In this guide 6 sections
AI can read an incoming email, form submission or chat message and assign it to a category, so the right team receives it without a person sorting it first. Definitions and messy real tests come first. Sensitive cases still go to people.
Defining useful categories
Short lists work better. It is tempting to mirror every internal team in a long list, but categories should map to who handles the enquiry and what happens next. Each needs a named owner. Without one, classification is only labelling.
Consider a regional broadband provider sorting billing, technical faults, new connections and plan changes, where “technical fault” works because a named team picks it up within an agreed time. “General query” routes nowhere. If two categories send enquiries to the same team with the same next step, merge them, because the split adds reporting work and no routing value.
Building a set of labelled examples first
Collect past enquiries before testing anything. Gather several hundred real enquiries across email, forms and chat, and have a person label each with the category a human handler would choose. That set does two jobs. It gives you a baseline to test the AI against, and the labelling itself often exposes category definitions that were less clear on paper than in practice.
Skipping this step is the usual mistake. Without a labelled baseline, a live pilot cannot tell you whether the categorisation is accurate or merely plausible, and misrouted enquiries will be the only evidence.
Testing where categories blur
Real enquiries rarely sort cleanly. “My WiFi keeps dropping and am I being charged for the days it was down” touches both a technical fault and billing, so test these overlap cases deliberately. Also test messages written in mixed languages or heavy local phrasing, because a system that scores well on clean examples often breaks down on exactly those.
The aim is not zero ambiguity. Some ambiguity is unavoidable, so the real question is how the system behaves when it meets a message it cannot place and whether that behaviour is safe.
Routing low-confidence and sensitive cases to people
Set a confidence threshold. Below it, an enquiry goes to a person for triage instead of being auto-routed on a guess, and sensitive categories always go to people whatever the score.
Sensitive means a complaint about billing fraud, a request about personal data, or any message that mentions a regulator, an ombudsman or a lawyer. Triage can be quick. Anything flagged as a possible complaint about unauthorised charges goes straight to a supervisor, never to a first-line queue, while low-confidence enquiries get a brief triage before routing.
Monitoring misrouting and updating the taxonomy
Accuracy is not measured once. A new plan launch can bring a wave of enquiries the original categories never anticipated, so a taxonomy that fit six months ago can quietly stop fitting. Track misrouted enquiries, meaning cases where the receiving team has to send them elsewhere, because they signal that the categories or the classifier need attention.
Review the taxonomy on a schedule. A short quarterly review of misrouted cases usually shows which categories need splitting, merging or better examples, and which owner needs a clearer next step. Launch is not the finish line.
Enquiry classification and routing flow
A flow diagram of an enquiry passing through a classification step that outputs a category and a confidence score, with low-confidence and sensitive cases forking into human triage. Beside it sits a routing table mapping every category to an owning team, an expected response time and the next step the owner takes when the enquiry arrives.
Open a blank worksheet to printQuestions owners ask
Can one enquiry have multiple categories?
Yes, the system should allow it. A message that is both a technical fault and a billing question is common, and forcing a single label means one concern gets lost unless multi-label routing exists.
How do we handle unknown topics?
Add an “unclear” category staffed by people. A message forced into the wrong bucket is often worse than one correctly flagged as unclassifiable, because it reaches the wrong team with no signal that anything is off.
How many categories are enough?
Fewer than you expect. Keep only the categories that lead to a different owner or a different next step, and add a new one only when misrouted cases show a pattern.