M1 · Free playbook
This is the playbook we follow on every paid M1 install — written so a small-business owner can run it themselves. Same day-by-day plan, same tools stack, same QA scenarios. Free. No email gate.
TL;DR
An AI receptionist for a small business is four things wired together: a voice platform, a phone number, a knowledge base, and a couple of tools (book-appointment, capture-lead). The whole stack costs about $40 to $70 a month and takes one engineer roughly six hours of focused work spread across seven days.
The hard part is not the technology. The hard part is the prompt — what the AI says, what it never says, what it does when someone yells, what it does when someone asks if it's human, what it does when someone's in pain. Most failed installs are prompt failures, not platform failures.
What follows is the day-by-day plan, the exact stack we use, the honest tradeoffs between doing it yourself and paying us, and the QA checklist that tells you it's actually working before you flip your business line over.
The 7-day plan
Each day is built to fit in a single focused block. Skipping QA is the most common reason an install goes sideways — don't skip QA.
Day 1
~90 min
Write down every question a customer ever asks you on the phone — hours, services, prices, insurance accepted, what you do when someone calls about an emergency, what you absolutely will not let the AI answer. This becomes the knowledge base. A Google Doc works fine. Most owners want to spend three days perfecting this; spend ninety minutes and ship it. You'll edit it weekly anyway.
Day 2–3
~30 min spread out
If you got it 80% done on day 1, the next two days are for catching the questions you forgot. Every time the phone rings with something the AI wouldn't have known, write it down. Add it to the doc that night.
Day 3–4
~90 min
Sign up for the voice platform. Create a new inbound receptionist agent. Pick a voice from the library — match it to your business (warmer for dental, calmer for plumbing, sharper for legal). Paste the greeting. Paste the system prompt (including the anti-patterns blocklist — don't skip this). Connect the knowledge base document. Wire two tools: book-appointment and capture-lead. Save. Deploy.
Day 4
~45 min
Buy a Twilio number (or use the one your platform provisions). Point that number at the receptionist agent's webhook. Test it from your cell — call the new number, ask one question from your KB, verify the AI answers correctly and the lead-capture email lands in your inbox. Then set your existing business line to forward to the new number when you don't answer in three rings.
Day 5
~60 min
Call your own AI twelve times. Easy question. Hard question. Question that's NOT in the KB (it should say "let me have someone confirm"). Booking happy path. Booking when the slot is taken. "Are you a real person?" (it must answer honestly). Emergency keyword (it should flag a hot lead). Angry-customer scenario. Industry-blocked topic (legal advice if you're a lawyer; medical advice if you're a clinic). Caller wants a human immediately. Caller hangs up mid-sentence. If any test fails, fix the prompt or the KB and re-run.
Day 6
~30 min
If you're in a regulated industry — medical, legal, financial, insurance — there are specific things the AI must never do. Diagnose. Recommend treatment. Confirm whether someone is a patient. Quote past appointment details without identity verification. Read your industry's safety section, paste the relevant rails into the system prompt, and run three adversarial test calls trying to get the AI to violate them. If it escalates instead of answering, you're good.
Day 7
~15 min/day, ongoing
Turn forwarding on for real. Day 7 watch the first calls live in the platform's dashboard. Day 8 review every transcript. Day 14 review hot-lead misses. Day 30 pull the report — total calls, percent resolved by the AI, percent escalated to you, total bookings, total hot leads. The 30-day report is what tells you if it's working.
The stack
This is the exact stack we use on paid installs. You can swap any layer for an equivalent — these aren't affiliate picks, they're what we'd pick if we were installing for our own business. Which we did, before we sold any of this.
Voice platform
$29–$49/mo subscription + per-minute usage
Bland gives us the lowest-latency telephony plus the open hook to bring our own LLM brain. Synthflow is the easier-to-configure default — a good choice if you'd rather not touch a webhook. Either works.
Phone line
$1/mo per number + $0.0085/min inbound
Provision a new business number in two minutes. Auto-recording, programmable routing, and the SMS layer for follow-up texts. Every voice platform integrates with it.
Knowledge base
Free
Boring on purpose. A Google Doc lets you edit your receptionist's knowledge from your phone at a stoplight. Most voice platforms re-sync the doc on a schedule.
Intake form
Free tier covers most small businesses
When the AI captures a lead, you want that data structured — not a free-form email. Tally is free, has webhooks, and the form embeds inside your CRM in minutes.
Brain (optional)
~$0.03–$0.10 per call extra
This is what makes Iris-built different from default-Bland or default-Synthflow. The voice platform's default LLM is generic; routing through Claude with the anti-patterns blocklist enforces the human-voice rules far more reliably. Adds about five cents per call.
All-in monthly
For a small business doing 50–300 inbound calls/month, expect roughly $40–$70 all-in. The voice platform subscription is the floor; per-call cost is the variable.
The non-negotiable
Almost every failed AI receptionist we've listened to commits the same handful of phrases. “I'd be happy to.” “Absolutely. That makes perfect sense.” “Could you please provide your full name and a brief description of your matter.” They're the AI tells customers learn to recognize in two seconds — and then hang up.
We catalogued ten categories of those tells with specific banned phrases and the human replacements we use instead. Before you deploy your agent, paste the full list into the system prompt under a “NEVER SAY” section. This single step does more for voice quality than picking a better-sounding voice model.
Read the full anti-patterns blocklist →DIY or paid — honestly
Most playbook PDFs end with “or just buy our thing.” This one is honest about when DIY is the better answer.
You like to be in your own tools. You're already comfortable with Twilio and SaaS dashboards. You can spare six hours over a week. Your business is small enough that one missed lead while you're tuning the system isn't catastrophic. Most owners under $200K revenue and any owner who already self-hosts other infrastructure should DIY.
You'd rather follow a voice on the line through every click than read a PDF and figure it out alone. You don't want strangers in your accounts — you want to do the work yourself with a guide. You have ninety minutes for one call. This is the middle path between DIY and full Build.
Your time is worth more than $150/hr and you'd rather pay for the install than learn the platform. You want it live in seven days, not seven weekends. You're comfortable handing us limited-scope account access (we never touch billing or credentials we don't need). Most businesses above $200K revenue land here.
QA — the twelve scenarios
Run these from your cell phone. If any of the twelve fails, fix the prompt or the KB and re-run that scenario. Don't turn forwarding on until all twelve pass.
DIY the whole thing free. Or pay $500 and Iris is on a call with you guiding every click. Or pay $997 and we do the entire install in your accounts.
Free
Download the PDF
This playbook as a PDF you can print and check off. Nothing else.
Get the PDF →
$500
Iris-Assist
You do the clicks; Iris is on the line. One 90-min call. You keep every credential.
Book a session →
$997 + $97/mo
Iris Build Pilot
We install it for you. Seven days. 30-day money-back guarantee.
See Pilot →