How AI Is Being Used in HVAC and the Skilled Trades in 2026

The short version

There are five jobs artificial intelligence (AI) tools are doing in the trades right now. The products overlap plenty, but the jobs are distinct:

One looks things up for you, like a fault code or a spec.
One walks you through the call while you are working it.
One helps you organize the options and how you present them. What is technically justified, and what to leave alone, stays your call.
One drafts the write-up of what you found, cutting down the notes you would be typing in the truck at nine at night.
One runs the office side, meaning dispatch, scheduling, and quoting.

Right now you buy those as five separate products, and that is the part worth paying attention to. Each one gets better when it already knows what the others know, because shared job and equipment context means nothing gets entered twice. Pinning down what the unit is makes your readings mean something. Your readings are the evidence behind the recommendation when the customer asks why. The write-up is made of all of it: the unit, the readings, the photos, and what you said out loud. That is why you are far less likely to rebuild it from memory at the kitchen table that night.

I was on a cooling inspection this summer where the numbers read like a low charge. Subcooling sat at 1.5°F with the outdoor temperature at 82°F, high side pressure was up, and a 15 degree split would not move. The evaporator coil looked clean from the front. It was packed on the back side, where you cannot see it without pulling a panel. Once it was cleaned and a pound of refrigerant went in, subcooling came up to 16°F and the split climbed to 18°F and kept going. The whole call depended on knowing to look somewhere you cannot see from where you are standing. None of it is license to skip the order, either: power stays off while a panel is off, airflow gets fixed before charge gets judged, and refrigerant goes in only against the manufacturer’s procedure for that unit. That is not the part of the job I thought software could help with, and it turns out it can: guided diagnostics can prompt the check you would not have thought to make, though you still have to pull the panel and look.

Where the hours in a service day actually go

The standard telling of the trades’ people problem goes like this: not enough young people come in, a wave of veterans retires, and the fix is to recruit harder.

The retirement half is real, but the age numbers are less dramatic than the telling suggests. The Bureau of Labor Statistics (BLS) put construction workers 55 and older at 22.7 percent in 2020, against 24 percent for the entire U.S. workforce. Recruiting is happening too: as of fiscal 2025, the Department of Labor counts 678,014 active registered apprentices across the registered apprenticeship system. What happens to people after they show up is the bigger, less-discussed problem.

The person best qualified to teach a green tech is often the one you most want out running calls, and mentoring takes experienced-tech hours even when the teaching happens on a live job. Time your best technician spends talking somebody through a diagnosis on the phone is time they are not on a call of their own. And when that call does not happen, the odds of a misdiagnosis go up, and so do the odds of the callback that follows it. Much of the tacit knowledge still moves through supervised field work, with one tech showing another. Classrooms and manuals carry the documented part. When the teaching falls to your best tech, the mentoring and the billing draw on the same hours.

Underneath all of that is a pile of work that uses up the day and never involves a tool:

Figuring out what you are looking at. Read the nameplate, get the age from the label or the serial, confirm fixed orifice or thermostatic expansion valve (TXV). That has to be settled before a single pressure reading means anything, because the metering device determines which charging procedure and targets apply under the manufacturer’s instructions for that unit.
Hunting for the right number. That means the charging chart, the target subcooling, or the fault code table for that exact model. Usually it exists somewhere. When it does, it can be behind a manufacturer portal that asks who you work for, and signing up mid-call is its own delay, often on a weak signal.
Getting hold of somebody who knows. You call the lead tech. They are on their own call, so you wait. When they call back you explain the whole thing from the beginning while the customer watches you do it.
Deciding what to tell the customer. This means turning what you found into options a homeowner will say yes to, without sounding like you are selling them the whole catalog.
Writing it all up. Figure a few minutes for a standard cooling inspection and more on a complicated diagnostic. The version that suffers most is the one done from memory in the truck, when you are already late to the next call. Typing on a phone is slower than a keyboard, and the jobs that produce the most to write down are where that costs you most.

None of that is the part of the job you signed up for, and all of it happens around the wrench time. That is where AI is proving useful today. It chips away at the mechanical part sitting in front of the decision. It does not make the decision; that stays with the technician. AI that runs equipment on its own is building-automation control, a different product category from anything on a service call.

My take

When AI comes up in this industry, somebody usually says it will hollow out the craft, and I think that is aimed at the wrong thing. Nobody is automating the part where you work out what is wrong and put your hands on it. What is getting automated is the time spent hunting for a number, and the phone call you made because there was no other way to get an answer at seven on a Sunday. Give a tech those two things back and more of the day can go to the part that pays.

The five kinds of AI showing up in the trades

Plenty of products have AI on the box now. Sorting them by what they do for you makes it easier to work out which belongs on your phone.

1. Looking things up

These answer questions about equipment. A question about a fault code, a wiring configuration, or a component spec gets answered from technical documentation. It is the easiest of the five to build, because much of that material was already written down by somebody, though not all of it and not for every unit. Bluon is probably the best-known standalone one in HVAC. ServiceTitan built Atlas into its own platform and draws on Bluon’s data for it. If your shop runs ServiceTitan with Field Pro set up, you may already have a version of this. It is not part of every ServiceTitan subscription. General assistants like ChatGPT get used this way too, and they are better suited to concepts than to model-specific equipment data.

Do not underrate it. Getting every manufacturer’s material into one place, instead of one app per brand with its own login, can save real time on real calls. What it will not do is think ahead of you. At its most basic, it answers the question you asked and stops there. Ask the wrong question and you get a clean answer to the wrong question. The answer may itself be wrong, which is why the vendors themselves tell you to verify against the manufacturer’s documentation.

2. Walking you through the call

This kind stays with you for the whole call instead of answering once and stopping. It works out what the specific unit is, takes you through the diagnostic sequence for that equipment in order, and folds your readings into the next step.

This is a different job from what a measurement app does. measureQuick reads your probes, profiles the equipment, flags faults, and suggests next steps. The company counts more than 100,000 technicians on it. It is built around the instruments and what they are reporting; this kind is built around the conversation, staying with you through the call with or without probes connected.

That is what we built Prentis to do. Photograph the nameplate or say the unit out loud, and it fills in the model and the specs that apply, with the refrigerant and metering device confirmed against the nameplate and the installed hardware. From there it walks you through the diagnostic step by step, and you answer however suits the job: out loud, typed, or by photographing what you are looking at. What it tells you comes from field-tested procedures across 146 makes and more than 18,000 indexed data points. It also asks how long you have been in the trade, so a first-year tech gets the detail they need and a journeyman gets the answer without the lecture.

This is harder to build than lookup for one reason. Being right in general is not good enough. It has to be right for the unit you are standing at, in the order you are doing the work.

3. Working out what to offer the customer

This is the easiest of the five to overlook. You found a bad contactor. What do you put in front of the homeowner? Just the contactor is one answer. The contactor plus the capacitor beside it that tested out of tolerance is a different answer, and the better one on a system that will give you trouble again in August. A hard start kit goes on the list of options when the compressor and the manufacturer’s literature require one.

Building that second option is a skill I have rarely seen taught anywhere formal. Trade school teaches you to find the fault. Laying out choices for a homeowner without feeling like a salesman mostly gets built on the job, and a tech who never got that practice presents the one repair and leaves the rest unsaid.

4. Writing up the call

These turn what happened into a record. Your photos, your measurements, and what you said out loud become a written summary of what you found, what you fixed, and what you recommended.

Documentation gets dropped when the day runs long, and plenty of contractors skip the service report entirely. Yet the report is your best evidence when a customer calls back angry three weeks later. Photos of a coil are your evidence in an argument about whether it needed cleaning. A shot of the carpet before you carried anything in prevents a different argument entirely. Prentis builds a timestamped close-out summary separating critical repairs from recommended work, ready before you pull out of the driveway.

5. Running the office side

Scheduling, routing, quoting, and inventory all carry AI features now. ServiceTitan is the biggest name of these, particularly for larger shops, and Service Fusion, Jobber, FieldEdge, and Housecall Pro all serve the same job from different angles. Plenty of shops get real value out of them.

They are a different product from anything above. Even so, shops evaluate them as if they competed with the field tools. A dispatch system’s real work is putting the right tech on the right call. That work is done by the time the tech pulls up to the house, even if the board keeps tracking the job after. The dispatch side cannot help once the panel is off. One note from the field: plenty of what these do assumes a connection. Offline support varies by platform, and basements and cinder block buildings are exactly where the signal drops out.

Knowing when to stop recommending

The half of the recommendation problem that gets ignored is knowing what not to recommend. There is a real difference between three options tied to the reason you were called and a list of ten things that reads to the customer like you came to sell them the whole system. The same judgment covers the bigger call, too, which is when a repair stops being worth making at all. A system on an obsolete refrigerant, out of warranty, and badly low on charge is a different conversation than a bad capacitor. A badly low charge also means a leak to find before anything else, since the guidance from the Environmental Protection Agency (EPA) is to locate and repair it rather than top the system off. Knowing which conversation you are in is worth more than any single measurement you took that day.

My take

This is the part I would watch most closely if I were a shop owner buying any of these tools. It is very easy to build software that recommends more, and much harder to build software that knows when to stop. Over-recommending does not just annoy the customer. It costs you the next call. A tool that generates ten line items because it can is not helping you. Ask what makes it hold back.

How the options compare when you are mid-call

Say you are mid-call, the unit is running, and you are not sure what you are looking at. Here is what you can reach for and what each one gives you.

What you reach for	Best for	Works with your hands full	Knows your specific unit	There at 9 p.m. Sunday
Calling a lead tech	Judgment calls and odd failures	Yes, if they pick up	Only what you describe or share	No
A general AI chatbot	Concepts and definitions	Partly, voice input exists	Only what you feed it	Yes
YouTube	Seeing a procedure you have never done	No, you stop working	No	Yes
Manual and spec lookup	Finding a published number	No	Yes, if you know the model	Yes
Connected measurement tools	Logged, shareable readings	Partly	Reads the system, not the model	Yes
A voice-first diagnostic copilot	Working a live call start to finish	Yes	Yes, once the unit is confirmed	Yes

YouTube gets its own line. It is free, it is on the phone already, and for watching somebody braze a joint or pull a blower wheel the first time, nothing beats it. What it cannot do is answer the question you have about the unit in front of you. You get somebody else’s unit, somebody else’s conditions, and a video that was right for that job and may not be right for yours. You scrub through the video hunting for the thirty seconds you needed, and the model on screen may not be the model in front of you. It is a training tool that techs are using as a diagnostic tool, and that gap is where bad calls come from.

Some of these answer a question and leave you on your own again. Others stay with you through the work. With both hands occupied you cannot type, scrub through a video, or hold a chart open. Anything that makes you stop working to use it tends to stay in your pocket, however good the answer would have been. And when the panel is live, stopping is the right call anyway.

What a service call looks like with a copilot running

Step 1. Pin down the equipment before you chase the complaint. Photograph the nameplate or say the make and model out loud. Before you touch a gauge, the unit, the refrigerant, the metering device, and the specs that apply are pinned down. The metering device is verified on the equipment itself, since many coils take either a fixed orifice or a TXV chosen at install. That is what makes every reading afterward mean something. It takes about thirty seconds to start a call.

My take

Almost every bad diagnosis I have looked at went wrong before anybody took a reading, because the tech was working off an assumption about the equipment instead of a confirmed fact. Charge a fixed orifice system by subcooling and you get subcooling readings that look like an answer and are not one. This is the least interesting step on the list and it prevents more mistakes than any measurement does.

Step 2. Work the sequence. The sequence runs disconnect, power, capacitor, amp draws, readings, in the order that applies to that equipment, with the safety steps kept where that order requires. How you answer back is your call. Speak it if your hands are full and you are working alone. Type it if the customer is standing there and you would rather not narrate. Photograph the meter and let the app read the measurement off the screen, then check the logged measurement against the display. The sequence is the same either way, and the point is that you are not stopping to go find the next step.

Step 3. Show it what you are looking at. Point the camera at your gauges, a wiring diagram, a code flashing on a board, or the component that failed. What it sees becomes context for the rest of the diagnosis instead of a separate question you have to go ask.

On another call this summer a homeowner had water coming through a ceiling from a system barely a year old. The evaporator coil had been installed backwards, the top of the A-frame facing the filter instead of the blower, so condensate was landing outside the drain pan. Nothing in the readings showed that. You find it by looking at the coil and knowing what right looks like.

Step 4. Get held up when you should be held up. It is built to hold you at a step until you confirm the safety work is done, so you get asked about the capacitor before you are working near it. It is also designed to say when it has reached the end of what it knows instead of filling the gap with a guess. The safety decisions still run on your judgment and your shop’s procedures, whatever the software does.

Step 5. Lay out the options. The options come out as critical repairs, recommended work, and the things you want on the record, sorted so you can walk a homeowner through a first, second, and third option instead of naming one price and hoping.

Step 6. Close it out before you leave. All of it becomes a customer-ready summary, timestamped, ready in the driveway.

What AI cannot do for you

My take

The habit that separates a senior tech from a green one is not being right faster. It is scanning the whole system before saying anything to the homeowner. A tech who finds a bad capacitor and stops has been technically correct and has still done half a job, because the contactor beside it was worn to the point of needing replacement and the coil was filthy, and that customer is getting a second call in August. Any tool that fixates on the first fault it finds is teaching the habit that produces callbacks.

It will not do three things. They are worth stating plainly, because they are easy to leave out of the marketing:

It does not make the call and it does not carry the risk. You are the licensed professional standing on that property: the diagnosis and the repair belong to you, and the safety decisions run on your judgment and your shop’s procedures, never on a suggestion from software. Any product whose marketing talks around that is promising something it cannot deliver.
It cannot know what you have not shown it. Anything that has never been given your unit, your readings, and your conditions is usually working from generalities, and a confident wrong answer can cost you more than no answer would have.
It does not replace time on equipment. Judgment comes from doing the work, and there is no version of this where a first-year tech and a twenty-year tech are the same on day one. What a good tool does is shorten the distance between year one and year twenty. It puts the right procedure in front of you the first time instead of the fourth, so more of your hours go toward learning the work rather than looking for it. It is not a substitute for the reps.

Why the trades are a harder problem than they look

Most software gets built for somebody at a desk with both hands free and decent lighting. That assumption routinely fails in this trade, since technicians often work in awkward, cramped, or poorly lit spaces, and a product that is excellent on a laptop can be useless in an attic in August.

There is a second thing about how people learn this trade. A lot of what gets taught early is procedure: on this unit, in this situation, do this. It comes alongside classroom instruction and supervised work next to somebody experienced. Procedure is a fine place to start, and repetition is how the basics take hold. But pairing the steps with the why is what makes them carry over when the system in front of you does not match the example you were shown. What a green tech is often short on is not steps. It is the reasoning that tells them which step applies here, and that part builds through instruction, supervised work, and time on a variety of equipment.

The third thing shapes the whole problem. Doing the work builds knowledge, but years on the job are not by themselves what make an expert. The work is also hard on the body. This trade carries one of the highest injury and illness rates of any occupation, and injuries tend to cost older workers the most, in dollars and in lost workdays. Plenty of the most knowledgeable techs have decades in the field, though tenure alone is a poor measure of expertise. The physical wear that came with those decades is preventable injury, not a cost the knowledge required. Retirements and career exits keep removing experienced people from hands-on work. Aging tends to lower physical capacity, while experience raises what a tech can do with what remains. Neither runs on a fixed schedule. Recruiting harder brings people in, but a new hire still needs years of instruction, mentoring, and time on equipment before the knowledge transfers.

Much of the knowledge was never scarce: standard procedures, published specs, and factory sequences are written down somewhere for a wide range of equipment. But not all of it ever got documented: unusual faults, field modifications, and the tacit judgment experienced techs carry often exist nowhere but in someone’s head. What is scarce is getting the part that is written down to the person who needs it in the minute they have to decide something. Our mission is to distill on-the-job trades knowledge into predictable outcomes for technicians.

If you are weighing any tool in this category, ask it four questions:

Does it know the unit you are standing in front of?
Does it work when your hands are full?
Does it know when not to recommend something?
Does it tell you when it does not know?

Ask those about our product, and about everybody else’s.

Frequently asked questions

What is AI actually being used for in HVAC right now?

AI is doing five main things: answering questions about equipment and fault codes, guiding a technician through a live diagnostic, helping build the repair options a customer chooses from, writing up the service call, and running the office side, which stretches from booking and dispatch through invoicing and marketing. Guiding the live call is the hardest of them, because it has to be correct for the one unit in front of you rather than correct in general.

Can AI diagnose an HVAC system?

AI can run the diagnostic sequence with you, hold the specs for your equipment, check your readings against the published targets it has for that unit, and tell you when a result does not fit what it expected. Check any number it reads off a photographed display against the instrument before you act on it. It cannot hook up the probes, since even the apps that pull readings straight from connected gauges rely on you to put them on. It also does not carry the responsibility. The licensed technician on site makes the call and does the work, and the manufacturer’s documentation is the authority over anything the software tells you.

Can AI help me build repair options for a customer?

Yes, building repair options is one of the more useful things AI does. Given what you found and what unit you are on, it can lay out a first and second option so the homeowner sees choices instead of a single number. Look for whether it also knows when to stop, because a tool that lists everything it can think of costs you the customer’s trust faster than it earns you a ticket.

Is AI going to replace HVAC technicians?

No, AI is not going to replace HVAC technicians. Somebody has to be on the property with hands on the equipment, and responsibility for the diagnosis stays with the licensed professional who is there. What it takes off your plate is the work around the job: identifying the unit, digging for specs, waiting on a call back, building the options, and writing everything up.

What is the difference between using ChatGPT and a diagnostic copilot?

A general chatbot starts with nothing on the unit you are working on, and whatever context it gets, you feed it yourself a piece at a time. A diagnostic copilot identifies the specific equipment, walks you through a sequence matched to it, takes photos of nameplates and gauges as context, and lets you answer by voice, by typing, or by camera depending on where you are and who is watching.

Do I need new tools or hardware to use AI on a service call?

No, you do not need new tools or hardware. Guided diagnosis and the write-up run on the phone already in your pocket, using the camera for nameplates, gauges, wiring diagrams, and fault codes. Connected measurement instruments are a separate category you can add on your own schedule.

Glossary

AI: artificial intelligence, software that answers questions or suggests steps from trained patterns.
Amp draw: the current a motor or compressor pulls while running, measured with a clamp meter.
BLS: the Bureau of Labor Statistics, the federal agency whose workforce age data is cited here.
Copilot: a voice or text assistant that follows a diagnostic sequence with you during a call.
EPA: the Environmental Protection Agency, which sets refrigerant handling rules.
Evaporator coil: the indoor coil where refrigerant absorbs heat from the air.
Fixed orifice: a metering device with no moving parts, charged by superheat rather than subcooling.
Nameplate: the equipment label carrying model, serial and electrical ratings.
Subcooling: the temperature drop of liquid refrigerant below its condensing saturation temperature.
Superheat: the temperature rise of refrigerant vapor above its evaporating saturation temperature.
TXV: thermostatic expansion valve, a metering device that adjusts to hold superheat steady.

Drafted with AI assistance and reviewed by the author.

← All posts