You're the owner. You already paid to make the phone ring. Then the front desk is busy or the shop closes, and the next caller gets voicemail. This install gives that caller a real next step: a staffed callback, a human-sent text, and a booking checked against the calendar.
This playbook covers the phone. Web forms, LSA/marketplace leads, and DMs are in Lead Intercept.
Count the leak for seven days. Write the script and the never-list. Separate overflow from after-hours. Review the results for four weeks. You are buying dependable coverage, not a robot personality.
Who it's for
You run HVAC, dental, or a local service business where phone inquiries become booked work. Low volume may need a better greeting and a callback queue. Higher volume may justify a human answering service. Either way, keep the number and the process under the shop's control.
What you get when done
One intake script, two coverage runbooks, human-sent text templates, checked calendar slots, a callback log and a fifteen-minute morning review. Every action has a named owner and backup. The greeting describes the coverage you actually staff.
Time and tools
Seven days to count, one afternoon to install, then fifteen minutes every morning. Use the existing phone system, an approved business AI workspace for drafts, and the calendar dispatch actually reads. Forward calls first; do not port the public number just to try a vendor. Confirm the service can keep drafts separate from sending and calendar writes.
Prerequisites
Write the service area, hours, visit types, durations, refusals and escalation contacts. Name the dispatcher or office manager who approves and sends messages, the person who approves bookings, and their backups. If no one covers nights, say when a human resumes work.
Safety & data
Use the shop's own approved business AI workspace, with a named administrator. A paid plan alone does not make customer data appropriate to upload. Before pasting, strip names, phone numbers, addresses, dates of birth, account numbers, card data, staff HR details and privileged material. Use synthetic or properly de-identified examples by default; initials alone are not de-identification. Keep the identity key in the source system. For PHI, the exact product, features and workflow must be approved and covered by a signed BAA where required; otherwise leave the data out. Legal confidentiality and other contractual duties still apply.
A free or personal chatbot is not where customer records go. Read the customer-data Field Note and install the one-page AI use policy.
AI drafts; the named dispatcher or office manager checks the current facts and sends each response. They approve each booking before it is committed. An owner-approved recorded greeting may describe the callback process when nobody is available. Clinical, legal and urgent safety decisions go to the qualified person; the model does not triage. Check applicable recording notice and consent requirements before recording calls.
Before any customer text: ask your phone/SMS provider which registration applies. US application-to-person business messages sent over 10-digit local numbers generally require A2P 10DLC brand/campaign registration; toll-free and short-code messaging use different processes. Registration is not consent. Record the required consent, respect opt-outs, and verify the provider's STOP/HELP handling with a test. Do not infer SMS consent from a missed voice call alone. Obtain and retain consent appropriate to the sender and message purpose; permission for one reply is not permission for recurring follow-up. If permission is unclear, use the approved human callback path instead.
Step 1 — Count missed calls for 7 days
Pick a week including a weekend. Open the phone system's call history and filter missed calls and voicemail; menu names vary. Keep contact details in the approved phone system. In a separate tally record date/time, random call ID, after-hours or overflow, broad inquiry category, voicemail left, callback time, and whether it became a booking. Use those fields consistently, including unsuccessful callbacks.
Count missed after-hours calls separately from overflow while open. Calculate callback delays from the original call, not from when someone opened the log. If data has not been approved for AI, use a synthetic example to test the prompt and do the real count locally.
Summarize this permitted, minimized seven-day call tally for a local [business].
Columns: call ID, time, after-hours/overflow, broad category, voicemail Y/N,
callback time, booked Y/N. No customer names or contact details.
[paste approved table or synthetic example]
Report total missed by coverage type; peak windows; voicemails; callbacks;
bookings from missed calls; median callback minutes; missing data.
Use only the rows. Never estimate revenue or infer a diagnosis.
Done looks like. A one-page tally with a credible baseline and a named coverage gap. Hypothetical: eleven missed calls, six returned and two booked. That is a measurement example, not a forecast.
Step 2 — Write the intake script
Open intake-script.md. Write the shop name, city, hours, area, allowed jobs, excluded jobs and real appointment durations. Add the callback promise you can staff. Clinical or safety questions go to your approved escalation path; the model may not improvise advice.
Draft a phone intake script for [shop], a [business] in [city].
Hours and coverage: [ ]
Service area: [ ]
Allowed jobs and visit types, with duration/buffer: [ ]
Jobs we NEVER do: [ ]
Escalation rules and named people: [ ]
Published starting fee, or “we do not quote on the phone”: [ ]
Tone: calm, local, short sentences.
Write: an after-hours greeting; an overflow greeting; intake questions;
a refusal line; and customer-message drafts for a named human to send.
No repair prices, diagnoses, invented services or arrival promises.
Do not send or book. Ask up to five questions where facts are missing.
Keep spoken blocks under 40 words. Use only approved facts.
Read the draft out loud. Use the vertical scripts below as examples, not as facts about your business. Put the final script in the desk binder and the answering service's instructions.
Done looks like. Greeting, questions, permitted actions, escalation and never-list fit on a page. The people covering the queue have read it.
Step 3 — After-hours vs overflow rules
After-hours: the approved greeting states that you are closed and gives the next staffed callback window. Capture only what your source system needs. Page an on-call person only under the shop's documented escalation rules. If no reviewer is working, messages and booking suggestions wait in the queue. Do not imply a person will text tonight when nobody is assigned.
Overflow: while open, the dispatcher checks the queue and aims for the written callback SLA, for example [15] minutes. The backup takes over when that person is busy. The calendar may suggest capacity; the dispatcher verifies visit type, duration, travel and buffers before offering a slot.
Read intake-script.md and rules.md. Turn our actual coverage into two runbooks:
AFTER-HOURS and OVERFLOW. For each list hours, queue owner, backup, callback
SLA, approved greeting, escalation conditions and who approves bookings.
If staffing or capacity is missing, flag NEED OWNER; do not promise coverage.
Prepare internal drafts only. A named person checks and sends every response.
Done looks like. Two short runbooks; the on-call person accepts the escalation conditions. The desk knows who takes over at lunch and at closing.
Step 3.5 — Missed call → text-back in 60 seconds
During staffed hours, treat 60 seconds as a response target, not an automated-send rule or a promise you cannot keep. The phone system alerts the queue owner; that person checks consent, the number and current coverage, selects an approved template, and sends it. If the queue cannot be staffed at that speed, publish a realistic target and measure it. Human callback target: [15] minutes during business hours.
Business-hours template, human sends: “Hi, this is [name] at [shop]. Sorry we missed your call. I can call you back within [staffed window]. Is this still a good number? Please don't text sensitive details. Reply STOP to opt out; HELP for help.”
After-hours template, only a staffed reviewer sends: “Hi, this is [name] at [shop]. We received your call. Our office opens [day/time], and a person will call by [real window]. Please don't text sensitive details. Reply STOP to opt out; HELP for help.”
With no after-hours reviewer, use the approved recorded greeting and queue a callback at opening. Do not send an AI reply unattended. Neither template contains prices, a diagnosis or a technician arrival time.
Before any customer text: ask your phone/SMS provider which registration applies. US application-to-person business messages sent over 10-digit local numbers generally require A2P 10DLC brand/campaign registration; toll-free and short-code messaging use different processes. Registration is not consent. Record the required consent, respect opt-outs, and verify the provider's STOP/HELP handling with a test. Do not infer SMS consent from a missed voice call alone. Obtain and retain consent appropriate to the sender and message purpose; permission for one reply is not permission for recurring follow-up. If permission is unclear, use the approved human callback path instead.
Done looks like. With an approved staff test number and documented permission, make one test call. The named reviewer sends the selected template, the phone receives it, and a human callback is logged. Check the provider's opt-out/help handling, record timestamps, and remove the test record from the business scoreboard. Test the unstaffed case too: it must wait.
Buying an AI receptionist: 8 questions
Ask the vendor to demonstrate the controls, not merely agree to them. If the product cannot support this playbook's human-send workflow, choose a human service or draft-only setup.
- Can I hear every recording we are permitted to record and retain?
- Where is our never-list enforced, and how do we test a refusal?
- How do live handoffs work when the first person is unavailable?
- Who owns the phone number if we leave?
- How long is data kept, where, and how do we delete or export it?
- For dental/medical use, will you sign a BAA covering these exact features and subcontractors?
- Can we verify real calendar availability without overbooking, with a person approving each write?
- How do we pause it in one click and restore the old greeting/forwarding?
Step 4 — Book into the real calendar with buffers
Write the allowed visit types, duration, buffer, hours and area beside the calendar. Use actual timing, not the software default. A maintenance may need travel; a new-patient slot is not interchangeable with a procedure. These are scheduling examples, not clinical rules.
Require human approval before every calendar write. If the tool cannot enforce approval, give it no write access. It can produce an internal suggestion. The dispatcher checks the source calendar, books the verified slot, and sends the confirmation through the approved channel.
Using only this OPEN-slot list and calendar-rules.md, suggest up to two slots.
[visit type | duration | buffer | hours | area]
No matching slot means NEED HUMAN, not a shorter duration or invented window.
Output: proposed slot, constraints checked, missing facts, confirmation DRAFT.
Do not offer, book, move, cancel or send anything yourself.
Done looks like. A synthetic inquiry remains pending until approval. The dispatcher approves a test booking on the working board, checks its buffer and removes it. Install Protect the Calendar for the complete rules.
Step 5 — Morning review of approved bookings and pending inquiries
Spend fifteen minutes at opening with today's calendar, overnight call log, voicemail and a morning-review sheet. For each item, the owner or dispatcher records keep, move, kill, or human callback. A correction reaches the customer from a person before their expected window. Review is an extra check; it never substitutes for approval before sending.
Review these permitted, minimized inquiry rows against intake-script.md.
Today: [date]. Rows: [approved fields or synthetic cases].
Recommend Keep / Move / Kill / Human callback, with one factual reason.
Flag invented prices, diagnoses, excluded jobs or missing information.
For Move/Kill, prepare a short correction draft for the dispatcher to send.
Never change a booking or send. Ask instead of guessing.
Done looks like. Two weeks of dated decisions, callbacks and corrections, including quiet days. The backup can run the same review.
Step 6 — Quality check: never invent a quote, never book a job you don't do
Keep the never-list beside the queue. Test a permitted request, an excluded job, a missing address, a request for a diagnosis and an unavailable slot. Use synthetic cases.
We never invent repair prices, offer a job outside the service area, diagnose, promise unstaffed arrival times, collect card data in an AI conversation, or argue with an angry caller. We never let the assistant send, book, move or cancel. A qualified person owns clinical, legal and safety escalation; use only their approved instructions.
Audit this script and these synthetic calls against our never-list.
Flag invented prices, diagnoses, excluded jobs, unsupported time promises,
missing handoffs and any suggestion that the assistant should send or book.
For each: quote the failing draft line, explain the risk, propose a correction.
Do not add facts or customer advice. A human approves all revisions.
Done looks like. The desk and vendor use the same version. Each stop/escalation case reaches the correct person; a failed test blocks launch.
Full scripts you can steal
All SMS examples below use the registration, consent and STOP/HELP checks in Step 3.5. Staff insert identifying details only in the approved messaging system, after the AI draft is complete.
These are drafts for human staff or recorded greetings approved in advance. A person reviews each AI-prepared reply and approves each booking before either is sent or committed. Put your name, your real coverage hours, your prices only if you actually charge them, and your zips. Use only callback commitments your team can meet. Hypothetical shops below are labeled. Do not ship them with someone else's city still in the brackets.
HVAC — phone
After-hours greeting: "You've reached [shop] in [city]. We're closed now. If this is no-heat, no-cool, or water at the unit, stay on the line and leave your name, number, address, and what's going on. A person will call during [staffed callback hours]. If it can wait, leave the same info and we'll book you a diagnostic during [hours]."
Overflow greeting: "You've reached [shop]. We're with another customer. Leave your name, number, address, and whether this is no-heat, no-cool, a leak, or a tune-up. We'll call you back within [15 minutes]. You can also text this number."
Intake order: (1) Name (2) Best number (3) Address and gate code (4) No-heat / no-cool / leak / maintenance / other, in their words (5) Any water or burning smell (6) Have they shut the system off (7) Are they in our zips.
Never say: a repair price, "it's probably the capacitor," "we can be there tonight" unless on-call said yes, "we beat any quote."
SMS first text: "This is [shop] — we got your call. Reply with your address and a short description (no-heat, no-cool, leak, or maintenance) and we'll get you on the board or have a human call you. We don't quote repairs by text. Reply STOP to opt out; HELP for help."
SMS confirmation: "[Name], you're on the calendar [day] [window] for a diagnostic at [address]. Tech will call from this number when they're on the way. Diagnostic is [your published fee or 'explained on site — we don't quote repairs until we see it']. Call [number] to change or cancel. Reply STOP to opt out; HELP for help."
HVAC — if they ask "how much"
"We don't quote a repair until a tech sees the system. We can book a diagnostic for [published fee], and we'll tell you the number before we do any work. Want that window?"
Dental — phone
After-hours greeting (clinician must approve the emergency line before use): "You've reached [practice] in [city]. We're closed. For an emergency, follow [the practice clinician's approved emergency instruction]. For other dental concerns, leave your name, number, and a short description. A team member will call during [staffed hours]. We don't diagnose or quote treatment on this line."
Overflow greeting: "You've reached [practice]. Both lines are busy. Leave your name, number, and whether you need to schedule, change an appointment, or ask about a visit. We'll call you back within [15 minutes]. You can text this number."
Intake order: (1) Name in the approved phone system; identity checks happen in the practice system (2) Best number (3) Existing patient or new (4) Schedule / change / concern — their words (5) A clinical concern follows the clinician-approved callback/escalation path; do not ask the model to triage (6) Insurance only as "bring your card"; do not verify benefits on this call.
Never say: treatment recommendations, "insurance will cover it," a crown price, "come in first thing and we'll pull it."
SMS first text: "This is [practice]. We got your message. Reply 1 to request scheduling, 2 to change a visit, or 3 for a staff callback. Please don't text health details. Reply STOP to opt out; HELP for help."
SMS confirmation: "[Name], you're scheduled [day] [time] for your appointment. Please arrive 10 minutes early with your ID and insurance card. Call [number] to change or cancel. Reply STOP to opt out; HELP for help."
Generic service shop — phone (plumbing, electrical, garage, lawn, you get it)
After-hours greeting: "You've reached [shop] in [city]. We're closed. Leave your name, number, address, and what you need. If it's an emergency we cover — [your list only] — say so at the start of the message. A person will call during [staffed hours]. We don't quote jobs we haven't seen."
Overflow greeting: "You've reached [shop]. We're on the other line. Leave your name, number, address, and a short description. We'll call within [15 minutes]. Text is fine."
Intake order: (1) Name (2) Number (3) Address (4) Job in their words (5) Is it on our list (6) Access notes.
SMS first text: "This is [shop]. Got your call. Reply with the address and what you need done. If it's something we don't do, we'll say so and won't book it. No prices by text. Reply STOP to opt out; HELP for help."
SMS confirmation: "[Name] — [day] [window] at [address] for [visit type]. We'll text when we're on the way. Call [number] to change or cancel. Reply STOP to opt out; HELP for help."
Worked example (hypothetical HVAC shop)
Hypothetical: Ridge Air in a mid-size city. Two trucks in cooling season. Google ads on. Front desk leaves at 5. Owner has been taking after-hours calls off his cell in the driveway and missing the ones that don't try his cell.
Week of counting: 14 missed or after-hours calls. 9 after-hours, 5 overflow (lunch, both techs on-site, desk on the other line). 8 left voicemail. They called 5 back the next day. 1 became a diagnostic. The rest had already called someone who answered.
They choose a human receptionist service for the hypothetical coverage need. They do not port the number the first week. They forward after-hours and overflow, with a person checking the information and approving responses and bookings.
Script facts they wrote for this hypothetical shop: zips they run; they don't do rooftop commercial; diagnostic is a published $89 and waived if they do the repair; they never quote replacements on the phone; no-cool with a house over a temperature they care about pages on-call; a person may approve maintenance in real slots with 20-minute drive buffers; change-outs require separate review.
They submit a synthetic diagnostic inquiry from the owner's phone at 8:12 p.m. A reviewer checks the proposed slot before approving the test booking for 10:30 the next morning with a buffer. They remove the test. A later no-cool inquiry arrives with an address transcribed incorrectly; the dispatcher checks the source, corrects it, approves the window, and sends the confirmation. The draft never becomes an unreviewed booking.
Two weeks later they are not "AI transformed." They are booking some of the overnight diagnostics instead of gifting them to whoever picked up. Effectiveness, Efficiency, Quality. Hypothetical walkthrough, not a client claim.
Operator Standard check
Trace: Separate seven days of missed calls into after-hours and overflow.
Choose: Pick coverage that fits the measured gap and name its human owner.
Build: Install the script, never-list, human-send queue and calendar checks.
Prove: Compare callback speed and bookings with baseline; review every correction.
Failure modes / what NOT to automate
Do not let AI send customer responses, book, move, cancel, or take other consequential actions. The named human checks the draft and takes the action. Do not let it quote (not a range, not "most customers pay") or diagnose. Do not approve jobs you don't do — rooftops, sedation, out-of-area addresses that look close on a map. Do not move a confirmed install or a reserved surgical block merely to make room for a new lead. Do not replace the morning review with a dashboard you don't open. Do not port your public number into a vendor you haven't run for two weeks on forwarding. Forward first. Do not skip the emergency page: water in the house, suspected gas, a patient who cannot breathe — human. Choose coverage based on the counted need and the approval process you can actually operate.
If it goes wrong (kill switch)
The owner or dispatcher can pause immediately. Disable the vendor forwarding/rule, restore the saved greeting, and forward to the staffed cell or human service. Stop sending pending drafts until checked. Keep the inquiry queue intact. A named person calls anyone given a wrong window: “We gave you incorrect information about [fact]. The confirmed next step is [verified step]. I'm sorry.” Do not blame the tool or promise a new slot before checking it.
Prove it: 4-week scoreboard
Use the same definition and observation window in every column. Fill the baseline before the install; leave unknowns blank, not zero. The thresholds below are decisions you set, not promised results.
| Metric | Baseline | Week 1 | Week 4 | Keep-Kill decision |
|---|---|---|---|---|
| Missed calls/week | [record] | [record] | [record] | Keep only if coverage fits demand; investigate a rising count. |
| % called back in under 15 staffed minutes | [record] | [record] | [record] | Set [target %]; change staffing if the target is missed. |
| Bookings from missed calls | [record] | [record] | [record] | Compare like weeks; keep only with accurate bookings and acceptable cost. |
Save it as a skill
Save the final prompt and never-list as skills/never-miss-another-call.md in the shop's shared drive, or in the approved project instructions your AI workspace supports. Keep one owner and one dated master copy. Copy this instruction block and replace every bracket before use.
Skill: Never Miss Another Call
Owner / backup: [dispatcher name] / [name]
Read company.md, customer.md, rules.md and intake-script.md, current versions.
Use only permitted minimized records. Ask when facts, consent or staffing are unclear.
Draft callback summaries and message options; never send, quote, diagnose or book.
Never-list: [paste shop's refusals, price limits and escalation rules].
A named human checks consent, coverage and calendar, then sends and books.
Pause: restore the saved greeting/forwarding; owner reviews pending work.
Log callback time, booking outcome and every correction; review each morning.
Version / reviewed by: [date] / [name]
Continue with Lead Intercept, calendar protection and human-sent review replies. Read the Field Notes on receptionist value and front-desk work.
FAQ
Should I buy a receptionist or DIY the SMS loop? Use the tally to decide whether you need human answering coverage or can handle the work internally. Either way, a person reviews AI-prepared replies and approves bookings before they happen. Low volume may fit an approved greeting, a callback queue, and staff-reviewed SMS drafts rather than a new platform.
What if the vendor wants to "sound friendly" and ad-lib? No. Friendly is fine. Ad-lib is how prices appear. Give them the never-list and listen to the first twenty recordings. If they cannot stay on script, fire them. An unsupported quote can cost more than the answering service.
Can it take payment or insurance details? Not in this playbook. Collect a card in person or on a link you already use after a human has set expectations. Insurance: "bring the card." Verification is a human. A model that stores card numbers in a chat log is a problem, not a feature.
Won't people hate talking to a machine? Do not assume how callers will react. This playbook uses AI to prepare work for people, with an approved recorded greeting when live coverage is unavailable. Keep the callback process honest and provide the human escalation you have actually staffed. Review complaints and missed handoffs instead of relying on a vendor demonstration.
Screen-recording notes (6–12 min, lo-fi)
Phone on a tripod. No smiling-family B-roll.
- 0:00–0:45 — Camera. "You already paid to make the phone ring. Then it goes to voicemail." Paper tally, not a slide.
- 0:45–2:30 — Synthetic missed-call log. Count after-hours vs overflow. Circle the peak hour. Keep real customer and patient records out of the recording.
- 2:30–5:00 — Step 2 prompt into ChatGPT or Claude. Delete a price the model invented. "That's the whole job — catching the lie."
- 5:00–7:30 — Calendar: show a proposed diagnostic with a buffer waiting for human approval, an excluded visit type, and a synthetic test booking that appears only after approval.
- 7:30–9:30 — Morning review. Three overnight rows. Keep / move / kill. Draft one SMS. Send nothing.
- 9:30–11:00 — Never-list on paper. Optional: HVAC SMS script on screen.
- 11:00–end — Close: "Let’s talk about your business. Start with a free introductory conversation. The $1,500 AI Workflow Sprint helps your team use AI on real work; implementation is scoped separately." Direct viewers to /start.