cx-complaint-root-cause
Use to find the causes behind a complaint population and track whether remediation actually stopped them recurring. Trigger for "root cause analysis on our complaints", "why do we keep getting the sam
SKILL.mdskills/compliance/cx-complaint-root-cause/SKILL.md
GitHubComplaint root cause and remediation
Complaint root-cause analysis has a characteristic failure: the analysis is done, the themes are named, actions are assigned, the report is filed — and the complaints keep arriving at the same rate. A year later the same themes are rediscovered by a different analyst.
The reason is almost never the analysis. It is that remediation is not tracked to outcome, so nobody notices when an action was completed and the cause was not removed.
Cause, not category
A complaint category is a reporting bucket. A cause is a mechanism you could remove.
Category: "payment delays" Cause: "transfers submitted after 15:00 miss the cut-off, and the product does not say so until after submission"
The second one names something a team can change. The first names something you can count. Most complaint reporting stops at the first and then wonders why the numbers do not move.
Get to causes bottom-up. Read a sample of the actual complaints in a category rather than trusting its label, and expect to find that one category contains three unrelated causes and that two categories share one.
Classify the cause by where it sits
The classification determines the owner and the remedy, and this is where the analysis earns its keep:
- Product or service defect — it did not work.
- Process design — it worked as designed and the design produces a bad outcome.
- Communication — the outcome was correct and the customer was not told, or was told in a way they could not act on. Reliably one of the largest and most fixable groups.
- Handling — the underlying issue was minor and the way it was handled created the complaint. Distinguish this carefully; a handling complaint about a product defect is still fundamentally a product defect.
- Expectation set at sale — they were sold something that does not behave as described. Route to sales and marketing; support cannot fix it.
- Customer error, with a design contribution — they made a mistake the design made easy. Almost never purely the customer's fault; look for the affordance that allowed it.
- No fault — the outcome was correct, well-communicated, and they are unhappy anyway. A legitimate category, and it should be small. A large "no fault" share is usually a sign the analysis is defending the firm rather than examining it, so check a sample of these specifically.
Complaints are the tip
Every complaint category has a much larger silent population who had the same experience and did not complain. Complaint propensity varies enormously by segment, channel, product and customer confidence — so complaint volume is a poor proxy for harm volume, and ranking causes by complaint count systematically under-weights issues affecting customers least likely to complain.
Wherever possible, size the affected population from operational data rather than from complaints: how many customers hit that cut-off, received that letter, saw that error. Then report both, and rank on the population.
This matters beyond accuracy. If the cause produced a detriment, remediating only the customers who complained is usually not defensible — proactive remediation of the affected population is the question, and it is one for compliance and legal, not for the analyst.
Track remediation to outcome, not to completion
This is the part that fixes the recurring-theme problem. For each cause:
- A named owner and a specific change, not "review the process".
- A measurable prediction: which complaint category should fall, by roughly how much, and by when. Writing the prediction down before the change ships is what makes the check possible.
- The date the change actually took effect, which is frequently much later than the sign-off date and is the date that matters for measurement.
- A verification date, deliberately after enough volume has accrued to tell a real fall from noise.
- The verification result, recorded — and a cause whose complaints did not fall goes back on the list rather than staying closed.
"Action completed" is not "cause removed." Track both states separately. The gap between them is the most useful metric this whole exercise produces, and almost nobody has it.
When checking whether complaints fell, check whether the rate fell rather than the count, and confirm the movement exceeds normal variation before declaring success. Complaint volumes are low, so a good month proves very little.
Traps
- Reclassification masquerading as improvement. A category that falls sharply immediately after a taxonomy change or a new triage rule has not improved. Check for definition changes before celebrating.
- A remediation that moves the complaints rather than removing them. Complaints falling in one category and rising in an adjacent one is a redirect, not a fix.
- Five Whys stopping at the person. "The agent didn't check" is a symptom. The question is why the process let an unchecked case through, and what made checking hard.
- Root causing individual complaints only. Per-case root cause is required for the response to that customer; thematic root cause is what reduces volume. They are different exercises and the first does not aggregate into the second.
- Small numbers. A category with four complaints does not support a trend. Report the count, suppress the rate, and do not run a programme off it.
- Complaints closed as "resolved" with no cause recorded. Common, and it removes those cases from every subsequent analysis. Report the share.
Guardrails
- Whether a cause produced a reportable failure, and whether proactive remediation is required, is a compliance and legal determination. Produce the cause, the affected population and the evidence; flag it clearly and do not rule on it.
- Do not attribute causes to named individuals in a thematic analysis. If a genuine individual performance issue emerges, it belongs in a separate, properly-controlled process.
- Cite ids; do not paste complaint content. Complaint files hold financial circumstances, health disclosures and vulnerability information, and these reports circulate.
- Do not let a complaint-reduction target become a complaint-recording target. If reducing complaints is an objective, the guardrail is the identification rate — a fall in recorded complaints with no fall in the underlying driver means fewer are being logged.
Present results to the user
- Causes, not categories, derived bottom-up, with the category-to-cause mapping shown where one category split or two merged.
- The cause classification, with an owner per class, and the "no fault" share checked rather than assumed.
- Affected population per cause, from operational data where available, alongside complaint count — and ranked on population.
- The remediation register: owner, specific change, predicted effect, effective date, verification date, verification result.
- Completed-but-not-verified and verified-but-not-fixed, as separate lists. The second is the one that explains recurring themes.
- Causes recurring from a previous analysis, named as such.
- What needs a compliance determination — reportability and proactive remediation.
- Data gaps — complaints with no recorded cause, categories too small to analyse.