Classification decides who handles a complaint, how fast, and whether it is reported. A complaint classification matrix fixes the rules: category, indicators, severity, reportability, owning team, when the deadline starts, and what an AI agent may do. Without explicit rules written down, an agent classifying at volume will drift.
Why classify complaints with a matrix?
Classification decides everything that follows: who handles the complaint, how fast, whether it is reported to a regulator, and which root cause it feeds. Classify inconsistently and the same grievance gets different treatment depending on who read it first, and trend data becomes noise. Good complaint handling starts with getting this step right.
An agent can classify fast and consistently, but only if the rules are explicit. People resolve ambiguity from experience; an agent needs indicators written down. The matrix is that writing-down, and it is also the basis for testing the agent: every row becomes a set of certification cases.
When should the agent use it?
On every inbound message that might be a complaint, in any channel. The first test is whether the message is a complaint at all. Many organisations define a complaint broadly, as any expression of dissatisfaction about the service, whether or not the customer uses the word. An agent that only picks up messages saying "complaint" will miss many.
Once classified, the deadline starts. Tracking it is a separate job, covered in how to track complaint deadlines with an AI agent.
How do you fill in each column?
- Category. Mutually exclusive at the top level. Use sub-categories for detail.
- Indicators. Words, facts or patterns that point to this category. Include examples of what does not qualify.
- Severity. Standard, high or critical. Critical usually means possible harm, vulnerability, discrimination or a data breach.
- Reportable. Whether this category feeds regulatory reporting, under your regulator's rules.
- Owner. The team that handles it.
- Agent may. Classify, acknowledge, gather facts, draft a response.
- Stays with a person. Usually the resolution decision and the final response.
Add one category many matrices lack: complaints about the AI agent itself. They need their own owner, because they are evidence about the agent's behaviour and should feed its certification.
The complaint classification matrix
| Category | Indicators | Not this category if | Severity | Reportable | Owner | Agent may | Stays with a person |
|---|---|---|---|---|---|---|---|
A filled example: a retail bank
| Category | Indicators | Severity | Owner | Agent may | Stays with a person |
|---|---|---|---|---|---|
| Fees and charges | Disputes a fee; says a charge is unfair | Standard | Complaints team | Classify, acknowledge, pull fee history, draft response | Decision on refund above the agent's limit; final response |
| Service failure | Delay, lost documents, no call back | Standard | Complaints team | Classify, acknowledge, build timeline | Final response |
| Fraud and scams | Reports being tricked; disputes a payment they made | High | Fraud team | Classify, acknowledge, open case | All decisions |
| Data protection | Wrong data, data sent to someone else, request ignored | Critical | Data protection officer | Classify, acknowledge | All decisions; breach assessment |
| Unfair treatment | Says they were treated differently because of who they are | Critical | Senior complaints handler | Classify, acknowledge | All decisions |
| Vulnerability | Mentions illness, bereavement, financial difficulty alongside dissatisfaction | High, added to any category | Specialist team | Flag, acknowledge | All decisions |
| AI agent conduct | Complains about what the agent said or did | High | Agent owner and complaints | Classify, acknowledge | All decisions; feeds re-certification |
Three sample messages, classified:
| Message | Classification |
|---|---|
| "Why was I charged 25 for a late payment? I paid on the due date." | Fees and charges; standard |
| "Your chatbot told me my payment was cancelled and it wasn't. I've now been charged twice." | AI agent conduct, with fees and charges as secondary; high |
| "Since my wife died I can't keep up with these letters, and now you've charged me again." | Fees and charges with vulnerability flag; high; specialist team |
The third message shows why vulnerability is a flag across categories rather than a category alone. The complaint is about a fee, but the handling must change.
How does classification feed root cause analysis?
Tag each complaint with a sub-category and the product or process involved. Consistent tags are what make patterns visible; whether an AI agent can do complaint root cause analysis covers what an agent can contribute there. Risk-rating matrices for customers work on a similar logic but serve a different purpose; compare the customer risk rating matrix. If you are also classifying your agents for regulatory purposes, that is a separate exercise, set out in the EU AI Act risk classification checklist. The wider boundary for compliance agents is in AI compliance agents: second line, never the last word.
Questions people ask
Can a complaint have more than one category?
Yes. Record a primary category for ownership and reporting, and secondary ones for analysis. Flags such as vulnerability apply on top of any category.
What should the agent do if unsure whether a message is a complaint?
Treat it as one. Under-recording complaints is a more serious failure than over-recording them, and a person can reclassify.
Who should review the agent's classifications?
A complaints team leader should sample them every week, with extra weight on critical categories and on messages the agent classified as not a complaint.