HVAC Call Script: Train Your CSRs and AI on the Same Script

Write the script as a set of rules, not a speech: one opening, qualifying questions in a fixed order, booking rules, routing for sales and emergency calls, and exact wording for safety situations. Your CSRs read it from a one-page card, the AI receptionist gets the same rules as instructions, and one named person owns the single document.

Most companies end up writing the script twice, a friendly version for the office and a separate prompt for the AI. Within a month the two disagree about what counts as an emergency, how much to say about price and when a caller goes to sales. Splitting the script into fixed rules and flexible wording stops that drift, because only the wording is allowed to differ.

Follow me on Instagram@sagnikteaches

Rules stay fixed, wording is allowed to differ

A rule is anything you would be unhappy to find two people had told a caller differently: which calls are emergencies, which questions come before a booking, what you will and will not say about price, who takes a replacement enquiry. Wording is simply how a rule is spoken. A CSR with fifteen years on the phones will ask "is the outdoor unit running?" differently from someone in their second week, and that is fine. An AI receptionist works the other way round. It needs the rule spelled out, and wherever you leave a gap it fills it with wording of its own.

Connect on LinkedInSagnik Bhattacharya

So the master document holds rules first and spoken lines second. The staff card is the rules plus friendly example lines. The AI version is the same rules, made explicit and cut to fit the product's settings. Jobber's AI receptionist page, for example, describes escalation keywords such as "emergency", "no heat" and "burst pipe", plus rules for what it should and should not do on a call. Housecall Pro's CSR AI page mentions scripts, hours, booking rules and a number it transfers callers to. Neither product reads your mind, so each needs your rules typed into its own fields.

Subscribe on YouTube@sagnikteaches

Write the source rule first and the two renderings second. If a rule cannot be turned into a line a CSR can say and an instruction an AI can obey, it is still too vague to be a rule. "Prioritise urgent calls" is a wish. "If anyone in the home is under two, over seventy-five, pregnant or relies on medical equipment, and the home has no heating or cooling, book the first priority slot and text the on-call lead" is a rule, and you can test it with a phone call.

A master script for a no-cooling call, filled in

Here is the skeleton for the commonest summer call, written for an imagined company with three CSRs, nine technicians and an AI receptionist covering evenings and overflow. Square brackets are the company's own facts; everything else is the rule. Each stage gives the rule, then the line a CSR says, then the instruction the AI receives.

1. Opening. Rule: name the company, say who is speaking (a person's name, or "the virtual assistant"), and let the caller finish the first explanation. Staff: "Thanks for calling [Company], this is [name], how can I help?" AI: open with "Thanks for calling [Company], you've reached the virtual assistant", then do not interrupt.
2. Safety screen. Rule: on any call that mentions a smell, a noise, an alarm, water or no heat, ask three safety questions before anything else. Any yes ends the normal script. Staff: "Before anything else: can you smell gas or burning, has a carbon monoxide alarm gone off, or is there water near anything electrical?" AI: ask all three, in that order, in one breath. On any yes, say the Tier 1 wording exactly and send the on-call alert. Ask nothing else.
3. Contact details. Rule: name, mobile number and service address, each read back; ask whether the caller owns the property or rents it. Staff: "Can I take your name and the best mobile number?" AI: capture all three, read the number and house number back digit by digit and wait for a yes, then ask owner or tenant.
4. What is happening. Rule: one open question, then four closed ones (when it started, what the thermostat shows, whether the outdoor fan is spinning, whether air is coming from the vents). Nobody asks a caller to open a panel, remove a cover or touch wiring. Staff: "Tell me what's going on, then I'll ask a few quick questions." AI: ask the open question, then the four closed ones. The only things a caller may check are the thermostat display, the filter and visible water.
5. System details. Rule: system type (furnace, air conditioner, heat pump, boiler, mini-split), rough age, brand if visible, plan member yes or no. Staff: "Roughly how old is the system, and do you know the make?" AI: accept "not sure" and move on. Never guess a brand or age.
6. Occupants. Rule: the vulnerable-occupant question from the table above, on every no-heat or no-cooling call. Staff and AI use the same sentence.
7. Tier and slot. Rule: Tier 1 never reaches this stage. Tier 2 (vulnerable occupant with no heat or cooling, or water leaking but clear of electrics) gets the first priority slot. Everything else is Tier 3 and gets the next standard two-hour window. Staff: "I'm putting you on our priority list; the earliest window is [window]." AI: offer only windows that appear in the calendar, state a window and never a time, and never promise same-day unless the calendar shows a same-day window.
8. Fee and terms. Rule: state the diagnostic fee and what it covers before confirming; repairs are priced on site; get a spoken yes to the fee. Staff: "The diagnostic visit is $[fee], and the technician prices any repair before starting." AI: use the same sentence word for word and record the yes in the notes.
9. Recap and close. Rule: read back name, address, window, problem and fee; say what happens next; give the safety net. Staff and AI: "If you notice a gas smell or an alarm goes off before we arrive, leave the building and call the emergency number, then call us."

Nine stages is the right size for one call type. Heating calls swap stage 4 for heating questions, tune-up calls skip stages 2 and 6, and a replacement enquiry skips everything after stage 3 and goes to the routing section below. Keep each call type as its own page inside the one document, so a CSR is never hunting through a manual while someone waits on the line.

Emergency wording that nobody improvises

Every call is sorted into one of three tiers within the first half minute, and the tier decides who handles it.

The Tier 1 lines are the one part of the script I would say word for word every time. What follows is a draft to start from. Your own safety lead and insurer should approve it, and you should compare it with your gas supplier's published advice, because their wording wins wherever it differs.

"I need you to do something before we talk about the repair. Please leave the building now, with everyone who is inside, and don't switch anything on or off on the way out. Once you are outside, call the gas emergency number or the emergency services. Please don't go back in until they say it is safe. I am alerting our on-call manager now, and a technician will come once it is safe to do so."

Three things about that wording are deliberate. It tells the caller what to do before it asks anything. It never reassures ("it's probably nothing") and never diagnoses ("sounds like a cracked heat exchanger"). And it promises no arrival time, because the technician cannot attend until the emergency service or supplier has cleared the property. For carbon monoxide, the public safety guidance I read for this tutorial says to move outside into fresh air first, not to go looking for the source, and to call the emergency number from outside. One government page lists headache, dizziness, weakness, upset stomach, vomiting, chest pain and confusion as the commonest symptoms, and warns that people who are asleep can die from it before they have any symptoms. Your script should not try to be cleverer than that.

The AI rules around Tier 1 are short. Say the wording exactly. If the caller says "it stopped" or "it's probably fine", repeat the key sentence once, then send the alert anyway. Never tell a caller it is safe to stay. Never continue into the booking questions. The same applies to a CSR who is tired at the end of a long day, which is the real reason to fix the wording instead of trusting judgement. The wider design of emergency handling for an after-hours service is covered in the tutorial on setting up an AI answering service that never misses an emergency.

Booking rules both sides can follow

Booking rules are where staff and AI drift apart fastest, because each gets small decisions wrong in a different way. Staff are generous: they squeeze one more call into a full day. An AI is literal: it offers whatever the calendar shows. Writing the rules as if-then lines settles both. These are the eight I would start with, with placeholders for your own figures.

• Windows, not times. Offer two-hour arrival windows only. Nobody promises "around 2:15".
• Fee before the booking. State the diagnostic fee ($89 in this imagined company, yours will differ) and whether it is credited against a repair, before asking for the caller's agreement.
• No repair prices, no diagnosis. "Sounds like the capacitor" is a diagnosis, and a caller will repeat it to the technician as fact.
• Same-day only from the calendar. A same-day slot exists if dispatch has released it. Neither staff nor AI offers one that is not showing.
• Tenants and landlords. A tenant can book a diagnostic visit. Repairs above $[approval limit] need the owner's name and number recorded before the visit.
• Daily cap. When a day reaches its cap of [number] standard calls, offer the next day and the waiting list. Tier 2 calls may go [number] over the cap.
• Plan members. A member gets the priority slot. If the AI cannot check membership, it records the number the caller gives and flags the job for the office to verify.
• Outside the service area. Decline politely and say which areas you cover, using your own list of postal codes or towns.

Test every rule by asking whether a stranger could break it by accident. "Offer the earliest slot" can be broken, because it does not say what to do when the earliest slot is a Tier 3 window and the caller is Tier 2. "Tier 2 callers get the first priority slot, never a standard one" cannot. Where your job software holds the calendar, check what it lets the AI see. Jobber's page says its receptionist books visits using your existing online booking settings, so those settings are part of your booking rules whether you wrote them down or not.

Routing sales calls so a replacement lead is not booked as a repair

The expensive mistake on an HVAC phone is not a wrong appointment time. It is a caller who rang for a replacement quote and was booked a diagnostic visit with a technician who has no brief to sell. The script needs a short list of trigger phrases and a destination for each. The table below is the short version; the full treatment, including how an AI receptionist books advisers, is in AI receptionist for HVAC companies: route sales calls, not just repairs.

There is one case that needs a double tag. A no-cooling call on a seventeen-year-old system is a repair first, because the family needs cool air tonight. But it is also a likely replacement lead, so the script tags the job "replacement potential" and the adviser calls after the repair. Staff are good at noticing this by instinct. The AI only does it if the rule exists: "If the caller gives a system age over [your threshold] years, book the repair as normal and add the tag." Pick the threshold from your own records of when repairs on your customers' systems stop making sense.

One more call type deserves its own line in the script: the caller who has had three visits and a system that still fails. That caller goes to the service manager the same day, by person, never to the AI, and the call is logged in the same place you track repeat visits. The tutorial on logging the cause of every HVAC callback shows a log that makes those calls countable.

Using AI to produce the staff card, the AI version and the test calls

Once the master document exists, an AI assistant can do the tedious conversion in a few minutes. Give it your rules and ask for three outputs, and forbid invention explicitly, because inventing helpful-sounding detail is what these tools do when they are not told otherwise.

You are helping me turn one master call script for an HVAC company into three outputs.

MASTER SCRIPT:
[paste the full master script here, rules and example lines]

OUTPUT 1: A one-page staff card. Keep my stages and numbering. Short spoken lines, friendly, no jargon.
OUTPUT 2: Instructions for an AI phone receptionist. Numbered rules. Every rule must say what to do and what never to do. Keep my exact wording for the Tier 1 safety text and the diagnostic fee sentence.
OUTPUT 3: Twelve test calls for the receptionist and the staff, each with a caller persona, the opening line, and what a pass looks like.

Do not add any rule, price, time, promise or question that is not in my master script. If something needed is missing, list it under QUESTIONS FOR THE OWNER instead of inventing it.

This is an excerpt of an illustrative reply to that prompt, with the problems a careful read turns up.

Rule 7: Offer the customer a technician within 90 minutes where possible. Rule 9: If the caller mentions a smell, you may suggest it is probably a minor leak and advise them to open a window. Rule 12: Ask how many square feet the home is before booking any repair.

All three lines are inventions. The 90-minute promise appears nowhere in the master script, and the AI has no way to guarantee it. The smell advice is the dangerous one: it replaces the Tier 1 wording with reassurance, which is the opposite of the rule. Square footage belongs to the sales route and would slow every repair call. The fixes are to delete all three, paste the Tier 1 text back exactly, and rerun the prompt with a sentence added: "Quote the Tier 1 wording from the master script character for character." Do this check on every output, every time, because a different run can invent different things. Reading the replies for invention takes ten minutes and it is the step people skip.

Product settings add one more conversion step. Some receptionists take free text, some use fields for greeting, services, escalation keywords and booking rules, and some limit how much instruction you can enter. Ask the assistant to compress the AI version to a stated word count if the product demands it, and keep the full version in the master document so nothing is lost. The tutorial on call scripts and escalation rules for an AI receptionist covers the generic version of this step for any trade.

Twelve test calls and a scoring sheet

Do not trust the script until you have rung it. Make the test calls yourself, or ask a friend, a family member or a CSR from another shift to play the callers. Each persona gets one scenario and one awkward habit, such as talking over the greeting or answering a different question from the one asked.

Score each call out of ten: two points each for the safety screen where it applies, the right tier or route, the required details captured, nothing forbidden said (price, diagnosis, promise) and a clean recap with a next step. Add two knock-out rules. A Tier 1 call that is not handled word for word scores zero. A forbidden statement caps the call at five. Replay the failures with whoever is responsible for the fix listening.

In an imagined first round, the staff member scored 104 of 120 and the AI 102 of 120, close enough to look fine. The knock-outs told the real story. The AI quoted "usually a couple of hundred dollars" on call 7 when the caller pressed, which is the exact sentence the rule forbids. It also offered a standard window to the baby-in-the-house caller on call 4 because the vulnerable-occupant question had been written as a suggestion, not a requirement. The CSR failed call 9 by booking the tenant's visit without recording the owner's details, and call 12 by promising "the manager will definitely call you within the hour", which dispatch could not guarantee.

The fixes went into the master document, not into the AI's settings first. Call 7 got a firmer line in the "Never say" column and a callback offer. Call 4 got the word "must". Call 9 got a mandatory owner field on the booking form. Call 12 got a promise the company could keep: same-day callback, no hour stated. Then the two changed versions were regenerated and calls 4, 7, 9 and 12 were replayed. A second round of the full twelve took the AI to 112 and the staff member to 113. Plan on half a day to write the first version, half a day for the first round of calls, and fifteen minutes a week afterwards.

Keeping one source of truth once the phones are live

The master document drifts the moment someone fixes a problem in the AI's settings and forgets the document. Prevent it with four habits. One person owns the document, and only that person edits it. Every version has a number and a date at the top, and the AI settings carry the same number in a notes field so you can see at a glance whether they match. Every change starts in the document and reaches the AI settings the same day, with a screenshot of the settings kept in the same folder. And changes are triggered by evidence: a complaint, a failed test call, a new service you now offer, never by someone's feeling that the script "sounds stiff".

Read real calls as well as test calls. For the first month, listen to or read about ten calls a week, five handled by staff and five by the AI, scoring them with the same ten-point sheet. After that, a monthly sample of ten is enough. The method for reviewing AI transcripts, including what to look for in the wording the AI actually used, is in how to review AI call transcripts for quality and compliance. If a product also records calls, tell callers, and check what rules on recording and on disclosing that a call is answered by AI apply where your customers are.

Seasons deserve a decision, not a second script. Keep one document and put seasonal switches at the top: the daily cap, the heating or cooling question set, the Tier 2 trigger, the date each switch changes. When the switch date arrives, the owner changes it in one place and regenerates the two versions. Companies that keep a summer script and a winter script end up with two documents to maintain and one of them stale. If you are still choosing between a built-in receptionist and a standalone one, AI receptionist for trades: built into your job software or standalone compares the options on the points that matter for rules like these.

Questions owners ask before putting AI on the phones

Should the AI tell callers it is not a person?

Yes. Put one line in the greeting, such as 'you've reached the virtual assistant', and let the caller ask for a person at any point in the script. Housecall Pro's product page leaves disclosure to you, and Jobber's help article encourages it without requiring it, so check your own greeting. If you have customers in the EU, the AI Act's chatbot transparency duty has applied since 2 August 2026, mainly as a duty on the vendor. Telling callers is the safer habit whatever the rules say.

How long should the master script be?

Short enough that the staff version fits one printed page, which usually means 400 to 600 words of rules and example lines. The AI version can run longer because it has to spell out what a person would infer. If either version keeps growing, you are probably writing wording where a rule would do.

Can the AI quote prices on an HVAC call?

Only the figures you have approved in writing, which for most companies means the diagnostic fee and any plan-member discount. Repair prices depend on what the technician finds, so both staff and AI should say that and offer a visit. Put the exact sentence in the script so nobody improvises a range.

How often should the script change?

Change it when a call went wrong in a way the script allowed, not on a calendar. Keep a version number and a short change log, update the AI settings the same day as the document, and replay two or three test calls after every change. Review about ten real calls a week for the first month, then monthly.

Further reads

• AI vs Human Answering Service for HVAC Firms — Decide which calls the AI takes and which stay with people.
• How Much Do Missed Calls Cost a Heating and Cooling Business? — Put a figure on the calls you lose today.
• HVAC Busy Season Checklist: Cope With the Summer AC Surge Using AI — Set capacity caps and call rules before the summer rush.
• HVAC Lead Generation: AI Call Tags Show Which Sources Book Replacements — Tag calls by source to see which book replacements.
• AI Receptionist for Tradespeople: What It Costs and When It Pays — Check what an AI receptionist costs against missed calls.
• HVAC Website AI Chatbot: Book Tune-Ups, Triage Without Overpromising — Apply the same rules to website chat.
• How to Set Up an AI Receptionist Without Losing Callers — Move the finished script into a receptionist without losing callers.
• AI Tools for HVAC Business Owners: Calls, Quotes and Maintenance Plans — Jobber's receptionist, Housecall Pro's CSR AI and ServiceTitan compared for calls, quotes and maintenance plans, with emergency rules and a four-van tally.
• Follow Up After Giving a Quote: An HVAC Replacement Sequence With AI — A day-by-day text and email sequence for HVAC replacement quotes, with consent and opt-out handling, stop rules and a fair way to track the close rate.
• Using AI in an HVAC Business: A Week-by-Week Setup Plan — A four-week HVAC rollout: calls, estimates, follow-up, then review, each week with one metric and a Friday owner checkpoint.
• HVAC Price Increase Letter: Maintenance Plan Members and One-Off Customers — Draft a plan renewal notice and a standard rate-change notice with AI, plus the sums behind the new price and what to say when customers push back.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Jobber's AI Receptionist product page and Receptionist help article, Housecall Pro's CSR AI page, and government carbon monoxide safety pages read for the emergency section (all read November 2026). Wording in the emergency section is a draft for your own safety lead to approve.