Five whys do not replace evidence: how to investigate a problem
Distinguish events, conditions and hypotheses, compare evidence and test a change without inventing a cause.
Article language: English

In this article
Why did the customer receive the wrong variant? Because an old file was selected. Why was the old file selected? Because an employee did not check. Why not? Because they were in a hurry. After a few questions, the company has a cause, a recommendation to be more careful and very little new knowledge. A quick discussion does not establish that its conclusion is right.
Imagine a print shop preparing personalised cards for welcome kits. The customer received a version bearing the name of a previous event. We are not examining an accident, printing safety or legal responsibility for the order. We want to establish what information could explain an ordinary service error and help choose a change suited to what actually happened.
“Why” can be a good beginning. Each answer remains a claim needing support. If each question rests only on the previous assumption, the result is a coherent story. It may be true. The fact that it can be told in five steps does not settle it.
Start with an event you can identify
“The team has lost control of files” judges a broad area. “A document marked variant A was used to produce the cards, although the customer approved variant B” describes an event to investigate. You must still establish whether that is what happened. The question should help locate material, not assign blame from the outset.
Distinguish the outcome from how it arose. The name of the previous event on a card confirms a content mismatch. By itself it does not say whether the error arose during file preparation, version selection, handover or interpretation of approval. “Old variant” may accurately describe a document without describing its path into production.
An initial error report to the team (in Polish) should enable an appropriate response and preservation of necessary information. Explaining causes is subsequent work. Do not require the reporter to provide a complete theory before accepting the signal. Their uncertainty about the cause does not erase what they actually observed.
In our example, a manager might first identify the batch, the relevant order and the approval text. Separately they record what remains to be established: which file was handed over for production and what its label meant at that moment. This reduces the risk that everyone starts explaining a different version of the same story.
An event, a condition and an assumption have different statuses
Sending a particular attachment can be an event. Two similarly named documents in one folder can be a condition. Believing that similar names caused the selection is an assumption. The first two may be supported by material; the third needs further investigation of the relationship. Keep this distinction even in a simple note.
HSE's 2004 HSG245 guide distinguishes immediate causes, deeper conditions and the organisational level. It concerns occupational safety investigations. We use only the distinction between explanatory levels, without transferring accident procedures or old UK legislation to this service example.
At the print shop, file selection can therefore be distinguished from how the folder was prepared and from who was responsible for marking the production version. That does not mean each was a cause of this event. They are possible places to investigate. Even when an answer sounds convincing, ask what supports it and what could contradict it.
“Haste” needs explanation too. Did the employee have less time than usual? Did a change occur in that period? Had exactly the same working method previously led to the right choice? Without such questions, the condition's name can become an elegant version of “it happened because it happened”.
Three possible explanations of the same outcome
First suppose the person preparing the work identified the correct document but an old copy was attached during handover. The hypothesis concerns the relationship between marking the finished variant and transferring the file. Evidence from that moment is needed. Looking at today's folder may not reveal what was available when it was sent.
Second, the team transferred the correct file according to its understanding, but customer approval concerned another preview. The problem would lie earlier. A new attachment-selection warning might not help: the employee could consistently follow a misunderstood decision. Check the connection between approval and the specific document.
Third, the customer approved B, the team transferred B, and someone later reopened A while preparing production. That is a different point at which the discrepancy arose. Do not label all three paths “version chaos” if the company wants a change suited to the problem. A broad label can hide the moment on which the next action depends.
Each story sounds plausible. None becomes evidence by sounding plausible. State what traces can distinguish them: the identified attachment, approval record and marking of the document handed to production. If material was not preserved, honestly state what can no longer be reconstructed.
What supports a hypothesis, and what weakens it?
In Bhanu Sood's 2018 NASA Goddard workshop material, the analysis separates evidence collection, hypothesis formation and verification, and later checking of a corrective action. The presentation concerns electronics reliability. We use its order of reasoning, not engineering recommendations for a print shop.

For the wrong-attachment hypothesis, record not only material consistent with it but also information that would rule it out. If the preserved transferred file contains the correct content, look for the discrepancy later. Do not rescue the first story by adding an invisible event no one can confirm.
Conversations with people doing the work can supplement records, but need not match in every detail. One person remembers sending; another remembers opening the document later. These may be different moments rather than a conflict of honesty. Name the source and confidence level. “The employee remembers variant B” has a different status from a preserved B document from a particular handover.
A similar successful order can be a useful comparison. Were file names the same? Did approval work differently? Was a later change accepted? Such contrast can reveal a condition worth investigating. One comparison still does not establish a causal relationship by itself. It can nevertheless weaken an easy claim that every instance of haste ends in the same error.
Choose data that answers a specific question. If the discrepancy arose between approval and handover, investigating the appearance of the whole company website may add nothing. Analysis scope should follow the gap, not the number of tools a team can run.
See what “check” means in ordinary work
“Check the file before production” sounds sensible but may repeat an existing instruction. Establish what material lets the employee recognise the correct variant. If two documents have similar descriptions and the customer's confirmation identifies neither, greater attention alone will not create the missing relationship.
That is why observing an expert's work (in Polish) can help. In the fictional shop, an experienced person may call the coordinator precisely to connect approval with a variant. If instructions omit that step, the company may wrongly assume checking means a quick comparison of names. Understanding that part helps describe the task accurately.
Do not reconstruct events as though the employee knew from the start that the document was wrong. After the event everyone can see that information. At the moment of selection, the person saw particular labels, messages and available cues. Try to establish that state of knowledge if you want to understand the action rather than merely judge its outcome.
If a hypothesis concerns interruption frequency or the share of cases needing consultation, sampling the structure of working time (in Polish) may be useful. That is a different question from explaining one event. A detailed history does not yet tell you how often a similar condition occurs across the company.
Several conditions can act together
Suppose further review of the fictional case confirmed three things: an old copy was easy to access, approval did not unambiguously identify a document, and the person taking over did not receive the change record. There is no need to choose one winner simply because a summary has a “main cause” field. Establish how these conditions related to the sequence and which relationships were actually confirmed.
You can ask: if the handover had clearly identified the approved variant, would the employee have had a basis to select document B? The answer may help choose a test point. It cannot rerun the past. If the employee's view or exactly what they received is unknown, preserve that unknown rather than pretending the question resolved it.
A useful result for a manager may contain several confirmed conditions and one unresolved relationship. This separates a change closing a visible gap from a claim that the full cause has been established. The company can act within a limited scope while honestly stating what the analysis did not explain. The absence of one dramatic answer does not invalidate the work.
A change should follow what was established
If a wrong attachment was confirmed, consider changing the handover method. If approval was ambiguous, improve its link to the variant. If an old copy was reopened, a different control point is needed. We are not choosing a ready-made solution for every print shop. The point is why the analysis result should affect the proposed correction.
After a change, check an ordinary flow and a case where the old problem could recur. Performing an instruction alone does not confirm that the necessary information was obtained. If a worker marked “checked” but still cannot identify the approved variant, an extra field did not solve the recognised difficulty.
A project retrospective (in Polish) can select an action and determine where to return to its result. Do not present the chosen correction as proof that the initial hypothesis was true. Deciding to test and explaining a past event rest on different bases. A sensible change can be tested even when part of the history remains uncertain.
AI can organise material, but cannot fill in missing history
A tool can compare accounts, flag conflicting dates or prepare questions. A fluent explanation does not establish that an event happened that way. If a model invents the employee's motive or a missing file handover, the analysis can become more coherent and less true at the same time.
In an automation-process review (in Polish), preserve sources and the limits of what is known. At Aeneia Audit, a specific incorrect AI result, input data and a preserved control trail can be a starting point. The work can help define tool and human roles for that use without promising to reconstruct information that is unavailable.
At the end of the analysis you should be able to say: this was confirmed, this was not resolved, and this change will be tested for a stated reason. The number of “why” questions may be three, five or more. More important is whether material appeared between answers that can distinguish a convenient story from a justified explanation.







