Complaint identification and classification

When to use it

Reach for this when someone says any of these — they are the phrases the skill itself triggers on:

• “are we identifying complaints correctly”
• “complaint detection”

How it works

The method, in the order the skill runs it. The full procedure — tables, worked examples and the edge cases — is in the skill itself.

1. The error that causes regulatory findings

   A complaint is a definition, not a feeling.

2. Step 1: get the definition, in writing, from compliance

   Do not draft it yourself. Obtain the definition that applies to your entity and licence, and hold it as the single source of truth.

3. Step 2: tune for recall, not precision

   This is the design decision that most distinguishes complaint detection from ordinary classification, and it is counter-intuitive to anyone with an ML background.

4. Step 3: build detection in layers

   Layer 1: structural signals. Cheap, high-precision, no model needed. Customer used the word "complaint" or "complain". Contacted an ombudsman or regulator. Third repeat contact on one issue. Escalation requested. Refund or compensation demanded. Solicitor or advocate involved.

5. Step 4: classify on separate axes

   Once something is a complaint, several independent things need recording. Collapsing them into one field is the most common structural mistake — it makes root-cause reporting impossible.

6. Step 5: treat vulnerability as its own detection problem

   Vulnerability signals frequently co-occur with complaints and carry separate, often stricter obligations. Detect them independently rather than as a complaint sub-category.

7. Step 6: build the audit trail before you need it

   Regulators ask how you decided, not just what you decided. Every complaint record should carry: The definition version applied, Which elements were met, with a verbatim quote for each, Who confirmed it, and when, Detection route, Timestamps, immutable, with the clock-start rule applied consistently, Every state change, retained.

8. Step 7: audit for what you missed

   Precision is easy to audit — review what you flagged. Recall is the hard part and the part that matters.

Related skills

• Complaint deadlinesTrack complaint deadlines in business days, and flag near-misses
• Vulnerability signalsSpot signs a customer is vulnerable, and check you acted
• Conduct risk monitoringSpot pressure, mis-selling and unfair treatment in conversations
• Complaint root causeFind what causes complaints, and check the fix held
• Customer outcome evidenceShow a board that customers are getting good outcomes
• Emerging harm scanSpot a new kind of harm before it has a category