Lead triage

The classifier returns a label from a vocabulary of five and an urgency from a vocabulary of three, or it returns an error. There is no "the model said 'Sales inquiry', close enough" — the workflow's Switch node matches exact strings, and close enough is exactly where a pipeline goes silently empty. Everything deterministic lives outside the prompt: normalisation knows that 0917 123 4567 and +639171234567 are the same caller, and the routing policy is a total function — every label-urgency pair, including invalid ones, routes somewhere, with the invalid ones routed to a person.

That split is the transferable part. The model is a component with a narrow contract, wrapped in validation, with business policy in code where it can be tested, reviewed and changed without wondering what else the prompt might do differently now.

The n8n workflow JSONs are not hand-maintained — a generator imports the classification prompt from the Python package and emits both flows, so the prompt n8n ships is byte-identical to the prompt the eval suite measured. A test regenerates the workflows and fails if the committed files differ; another walks every connection and fails on a dangling node reference; a third asserts the production flow carries a credential slot and no actual key.

Two variants come out of the generator. The local one runs end to end with zero accounts — a mock Gemini server answers the same wire format and a recording sink stands in for the spreadsheet — which is the same discipline as the backend repositories, where every suite runs with no broker and no Docker. The production variant swaps in real Gemini and Google Sheets appends, with an urgent-lead escalation email.

Triage is the automation with the best ratio of value to risk: it never answers a customer, so a wrong classification costs a mis-sorted queue entry rather than a wrong promise. The design puts every consequential decision — what escalates, what archives — in code a reviewer can read, and reserves the model for the judgment call software is genuinely bad at.

The same skeleton — webhook in, model answers a narrow question, validated answer drives tested routing, rows land in a sheet — is the shape of most business automation worth building: document intake, ticket routing, order-status inquiries. Swapping the vocabulary and the sinks is configuration; the discipline is the product.