also clarify whether starting preparation earlier is permissible from a domain and organizational perspective. This leads to the hypothesis that clearly defined parallel preparation reduces waiting time without compromising test quality or traceability.

The change is approved for a few suitable runs. Preparation start and completion, test time, waiting time, deviations, rework, and result continue to be recorded as events. After the limited period, the process paths are compared again. A shorter waiting time without observed disadvantages initially provides a signal for further review. A few runs are insufficient either to provide reliable evidence of effectiveness or to rule out rare negative consequences. Before extending the change, the responsible people define the additional observation period, case base, and comparison conditions needed. If the expected effect does not occur, they examine both the hypothesis and the evidential value of the trial; an inconclusive finding does not yet refute the hypothesis.

The learning value does not lie solely in a lower number. It lies in the traceable connection between the initial observation, the decision, the controlled change, and the observed effect.

A Learning Signal Is Not Yet an Insight

A learning signal can take many forms: a process step repeatedly takes longer for one material group. A particular sequence occurs more often in successful batches. An above-average amount of rework arises after a handoff. A test is repeated with different frequencies under similar initial conditions. Such observations direct attention to a limited segment of the process.

The signal does not yet explain why the difference arises. It can be influenced by the material, a machine configuration, a role's qualifications, a different SOP version, a measurement method, or a context not yet considered. Several factors can act simultaneously. An observed relationship must therefore be treated neither automatically as a cause nor immediately as an instruction for action.

Research on action-oriented process mining likewise distinguishes between process diagnostics and the actions based on them. It shows that the transition from insights to actions requires a distinct level of design.^[2] This boundary is particularly important in regulated processes: an algorithmic signal must not lead to a change in execution without substantive review.

De Leoni distinguishes two forms of process enhancement: process extension adds data, resource, or time perspectives to an existing model; process improvement changes the model to better reflect reality or support valid, better-performing executions.^[7] Adding observed waiting times to a model therefore makes the analysis more informative, but does not itself shorten those waits. The proposal to start preparation earlier is a separate change whose effects remain to be tested.

Recurring Structures Lead to a Learning Question

Process patterns condense local sequences, branches, concurrency, or loops that recur in comparable executions. Process learning uses such patterns to formulate a specific question: Under which conditions does the pattern occur, which outcome accompanies it, and which alternative explanation must be ruled out?

A good learning question is narrower than a general improvement goal. “How can we produce faster?” is too open. “Does changing the order of test A and preparation B for product group X reduce waiting time without causing additional rework?” can instead be tied to a defined context, a change, and observable target measures.

This turns an observation into a testable hypothesis. The hypothesis remains explicitly provisional. It should state the scope considered, the assumed relationship, the expected effect, and possible counterindicators. This also allows a negative result to be retained as useful learning.

The responsible roles check permissibility, missing information, and alternative explanations. Perhaps the shifts also differ in staffing levels. In that case, the observed difference cannot readily be attributed to the sequence.

Comparability Must Be Established Before Assessment

Process paths are rarely comparable on the basis of their activity sequence alone. Two batches can show the same sequence yet have been produced under different conditions. Product variant, starting material, equipment, tool, room state, shift, SOP version, lot size, or test method can change the result. Process learning therefore requires a deliberately defined basis for comparison.

This basis defines which characteristics must be identical or at least sufficiently similar. It also documents which differences are allowed. This makes it clear whether an apparent effect actually coincides with the change being tested or merely reflects different starting conditions.

Process Mining supports this phase by collecting, connecting, filtering, and comparing real process paths. It can highlight relevant groups and recurring differences. The research field also includes predictions, recommendations, and the integration of actions. In the 420+ learning process described here, the responsible roles review substantive permissibility and decide whether to implement a measure.

In the architecture described, Process Intelligence refers to support for this assessment.

Changes Are Tested Within a Limited, Controlled Scope

A learning cycle needs an identifiable intervention. If several parameters are changed simultaneously, it is difficult to determine afterward which part influenced the observed effect. The change is therefore described as narrowly as possible: What is being changed, for which process segment, during which period, for which products, and by which authorized role?

A limited scope does not necessarily mean a scientific experiment. In operations, it can consist of a pilot group, a defined product family, an approved time window, or a few controlled runs. What matters is that the initial situation and the changed condition remain distinguishable. Safety, quality, and compliance boundaries continue to apply unchanged.

The timing of the assessment is also defined in advance. Assessing too early can confuse start-up and familiarization effects with a lasting effect; assessing too late can overlook intervening changes. The appropriate period depends on process frequency, product cycle, and risk. Many observations under nearly identical conditions do not extend the comparison base to new operating conditions.

The documentation also preserves who approved the trial and how affected employees were informed. A measure that appears clear on paper can create new handoffs, additional effort, or workarounds in actual execution. Feedback from the work therefore complements the measured process data. It is treated as a separate set of observations and is not retroactively fitted to a desired result.

Effects Are Observed Again in Actual Execution

The change is followed by a new observation phase, not an abstract retrospective exercise. Data generated during execution show whether the expected effect occurs in the defined context. The same comparison criteria as before should be retained as far as possible.

A result can support the hypothesis, contradict it, or remain inconclusive. All three outcomes matter. A refuted hypothesis prevents a plausible but inaccurate explanation from being reused later as supposed certainty. An inconclusive result shows which data, controls, or additional runs are missing.

Assessment distinguishes immediate from delayed effects. A changed sequence can reduce waiting time immediately, while consequences for the error rate or maintenance needs may appear only later. Target measures are therefore not considered in isolation. Time reference, affected objects, and possible downstream processes are part of the statement. Where an effect can be measured only indirectly, that uncertainty also remains documented.

Action-oriented process mining describes technical approaches for continuously monitoring process states and tying actions to detected conditions.^[3] Van der Aalst and Carmona emphasize in the Handbook’s closing chapter that a report alone does not improve a process: implementation requires change management and, where appropriate, automation. They therefore argue for process mining as an ongoing activity rather than an isolated project.^[8] In the trial described here, this means continuing to observe the changed preparation step after its introduction; an initial favorable comparison does not establish a lasting effect.

Learning Remains Tied to Time, Version, and Context

Processes change. New materials, equipment, products, regulations, or working methods can cause a previously observed pattern to occur less frequently, more strongly, or in a different form. Process mining research refers to such changes in observed process behavior as concept drift.^[4]

A confirmed experience must therefore not be stored as a timeless truth. It needs a scope of applicability: For which product group, SOP version, equipment configuration, and period was the effect observed? Which measures supported the assessment? Under which conditions must the statement be reviewed again?

Versioning preserves what applied at a particular time. When new experience is confirmed, it does not replace earlier experience without a trace. It supplements or limits its validity. This makes it possible to understand later why a rule was changed and which observation underlies the new version.

An Open Review Status Must Also Remain Retrievable

Process memory preserves a case and its assessment status. Process learning can extend that status through a reviewed change and new observations. An open question or a refuted assumption likewise remains retrievable as such.

For the waiting-time question, this is particularly useful when a later group wants to change the same preparation. It should be able to see what has already been tested and which uncertainty remained at the time.

What Process Learning Does Not Provide

Process learning does not guarantee improvement. A carefully reviewed change can remain ineffective or show unexpected disadvantages. Even extensive event data do not eliminate measurement errors, incomplete contexts, or unsuitable comparison groups. Quality and evidential value depend on the actual data foundation and substantive review.

Process learning does not replace quality management, change control, validation, risk assessment, or formal approvals. It provides a structured connection between these decisions and observed execution. The required procedures continue to depend on the product, process, risk, and applicable regulations.

Temporal proximity likewise does not establish causation. If a better result occurs after a change, other conditions may be involved. A sound assessment therefore requires repetition, comparability, counterindicators, and a willingness to revise an earlier assumption.

Regulatory finding

Finding: The FDA found that a sterile drug manufacturer routinely recorded microbiological results outside established limits in aseptic ISO 5 areas from 2023 to 2025. According to the letter, production continued without adequate corrective and preventive action (CAPA) to address the persistent microbiological contamination.^[6]

Assessment: Process learning connects observation, assessment, a reviewed change, and renewed observation. Recurring findings become a source of learning when a resulting change is implemented and its effect is assessed against subsequent observations. This CGMP finding illustrates the gap between repeated observation and an adequate response; it does not establish that no action at all was taken.

Review question: For each action derived from a recurring finding, is the observation needed to demonstrate its effect documented, and is that observation subsequently reviewed?

Letter dated 4 September 2026 · Source checked on 18 September 2026.

This section presents the selected regulatory finding as stated at the time of the letter. Company responses and subsequent developments are not assessed here; this account does not describe the company’s current compliance status.

Less Waiting Time Does Not Yet Answer the Entire Learning Question

In the example, earlier test-equipment preparation would warrant further pursuit only if test quality, additional effort, and other measures of potential adverse effects are also adequately investigated. A few favorable runs do not rule out rare disadvantages.

The learning value therefore lies in a limited, verifiable statement: Within the scope investigated, what supports the change, what argues against it, and what remains open? The next decision can then build on the actual review status.