AI agent · Customer service · Sort & route
Complaint classification and routing
Complaints arrive through every channel and wait to be read before anyone can act. The agent classifies each one against your definition, routes it to the team that owns it, and keeps the unclear and sensitive ones for a person.
Typical volumes for this process, not a client figure.
Complaints land in a shared inbox; someone reads each one, picks a category and forwards it.
Classified, routed with the clock started; sensitive and unclear ones go to a person.
Where the time goes today
Complaints reach you through every door at once: the complaints mailbox, the general contact form, chat transcripts, call notes typed by the contact centre, scanned letters, and the occasional message forwarded by a director. A triage team reads each one, decides whether it is a complaint or an ordinary query, finds the customer and the order, picks a category and forwards it to the team that owns the problem. Where your rules set a response deadline, it runs from arrival, not from opening.
Most of the time goes on context, not on the decision. The triager searches the customer record by email address, then by name, then by the order number buried in the third paragraph. Long threads must be read to the end to learn whether this is a new complaint or the fourth chaser. A message that raises two issues, say a late delivery and a rude driver, is forwarded to one team and the second issue disappears.
The costly errors are quiet ones. A complaint logged as a query never gets a formal response and never reaches the root-cause figures. A misrouted case bounces between billing and logistics while its deadline runs. A customer who mentions a bereavement or financial hardship is handled like everyone else, because the triager was reading for the category, not for the person.
How the agent works
- Collect every channelThe agent watches the mailbox, web forms, chat exports, call notes and the scanning queue. Each message is logged once, with its channel and its time of receipt, which is the time any deadline runs from.
- Find the customerIt matches the sender to an account and an order using email address, phone number, references quoted in the text and names. If that customer has an open case on the same issue, the message is linked to it instead of opening a new one.
- Apply your definitionIt tests the message against your written definition of a complaint and records which part of the definition was met, quoting the sentence that meets it. Messages clearly outside the definition go to the service queue as queries.
- Classify and flagIt assigns category, product and issue from your taxonomy, splits a message raising several issues into linked parts, and looks for the markers your policy names: vulnerability, legal action, contact with a regulator or the press, safety, financial loss.
- Route and timeIt assigns the owning team from your routing table, sets the response deadline from your rules, and sends an acknowledgement from an approved template where you allow it.
- Hand over the restMessages it cannot classify with confidence, and every message carrying a vulnerability or legal marker, go to the triage lead with the reason and the passages it relied on.
What stays with a person
The triage lead keeps the boundary cases. Whether an angry message about a price rise is a complaint or feedback depends on your definition and sometimes on context the text does not give. Vulnerability markers always reach a person, whatever the confidence score, because treating a vulnerable customer as routine costs far more than the minute it takes to look. The same applies to threats of legal action and anything that mentions a regulator or the press.
The agent does not decide redress, write the substantive response or close a complaint; those stay with the owning team. Nor does it log a message as not a complaint without leaving a trace: those decisions are sampled every week, because under-recording is the error nobody notices until an audit finds it.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Complaints mailbox, web forms, chat transcripts, call notes and scanned letters | A complaint case with category, product, issue and the quoted text supporting each |
| Customer accounts and order history in the CRM | Follow-ups linked to the existing case instead of new duplicates |
| Open cases in the complaints system | Acknowledgements from approved templates, stating the case reference and deadline |
| Your written complaint definition and category taxonomy | A queue for the triage lead, each item with the reason it was held back |
| The routing table: which team owns which category and product | A daily list of cases whose deadline falls in the next two working days |
| Your response-deadline rules and acknowledgement templates |
Controls that come with it
- A confidence threshold per category; below it, a person classifies
- Vulnerability, legal and regulator markers always go to a person, whatever the confidence
- A weekly sample of messages logged as not a complaint, reviewed by the triage lead
- Acknowledgements only from approved templates; the agent writes no free-text reply
- Each decision recorded: what was read, which rule applied, which team received it
- Reclassification in one step, which re-routes the case and keeps the original receipt time
How you know it works
- Time from receipt to the correct owning team
- Share of cases re-routed after first assignment
- Share of sampled not-a-complaint decisions overturned
- Response deadlines missed, by category
- Duplicate cases opened for the same issue
Is your process ready?
- Written rules: you need a written complaint definition and a taxonomy people actually use. A list of sixty categories where a dozen carry the volume should be cut back before the build.
- Systems: the case system must let software create and update cases, and your vendor licence must permit it.
- Cheap check: the receiving team accepts or rejects each routing within a day, so every classification is checked as a side effect of normal work.
- Volume: several hundred complaints a month across channels repays the build. A handful a day is better served by a well-organised shared inbox.
- One description: if billing and the contact centre disagree on what counts as a complaint, settle it on paper first. The agent applies whichever definition you give it, consistently.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- Broken email threading turns every reply into a new complaint unless matching uses the customer and quoted references rather than subject lines.
- The routing table goes stale after a reorganisation, and cases flow to a team that no longer exists.
- Handwritten letters and poor scans are read with low confidence; they need a person, not a guess.
- Acknowledgement templates that promise a call-back create a commitment the owning team never made.
Questions we get
Does the agent reply to the customer?
Only with an acknowledgement, and only from templates you approve. It confirms receipt, gives the case reference and states the response deadline your rules set. It does not write the substantive answer, offer redress or apologise for something specific, because those depend on facts the owning team has not yet established. You can switch acknowledgements off entirely and use the agent for classification and routing alone.
What happens when one message contains several complaints?
The agent splits it into linked parts, one per issue, each with its own category and owning team, and keeps the original message attached to every part. The customer receives one acknowledgement listing the issues. Where your rules require a single response per customer, the parts are grouped under a lead case owned by the team with the most serious issue.
How does it handle vulnerable customers?
Your policy names the markers: bereavement, illness, disability, financial difficulty, a third party writing on someone's behalf. The agent looks for them in the text and in any flag already on the customer record. When it finds one, the case goes to a person with the passage highlighted, whatever the classification confidence. The agent does not judge how vulnerable someone is; it makes sure a person sees it.
What do we need before starting?
A few hundred past complaints with the category and team they should have had, confirmed by your triage lead rather than copied from the system, since the history contains the errors you want to remove. That set measures the agent before go-live and after every change. You also need the written definition, the taxonomy and a current routing table.
Want this agent on your process?
Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.