This week the most useful thing I read about physical AI didn't come from a robotics lab. It came from a building owner, in a smart-buildings newsletter, describing a night that cost them a lab's worth of research.
The setup is ordinary, which is exactly why it matters. A bank of air handlers. The freeze-protection stats tripped overnight and forced an emergency shutdown to keep the pipes from bursting. The control sequences running those units were written in the 1990s. The software supervising them reached end of life in 2018. And when it failed, no alarm reached a single human being — because, as the owner put it, nobody is capable of watching the old system that closely. By morning the rooms were frigid and the experiments inside were gone.
The building wasn't neglected. It had modern laboratories and live, funded research. What was neglected was the nervous system underneath all of it: the building automation system that everyone agreed should be upgraded and nobody had upgraded.
Everyone wants the upgrade. Almost nobody buys it.
Here's the part that should interest anyone who sells into buildings. In that same report, roughly 30% of the building owners interviewed this year said upgrading their controls was a top priority. Not a nice-to-have. A top priority. And it still doesn't happen.
It doesn't happen because a full building-automation replacement isn't a purchase, it's a treaty. Facilities has to sign off on new control logic they'll have to live with for fifteen years. IT security has to bless a new system touching operational equipment. The energy team wants measurement and verification so they can prove the savings. Capital planning has to find the money and defend it against the roof and the parking lot. Four departments, four incentives, one live building that can't go offline while they argue.
You can see the same caution in how the most sophisticated operators behave. In the same roundup: Morgan Stanley won't even run active vulnerability scans on its building-management systems, because an active scan risks knocking operational equipment offline — they use passive pattern-recognition instead. Amazon spent six months on stakeholder mapping and measurement design before an optimization AI was allowed to change a single setpoint. Touching a live building is heavy. Everyone who does it at scale already knows that.
The thing failing at 2am is a watching problem, not a controls problem
Go back to the frozen lab. Strip away the politics and ask what actually went wrong that night. It wasn't that the sequences were from the 1990s. Plenty of 1990s sequences run fine for decades. It was that when this one failed, nobody was positioned to notice. The alarm path assumed a human was watching a screen, and no human was.
That's not a reason to replace the BAS. It's a reason to put something next to it whose only job is to watch.
This is the machine I actually install. A sump pump in a basement in Watertown, monitored for over a year by one off-the-shelf vibration-and-current sensor feeding a small model on a small local box. In a 40-device building in Northampton, the same idea across more equipment. The box doesn't run the building. It doesn't issue commands. It learns how one specific machine normally behaves — the vibration signature, the current it draws on a hot day versus a cold one, how it starts, how it settles — and it pages a person the moment that machine drifts away from its own normal.
A monitor needs no committee
Here is why the read-only monitor is a different kind of purchase than the BAS reboot, and why it closes when the big project stalls.
It doesn't touch the control logic, so facilities doesn't have to bless a new fifteen-year system. It's read-only and never on the control network, so security isn't scanning anything or exposing anything — it's the passive posture Morgan Stanley already decided was the safe way to observe a live building. It changes no setpoints, so there's no energy-team M&V fight about who gets credit for savings. And at $99 to $199 a month with hardware under $3,000, it's an operating expense a facilities manager can approve without a capital committee.
Four departments have to agree to replace the BAS. Zero of them have to agree to watch one machine. One sensor on the air handler that scares you most is a decision one person can make this week, while the reboot stays exactly where it's been: in committee.
| Full BAS reboot | Read-only monitor | |
|---|---|---|
| Who must approve | Facilities, IT security, energy, capital planning | One facilities manager |
| Touches control logic | Yes — replaces it | No — read-only, off the control network |
| Timeline | Multi-year | Installed this week |
| Cost shape | Six-figure capital project | $99–$199/mo, hardware under $3,000 |
| Catches the 2am silent failure | Eventually, if it ships | Now, on the machine you point it at |
This isn't a replacement for the reboot. It's the thing you do while it stalls.
I'm not arguing against upgrading building controls. Old systems should get modernized, and the 30% of owners who want it are right to want it. But that project moves at the speed of four departments agreeing, and the freeze stat doesn't check the capital calendar before it trips.
A monitor is what you put in the gap. It buys you eyes on the equipment most likely to fail expensively, starting the week you install it, without asking anyone to rip out anything or take on a scrap of new risk. If the big reboot happens in three years, great — the monitor rides alongside the new system just as happily as the old one. And if it doesn't happen for three years, you weren't blind for three years.
The lab in that story lost its research because a machine failed quietly and nobody was watching. Everything about that night was preventable, and none of the prevention required winning the argument about the BAS. It required one box that was awake, and one person who'd get the message.
Your building automation upgrade can stay in committee. Your equipment can still be watched.
Each critical asset — your air handlers, your pumps, your compressors, your boiler — gets one off-the-shelf sensor and a small model on a local box. It learns that machine's normal in about two weeks, then watches for the drift that shows up before a failure and puts it in front of a person who knows the equipment. Read-only, never on your control network, nothing leaves the building, nothing to approve but the sensor itself. $99 to $199 per month against a $125 to $208 market rate, hardware under $3,000, installed this week. Start with the one machine that scares you most.
See how it worksSources: Nexus Labs “Smart Building Signal — Owner Signal: The BAS Reboot, A Cross-Departmental Problem” (July 2026), including the frozen-air-handler anecdote (1990s control sequences, supervisory software at end of life in 2018, no alarm reaching a human, lab research lost), the finding that roughly 30% of interviewed building owners rank a controls upgrade a top priority, the cross-departmental “people problem” framing, Morgan Stanley’s use of passive pattern-recognition baselining over active vulnerability scanning on building-management systems, and Amazon’s six months of stakeholder mapping and measurement-and-verification design before its Brainbox AI pilot changed a setpoint. Small commercial HVAC predictive-maintenance pricing of $125 to $208 per month via Oxmaint 2026 benchmarks. Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-25.md.