Every enquiry scored and sent to the right team

An enquiry arrives, the AI (Claude) reads what the person actually wrote, and plain code decides how warm it is and who should pick it up. The team gets the alert, and the tracker behind it feeds a dashboard anyone can open.

Working demo, run 23 Sep 2026 on 8 sample enquiries: 5 routed to a team, 1 held for nurture, 1 blocked as a repeat, 1 sent for a person to read. All 8 accounted for.

8 → 8Sample enquiries in, all accounted for
51/51Automated checks passing on the live code
36Steps in the finished build
2 hResponse target set for a hot lead

How it works

Five steps, from the moment someone writes in to the moment the right person has it.

The workflow on the n8n canvas, left to right
  1. Someone writes in

    A form, a website, or anything that can pass the details on.

  2. It is checked for repeats

    The same person clicking twice is recognised and stopped before any work happens.

    Second click
    Same enquiry, seconds later
    Result
    Repeat — stopped, nobody alerted twice
  3. The AI reads what they wrote

    It returns six facts and nothing else: what they want, company size, budget, timing, their need, a summary.

    Message
    “Q4 budget approved, live within 6 weeks”
    Read as
    New business, 1,200 staff, budget confirmed
  4. Plain code scores and routes it

    Not the AI. The facts become a score out of 100, a band, a team, an owner and a response target.

    Score
    85 of 100 — hot
    Goes to
    Enterprise sales, 2-hour target
  5. The team is told, the tracker updated

    The owner gets the alert with one-click status links. The tracker row updates, and the dashboard shows it straight away.

How it’s built

Three decisions that make the difference between a demo and something a sales team can rely on.

The AI reads. The code decides.

The model never picks a team. It answers six questions from fixed lists, and plain code turns those answers into the score, the owner and the response target. So the same enquiry always lands in the same place, every rule can be read and changed by the business, and an odd answer from the model cannot send a lead somewhere strange — anything outside the list counts as unknown and goes to a person.

The repeat check closes before the AI starts.

The tracker row is written before the model is asked anything. Reading a message takes a few seconds, and a second click inside those seconds is exactly when a duplicate slips through. Claiming the enquiry first makes that impossible, rather than unlikely.

Teams and rules live in one place.

Teams, addresses, members, the scoring weights, the hot and warm thresholds and the response targets all sit in one settings block. Adding a team is one line, not a new branch on the diagram, so the build does not grow every time the sales structure changes.

Also built

A dashboard anyone can open

Totals, a day-by-day chart and a “needs attention” list: leads past their response target, leads waiting on a human, and any that got stuck part-way. Built from the tracker on every load, so there is nothing to keep in sync.

Status links in the alert

Contacted, meeting booked, proposal sent, won, lost — one click in the alert, no logging in anywhere. Tested with a valid lead, an unknown one and an invalid status; all three answered correctly.

Nothing disappears quietly

An enquiry nobody can read is tagged with the reason and sent to a person. An empty delivery still produces a summary. A row stuck mid-way shows on the dashboard as stuck — the visible sign of a run that died.

What was verified

Counted from the real run, not inferred from a green tick.

The build validated clean: 36 steps, no errors, one expected warning for the switched-off email step standing in for the real one. It ran live on 23 September 2026 on eight sample enquiries. Counts read step by step: 8 in, 7 past the repeat check, 6 claimed, 6 scored, 4 alerted and 2 held, 8 accounted for at the end. Nothing was dropped.

Enterprise enquiry → enterprise sales, hot, 85. Agency → smaller-business sales, warm, 50. Hobbyist → nurture, cold, 20. An agency pitching its own services → disqualified. The double click → repeat, skipped. A broken email address → held for a person to read.

Four more cases ran on the live build: an empty delivery, four malformed entries each tagged with its own reason, the same enquiry sent twice across separate runs, and the AI being unavailable — which falls through to a person rather than guessing. Alongside that, 51 automated checks pass against the code taken straight out of the deployed build, so the tests cannot drift from what is running.

Not verified: a real email send, and a submission typed into the form in a browser — the form is live, but the demo drives the same path directly. The demo uses Claude (Anthropic); swapping in OpenAI is one step on the same slot. The AI also read “Q4 budget, live within six weeks” as this quarter rather than this month; the lead still scored hot at 85, which is the kind of small misreading the scoring absorbs.