Free playbook · no email required
Every rule below was written after a real call went wrong. We spent seven months running a working laboratory's entire inbound flow — their phone numbers, their PBX, 2,717 calls in a typical month. This is what actually separates an AI that books the appointment from one that makes callers ask for a human.
Works with any AI phone agent. Nothing here is SIPmind-specific.
This is the single highest-impact rule, and almost every failing agent breaks it. A voice model streams audio as it generates — so a three-sentence answer is twelve to eighteen seconds of talking. The caller doesn't hear "thorough". They hear a machine that won't let them speak, and they interrupt or hang up. A chat reply can be long. A spoken one cannot.
GOOD
"We're open until six today. Would you like to come in this afternoon?"
BAD
"We're open Monday through Friday from nine to six, Saturday from ten to four, and we're closed on Sunday. We also have a second location that has different hours, and during holidays the schedule may change, so…"
Practical test: read the agent's answer out loud. If you run out of breath, the caller ran out of patience.
An AI that lacks a booking tool will still happily say "You're all booked for Thursday!" — because that's what a helpful receptionist says. Nothing was booked. The customer shows up and nobody is expecting them, and that costs you more than the missed call ever would.
Put this in your prompt, verbatim: never state that an action was completed — booked, sent, cancelled, transferred, registered — unless the tool that performs it returned success. Never invent a confirmation number, a link, or a time. If it can't do the thing, it should say what happens next instead: "I'm noting this down and a colleague will confirm within the hour."
Never let it spell out numbers on its own. Ask a model to say a phone number and it drifts — inventing a digit that doesn't exist. Write the spoken form out yourself, once, and have it read that verbatim. If the caller doubts it, repeat the same words and offer to connect them. Never argue about how many digits a number has.
Never say "write to us" on a phone call. They're holding a phone to their ear. Say "tell me" and "say it out loud."
Greet exactly once. After a pause or background noise, agents love to restart with a fresh hello. To the caller it sounds like the line reset. Say "I'm listening" instead.
Answer the question that was asked. "What's your address?" gets the address — not "which service are you interested in?" If you need one detail, ask for that one detail.
Don't read the catalogue. Offer one or two options and ask a question. A list of eleven services is unlistenable and it moves nobody forward.
Never guess garbled audio. A misheard city or name stated as fact is worse than asking again. Re-ask in the same language.
If they ask for the cheap or free option, explain how to get it first. Ending on a bare "no" loses a customer who was entitled to a yes.
Everyone worries their AI won't transfer often enough. In practice the expensive failure is the opposite. We changed a prompt and switched on an eager auto-transfer rule, and watched our own handoff rate climb from 28% to 41% in days — meaning hundreds of calls the AI could have finished were dumped on staff who were busy. Handoff rate is a metric to watch weekly, not a setting to max out.
Transfer when: the caller asks for a person, is angry, needs a decision
only a human can make, or the question is outside what the AI actually knows.
Don't transfer when: it's a routine question with a
known answer, or when the caller just said "no, that's all" — that's the end of a good call, not a
reason to escalate.
Every underperforming agent we've audited was missing at least three of these. Score yours honestly.
The facts about your business — services, prices, hours, addresses — should be the only part you rewrite per location. Everything above is behaviour, and behaviour is the same everywhere.
Day 1 — Let it answer only what nobody is answering. After hours and the second line while staff are busy. Don't replace anything that already works; add cover where the silence is. That's where the return is, and it makes the first week impossible to lose.
Day 2 — Write the greeting yourself. On clean VoIP lines (WhatsApp calls, most SIP trunks) there's no line noise for the model to detect, so an agent told to "greet the caller first" may simply sit in silence until the caller speaks. An explicit greeting removes the whole class of problem.
Day 3 — Read ten transcripts end to end. Not metrics. Transcripts. You will find the two sentences that are costing you callers, and they're never the ones you expected.
Day 5 — Check the handoff rate. If it's above roughly a third, something in the prompt is teaching it to give up.
Day 7 — Ask what it should have known. Every "I don't have that information" in a transcript is a missing line in your facts section. That list is your knowledge base.
It won't close every call. Our best measured stretch was 71.4% finished with no human — 2,620 calls over six weeks, counted on the client's own PBX log rather than our telemetry. The rest needed a person, and that's the correct outcome.
Booking is not conversion. An AI can take the appointment. Whether people actually show up is a business problem, not a model problem.
It will be wrong in public sometimes. Plan for that: a clear handoff path, and someone reading transcripts weekly for the first month.
SIPmind is the product this playbook came out of — an AI employee that answers your phone with a real voice, replies on WhatsApp, Telegram and Instagram, books appointments and keeps the CRM. Keep your number and your phone system; nothing gets ripped out.
Or send a DM to @inndmitrii and describe your business — the AI itself will tell you what it would take off your plate.