This week Siemens and Arm described their version of the future of predictive maintenance, and it is more ambitious than just predicting failures. They put generative AI models directly onto Armv9-based sensors, bolted to the motors and conveyors and actuators on the factory floor, running locally with no trip to the cloud. The headline capability is not detection. It is action. When the system sees trouble coming, it "can automatically adjust machine parameters in real time, slowing the motor, balancing loads, or triggering a targeted cooling cycle." The AI does not tell a human the motor is sick. It reaches into the motor and changes how it runs.
That is where the whole field is pointing right now. The trade press calls it agentic predictive maintenance: a system that does not just predict a failure but acts on the prediction on its own, rerouting production or throttling a machine with no human in the loop. EE Times ran a piece this week noting the industry has not even agreed on where the intelligence should live, with automation vendors guarding deterministic control and chip vendors stuffing ever more inference into the sensors themselves. It is a real and unsettled fight. And for a Siemens production line, with a PLC, an integrator, a safety case, and an operator on every shift, letting the AI close the loop and drive the machine may well be the right answer.
For a sump pump in a basement, it is exactly the wrong answer. And refusing to do it is one of the most valuable design decisions we make.
The moment your AI drives the machine, you own the machine
Think about what changes the instant you let an AI take an action on a physical asset instead of just reporting on it. Before, the worst thing a bad model could do was send a false alert, and a human would shrug and move on. After, a bad model can slow a pump that needed to run, throttle a compressor that was fine, or trigger a cooling cycle that masks the real problem. You have created a brand-new failure mode where the monitoring system itself can break the equipment it was hired to protect.
And you have quietly taken on the asset's entire safety case. The day that motor was specified, somebody decided how it should behave, what its limits are, and what is allowed to change its operation. The float switch, the thermostat, the deterministic PLC, the manufacturer's controller, those are boring and they are reliable and somebody already certified them. The moment your AI is allowed to override or adjust any of that, you are now part of that safety story, and you owe the customer an explanation every single time the system intervenes. For a small operator putting cheap sensors on someone else's building, that is a staggering amount of liability to absorb in exchange for a feature nobody asked for.
Our sump pump system never touches the pump
Our sump pump system in a Watertown basement has been running for a long time, and in all that time it has never once told the pump what to do. It cannot. By design, it has no path to the motor. The pump is still controlled by the same dumb, reliable float switch it shipped with. What our system does is watch: it learned what a healthy pump cycle looks like, and it notices when today's reality drifts from that baseline, a cycle running long, a frequency creeping up, a pattern that does not match the pump's own history. When it sees drift, it does exactly one thing. It raises a hand and tells a human.
That is the entire product, and the restraint is the point. The homeowner's existing control keeps controlling. The model's only job is to be the early-warning layer that boring hardware never had. If our detector is wrong, the cost is a text message that turned out not to matter. If it is right, somebody gets to the pump before the basement floods. There is no version of our system's failure that ends with our software having broken the customer's equipment, because our software was never allowed to operate the equipment in the first place.
Contrast that with the agentic version. To let the AI slow the pump, we would have to wire into the control path, certify that we will not make things worse, handle the case where the model misreads a normal heavy-rain cycle as a fault and throttles the pump exactly when the basement needs it most, and explain all of it to a homeowner who just wanted to know if the thing was about to die. We would be trading a clean, defensible product for a tangle of liability, all to automate a decision a human is perfectly happy to make.
Forty observers, zero drivers
The objection is always that a single pump is too simple, and that a real building needs the AI to actually run things. Our 40-device site in Northampton is the test of that, and it holds. Forty assets that matter, forty small detectors, each trained on its own history, each watching its own signal. Not one of them actuates. Not one of them is wired to drive the thing it watches. The building did not become controllable by our software; it became observable. The existing controls, the thermostats and the timers and the manufacturer logic, keep doing the driving. We keep doing the watching.
That separation is what makes the system safe to deploy fast and cheap across a whole building. Every actuator you add is a new safety case, a new failure mode, a new conversation with the owner about what the AI is allowed to break. Every observer you add is just another cheap sensor and another baseline. One of those scales by repeating a small, safe pattern forty times. The other scales by signing up for forty liabilities. For a small operator, only one of those is a business.
What ICRA reminded everyone this week
There is a nice coincidence in the other big robotics story of the week. The editor of Weekly Robotics went to ICRA 2026, the field's flagship conference, expecting a floor full of humanoids and reasoning demos. Instead, robot hands and tactile sensing took something like a third of the exhibition. The value, even at the frontier, kept concentrating in the sensing, in the boring layer that perceives the physical world accurately, not in the brain that decides what to do about it.
That is the same lesson from the other direction. The expensive, glamorous, dangerous part is the decision and the action. The durable, valuable, underrated part is sensing the asset well and reporting it clearly. For a small building, you can skip the expensive part almost entirely. Sense the pump well. Learn its baseline. Raise a hand when it drifts. Let the float switch, the thermostat, and the human keep driving, because they have been driving safely the whole time.
Watch the asset. Let the asset's controls control it.
The agentic predictive-maintenance wave is real, and on a factory line with the staff and the safety case to back it, letting the AI close the loop may be exactly right. But a plumber, a property manager, and a small building owner are not running a Siemens line. They are running a specific machine that needs to be watched cheaply and reliably for years, and that already has a controller they trust.
For that world, the strongest product is the one that knows its place. Put a small, specific detector on the asset, train it on the asset's own history, run it local, and have it do one job: watch, and raise a hand when healthy turns into trouble. It never grabs the wheel. That is not a limitation we apologize for. It is the feature: explainable, bounded, and incapable of breaking the thing it was hired to protect. Let the frontier teach the machines to drive themselves. We will keep watching the pump.
We watch your equipment. We never take control of it.
Each asset that matters gets a small, dedicated detector trained on its own history, watching for the moment healthy turns into trouble, and alerting your team. It runs local, with no cloud and no per-token bill, and it never touches the controls. Your existing systems and your people stay in charge. $99 to $199 per month, hardware under $3,000.
See how it worksSources: Arm Newsroom, "Siemens Reinvents Factory Reliability with Edge AI-Driven Predictive Maintenance" (Armv9-based AI sensors, SIMATIC S7-1500 / IoT2040, autonomous real-time parameter adjustment), 2026; EE Times, "Edge AI Is Forcing a Rethink of Predictive Maintenance Architecture" (deterministic control vs. inference-in-sensor; agentic PdM), 2026; Weekly Robotics #363, ICRA 2026 field report by Mat Sadowski (tactile sensing and robot hands ~30% of the floor), 2026-06-09; field deployments at The Intersecto Watertown sump pump site and Northampton 40-device building.