Ask ten HVAC technicians what they think of AI and you will hear two answers: "it makes things up" and "it saved me an hour last week". Both are true. The difference is almost always what the AI is connected to. A chatbot that only knows the public internet guesses. A tool that reads your building's live points, your manuals and your past jobs gives answers you can check.
This article looks at where AI is actually useful in HVAC today, where it is heading over the next few years, and what technicians, controls contractors and building owners should watch for.
From generic chatbots to tools that know your equipment
The first wave of AI in our trade was the general chatbot. It can explain what a reversing valve does or what a BACnet BBMD is, but it has never seen your rooftop unit, your controller's point list or the fault you fixed on the same model last winter. When it doesn't know, it fills the gap with something that sounds right.
The second wave, the one arriving now, connects AI to specific, trusted sources:
- the company's own solved service calls and notes;
- manufacturer documents: installation manuals, fault-code tables, wiring diagrams;
- live data from the building: BACnet objects, Modbus registers, trends and alarms.
Answers then come with a source you can open, and they get better every time the team adds a solved case.
Example: field assistants that learn from your own jobs
A good example of this direction is Brixon, an AI field assistant built for HVAC and electrical technicians. Instead of searching the internet, it answers from the team's own documents and from problems the team has already solved, and shows where the answer comes from. A technician describes the problem on site (with a photo if needed), and when the job is solved it becomes part of the company's knowledge for the next person who meets the same fault.
That matters because the most valuable knowledge in a service company is usually in the heads of two or three senior technicians. When an assistant captures how they solved a recurring VFD fault or a stubborn heat pump defrost problem, a newer technician can fix it the first time, and the knowledge stays when someone retires.
Free field tools that pair well with ControlsLens
Brixon also offers a set of free field calculators with no account needed. Several of them are exactly what you reach for when a value read with ModLens or BacLens looks wrong and you want to know if the problem is the sensor, the scaling or the equipment:
- 4–20 mA and 0–10 V signal scaling: check that a raw analog value in a Modbus register or BACnet analog input matches the engineering units you expect.
- Thermistor and RTD lookup: compare a temperature on the BMS with the resistance you measure on a 10K Type II/III or Pt100/Pt1000 sensor.
- Mixed air temperature and % outdoor air: verify an economizer by comparing outdoor, return and mixed air points with the damper command.
- Psychrometric chart: plot return and supply conditions from the BMS humidity and temperature points.
- Fan laws and VFD speed: sanity-check the speed, frequency and airflow a drive reports over Modbus.
- Hydronic GPM, BTU and ΔT and chiller tons: confirm the flow and energy values from a BTU meter or chiller controller.
- Error code lookup: find what a fault code read from a controller or drive means before you go to the manual.
A practical workflow: read the live value with ModLens or BacLens, check it against one of these calculators, and if something still doesn't add up, ask the field assistant with the readings in hand.
AI on the building automation side: live data, not guesses
Field assistants work from documents and past jobs. On the controls side, the big change is AI that reads the building itself.
Alarm diagnosis
An alarm says what is wrong ("zone too warm", "pump commanded on, no status"). The technician still has to find why. AI is good at the first pass of that investigation when it can see the data around the alarm:
- every point of the equipment and how it trended over the last hours;
- the other alarms on the same unit;
- recent operator actions and overrides left in place;
- past service tickets on that equipment.
With that context, an AI can say "the damper is commanded at 100 % but airflow stays at the minimum: likely a failed actuator or a broken linkage; check the actuator first" instead of a generic checklist. In BacLens, the "Diagnose with AI" button on an alarm does exactly this, and the answer goes onto the service ticket so the technician sees it before driving to site.
Point mapping, graphics and control code
Large BMS jobs spend days on repetitive work: figuring out which of 3,000 cryptically named points is the supply air temperature of AHU-3, drawing graphics, writing sequences. AI shortens that work a lot, as long as a person checks the result. We compared what different models do on this kind of job in Which AI Model for BMS Work?.
Clean data comes first
None of this works on bad data. A temperature read with the wrong byte order or scale (see Modbus float byte order), a device that never answers, or a point named "AI-17" with no description will mislead an AI just as it misleads a person. The unglamorous work of commissioning, verifying addresses, scaling and units, is what makes AI useful later. Tools like ModLens for Modbus and BacLens for BACnet exist for exactly that step.
What the next few years probably look like
Predictions age badly, so here are trends that are already visible rather than promises:
- Assistants that span the whole job. The same assistant that helps on site (Brixon-style, from documents and past jobs) will increasingly be able to look at the live BMS data for that equipment, so the answer combines "what we did last time" with "what the unit is doing right now".
- Fewer nuisance alarms. AI is good at spotting alarms that fire because a threshold is too tight or a delay too short, and at grouping a flood of alarms under one root cause.
- Faster commissioning and retro-commissioning. Automatic point recognition, functional test scripts and reports written for the owner, checked and signed by the technician.
- Energy and peak management that explains itself. Not only "shed stage 2", but why, how much it saved and what comfort impact it had.
- Local and private options. Many sites cannot send building data to the cloud. Models that run on a local PC or server (for example through Ollama) are getting good enough for a lot of everyday work.
What does not change: the technician stays in charge
Building systems are physical and sometimes safety-related. An AI that suggests is helpful; an AI that writes to a chiller or a fire damper without a person checking is a liability. Good tools are built around that:
- show the evidence behind every suggestion (values, trends, source documents);
- run new control code in a test mode first, where nothing is written to equipment;
- keep a log of what was changed, by whom, and why;
- let the technician override or reject every suggestion.
The technicians who will do best are not the ones who avoid AI, nor the ones who trust it blindly, but the ones who use it to get to the right question faster and still check the answer with a meter, a trend or a site visit.
FAQ
Will AI replace HVAC technicians?
No. Diagnosis on site, safe work on live equipment, judgment about what to repair and talking with the customer all stay human. AI removes part of the searching and paperwork, so technicians spend more of their time on the work only they can do.
What is an AI field assistant for HVAC?
A tool a technician uses on site, usually from a phone, to ask about a fault, a part or a procedure. The useful ones, such as Brixon, answer from the company's own documents and solved jobs and show their sources, instead of guessing from the internet.
How does AI help with BMS alarms?
By reading the equipment's live points, trends, recent overrides and past tickets around an alarm and proposing the most likely causes with what to check first. The technician confirms on site. Nuisance alarms (thresholds too tight, delays too short) are also easier to spot.
Is it safe to let AI control HVAC equipment?
Only with guard rails: suggestions shown with their evidence, new code tested without writing to equipment, every change logged, and a person approving anything that writes to a live system. Safety-related equipment should stay under normal, verified controls.
Do I need cloud AI, or can it run locally?
Both exist. Cloud models are usually stronger on hard tasks; local models (for example through Ollama on a site PC) keep data on site and are often enough for everyday mapping, short scripts and questions.