AI in building automation: what's real vs. hype in 2026
Walk any trade show floor and you'll see "AI" on every booth. Some of it is real, shipping, and useful in a mechanical room. Some of it is a rules engine with a new label. Here's how to sort it.
Five kinds of "AI" you'll actually see
1. FDD with smarter prioritization
Most analytics products still run on engineering rules: if the heating and cooling valves are both open, flag it. What's new is using machine learning to rank, group and explain those faults, so a tech gets five things to fix instead of five hundred alerts. This is mature and useful. The underlying rules matter more than the AI label, so ask to see the rule library. See our FDD guide.
2. Supervisory optimization on top of the existing BAS
These products read building data, predict loads, and write setpoints back to your existing BMS: supply air temperature, static pressure, start times and so on. BrainBox AI is one well-known example; Trane Technologies completed its acquisition of the company in January 2025. This category is real, but it needs write access to your controls. That's a much bigger deal than read-only analytics (see the red flags below).
3. Physics-based or "autonomous" control
A different approach replaces programmed sequences with a model of the building's systems that decides how to operate them. PassiveLogic describes its platform as physics-based AI that connects design, HVAC controls, commissioning and operations, running on its own Hive controllers. It's ambitious and interesting, and usually a new-construction or major-retrofit decision rather than a bolt-on.
4. AI assistants for techs
Tools like Aquant (conversational troubleshooting) and XOi (AI-captured job documentation) help the person in front of the equipment. They're judged on whether they save a tech time on a real call. Ask for a trial on your own equipment types, not a canned demo.
5. Data plumbing: tagging and semantic models
The least flashy and arguably most important category. AI and analytics only work if software knows that AHU1_SAT is the supply air temperature for the unit that serves the third floor. Open efforts like Project Haystack and Brick standardize that. Companies like Mapped sell automated normalization. ASHRAE's proposed Standard 223P aims to unify the approach, but as of this writing it was still a proposed standard: its second public review ran April 24 to June 8, 2026. Be wary of any spec or sales pitch that treats it as final.
Red flags
- Savings claims with no baseline. "Up to 30%" means nothing without an M&V method. Ask how savings will be measured in your building, against what baseline, adjusted for weather and occupancy.
- Write access without a plan. If a product writes setpoints, ask about limits, override priority, how to release control back to the BAS in one step, and what happens when the cloud connection drops.
- Vague data requirements. A serious vendor can tell you exactly which points, trend intervals and history they need. "Just connect us and the AI figures it out" is a warning sign.
- Cybersecurity hand-waving. Ask how they connect (outbound-only? VPN? a gateway on the OT network?), who can reach your BAS remotely, and whether they've been through your IT team's review before.
- Data you can't take with you. Your trend history and fault records should be exportable in a usable format.
Questions that separate real from slideware
- Show me a fault or optimization result from a building like mine, with the actual trend data.
- What does your product do when a sensor fails or reads garbage?
- What exactly is the "AI" doing that a well-written rule couldn't?
- Who on your team reviews results, and who on mine has to?
- What's the contract length, and what happens to my data if I cancel?
Heading to Chicago in January? We put these and more into a printable AHR Expo question sheet.
