Human error is not a root cause
Inadequate investigation is one of the most consistently cited findings in FDA inspections, and the problem is rarely that firms do not investigate. A walk through the mechanism stage by stage, the decisions each stage forces, and the defects that get records sent back.
If you read enough Form 483s and warning letters, the same sentence keeps appearing in different clothing: the investigation failed to determine a root cause. It is one of the most consistently cited findings in drug GMP inspections, and it is almost never because the firm did not investigate.
It is because the firm stopped at the first explanation that fit.
What the regulation actually asks for
21 CFR 211.192 is short, and worth reading in full rather than in summary. Three things in it do most of the work:
- Any unexplained discrepancy, or any failure of a batch to meet its specifications, requires a thorough investigation, whether or not the batch has already been distributed.
- The investigation must be recorded in writing, and the record must include the conclusions and the follow-up.
- The investigation must extend to other batches and other products that may have been associated with the failure.
Notice what is not in there: any permission to close an investigation because the deadline arrived. The standard is thoroughness, and thoroughness is assessed from your written record by somebody who was not in the room.
Vioela’s note: How your investigations actually get chosen for review: nobody reads them at random. An investigator asks for the log, the list of every deviation, out-of-specification result and complaint since the last inspection, and then picks. The picks are predictable. Anything closed as human error. Anything with no root cause. Anything that repeats. Anything open for a very long time, and anything closed in the weeks just before the inspection. Sort your own log that way first. If a pattern jumps out at you in ten minutes, it will jump out at them.
The mechanism, end to end
Before the critique, the map. Quality systems use different names for these stages, and some merge or split them, but the sequence is close to universal. Each stage exists to force one decision, and most weak investigations are weak because a stage was treated as a field to fill rather than a decision to make.
| # | Stage | The decision it forces |
|---|---|---|
| 1 | Record | Is this a quality event, and what is the date it was identified? |
| 2 | Triage | Is it GxP, what type is it, who owns it, has it happened before? |
| 3 | Contain | What must stop moving, and what must stop running? |
| 4 | Assess risk | How bad, how likely, how detectable, and what path does that put it on? |
| 5 | Bound | Which other lots, lines, dates and products are in scope? |
| 6 | Investigate | What caused it, and what did we rule out? |
| 7 | Disposition | What happens to the affected product, and on what justification? |
| 8 | Correct | What fixes this instance, and who verified that it did? |
| 9 | CAPA | What stops recurrence here, and occurrence elsewhere? |
| 10 | Effectiveness | How will we know it worked, measured how, over what period? |
| 11 | Approve | Does the record support every conclusion in it? |
Four points where the path is actually decided
Containment is two questions, not one. Product containment stops material moving: hold, quarantine, stop-ship, recall assessment. Process containment stops the thing recurring while you investigate: pause the step, add a temporary inspection, revert to a manual check. Answering one and calling it containment is a common gap. So is treating “we opened a record” as containment, which is documentation, not control.
The risk score chooses the route. Most systems rate severity on several axes rather than one, typically patient safety, product quality, regulatory exposure and process impact, then combine that with likelihood and detectability. The resulting score decides whether the event can close as a local nonconformance or must escalate into a full CAPA investigation. Two rules matter more than the arithmetic. Severity is a property of what happened, so a passing downstream result does not lower it. And the detectability rating should name the control that actually caught it, because “we found it” is not a detection argument.
The recurrence search sets the scope. Search for similar events before you conclude anything, run the window back from the date you searched rather than forward to the date of the event, and link what you find. A search that stops at the occurrence date quietly excludes everything between occurrence and detection, which is usually the interesting part. If the search returns matches, the event is recurring, and the record has to say so.
Bounding is where most investigations under-reach. Occurrence and detection are different dates. Everything processed in between is in scope until you show otherwise, and if the task is performed several times a day, “one batch” is a conclusion that needs evidence behind it.
Corrections, corrective and preventive are three different things
They get collapsed constantly, and the collapse is visible to anyone reading.
- A correction fixes this instance. Rework the batch, amend the record, retrain the person who made the error.
- A corrective action removes the cause so it does not happen again here.
- A preventive action addresses the same cause everywhere else it could occur: the other line, the other site, the other product using the same step.
A CAPA consisting only of corrections is the most common weak plan in circulation. It is also the easiest to spot, because nothing in it would have prevented the event it was raised for.
Watch the clocks
Due dates usually run from the date the event was identified, not from the date somebody opened the record or finished triage. That matters more than it sounds, because when a nonconformance escalates into a CAPA, the CAPA often inherits the original clock. Time has already been spent before the investigation properly starts.
If the date cannot be met, every system has a mechanism for it: an extension request, or an interim update recording what has been done and what remains. Use them. A documented extension with a reason is a normal part of a healthy quality system. A conclusion that arrives on the due date because the due date arrived is the thing investigators are trained to notice.
Does this sequence match what FDA actually asks for?
It is a fair question to ask of any framework, so here is the mapping. Nothing in the sequence above is invented. Each stage exists because something requires it.
| Stage | What requires it |
|---|---|
| Record, triage | 21 CFR 211.192: any unexplained discrepancy or batch failure triggers an investigation. 211.22(a): the quality unit owns it |
| Contain | 211.42(c) and 211.113(b): prevent contamination and mix-ups. 211.142: quarantine pending disposition |
| Assess risk | ICH Q9 on quality risk management. 211.192 requires the investigation to reach the product impact question |
| Bound | 211.192 in terms: the investigation “shall extend to other batches” that may have been associated |
| Investigate | 211.192 sets the standard at thorough. For laboratory results, FDA’s 2022 OOS guidance sets out the Phase I and Phase II structure |
| Disposition | 211.165(f): batches failing specifications are rejected. 211.192 requires the written conclusion |
| Correct, CAPA | 211.22 and ICH Q10, which makes CAPA a named element of the pharmaceutical quality system. For devices, 21 CFR 820.100 and now QMSR |
| Effectiveness | ICH Q10 again, and 820.100(a)(4) for devices, which requires verifying that the action is effective rather than merely done |
| Approve | 211.192: the investigation “shall be recorded in writing” and include conclusions and follow-up |
The useful observation is that the regulation is thinner than most people assume. 211.192 is a few sentences. It does not prescribe tools, templates, or a scoring matrix. What it prescribes is thoroughness, a written record, and extension to other batches.
Everything else in your procedure, including the risk scoring, the triage categories and the field structure, is your company’s chosen method of demonstrating those three things. That is worth knowing, because it tells you what an investigator is actually testing. Not whether you followed your template. Whether the record in front of them is thorough, written down, and honest about scope.
Vioela’s note: One rule from the laboratory side that catches people. When a test result fails, you may not simply test again and keep the passing answer. FDA’s guidance on out-of-specification results expects the laboratory to look for an assignable error first, and any retesting to follow a plan written in advance, including how many retests are allowed. Averaging a failing result with passing ones to produce a passing average is specifically warned against. Investigators have a name for the alternative: testing into compliance.
The failure mode: stopping at the symptom
The most common ending to a weak investigation is “human error,” usually followed by retraining.
Human error is a finding, not a cause. The useful question starts where that sentence stops:
- Why was the error possible? What in the process allowed a wrong action to be taken and accepted?
- Why was it not caught? What control should have detected it, and why did that control not fire?
- Why did it happen now? What changed, or what has always been fragile and only failed this time?
An operator who misreads an ambiguous display has given you a design problem. An operator who skips a step that the batch record puts three pages away from the task has given you a documentation problem. An operator who was trained, qualified, rested, and still performed the step wrong, in a process where nothing else was unusual, is the rare case where the cause really is individual. Rare, and you need evidence for it like anything else.
Retraining a person for a system fault does not change the probability of recurrence. The next investigator will see the repeat and will know exactly what happened.
What a defensible root cause looks like
Treat the investigation as a hypothesis-testing exercise, not a search for a culprit.
Define the problem precisely. What failed, where, when, on which equipment, in which lot, against which acceptance criterion. Vague problem statements produce vague causes.
Generate the plausible causes before you test any of them. A fishbone or a fault tree is useful here, not because the diagram is impressive but because it forces you to list causes you would rather not consider. Is/is-not analysis earns its place too: what is affected, what is not, and what distinguishes the two.
Test each hypothesis against evidence. Batch records, trend data, calibration history, EM results, maintenance records, raw data, interviews. Each candidate cause is confirmed or excluded, and on the strength of something specific.
Write down what you ruled out, and why. This is the step most often skipped and the one that most clearly separates a credible investigation from a plausible story. An investigator reading a record that only names the surviving hypothesis has no way to tell whether you considered the others. A record that shows five candidates, three excluded on data, one excluded on timing, and one confirmed, reads as science.
Say how confident you are. Where the evidence supports a most-probable cause rather than a definitive one, say so in those words and explain what would have been needed to be certain. A stated most-probable cause with sound reasoning is defensible. A definitive cause asserted on thin evidence is not, and it is far easier to attack.
A worked example: the same event, two investigations
Take an illustrative case, of a kind FDA documents regularly. On an aseptic filling line, environmental monitoring recovers organisms in the surrounding classified area, and over following weeks within the critical zone. Records from the same period show intermittent environmental excursions during production. The investigation runs a 5 Whys.
Here are two chains built from the same facts.
| The reported 5 Whys, ending in human error | What was actually happening | |
|---|---|---|
| 1 | Why was contamination found? The area was not in control. | Why did the critical zone see contamination? Lower grade air reached it. |
| 2 | Why? There were environmental excursions. | Why did it reach it? Airflow over the critical zone was not unidirectional during production. |
| 3 | Why? Operators were working at the line. | Why not? Air swept toward a return rather than down across the zone, drawing in air from the surrounding area. |
| 4 | Why? They have to reach in to clear stoppages. | Why was this not known? The smoke study was run without the interventions that happen in every batch. |
| 5 | Why? Aseptic technique training was insufficient. | Why not? Nobody had recorded how often those interventions actually occur. |
| Reported cause | Operator practice | Airflow design, never qualified against real operations |
| CAPA | Retrain, add signage, audit technique | Repeat smoke studies in operation including the real interventions, correct the airflow, remove the need to reach in |
| Next year | The excursion returns | It does not |
Both chains are internally coherent. Only one of them explains the facts.
The three questions that separate them
Does the cause explain everything that happened? Technique is a story about the surrounding area, where people work. It does not by itself explain recoveries inside the critical zone, which is supposed to be protected from exactly that. A cause that accounts for part of the event is a contributing factor, and calling it the root cause closes the investigation early.
Was the correlation actually tested? This is the step most often skipped. The excursions and the contamination recoveries are both timestamped, as are production activities. Nobody has to speculate about whether they coincide. Plot them against each other. If they do not line up, the theory is wrong and should be excluded in writing. If they do, you have an association worth pursuing to mechanism. Either outcome advances the investigation. Asserting the link without checking advances nothing.
Did anyone read the engineering records? Air balance reports, commissioning data, equipment extraction rates, maintenance history on fans and dampers. Quality investigations tend to stay inside quality documents. The explanation for a physical event is usually sitting in a physical record that nobody on the investigation team has opened, because it belongs to a different department and is written in a different vocabulary.
Airflow is not the only candidate
Replacing one confident answer with another confident answer is the same mistake wearing a lab coat. Airflow is a strong hypothesis here, not a conclusion. At least two others deserve testing before anything is written down.
The organism tells you where to look. Identify to species before you theorise. The recovery is evidence about its own origin, and it narrows the search faster than any workshop:
| What was recovered | Where that usually lives |
|---|---|
| Gram negative rods | Water. Drains, condensate, damp residues, wet cleaning equipment |
| Spore formers | A cleaning and disinfection gap. Check sporicidal rotation and contact time |
| Skin and nasal flora | People. Gowning, technique, personnel flow |
| Mould | Ingress, fabric, humidity, construction or maintenance work |
If the organism is a water dweller and your theory is door openings, the theory is in trouble before you test anything else. That single check costs nothing and routinely redirects an investigation.
Something changed. Pull change control and maintenance records for the weeks before the first recovery, not just the day of it. Filter changes, damper adjustments, AHU servicing, a new disinfectant lot, a modified gowning step, a different supplier of a component, a software update to building management. Events that begin after a change usually began because of it, and the change record is the cheapest place to look.
A useful discipline: write down your candidate causes before you start gathering evidence, and record which ones the evidence kills. An investigation that names one hypothesis is indistinguishable from a guess that happened to be written up.
How to interrogate a control
Once you have a candidate cause, the question is not only what happened but which control should have stopped it, and which should have caught it. Those are two different controls and both are in scope.
Controls fall into three classes, and the class matters more than the wording:
- Engineering controls are physical. Pressure cascade, unidirectional airflow, isolators and barriers, interlocks, filtration, segregation, automation. They work whether or not anyone remembers them.
- Administrative controls are instructions. Procedures, training, gowning steps, cleaning schedules, line clearance, door discipline. They work only when a person performs them correctly, every time, under pressure.
- Detection controls tell you something went wrong. Environmental monitoring, alarms, in-process checks, batch record review, trending.
For any control in any of those classes, five questions:
- Does it exist, specifically? Name it. “We have procedures” is not a control.
- Was it qualified for the condition that actually occurred? A room balanced at rest is not qualified for dynamic operation. A disinfectant validated against vegetative organisms is not qualified against spores. Most control failures are scope failures, not breakdowns.
- Was it functioning at the time? Calibration status, alarm history, maintenance records, whether a sensor was in range or an alarm had been silenced.
- Would it have fired? If a detection control did not alert you, ask why. The limit may be set above the condition, the sensor may be in the wrong place, the sampling frequency may be too low to catch an intermittent event, or the alert may go somewhere nobody reads.
- Does it depend on a person remembering? If the answer is yes, it is administrative, and you should expect it to fail eventually.
The rule that follows from this
A CAPA that moves a control down the hierarchy has weakened the system.
If an engineering control failed and your corrective action is a procedure and a training session, you have replaced a physical barrier with a human promise. That is a downgrade, and it is the single most common reason the same event returns a year later with a new record number.
Ask it directly in the CAPA review: is this action stronger than the control that failed, or weaker? If weaker, say why that is acceptable, in writing, and expect to be asked about it.
On talking to people
Interviews matter, but not as a group exercise. Read the records first, so you can test what you are told against what was written down. Talk to people separately, because the second person in the room rarely contradicts the first. Ask what actually happens rather than what the procedure says, and ask the maintenance and engineering staff as well as the operators. In a contamination investigation, the person who services the air handling unit often knows something the quality team has never been told.
Why the weaker chain is so attractive
It is not stupidity. The behavioural chain is faster, it needs no other department’s cooperation, it produces a CAPA that can be implemented immediately, and it closes within the due date. The mechanism chain requires engineering time, probably a requalification, and an uncomfortable conversation about whether the room was ever correctly balanced.
A system under schedule pressure will select the first chain every time unless somebody with authority asks the three questions above.
This is not hypothetical: a public example
FDA’s warning letter to Pharmathen International, dated 27 May 2026, documents the pattern precisely.
| What the firm did | What FDA said |
|---|---|
| Attributed recurring sterility and media fill failures to operator error | “your investigation did not include in-depth analysis of the conditions that create the circumstances for human error” |
| Attributed recurring gram negative recoveries to room disinfection practices | FDA called it “highly atypical to find gram-negative microbes in an aseptic processing room environment” and asked for a thorough investigation of the recurrence |
| Treated the failures as behavioural | FDA separately found that “This problematic airflow pattern in your RABS creates significant potential for microbial contamination” |
| Responded with operator-focused corrective actions | FDA asked for “in-depth, root-cause analysis to identify effective CAPA including design remediations” |
Note the shape of it. The firm was not wrong that operators were involved. FDA’s own investigators observed an operator blocking first air and failing to discard exposed vials. The finding is not that the behaviour did not happen. It is that the investigation stopped at the behaviour and never asked what conditions made it likely, while a physical cause sat in the airflow pattern of the same room.
The phrase worth committing to memory is FDA’s: the conditions that create the circumstances for human error. That is the standard. Not whether a person made a mistake, but what made the mistake available.
The same enforcement cycle makes the other half of the point. A letter to Hanlim Pharm, dated 2 October 2026, describes airflow that “enables the entrainment of lower-quality air from the surrounding lower-classified” area, and smoke studies that “failed to sufficiently evaluate airflow directionality.” That is the mechanism chain above, found by a regulator rather than by the firm.
The patterns that get investigations sent back
Reviewing other people’s investigation records, the same defects recur across companies and product types. None of them are exotic.
A 5 Whys that does not chain. Two failures are typical. The first Why does not answer the question the problem statement actually asked, so the chain is pointed in the wrong direction from the start. And the final Why restates the immediate cause rather than explaining why the control that should have caught it did not. If your last Why is a rephrasing of your first observation, you have written a list, not a causal chain.
Naming a tool without showing it. The record says “5 Whys” or “fishbone” in a field, and the analysis itself exists only in somebody’s notes or a Word file that never made it in. The tool is evidence. If it is not in the record, it did not happen.
A recurrence flag that contradicts the narrative. The similar-event search returns matches, the text calls the issue recurring, and the recurrence field says no. Both cannot be true, and whichever is wrong, the reviewer now distrusts the rest. Related: search windows that run to the date of the event rather than back from the date you searched, which silently excludes everything between occurrence and detection.
Severity softened by downstream results. A limit was exceeded, but release testing passed, so the severity rating comes down and the disposition rests on the passing result. This gets cited reliably, and most procedures prohibit it explicitly. Detection after the fact does not reduce the severity of what happened; it is a separate axis.
Numbers nobody recomputed. A transcription error of two milligrams credited with a one percentage point shift in a result. These are easy to catch and embarrassing to leave in, because an investigator who finds one arithmetic error stops trusting every other number in the record.
Text from a different investigation. Copy and paste from a similar event, with the other event’s batch, dates and problem statement still in it. It happens under deadline pressure, and it tells a reviewer exactly how the record was produced.
Conflicting accounts of who did what. The narrative names one person as performer, the interview notes name them as verifier, and the training action is assigned to a third. If the record cannot establish who performed and who verified, it cannot support a cause that rests on either.
Mandatory fields left empty. Investigation details, bounding, risk assessment participants, summary. An investigation can be sound in substance and still be unapprovable because the record does not contain it.
The two parts people skip
Scope extension. 211.192 requires it explicitly, and it is routinely treated as a formality. If a cause is real, ask honestly which other batches, products, lines, or time periods it could have touched. The answer “this was isolated” needs the same evidentiary support as any other conclusion. An investigation that never looked beyond the batch in front of it is an investigation that has not finished.
Effectiveness. A CAPA is not effective because it was implemented. It is effective because the thing stopped happening, measured against something you defined in advance, over a period long enough to mean something. Define the metric and the window while you are writing the CAPA, not when the effectiveness check falls due.
How this reads from the other side of the table
When an investigator pulls your investigations, they are usually testing one thing: does this quality system find and fix its own problems, or does it process paperwork?
The signals that answer that question are visible fast. Repeat deviations with the same cause across different investigations. CAPAs that are all retraining. Effectiveness checks that are a signature and a date. Root causes that are restatements of the problem. Timelines where the conclusion arrives on the day the record was due to close.
None of those require an expert to spot, and each of them suggests the same thing: that the investigation existed to close the record.
Vioela’s note: If an inspection does end with observations about your investigations, the response matters as much as the finding. The agency considers a written response received within 15 business days before deciding what to do next. The responses that work have a recognizable shape: they fix the specific instance, then show a search for the same weakness elsewhere, then address why the system allowed it, with dates and owners. The responses that fail defend the original investigation, or promise retraining. A promise to retrain, in reply to a finding that your investigations blame people, proves the finding.
Get this right and the reward is bigger than an inspection outcome. A quality system that genuinely finds causes stops producing the same deviation, which is the only version of this work that is cheaper than doing it badly.
The rest of the series
This piece is the method. Two companions:
- The investigation they should have run takes one event end to end, built on the Pharmathen letter. It reconstructs how an investigation reaching “operator error” typically runs, then rebuilds it step by step, including who you would need to talk to, in what order, and what each conversation rules in or out.
- What stands between you and a sterility failure is the reference list: every control that keeps sterile product sterile, sorted by class, with how each one fails and what FDA has written about it.
Sources
- 21 CFR 211.192, production record review and investigations
- FDA, Investigating Out-of-Specification Test Results for Pharmaceutical Production, 2022
- FDA warning letter, Pharmathen International S.A., 27 May 2026
- FDA warning letter, Hanlim Pharm Co., Ltd., 2 October 2026
Warning letters state FDA’s position at the time of writing. They are not adjudicated findings, and firms respond to them.