How to Write an AI Enquiry-Classification Rubric

AI

The AI routes a message to sales because it mentions pricing. The sender is an existing client disputing an invoice. A familiar word has produced a confident route to the wrong team.

A man arranging coloured sticky notes on an office wall.

Photo by Vitaly Gariev on Unsplash.

An enquiry-classification rubric defines categories, evidence, priority rules and an uncertain route. Build it around the decision the receiving team needs to make, rather than a loose list of keywords.

Make categories operational

Each category should lead to a useful action or owner. If two labels always receive identical treatment, ask whether the distinction adds value.

For an illustrative business, new-service enquiries, delivery issues and billing questions have different owners. Define categories using the sender's purpose and relationship with the business.

Include boundaries

For each category, record:

  • What the sender wants to achieve.

  • Evidence supporting this route.

  • Similar cases belonging elsewhere.

  • Missing information preventing a reliable decision.

  • Owner resolving uncertainty.

Include cases where familiar words mislead. Pricing in a complaint does not necessarily mean a sales opportunity. A request for an update might concern delivery rather than changing contact information.

Decide how mixed requests work

One message may request another service and raise an unresolved issue. Choose whether the workflow assigns one primary owner, creates linked tasks or sends the case for review.

Make that a business rule before asking AI to apply it. Do not lose a secondary request merely because the output template permits one label.

Write a short example showing the expected route and what the receiving owner must do with the other request.

Evaluate mistakes by consequence

Review category-level results. A high overall score may conceal weak performance on a less common but important request type.

Record misroutes, unnecessary escalations and missed secondary requests separately. Refine boundaries from those patterns. Avoid adding another label whenever one example is awkward; clarification may be the correct response.

Separate routing from commitments

Classifying a message does not authorise a refund, delivery promise or response to the client. Define later actions independently with appropriate approval boundaries.

Related guidance: guide to mapping the work before changing it.

Hili Consulting can help build an enquiry rubric reflecting your service responsibilities and the exceptions the team must handle.

Previous
Previous

How to Keep AI-Drafted Client Emails Under Meaningful Approval

Next
Next

How to Check AI-Extracted Fields Before Updating Records