Using AI in an HVAC Business: A Week-by-Week Setup Plan

Spend week one on calls (an after-hours script and emergency rules), week two on estimates drafted from technician notes, week three on follow-up for open estimates and maintenance renewals, and week four on review. Give each week one metric and one owner checkpoint, and begin about six weeks before your busy season starts.

The order matters because each week feeds the next: the call log shows which jobs come in, those jobs become estimates, and the open estimates become the follow-up list. Change one process a week and keep a person reading every AI draft. Trying all four at once makes it impossible to tell which change helped, or which one upset a customer.

Follow me on Instagram@sagnikteaches

Before week one: an owner, a baseline and a scoreboard

Name one owner for the month. In a small company that is usually the office manager or the dispatcher, not a technician and not you. The owner needs a couple of free hours on most days in week one. Your own job is a twenty-minute checkpoint every Friday: read the metric, read a handful of real messages, and decide what changes on Monday.

Connect on LinkedInSagnik Bhattacharya

Pin the plan to a single page so nobody has to ask what this week is about.

Subscribe on YouTube@sagnikteaches

Pull four weeks of history before you change anything. The phone log gives inbound and unanswered calls, your field-service software gives visit and estimate dates, and an export of open estimates gives the age of each. If a number is not available, count by hand for two weeks. A rough baseline is far better than none, because week four is a comparison.

Take a six-technician company heading into its first cold month. Its four-week history reads: 240 inbound calls, 31 unanswered (13%), and only 9 of those 31 returned within the hour; a median of 2.6 days from visit to estimate; 18 open replacement estimates older than a week; and 11 approvals from the last 52 estimates sent, about 2 in 10. These figures are illustrative. Yours will differ, but the shape of the table will not.

A quick sum shows why calls come first. If a quarter of those 31 missed calls would have become a $380 repair visit, that is 31 × 0.25 × $380 = $2,945 of work a month going to whoever answered. Swap in your own average ticket and booking rate; the arithmetic stays the same.

Week one: stop losing calls, starting after hours

Start with the hours when nobody is in the office. Daytime calls reach a person. The 10 pm no-heat call reaches voicemail, and that caller is already ringing your competitor. The week has five jobs.

1. Sort a month of calls into five buckets: emergency, repair booking, replacement enquiry, maintenance agreement and everything else. Paste call-log notes or voicemail transcripts into an AI assistant with names and addresses removed, ask for counts and three typical examples per bucket, then spot-check ten calls yourself. You are looking for the bucket that arrives out of hours more than the others. It is often emergencies, but check.
2. Write the emergency rules in plain words (the table below).
3. Build the one-page call sheet that the assistant, and any new office hire, will work from.
4. Pick the call-handling route that fits your size and budget.
5. Run ten scripted test calls before a single real customer reaches it.

The emergency rules, written by you

This is the part nobody should leave to the software. Each row ends in a person or in words you wrote, never in the assistant improvising. The table is a filled-in illustration; yours will differ.

The wording for the first row should come from your gas supplier or fire service, since they publish it. The assistant reads it out; it does not paraphrase. For the full set-up of an assistant that handles these rows, see setting up an AI answering service that never misses an emergency, and for training staff and the assistant on one script, the HVAC call script tutorial.

The one-page call sheet

Ask an assistant to turn your facts into the sheet, and tell it not to guess. The instruction to write [ask owner] when something is missing is what keeps invented prices out.

You are helping me write a one-page call sheet for the assistant that answers our phone after hours. We are an HVAC company with [number] technicians.

Use only the facts below. If a fact is missing, write "[ask owner]" instead of guessing.

Facts:
- Hours for booked work: [hours]
- Call-out fee: [amount and what it covers]
- We service: [equipment types and brands]
- We do not: [for example commercial refrigeration, plumbing]
- On-call number: [number], used only for the emergency rules below
- Emergency rules: [paste the table]

Write:
1. A greeting of about 60 words that says this is an automated assistant.
2. The five questions to ask every caller, in order.
3. What to say for each emergency rule.
4. Five things the assistant must never say, including arrival times, prices not listed above and any diagnosis.
5. How to end every call.

Choosing the call-handling route

There are three routes, in rising order of effort. Pick the lowest one that covers your emergency rows.

• Missed-call text-back with voicemail transcription. The cheapest start. Many business phone systems offer both, so check yours. The text promises a call back and nothing more.
• An AI receptionist inside your field-service software. Jobber's Receptionist answers calls and texts around the clock, lets you set urgent keywords such as "no heat" that trigger a call transfer or a text alert, and keeps a summary, transcript and recording of each conversation. Its help article listed it at $29 a month for 30 conversations, then $0.79 each, on selected plans, so check the current page and your plan.
• A standalone AI answering service. Smith.ai's AI receptionist lists a free plan of 25 calls a month, then $3 a call, with paid plans from $150 a month. It answers 24/7, gives call summaries and routing, and can book through a scheduling tool.

Whichever you pick, ask three questions before paying: can it transfer an emergency to my phone, can I choose the words that trigger that, and where do the caller's details land (my software, an email, a text)? A fourth question is about notice. Check the rules that apply to you on telling callers they are speaking to an automated assistant and on recording calls.

Ten test calls

Ring it yourself, or have a colleague do it, with these ten and write down what happened.

1. "My furnace is making a banging noise." It should book and not diagnose.
2. "I can smell gas." It should read the safety message and nothing else.
3. "How much for a new system?" It should offer an in-home estimate visit, not a price.
4. "Do you work on [brand]?" It should answer only from the sheet.
5. "I am a landlord and my tenant has no heat." It should follow the emergency row.
6. "I think I have a plumbing leak." It should decline politely.
7. "Your technician left a mess yesterday." It should take the details and pass them to the owner without apologising on the company's behalf.
8. "Is the technician on his way?" It should not promise an arrival time.
9. "Are you a real person?" It should say no.
10. "Can I speak to the owner?" It should take a message unless a rule says transfer.

Go live for after-hours calls only, and leave daytime calls with your staff until week two has gone smoothly.

What a failed test looks like

Below is a call summary of the kind an owner might receive. It is illustrative, with one line that should never have appeared.

Caller: [first name], [number]
Issue: Furnace not heating, thermostat blank, clicking noise from the unit
Urgency: Caller says an elderly parent is at home and it is below freezing outside
Action: Transferred to on-call number at 21:42, no answer, text alert sent
Promised: "A technician will be with you within the hour."

The last line breaks the rule against arrival times. The fix is two-part: add "never state or imply an arrival time; say the on-call technician will call back within 15 minutes" to the call sheet, then rerun test call 8 until it passes three times in a row. One slip of this kind is a prompt fix. Two in the same week means pausing the after-hours assistant until the sheet is repaired.

The week-one metric is the share of inbound calls answered or returned within 15 minutes. At the Friday checkpoint, read the transcripts of every transferred call plus five others, count the promises you could not keep, and write down one change for Monday.

Week two: estimates drafted from technician notes

Estimates stall at the desk, not in the field. The technician has the information on Tuesday, and the proposal is sent on Friday because writing it is a chore. Week two shortens that gap without letting the assistant make any technical decision.

Fix the note format first

Give technicians five lines to fill in, by typing or by dictating into their phone:

1. Equipment: brand, model, serial number, approximate age.
2. Findings: what failed, what is worn, what is unsafe.
3. Measurements: duct and return sizes, line-set length, anything the load calculation needs.
4. Site constraints: attic or crawl-space access, electrical capacity, permit needs.
5. Photos taken: data plates, outdoor unit, indoor unit, thermostat, electrical panel.

Dictated notes are messy, and that is fine. The assistant's first job is to tidy them into the five lines, not to improvise missing ones.

The drafting prompt, and what comes back

Draft a customer-facing estimate summary from the technician's notes and the price-book lines below.

Rules:
- Do not choose equipment size or model. Copy them from the notes.
- Use only the prices in the price book.
- If the estimator needs a line that is not in the notes, list it under "Questions for the estimator" instead of inventing it.
- Plain English, no jargon. Three options labelled Good, Better and Best.

Technician notes: [paste]
Price-book lines: [paste]

Suppose the notes describe a 14-year-old split system with a locked-out compressor and a leaking indoor coil, and the customer wants a replacement quote. The draft comes back with three tiers at, in this illustration, $7,900, $9,600 and $12,400. It reads well, which is exactly why it needs an estimator's red pen. In an illustrative review, these are the corrections:

• The draft kept the old unit's 3-ton size. The estimator replaces it with the size from the load calculation, because matching an old nameplate repeats whatever mistake the original installer made.
• There was no line for permit and inspection fees, disposal of the old equipment and refrigerant recovery, a new pad, or the electrical disconnect. The estimator adds them from the price book.
• The draft wrote a "10-year warranty" line that came from the model's general description. The estimator replaces it with the manufacturer's actual terms for that model and the labour terms the company really offers.

Those three corrections become your estimator checklist, taped beside the screen. Every proposal passes the same six checks before it is sent: size from the load calculation, a price-book price on every line, a permit and inspection line, a disposal and recovery line, warranty wording copied from the manufacturer, and a plain list of exclusions. For a fuller treatment of tiered quotes, see the worked good-better-best replacement quote.

The sum that justifies week two

Say writing an estimate by hand takes 45 minutes and drafting plus checking takes 20. At 12 estimates a week, that is 12 × 25 = 300 minutes, or five hours, returned to the office. Treat those minutes as a hypothesis to test, not a result: time ten real estimates in week two and replace my numbers with yours.

The week-two metric is hours from site visit to estimate sent, measured as a median so one odd job does not distort it. At the checkpoint, ask which lines the estimator corrected most often, and add that point to the prompt as a rule.

Week three: follow-up that stops when the customer answers

An estimate that nobody follows up is a lead you paid for and dropped. Week three builds a small, polite routine for open estimates and for maintenance agreements that are about to lapse.

Rules before messages

Write the rules down first, because the assistant will follow whatever you give it:

• Three touches at most: a text two days after the estimate, a phone call on day five, an email on day ten.
• Everything stops the moment the customer replies, approves or declines.
• No invented urgency. "This price is held until Friday" is allowed only if it is true.
• No discounts unless the owner approved that specific one.
• Keep a do-not-contact list, honour "stop" at once, and only text people who have agreed to be texted.

Export the open estimates with their age and value, oldest and largest first. In the illustrative company that is 18 estimates, with the biggest worth about $11,000. Paste the list without names and ask for the three-touch sequence for each estimate type (repair, replacement, indoor air quality). The detailed sequence is in the HVAC quote follow-up tutorial; here the aim is to get the routine running.

Write touch 1 (a text of under 300 characters) for an open HVAC replacement estimate.
Use [first name] and [your name] as placeholders. Mention that the estimate was sent on [date], offer to answer questions, and give a way to reply STOP.
Do not mention a discount, a deadline or a competitor.

A good first draft looks like this: "Hi [first name], it is [your name] from [company]. I sent your replacement estimate on [date] and wanted to check it arrived and that the options made sense. Happy to talk it through or adjust anything. Reply STOP and I will not text again." What you would fix in a weaker draft is anything that adds pressure. "Only two slots left this month" and "prices rising next week" are claims you cannot prove, and customers can tell.

Maintenance agreement renewals

List agreements ending in the next 60 days, with the date of the last visit. Renewal messages should name that visit, because "we last looked at your system on [date]" reads as service, while a generic reminder reads as a bill. If your records do not hold the last-visit date, leave the line out; do not let an assistant guess it. More on the renewal side is in the service plan reminders tutorial.

When a job is finished, ask for a review. Google's review rules ban incentives and asking only the customers you expect to be happy, so the safe habit is the same message to every customer. Never write the review text for them.

The week-three metric is the share of open estimates with a follow-up logged in the last three days. In the illustration, 17 of 18 are touched by Friday. Three customers reply, one approves, and two ask to be left alone until next month. Putting those two in the calendar for next month is the checkpoint's real job: confirm that the stop requests were honoured.

Week four: score the month and decide what stays

Week four adds no new automation. It puts the baseline next to the results and makes three decisions: keep, fix or drop. Here is a filled-in review for the illustrative company.

The edit rate is a column you add to your log in week two: for each draft, tick sent unchanged, edited lightly or rewritten. It tells you whether the prompt is improving. A rewrite rate above half means the prompt or the reference material is wrong. An edit rate near zero for a month means either excellent drafts or nobody reading them, so check which by sampling five.

Estimates approved per 10 sent is the headline measure, but four weeks is a small sample. If the cohort moves from 2 in 10 to 3 in 10 on 24 estimates, that is a signal worth keeping the routine for, not proof. Repeat the count at eight and twelve weeks.

Here are our baseline figures and four weeks of results (table below).
For each change (calls, estimates, follow-up) tell me: did the metric move, what else could explain the movement, what extra data would settle it, and should I keep, fix or drop the change?
Do not recommend new tools.

[paste the table]

Finish week four by writing the fallback: who answers calls if the assistant goes down, who sends estimates if the drafting prompt fails, and who can switch each automation off. Put those names on the page with the plan. For what to add next, start tagging every inbound call with its source, so you can see which sources book replacements rather than repairs.

Fitting the four weeks around your busy season

Run the plan so that week four ends two to three weeks before the first crush. That leaves time to fix what you find and rehearse the fallback.

A company with a cooling peak and a heating peak should run the plan before the first one, then use the quieter weeks between them for the second review. The busy-season checklist in the summer surge tutorial covers what else to prepare, and if your peak is already here, run week one alone.

Five slips that show up in a first month

1. An arrival time promised. It appears in the transcript as "within the hour" or "this afternoon". Fix it in the call sheet and keep test call 8 in your monthly rerun.
2. A price quoted from memory. The assistant says the call-out fee is $89 because an old page said so. Keep one current price list, paste only from it, and date it.
3. The same lead twice. The receptionist creates a request in your software and also texts the office, and two people ring back. Choose one place where new leads land.
4. Nobody reads the drafts any more. The edit rate falls to zero by week three. Sample five drafts each Friday, including ones that went out unchanged.
5. Customer details in a personal AI account. A technician pastes an address into a free consumer app. Business plans such as ChatGPT Business and Claude Team do not train on business content by default, but check the setting rather than assuming, and use the business account for anything with a name in it.

None of these needs better technology. Each needs the owner to look at real output once a week, which is the whole point of the Friday checkpoint. If your first month goes well, pick one more automation for month two and run it through the same four-week pattern.

Questions HVAC owners ask before starting week one

Do I need new software to start week one?

Not necessarily. A phone line with voicemail transcription, a shared spreadsheet and one paid AI assistant account cover weeks one to three. Field-service software with a built-in receptionist can replace the phone step, but check which plan includes it. Add software only when a metric shows the manual version is the bottleneck.

How many office hours does the plan take?

Budget about two hours on most days in week one, roughly an hour a day afterwards, plus a twenty-minute owner checkpoint on Fridays. Those are planning figures, not measurements, so log your own time in week one and adjust. If the owner cannot spare the checkpoint, slip the plan by a week instead of skipping it.

Should technicians use AI on the job?

Give technicians one job at first: dictate notes in the fixed five-line format so the estimator can draft from them. Diagnosis, equipment sizing and safety calls stay with the technician and the manufacturer's documents. An assistant cannot see the equipment, so it should never decide what is wrong.

What if my busy season has already started?

Run only week one, after hours, and leave the rest until the rush ends. Calls are the change that pays back fastest and needs the least staff time. Estimates and follow-up need someone checking drafts every day, which is exactly the capacity a busy office does not have.

Further reads

• How Much Do Missed Calls Cost a Heating and Cooling Business? — Put your own numbers into the missed-call sum before week one.
• AI Estimating Software for HVAC: Replacement Quotes and Duct Takeoffs — Compare estimating software once the manual week-two routine works.
• HVAC Lead Generation: AI Call Tags Show Which Sources Book Replacements — Tag calls by source to see which ones book replacements.
• HVAC Callbacks: Log the Cause on Every Job and Spot the Pattern With AI — Log the cause of every callback to feed your month-end review.
• AI Tools for HVAC Business Owners: Calls, Quotes and Maintenance Plans — A wider tool list if you outgrow this four-week plan.
• Six AI Automations for HVAC Companies, Ranked by Effort — Rank what to add in month two by effort and payback.
• Van Stock Management for HVAC and Electrical Vans: Reorder Levels With AI — A good month-two project: set van stock levels from job history.
• AI Marketing for HVAC: A Worked Spring and Autumn Tune-Up Campaign — Run a tune-up campaign once calls and follow-up are steady.
• ChatGPT for HVAC: 10 Prompts for Plans, Replacements, Reviews — Ten prompts with filled examples for maintenance plans, replacement options and review replies, plus the customer-data rules and send-check.
• HVAC Technician Job Description: Draft It With AI, Test Fairly — A master job description, an AI-drafted advert, an honest on-call rota, and a one-hour practical test with a scorecard that treats applicants alike.
• How to Price HVAC Jobs: Test Flat Rate Against Hourly With AI — A 120-job replay compares flat rate with hourly by job type, sets flat prices from your own costs and a minimum call charge, and warns against copied prices.
• How to Use AI in HVAC Design: Options, Duct Checks and Proposals — Contractor-side AI steps around HVAC design: survey notes to input sheet, duct questions, good-better-best drafts and a proposal audit against the load report.
• Refrigerant Log: Record Leak Checks and Top-Ups on Every HVAC Job With AI — Per-job fields for leak checks and top-ups, a voice-note routine and an AI prompt that structures the entry, with a worked cold-room repair.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Jobber Receptionist product page and help article; Smith.ai AI Receptionist pricing page; OpenAI and Anthropic business-plan data-use terms. Checked October and November 2026.