JEV sorts n8n inquiries by meaning, not keywords

The setup here is deliberately narrow: AI handles only the middle step of reading meaning and choosing from candidates, while the entry form and the downstream recording stay as ordinary n8n nodes. That means an existing automation does not need to be rebuilt — a single judgment box can be inserted into the flow. The article's example stops at writing a row into a Data Table rather than performing the actual booking change or cancellation, keeping the original text and the AI's decision on the same line so a person can review before anything irreversible happens.

The distinction from prompting an LLM for classification is structural. JEV's answer is always one of the supplied candidates, so no out-of-set wording or explanation comes back, and each answer carries a confidence score that can route uncertain items elsewhere. It does not generate text at all, so it is unsuited to writing replies or summaries. The article is explicit that choosing the wrong candidate still happens, citing the official Agent Skill line that typed output guarantees the interface, not truth.

The confidence score behaves less like a safety net than one might hope. Fallback catches answers below the threshold, but a sentence containing both a change and a cancellation scored 0.96 and passed through the normal branch, which the article reads as evidence that confidence measures the concentration of a probability distribution rather than overall workflow correctness. The stakes therefore hinge on where the human check sits: for teams, the practical reading seems to be that reversible reception and recording can be delegated while cancellations or refunds stay behind a person, and the 0.65 threshold is presented as one trial setting rather than a recommended value.