Seventy Percent of These Pilots Fail. Almost None of Them Fail in the Model.

A number went around the predictive-maintenance world this week that should stop any small building owner in their tracks, and then, once you understand it, should make them feel a great deal better.

The number is seventy percent. That's roughly how many predictive-maintenance pilots fail to make it into production. You buy the sensors, you run the trial, you build the model, and seven times out of ten the whole thing quietly dies before it ever becomes the thing you actually rely on.

Here's the part that matters, and it's the part the trade press finally said out loud. When these pilots fail, the cause is almost never the sensor, and almost never the model. The sensor works. The model works. What fails is the architecture between them. The data pipeline. The gateway hops. The cloud round-trip. The central platform that has to swallow a thousand assets across a dozen sites and reason over all of them without ever dropping a beat. The failure lives in the middle.

I build physical AI for small buildings. A sump pump in a basement in Watertown. Forty devices in a building in Northampton. And when I read that the seventy percent fails in the middle, the first thing I thought was: my builds barely have a middle.

Where the pilot actually dies

Picture the kind of project that produces that seventy percent. A facilities team with hundreds of motors, pumps, compressors, and rooftop units spread across many buildings. They put sensors on a meaningful slice of them. Now all that data has to go somewhere, get normalized, get time-aligned, get pushed up to a central analytics platform, get correlated, get turned into alerts, and get routed back down to the people who can act. Every one of those arrows is a place to fail. A gateway drops offline. A site's network changes. The cloud ingestion schema shifts. Two vendors' timestamps disagree. The platform that was supposed to unify everything chokes on the eleventh data format.

None of that is the sensor's fault, and none of it is the model's fault. The sensor on motor number 412 is reading vibration just fine. The model would catch the bad bearing if it ever got clean data. But between the sensor and the model is a sprawling integration layer, and that layer is where the project goes to die. The bigger and more centralized the deployment, the more middle there is, and the more middle there is, the more of it breaks.

A single asset barely has a middle

Now look at the sump pump in that Watertown basement and count the arrows. There's a vibration sensor and a current clamp on the pump. There's a small edge box on the wall, a few inches away, running a model that learned what that one pump sounds like when it's healthy. When the pump's normal slips, the box notices and sends an alert. That's the whole architecture.

Sensor, to edge box, to model, to alert. All of it in one room. On one asset. With no cloud in the path it depends on and no central platform trying to correlate it against nine hundred other machines. The thing that kills seventy percent of enterprise pilots, the fragile integration layer between sensor and model, is the exact thing this build doesn't have. There's no pipeline to break, because the sensor is six inches from the model. There's no multi-site sync to drift, because there's one site. There's no cloud schema to shift under me, because nothing I rely on lives in the cloud.

The trap in "most of this stuff fails"

Here's the conclusion a small building owner reaches from a headline like "seventy percent of predictive-maintenance pilots fail." They think: the technology isn't ready. It mostly doesn't work yet. I should wait until the failure rate comes down.

That's backwards, and the reason is exactly where the failures happen. The seventy percent isn't telling you sensors can't hear a failing bearing, because they can. It isn't telling you models can't catch drift, because they do. It's telling you that the integration layer at centralized scale is brutally hard, which is a problem you only have if you build like an enterprise. The owner watching one pump is not running the kind of project that fails seventy percent of the time. They're running the kind that has almost no middle to fail in the first place. The failure rate you read about is a property of centralized scale, not of the thing that would sit on your wall.

The frontier is quietly moving toward your basement

The most telling thing this week wasn't just the failure number. It was the fix the big vendors are converging on. Siemens, working with Arm, started shipping AI-capable sensors and controllers that run the models on the asset itself, with a small model on the sensor, a larger one on the gateway, and little to nothing in the cloud on the hot path. Read that carefully. The industry's answer to "the middle keeps failing" is to shrink the middle, push the intelligence down to where the sensor is, and pull the cloud out of the loop.

That is the architecture a single-asset edge build has run from day one. The detector lives on the wall, next to the machine. The decision is made locally, in milliseconds, with no round-trip. There was never a sprawling pipeline to collapse, because the design started where the frontier is only now arriving: keep it local, keep it small, keep the model next to the sensor. The big players are spending real money to delete the exact middle that a fleet-of-one never had.

What this means if you own a building, not a fleet

When the whole industry spends a week admitting that most predictive-maintenance pilots fail, it's easy to hear that as "the technology isn't ready, sit tight." But that failure rate is about the plumbing of centralized, multi-site deployments, not about whether a sensor can hear your pump dying. It can. The hard, failure-prone part is the middle, and the middle is something you only build when you try to run a thousand assets through one brain.

Your building isn't a fleet, and it doesn't need to be wired like one. It has a handful of assets that have been quietly narrating their own health the whole time with nobody listening: the pump, the boiler, the compressor, the rooftop unit. The twenty percent that cause eighty percent of your trouble. Give each one its own small watchman, sensor a few inches from the model, decision made locally on the wall, no central platform and no cloud round-trip to fail. You don't get the seventy percent failure rate, because you never built the thing that fails seventy percent of the time.

Sources: edge-AI predictive-maintenance coverage reporting that roughly 70% of predictive-maintenance pilots fail to reach production, with the cause "almost never the AI model or sensor hardware" but the architecture between them, and a vendor shift toward distributed inference (small models on sensors, larger models on gateways, cloud out of the hot path) (EE Times, "Edge AI Is Forcing a Rethink of Predictive Maintenance Architecture," and 2026 PdM trade guides); Siemens and Arm shipping Armv9-based AI sensors and controllers (SIMATIC S7-1500 PLCs, SIMATIC IoT2040 edge nodes) that run models directly on the asset rather than in the cloud (Arm Newsroom, 2026); industrial IoT sensor hardware down ~85% since 2019, vibration-monitoring nodes under $50 (2026), crossing positive ROI on assets worth $5,000+; edge AI market $24.91B (2025) projected to $118.69B (2033) at ~21.7% CAGR (Grand View Research); Japan Airlines deploying Unitree-platform humanoids at ~$15,400/unit and Figure logging 1,250+ hours / 90,000+ parts at BMW Spartanburg (humanoid trade coverage, June 2026). Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-06-25.md.