HVAC Website AI Chatbot: Book Tune-Ups, Triage Without Overpromising

An HVAC website chatbot should ask about the equipment, the symptom and whether anyone is at risk before it offers a time, answer only from your own tune-up scope and price page, and pass anything involving gas, carbon monoxide, burning smells or vulnerable occupants straight to a person. It requests slots; it never promises arrival times or repair prices.

Most overpromising comes from three gaps rather than from the AI being reckless: the bot was never told what a tune-up includes, it cannot see the diary, and nobody wrote down what it must refuse. Picture a four-van heating and cooling firm whose bot tells a customer on a freezing Sunday night that "a technician will be with you shortly". The on-call technician is already on another job, the customer waits, and the one-star review blames the company, not the software. The fix is a short rulebook plus ten test chats.

Follow me on Instagram@sagnikteaches

The answer list and the refusal list

Write two lists before you open any chatbot tool. The answer list holds facts you control and can keep true. The refusal list holds everything that needs a technician's eyes, a manager's decision or a promise you cannot guarantee. Together they become the knowledge file the bot reads, and the FAQ-building tutorial covers how to turn them into clean question-and-answer pairs.

Connect on LinkedInSagnik Bhattacharya

What it answers, from your own pages

• The towns you serve, as a list it checks the visitor's address against.
• Opening hours, and how the after-hours line works.
• What a tune-up includes, what it costs and what it does not cover. Say plainly whether refrigerant top-ups are included, because an "AC tune-up" customer can easily assume they are.
• Membership or plan benefits, if you sell them.
• The brands and equipment types you service, and the ones you don't.
• How booking works: request, confirmation text, technician window.
• Payment methods, and financing only if you really offer it.

What it refuses, every time

• Naming a cause ("it's probably the capacitor").
• Quoting a repair or replacement price before a technician or estimator has seen the equipment.
• Promising an arrival time, "same day" or "within the hour", beyond what a dispatcher has confirmed.
• Deciding warranty cover. That depends on the manufacturer's terms, the installation date and registration.
• Instructions that involve opening panels, touching wiring or refrigerant lines, relighting pilot lights or bypassing safety switches.
• Saying a system "should be replaced", or making any claim about energy bills.
• Comparing your firm with other companies.

Sorting visitors into four tiers in the first minute

Triage here means deciding how fast a person should look, never what is wrong. Four tiers are enough, and the bot's behaviour changes sharply between them.

Subscribe on YouTube@sagnikteaches

The test for tier 1 is one question: could the visitor be hurt in the next hour? If the answer might be yes, the bot's only job is the safety message. Here is wording to adapt:

If you can smell gas, hear hissing, or a carbon monoxide alarm is sounding, or anyone feels dizzy, sick or has a headache: leave the building now with everyone, leave the door open, and don't switch anything on or off. Once you are outside, call the emergency number or your gas supplier's emergency line. Please don't wait for us. When it is safe, call our office and we'll arrange an inspection.

That follows widely published utility and safety guidance: for gas, leave on foot and call from a safe distance without touching switches or flames; for carbon monoxide, get everyone into fresh air and then call the emergency number. Have your senior technician approve the final text, and show the same message on every page where the chat widget loads, not only inside the chat.

A tune-up booking flow in eight questions

Each question has to earn its place: it changes what the technician brings, decides whether you can serve the address, or lets the office confirm a time. Eight is a sensible ceiling, because every extra question hands the visitor another reason to leave.

1. Which service? Tune-up, repair, new-system estimate or something else. This routes the visitor into the right flow.
2. What equipment, and which fuel? Furnace, boiler, heat pump, air conditioner, mini-split or not sure; gas, oil, electric or not sure. This sets the technician's tools and whether you service it at all.
3. How many systems, and roughly how old? It sets visit length and which price line applies.
4. Which town? Checked against your service list. If it is outside, the bot says so politely before collecting any personal details.
5. Which days and times suit? Offer day-part windows such as "weekday mornings". Exact times only when the bot reads live availability.
6. Access notes. Gate or key safe, pets, attic or basement access. Optional.
7. Name, mobile number, email. Include consent to a text confirmation.
8. Anything else the technician should know? Free text. This is where the bot catches "the thermostat has also gone blank".

Here is how the flow reads end to end. It is an illustration, and every bracket is a placeholder.

Bot: Hi, I'm the AI assistant for [Company]. If you smell gas, a carbon monoxide
     alarm is sounding, or you see smoke, leave the building and call the emergency
     number now. Otherwise, how can I help?
Customer: Can I book a tune-up for my furnace?
Bot: Yes. Is it gas, oil or electric, and about how old is it?
Customer: Gas, maybe 9 years.
Bot: Thanks. Which town is the property in?
Customer: [town]
Bot: We cover that. Which days and times suit you best?
Customer: Weekday mornings.
Bot: Any access notes, like a gate code, pets or an attic hatch?
Customer: Dog, friendly.
Bot: Last bit: your name, mobile number and email, and OK to text you the confirmation?
Customer: [first name], [mobile], [email], yes.
Bot: Thanks, [first name]. I've sent your tune-up request to our office: gas furnace,
     one system, weekday mornings, dog on site. Someone will confirm a time by text
     during office hours (Monday to Friday, 7:30 to 5:30). It is not a confirmed
     booking until you get that text. Anything else?

The last message does the heavy lifting against overpromising. It says "request", names who confirms, says when, and states in words that the booking is not yet confirmed.

Request, offer or confirm: three levels of promise

Start at request and move up one level only after a month of transcript reviews with no wrong promises. Some field-service tools build the upper levels in. Jobber's AI receptionist, for instance, can offer to book a visit using your online booking settings, or create a work request for you to review, but it answers calls and texts rather than website chat.

No-heat and no-cool: three questions and a hard stop

These are the chats where overpromising does the most damage, because the visitor is cold or hot and wants a time. Ask in this order and stop as soon as a tier is clear:

1. Is there any smell of gas, a carbon monoxide alarm, a burning smell, smoke or sparks? Yes means tier 1 and the safety message.
2. Is anyone at home vulnerable, for example an infant, an older person or someone unwell or on medical equipment?
3. Is the weather extreme enough to be dangerous: freezing for no heat, a heatwave for no cooling?

A yes to the second or third question makes it tier 2: the bot gives the live office or on-call number, or alerts the dispatcher, and says only what is true, for example "I have flagged this as urgent and passed it to our dispatcher, who will call you". Two no answers make it tier 3, a normal fault request.

Safe checks the bot may suggest

Every self-check you allow is a safety decision, so get your senior technician to approve each line in writing, and delete any they don't like. Lines a firm might allow once its senior technician agrees:

• The thermostat is on the right mode, set to call for heat or cooling, and has working batteries if it uses them.
• The service switch near the equipment is on.
• The breaker for the system hasn't tripped. If it has, reset it once; if it trips again, stop and book a fault request.
• The filter is not visibly blocked, only if it can be reached without tools.

The hard stop is the other list: never advise opening the equipment, touching wiring, relighting a pilot light, adjusting a gas valve, adding refrigerant, bypassing a float or safety switch, or running a system that is leaking water onto electrics.

What a risky reply and a safe reply look like

The first reply below is the kind of thing an unconstrained bot produces. The second is what the rulebook should force.

Risky: That sounds like a failed capacitor. You can swap one for about $20, and we can usually be there today if you'd rather not.

Safe: I can't tell what's wrong from here, and I don't want to guess. Two quick checks first: is the thermostat set to cool and below the room temperature, and is the system's switch on? If neither helps, I'll send your details to the office as a fault request and the team will confirm a time by text.

Price wording that is still true when the invoice arrives

A chatbot invents a price when the page it reads has none. Give it a short facts sheet, and for anything that isn't a fixed price, give it the rule for how the price is set instead of a number. The sheet below uses placeholder figures; replace every one. If you are still working out the call-out fee itself, the service-call fee tutorial builds one from your costs.

PRICE FACTS (example wording, replace every figure with your own)
Tune-up, heating or cooling, one system: $149
Second system at the same visit: $109
Diagnostic visit for a fault: $89, credited if you approve the repair on the day
After-hours or weekend visits: extra charge applies; the dispatcher confirms before dispatch
Repairs: a technician diagnoses, quotes before starting, you approve before any work
Replacement: free in-home estimate, no obligation
Not part of a tune-up: refrigerant top-ups, repairs, parts
Payment: card or bank transfer after the visit

The second row catches a common slip. "Free estimates" is true for replacements and false for a repair diagnosis, so the bot must be told the difference or it will merge them.

Hand-off rules: when the bot stops, and what the person receives

A hand-off should be a rule the bot cannot talk its way out of. Hand over to a person when:

• the visitor asks for one, in any wording;
• any tier 1 word appears;
• the visitor rejects two answers in a row ("that's not what I asked");
• the topic is warranty, a complaint, a refund or an invoice;
• the visitor manages several properties, such as a landlord or property manager;
• it is tier 2 and the office is closed, in which case the bot gives the on-call number.

Out of hours, the bot says what really happens: "Our office opens at 7:30. For no heat in dangerous weather, call [on-call number]." It never says "someone will call you back shortly" without a time, and the human hand-over tutorial goes deeper on timing and wording. Whatever the trigger, the person should receive a summary so the customer never repeats themselves:

HAND-OFF SUMMARY
Tier: 2 (urgent)
Name / mobile: [first name] / [mobile]
Town: [town]
Equipment: gas furnace, about 9 years old, one system
Symptom: no heat since last night; thermostat display on; no error code
Safety flags: none reported; elderly parent at home
Self-checks done: thermostat mode and set point, service switch
Told the customer: urgent request passed to dispatcher; office number given
Contact preference: phone call

Where the rules go in the tool you choose

The rulebook is the same everywhere; what changes is where you paste it and how hand-off works. There are three families of tool, and the figures below are as of November 2026, so check the pricing pages before you decide.

• Website chat builders. Tidio's Lyro AI agent costs $39 a month for 50 conversations on monthly billing ($32.50 a month billed yearly), and the first 50 conversations are free (a one-off allowance, not monthly). A conversation counts once the AI has replied at least once, however many messages follow. Chatbase has no free plan: Hobby is $40 a month ($32 billed yearly) for 700 message credits and Basic is $150 ($120 billed yearly) for 4,000. Chatbase passes hand-offs to a person through its Helpdesk inbox, which the pricing page lists from Basic upwards, not on Hobby. Every rule in the previous section depends on hand-off, so check the plan includes it before you buy.
• Field-service software with its own AI receptionist. Jobber's is a $29 add-on covering 30 conversations, then $0.79 each, according to its help article of late October; its pricing page shows only the $29. It covers phone and text, not your website, but its urgent-keyword setting, where words such as "no heat" put a caller through to you or send you a text alert, is the phone version of the tier 2 rule above. Housecall Pro's CSR AI has no published price, and in October its help article and its product page disagreed on whether it books jobs or only takes details, so test it with chats 1 and 7 below before trusting it.
• A general assistant as your test bench. Create a project in ChatGPT or Claude, paste the rulebook as instructions and the price facts as a file, and run the ten chats there before copying the instructions into the live tool. Mistakes cost nothing in a test project.

Check what unit each tool bills in. Tidio and Jobber count a whole conversation as one unit, however long it runs, while Chatbase sells message credits, so ask how it counts a long chat. Then price the tool at your real volume, not the entry tier. At about 120 chats a week, the volume in the review example below, a firm needs roughly 520 AI conversations a month; on Tidio's monthly billing that means the 750-conversation tier at $525 a month, because the 500 tier ($350) would run out before the month ends. Whatever the billing unit, a cap such as "ask at most eight questions" spares the visitor's patience. Here is a skeleton for the instructions field. Treat it as a starting point and keep your own wording for the safety message.

You are the website assistant for [Company], a heating and cooling firm.
Say you are an AI assistant in your first message and offer a person at any time.

ANSWER ONLY from the attached PRICE FACTS and SERVICE NOTES. If the answer is not
there, say you don't know and offer to pass the question to the office.

SAFETY FIRST. If the visitor mentions gas, a carbon monoxide alarm, a burning smell,
smoke, sparks or water near electrics, show the SAFETY MESSAGE exactly and ask no
booking questions.

TUNE-UP FLOW. Ask at most eight questions, one at a time, in this order: service,
equipment and fuel, number and age of systems, town, preferred days, access notes,
contact details, anything else. Finish by saying this is a request, not a confirmed
booking, and that the office confirms by text during office hours.

NEVER: name a cause, quote a repair or replacement price, promise an arrival time,
decide warranty cover, advise opening equipment, or compare us with other firms.

HAND OFF to a person, with the HAND-OFF SUMMARY, when asked, after two rejected
answers, for warranty, complaints, refunds or invoices, and for landlords with
several properties.

Ten sample chats to run before it goes live

Type each message below into the bot in a test project, then judge it against the pass and fail columns. The wording of an AI's reply varies from run to run, so run every chat three times.

My suggested pass mark: chats 2, 3 and 4 must pass all three runs, with no exceptions, because they protect people. The rest should pass at least two of three runs. Every failure becomes a new line in the instructions, and then you rerun the whole set, since fixing one rule can quietly break another. After launch, add three more chats taken from real transcripts, which is where the surprises show up.

Reading transcripts after launch

For the first month, read twenty chats a week: ten random ones and ten that ended in a hand-off or a safety stop. Search the week's chats for "today", "shortly", "guarantee", "free" and "warranty", because each marks a promise the bot should not have made. Then count four things per hundred chats: requests, hand-offs, safety stops and wrong promises. The target for wrong promises is zero, and the chatbot performance tutorial shows how to track the other three over time.

Here is what a first review can turn up, with invented numbers for a small firm. In week one, 120 chats produced 31 tune-up requests, 9 hand-offs and 2 safety stops. The review turned up six chats where the bot said "today" because the website FAQ said "same-day service available" without mentioning that this applies only in office hours. The fix was one rule, "never use the words today, tomorrow or shortly; refer to the dispatcher", and one corrected FAQ line. Three visitors had also asked what a replacement costs, got a vague deflection and left, so the firm added a replacement-estimate flow that books an adviser visit instead of dodging the question.

Keep the opening line honest as well. If you have customers in the EU, the AI Act expects people to be told they are talking to an AI, and the chatbot disclosure tutorial covers wording. The first message in the transcript above does it for every visitor, wherever they are.

Questions HVAC owners ask before switching a chatbot on

Can an HVAC chatbot book straight into my diary?

Only if it has a two-way link to your scheduling software, and even then I would start with request-only mode. In request mode the bot collects details and the office confirms by text. Once a month of transcripts shows no wrong promises, let it offer real windows from the diary, and only later let it write bookings.

Should the chatbot tell visitors it is an AI?

Yes. Say it in the first message, in plain words, and give a way to reach a person. If you have customers in the EU, the AI Act expects people to be told they are talking to an AI, and one honest opening line covers every visitor wherever they are. It also spares you the awkward moment when a customer works it out alone.

Should the bot ever quote a repair price?

Not unless the price is fixed and published, such as a diagnostic fee. Repair prices depend on the fault and the parts, so the bot should explain how pricing works: a technician diagnoses, quotes before starting, and the customer approves first. A guessed range becomes a promise in the customer's mind.

How many transcripts should I read each week?

Twenty is enough for a small firm: ten at random and ten that ended in a hand-off or a safety stop. Search the whole week for words like today, shortly, guarantee and free, because those mark promises the bot should not have made. Add a rule for every miss, then rerun the ten test chats.

Further reads

• Chatbot Guardrails: Stop AI Promising What You Don't Offer — More guardrail wording for stopping a bot promising things you don't offer.
• How to Test a Customer Chatbot Before It Goes Live — A fuller pre-launch test plan once your ten chats pass.
• Setting Up an AI Answering Service That Never Misses an Emergency — Covers the phone side of out-of-hours emergencies.
• HVAC Call Script: Train Your CSRs and AI on the Same Script — Train office staff and the bot on one shared script.
• AI Receptionist for HVAC Companies: Route Sales Calls, Not Just Repairs — Route replacement-sales calls differently from repair calls.
• AI Service Plan Reminders for Heating and Cooling Customers — Turn tune-up bookings into recurring plan reminders.
• AI Marketing for HVAC: A Worked Spring and Autumn Tune-Up Campaign — Fill the tune-up diary before each season starts.
• HVAC Lead Generation: AI Call Tags Show Which Sources Book Replacements — Tag calls with AI, follow each booked job to its outcome, and compare replacement sales by source, with a worked quarter across four HVAC lead sources.
• AI Chatbot for Plumbers: Answer Questions, Collect Photos, Refuse Jobs — Set the scope of a plumber's website chatbot: three lanes, an approved answer sheet, price rules, photo requests, refusals, handover and a 20-message test.
• HVAC Busy Season Checklist: Cope With the Summer AC Surge Using AI — An owner's checklist for the summer AC surge: capacity rules, four triage tiers, overflow answering, text-back, deferred tune-ups and parts prep.
• Heating Engineer Website: What the AI Chatbot and Form Should Ask — The questions that stop wasted boiler visits: make and age, fault wording, a data plate photo, urgency, and when the chatbot must hand over to a person.
• Website Builder With AI Chatbot: Which Suits a Trades Business? — Compare Wix Smart Chat, Squarespace add-ons, Tidio and Chatbase for a trades site: lead capture, handover to your phone and a 30-minute test.
• AI Chatbots for Websites: What a Chimney Sweep Should Let Them Handle — Let the chatbot book sweeps and read your price list. Keep safety diagnosis away from it, and send carbon monoxide worries to a fixed message and a person.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Tidio and Chatbase pricing pages, Chatbase's helpdesk page, Jobber's Receptionist page and help article, and public safety guidance from gas utilities and safety bodies on gas leaks and carbon monoxide (checked November 2026).