Stage 1 — the call goes unanswered
Nothing changes about how your phone rings. We add conditional call forwarding to your existing line: if a call isn't picked up within a set number of rings, it forwards to a number we've registered for your business. That forward is what tells the system a call was missed.
This matters because it means we are never in the path of a call you do answer. If someone picks up on ring two, the automation never knows the call happened. There is no scenario where we intercept a live customer.
Some phone systems handle forwarding differently — a hosted VoIP PBX, a mobile carrier and an old analogue line each need a slightly different setup. Establishing which one you have is the first thing we do on the audit call, because occasionally it's the thing that makes a build harder.
Stage 2 — the first text, inside a minute
The caller gets a message that names your business, apologises for missing them, says who is writing, and asks one open question. Target is under 60 seconds; in practice it's usually well under that, because the whole thing is event-driven rather than polled.
Hi! Radiance MedSpa here — sorry we missed you. This is Ava at the front desk. What were you hoping to book?
Four deliberate choices in that message:
- It names the business first, so it doesn't read as spam from an unknown number.
- It gives the assistant a name. People converse better with a name than with a brand.
- It acknowledges the miss rather than pretending the call didn't happen.
- It asks exactly one question. Two questions halves your reply rate.
You write this with us and approve it before launch. Nothing goes out in your business's name that you haven't read.
Stage 3 — the conversation
When they reply, the AI handles it. It is scoped tightly, and the scope is the most important design decision in the product. It is allowed to:
- Answer questions about your services, using the menu and price ranges you gave us.
- Quote prices as ranges when the real answer depends on an assessment.
- Ask qualifying questions — what they need, when, whether it's urgent, whether they're an existing customer.
- Check genuine availability and offer specific times.
- Handle the ordinary friction: wrong number, "just browsing", "call me instead", "how much is it really".
It is not allowed to:
- Give clinical, legal, financial or safety advice of any kind.
- Invent a price, a time, a policy or a capability that isn't in its brief.
- Claim to be a human if asked directly. If someone asks, it says it's an assistant and offers a person.
- Argue, upsell aggressively, or keep messaging someone who has gone quiet beyond the agreed sequence.
- Discuss anyone's records, history or account details.
The escalation rule is the safety net. Anything outside scope triggers a handoff: the customer gets an honest "that's one for our team rather than me, I've flagged it and someone will call you", and your staff get an alert with the full transcript. You can see this modelled in the dashboard demo — the "Needs you" conversation is a caller asking whether a treatment is safe with their medication, which the AI correctly refuses to answer.
Stage 4 — the booking
Once they pick a time, the system writes it into your calendar, sends a confirmation with the details, and — if you take deposits — sends a payment link to hold the slot. Deposits are optional and we'd usually suggest turning them on: a booking with money attached shows up far more reliably than one without.
It respects real constraints: your opening hours, service durations, buffer time between jobs, which staff member can do what, and how far ahead you'll let people book. Those come from the intake form on day one.
Stage 5 — the write-back
The contact, the transcript, the source and the outcome all land in your CRM, tagged so you can filter for them later. If you don't have a CRM, we can put bookings in a calendar and contacts in a spreadsheet — that's a perfectly reasonable setup for a small business and we won't try to sell you a platform you don't need.
Stage 6 — the ones who don't reply
Most conversations don't convert on the first exchange, and this is where most automations quietly stop. A short, finite follow-up sequence runs — typically day 1, day 3, day 7, then a final check at day 14 or 30. Every message carries opt-out language. When the sequence ends, it ends; the system does not keep poking people forever.
The same machinery handles no-shows and cancellations: a missed appointment triggers a polite reschedule offer, which is one of the highest-return automations in the whole set because those people already wanted to come.
The five-day build
| Day | What happens | What we need from you |
|---|---|---|
| Day 1 | Intake. We start carrier registration immediately because it's the slowest step. | Service list with price ranges, top questions you get, tone of voice, calendar and CRM access |
| Day 2 | Plumbing: messaging number provisioned, call forwarding configured, calendar connected. | A short call with whoever administers your phone system |
| Day 3 | The agent: scripts written, qualifying flow built, escalation rules set, opt-out handling wired. | Nothing — we're working |
| Day 4 | Testing. We run ten scenarios including the awkward ones: angry caller, wrong number, out-of-scope question, someone replying three days late, a double-booking clash. | You read and approve every script |
| Day 5 | Go live plus a 20-minute walkthrough for whoever is on the desk. | 20 minutes of your team's time |
Total time from you: roughly 45 minutes of calls plus however long it takes to read the scripts. Everything else is ours.
After launch you get a monthly report — calls recovered, conversations held, appointments booked, revenue attributed — and we rewrite whichever message performed worst. That rewrite loop is the actual product over a twelve-month period. Automations that nobody tunes decay.
What happens when something breaks
Because it will, occasionally. Here's the honest version:
- A message fails to send. Carrier filtering, a bad number, a landline. It's logged, retried where sensible, and surfaced in your dashboard rather than silently dropped.
- The AI misreads someone. It happens. That's what the escalation path and your transcript access are for — you can read every conversation and jump in mid-thread at any point.
- Your calendar is out of sync. Bookings are written as holds against live availability, so a clash surfaces immediately rather than double-booking someone.
- You want it off, right now. One switch in the dashboard, or a message to us. It stops sending immediately.