AI enquiry classification should help the right team take over the right work, not merely attach a label. Define category boundaries, mixed requests and a review route for uncertain records together. The worked messages below show what to test before enabling automatic routing.
Build categories around customer needs
Build categories around distinctions that change the customer need and team decision. Give each a definition, an included example and an excluded example. Review the other category for missing definitions or unclear explanations.
One message can contain several tasks. An existing customer may ask for a new project and support in the same enquiry. Do not lose the support request by sending everything to sales. Agree with the teams which follow-up should be created for primary and additional needs.
Provide relevant context without treating guesses as records
The text may need support from the page it came from, a selected service or a verified customer relationship. Record which input fields are available and where they originate. Do not include unrelated personal details just because they exist in another system.
A request to add something to an existing package is ambiguous without account context. AI may suggest an expansion enquiry, while selecting the account owner still requires a reliable customer match. Interpreting the message and identifying the correct customer record are different tasks and should be evaluated separately.
Separate single labels from mixed needs in worked messages
These messages were written for testing; they are not customer records. Agree the expected routes with the receiving teams. The model’s own confidence score is not proof of correctness; evaluate the decision against observed errors in relevant examples.
- Who reviews a request that cannot be classified?
- Who owns an enquiry involving several teams?
- How can a recipient correct a mistaken route?
- What happens to records arriving outside working hours?
- Which facts are verified before an automatic acknowledgement is sent?
| Example message | Expected distinction | Follow-up |
|---|---|---|
| I need a quote for a new website. | New project / web | Draft record for the relevant sales owner |
| My current site is down; let’s discuss a new module too. | Support + additional project | Preserve both needs without losing the support task |
| I want to change my plan. | Customer and plan context missing | Ask for relevant clarification before record matching |
| I am not sure what we need yet. | Unclear requirement | Clarification or colleague review |
| Can I speak to accounts? | May be outside defined service categories | Provide the appropriate route instead of forcing a sales label |
Measure what reaches the team as well as the label
Record the expected category and next task for test messages with the team first. When comparing AI suggestions, distinguish unclear category definitions, missing context and incorrect interpretation. Do not use every example both to adjust the system and to measure its performance; test new examples too.
Track the wrong team, manual reassignment, unowned records and first meaningful response alongside label accuracy. A correct category with no follow-up record is not a completed route. Update the examples and destination rules when adding a service.
Version category definitions. Combining two categories can change what a period percentage measures. Explain how historical records are remapped or state the boundary of any comparison in the report.