How to Train Your AI Receptionist

Before you can train one, it helps to understand the loop. An AI receptionist does five things on every call, and every one of them is a decision point you control. First, it answers the call and delivers a greeting — the first three seconds where a caller decides whether they trust the voice on the other end. Second, it identifies intent by listening to the caller's opening sentence and mapping it to a known call type: booking, quote, status check, emergency, callback for a specific person, or something weirder. Third, it qualifies by asking the right follow-up questions for that intent — not all questions, the right ones in the right order. Fourth, it takes action: it books the slot in your calendar, creates a CRM contact, pages the on-call tech, or drops a message in the callback queue. Fifth, it confirms what it did out loud and disconnects. That five-step loop is the entire job. Every other feature — call recording, SMS follow-up, voicemail transcription, after-hours routing — is a variation on one of those five beats. When your flow feels off, it is almost always because one of those five steps is weak.

The biggest shift for most owners is mental. You are not writing a script. A script tries to predict every word the AI will say and ends up sounding robotic when the caller goes off-plan. You are writing a decision tree: if the caller says X, go to branch A; if they say Y, go to branch B; if neither, fall back to a clarifying question. Think in 'if this, then that' plus a fallback for when neither 'this' nor 'that' shows up. Start with the three most common call types for your business and ignore the long tail for now. You will expand coverage as you learn which edge cases actually happen versus which ones you imagined on a whiteboard. A flow that handles eighty percent of calls cleanly is better than a flow that tries to handle one hundred percent and fumbles every third one because the logic is too deep to maintain. Build narrow, ship, listen, widen.

“

A script tries to predict every word the AI will say and ends up sounding robotic when the caller goes off-plan.

**** — **** — **** — **** — **** — **** —

**Step 1: Capture your top 5 call types** You cannot train a flow for calls you have not characterized. Spend an hour listening to twenty recent calls from your voicemail, call log, or recorded line. If you do not record calls, sit down with whoever answers the phone and have them rattle off a week's worth from memory. Cluster each call into a type: new booking, quote request, status check on an existing job, emergency, complaint, vendor or supplier call, job applicant, wrong number. Tally them. You will almost always find that three types account for about eighty percent of inbound volume. Those three are where you focus your flow. The rest either do not need a smart flow (hang up on wrong numbers) or happen rarely enough that a simple 'let me take a message' branch is fine. _Example:_ A plumbing shop runs the numbers on forty inbound calls: 18 new bookings (45%), 9 quote requests (22.5%), 6 emergencies (15%), 3 status checks (7.5%), 2 vendor calls, 2 wrong numbers. The flow gets built for bookings, quotes, and emergencies. Everything else falls through to a message-taking fallback. **Step 2: Write the greeting** The greeting is the most overworked and most under-thought part of the whole flow. Lead with your business name, not with a throat-clearing phrase like 'thanks for calling.' Callers dialed the number on purpose — they know they called. Next, name the two or three things the AI can actually help with, so the caller can self-route without being asked. Keep it under twelve seconds. Avoid 'how may I direct your call today' — it is too formal for service trades and frankly too formal for most businesses. The goal is for the caller to hear, within the first five seconds, both who they reached and what they can do next. _Example:_ Good: 'Hi, this is the AI receptionist for ABC Plumbing. I can book you an appointment or get you a quote — what's your call about?' Bad: 'Thank you for calling ABC Plumbing, your trusted partner in residential and commercial plumbing services. How may I direct your call today?' The second one burns ten seconds and gives the caller nothing to work with. **Step 3: Build the qualifier branches** Each call type gets its own qualifier branch — a short, ordered list of questions the AI asks once it has identified intent. Order matters. Ask the question that most narrows the path first, then the logistical questions, then confirmation. For bookings, that usually means: what's the issue, what's the service address, and when do you need us. For quotes, the issue is broader so the first question is broader: what are you trying to do, followed by sizing questions (square footage, number of units, scope), followed by timeline. For emergencies, flip the order — triage severity before anything else, because the answer changes what happens next. A water leak that is actively running means you dispatch now; a water leak the caller already shut off can wait for a morning booking. _Example:_ Booking branch (plumbing): AI asks 'what's going on?' Caller says 'my kitchen sink is backed up.' AI asks 'what's the service address?' Caller gives it. AI asks 'when do you need us — today, tomorrow, or later this week?' AI books the slot and confirms. Emergency branch: AI asks 'is there water actively flowing or leaking right now?' If yes, AI says 'I'm paging our on-call tech now, please stay near the phone' and fires the SMS. If no, AI routes to the booking branch. Quote branch (roofing): AI asks 'what are you trying to do — full replacement, repair, or just an inspection?' Then 'roughly what's the square footage or how many stories?' Then 'what's your timeline — this month, this year, or just exploring?' Each branch has three to four questions, not ten. **Step 4: Integrate your calendar and CRM** A flow that cannot actually book a job is a glorified answering machine. Connect the calendar your shop actually lives in — Google Calendar, Outlook, or the scheduler inside ServiceTitan, Jobber, Housecall Pro, or whatever field service platform you use. Define bookable slots honestly: block out lunch, block out the hour you reserve for morning staff meetings, block out the afternoon your lead tech is always buried. Set the earliest available slot to respect your dispatch reality — if you need two hours of lead time for a tech to get there, do not let the AI offer appointments in the next ninety minutes. On the CRM side, tag every captured contact by call type so your morning follow-up workflow can route leads differently from emergencies and status checks. A new quote lead should trigger an email from your estimator; a completed booking should trigger a confirmation text. These integrations pay for themselves the first week. _Example:_ A Jobber integration flow: AI identifies a booking, collects the address, offers three available slots pulled from the Jobber calendar ('Tuesday 10am, Tuesday 2pm, or Wednesday 9am?'), caller picks Tuesday 2pm, AI creates the job in Jobber with the address and issue description, and tags it with source 'AI-phone-booking'. The dispatcher sees it the next time they refresh the board. **Step 5: Write fallback and escalation rules** The single biggest reason AI receptionists fail in the real world is that teams forget to write the escalation path. Every flow needs an answer to 'what happens when the AI does not know what to do?' — and that answer cannot be 'I don't understand, goodbye.' That is how you lose customers. The rule is simple: if the AI hits a dead end, it says something like 'let me text the shop, someone will call you back within thirty minutes' and fires an SMS to the on-call person via a Twilio webhook or the equivalent in your stack. Everything the caller said up to that point — transcript, phone number, intent guess — goes into a callback queue that a human reviews on a schedule. The caller hangs up knowing a human is about to call them. That is a graceful fallback. Build this branch before you go live; do not leave it for later. _Example:_ Caller asks about a warranty on a water heater installed three years ago. The AI does not have warranty lookup wired up yet. Instead of guessing, it says: 'I'll have the office pull your warranty details and call you back within the hour — is this the best number to reach you?' It confirms the number, tags the callback as 'warranty-lookup', and texts the office manager. Thirty minutes later, a human calls back with the answer. The caller never felt stuck. **Step 6: Test with 20 calls** Before you turn the AI receptionist loose on real customers, put it through twenty test calls. Call it yourself from five different phones — your cell, a landline, a VoIP line, your spouse's phone, a cheap prepaid. Different carriers and handsets surface different audio quirks. Have a friend or a family member call and deliberately stress-test it: talk over the AI, change their mind mid-sentence, mumble the address, give a partial answer. Try edge cases: an angry caller venting before stating the problem, a Spanish speaker if you serve that market, a very short caller who just says 'appointment' and waits. Write down every place the AI fumbled and fix the flow before launch. This is the single highest-leverage hour of setup work. _Example:_ Test call reveals that when the caller says 'I need someone out ASAP,' the AI routes to the general booking branch instead of the emergency branch because 'ASAP' was not in its intent list. Fix: add 'ASAP', 'right now', 'urgent', and 'today' as emergency-adjacent phrases that trigger the triage question. Another test call reveals the AI cannot understand addresses spoken with unusual street names — fix is to always repeat the address back and ask for confirmation. **Step 7: Iterate weekly for month one** Every Monday for the first four weeks, sit down with ten recorded calls from the previous week and listen to them with a pen in hand. Note every place the AI misunderstood the caller, gave a clunky answer, failed to book a bookable job, or handed off to humans when it could have handled the call itself. Those notes become your change list for the week: adjust prompts, add intent phrases, tighten branch logic, rewrite a confusing question. By week four, the flow should feel stable — you will still find things to tune, but you will not be making structural changes every Monday. That rhythm is what separates teams whose AI receptionist keeps getting better from teams whose flow stagnates and gets abandoned six months in. _Example:_ Week one Monday review: four of ten calls had friction — the AI said 'sorry, can you repeat that?' more than twice on two calls, and it offered a booking slot during a blocked lunch hour on one call. Fixes: lower the 'repeat that' threshold and add the lunch block to the calendar. Week four Monday review: nine of ten calls clean, one weird one where a caller kept switching between English and Spanish mid-sentence. That goes on the backlog for the bilingual upgrade.

Your flow will suck in week one. That is normal and it is not a sign the technology is broken. It is a sign you are doing the job honestly — you cannot know what callers will throw at you until real callers start throwing things. Ship the week-one version anyway, listen to what happens, and fix the top three friction points every Monday. By week four you will be booking calls you used to miss, capturing leads that used to hang up on voicemail, and handing off cleanly to humans when the AI hits its limits. That is the difference between an AI receptionist that becomes part of the business and one that gets turned off after a month because nobody tuned it. Build the flow, test the flow, iterate the flow. The rest takes care of itself.