← All builds

Build · AI receptionist

June, the missed-call text-back agent

A call goes unanswered. June texts the caller back in seconds, works out what they need, and books the job. Live on my website as a public demo. Redeploys for a real client in four swaps.

Try the live demo

The problem she closes

Industry reporting puts missed calls at a quarter or more of a home service company's inbound. Almost nobody leaves a voicemail; they call the next contractor on Google. June's job: get a text into that caller's hands before the competitor picks up.

Three phone screens: missed-call text-back, booking conversation, and booked confirmation
One conversation, start to finish: text-back 9 seconds after the missed call, two questions, real booking windows, job booked. Nobody picked up a phone.

How I built it

The website originally carried a scripted fake of this conversation. I replaced it with the real agent: a prospect who can type anything at your demo and watch it hold up trusts you more than any recording. That decision forced the engineering to be honest.

  1. Stand up the agent. An n8n workflow with Claude as the brain, session memory per caller, and two tools: book_job and request_callback. A business-brief node feeds it the shop's facts so it never invents a price.
  2. Harden the output. A cleanup pass strips the model's hidden scratch-work so reasoning never leaks into a text. A formatting node enforces the texting voice mechanically instead of hoping the model remembers.
  3. Fix what testing exposed. An early run treated "2 crews" as "2 missed calls a day", so the prompt now banks only numbers the caller actually said.
  4. Run personas until the failure modes run out. The skeptic who wants a human gets a callback path. The gas-leak message gets emergency triage before any booking talk. STOP is honoured mid-conversation, because Canadian SMS law does not care that your bot was mid-booking.
Annotated n8n canvas of the June workflow
The production workflow, annotated: intake, business brief, agent with tools, transcript log, cleanup, reply. The four go-live swaps are written on the canvas.
June triaging a gas leak message
Judgment built in: a gas smell gets the utility's emergency line and a safety check, before any booking talk.
Architecture diagram from missed call to booked job
Missed call to booked job in four stages. Redeploying for a real client swaps the business brief, phone line, booking software, and reply channel.

What the build taught me: demos win when they are allowed to fail in front of the buyer and don't. Conservative math and visible guardrails beat impressive claims, because owners have been burned by lead-gen promises before. And the boring parts, transcripts, structured records, opt-out handling, are what turn a chatbot into something a business will attach to its real phone number.

Why this matters for selling AI