What classification actually does
Every enquiry arriving in your business gets read and sorted before anyone can act: what is it about, how urgent is it, who should handle it, is it worth pursuing. Doing that automatically at the point of arrival compresses first response time from hours to minutes without adding headcount.
Design the categories from your own history
Take a few hundred real enquiries and sort them by hand first. The categories that emerge will be more useful than the ones you would have invented, and they double as the test set for whether the automated version is any good.
- Six to ten categories is usually enough; more than that and the boundaries blur
- Include an “unclear” category — forcing a guess is how errors get hidden
- Separate topic from urgency; they route differently
- Record who ended up handling it, so routing can be checked against reality
Use confidence, not certainty
A classification with low confidence should go to a person, not to a queue nobody watches. In practice this means three routes: confident and routed automatically, uncertain and flagged for a quick human decision, and clearly out of scope.
The design question is not how often it is right. It is what happens on the occasions it is wrong.
Keep the human loop visible
When someone re-routes an enquiry, that correction should be captured. It is both the safety net and the improvement mechanism, and it gives you an honest accuracy figure instead of a vendor’s.
Measures worth agreeing before you start
- Median time from arrival to first response
- Proportion routed correctly on first attempt
- Volume of enquiries needing manual triage per week
- Enquiries with no response after 24 hours — the number that should be zero
