This week a manufacturing engineering guide on edge AI for predictive maintenance did something most physical AI writing refuses to do. Instead of selling the model, it counted the hours. It broke a real deployment down into where the effort actually goes, and the numbers are the most honest thing I have read about this field all year.
Here is the breakdown it gave for getting an edge-AI predictive-maintenance system running on a factory floor:
| Where the work goes | Share of effort |
|---|---|
| Sensor normalization | 24% |
| On-device inference (the model) | 22% |
| Maintenance-workflow integration | 20% |
| Alert routing | 18% |
| Model monitoring | 16% |
Read the second row again. The model, the part of the project that gets the conference talks and the press releases and the funding rounds, is 22% of the work. The other 78% is plumbing. As the guide put it: "the model may be compact, but the surrounding system still needs trusted inputs, deterministic deployment, explainable alerts, safe failover, and integration with the maintenance workflow."
If you have ever shipped one of these systems, that 78% is not a surprise. It is the entire memory of the project.
The sump pump model was a weekend. The system was a year and a half.
We have been running an edge AI system on a sump pump in a Watertown basement since 2024. The model at the heart of it is a small residual autoencoder that learns what a normal pump cycle looks like and watches how far reality drifts from that. It runs in thirty-eight kilobytes on a microcontroller. Building it took a weekend.
The weekend is not why the system is still running today. Everything in that 78% column is why.
Sensor normalization came first, and it was humbling. A raw water-level reading is not a fact, it is a noisy voltage that means different things at different temperatures and on different days. Before the model saw a single cycle, we spent real time learning what that specific sensor's clean signal looked like and how to catch it when it started lying. Then alert routing, which is its own discipline: an alert that fires too often gets muted, and a muted alert is the same as no system at all. We tuned for months to land on alerts a human would actually trust. Then integration, the least glamorous part of all, which meant making the system fit how the building is actually checked on and acted upon, not how a demo imagines it would be. And then monitoring the model itself, because a model that was right in March can quietly go stale by September if nobody is watching it.
Two stories this week said the same thing
The manufacturing guide was not the only signal. Over in Weekly Robotics #362, the project everyone passed around was a developer turning a ten-euro toy robot into an AI system by bolting on a four-dollar microcontroller and offloading the heavy thinking to a nearby laptop. Cheap dumb node, smart compute nearby, no cloud. The lesson buried in a fun hack is the same lesson in the boring manufacturing table: the intelligence is the commodity now. The architecture around it, what runs where, what talks to what, what happens when a part fails, is where the actual engineering lives.
Even NVIDIA's big National Robotics Week push, under all the foundation-model announcements, kept circling the same point. The headline-worthy news was that their systems now run inference locally, on the device, at the asset, with no cloud dependence. The frontier is spending enormous effort to move the smart part down to where the boring part already lives. In a building, the boring part is all you ever had, and that turns out to be an advantage.
Why 78% is the small operator's moat, not its burden
It is tempting to read that effort table as bad news for a small shop. The model is the fun part, and it is only a fifth of the job? The rest is grunt work?
It is exactly backwards. The model is the part you cannot win on. You are not going to out-train NVIDIA, and you do not need to, because a good-enough small model is now a download. The 78% is the part a generic vendor cannot ship you in a box, because it only exists in your building. Nobody can normalize your sensors from a datacenter. Nobody can route your alerts to the right person on your team from a slide deck. Nobody can integrate with how your maintenance actually happens without standing in the boiler room.
That is the forward-deployed-engineer advantage, scaled down to a small commercial building. The big robotics labs are spending fortunes to make the model better. The durable business is being the person who does the other 78% honestly, on site, for a building too small to ever get a six-figure automation contract. The commodity got cheap. The plumbing is still worth money, and it is worth more to the person standing next to the pump than to anyone selling a model.
What to do with this
If you are evaluating a monitoring system, or building one, the effort table is a buyer's checklist in disguise. Ask the three questions that cover the 78%:
- How do you handle a sensor that drifts? If the answer is "the model is very accurate," they are selling you the 22% and ignoring the part that breaks. Normalization and trusted inputs come before the model, not after.
- Who gets the alert, and how do you keep it from crying wolf? Alert routing is 18% of the work for a reason. A system that fires on every blip gets muted within a week, and a muted system protects nothing.
- What happens when the model goes stale? Equipment ages, seasons change, baselines move. If nobody is monitoring the monitor, the system slowly stops being right and nobody notices until the basement floods.
The model is the part everyone talks about because it is the part that is easy to talk about. The job is the other 78%, and it is the same job whether you are running a factory floor or a single sump pump in a closet. We have been doing that job in basements for two years. The manufacturing world just put a number on it.
We do the 78% the box can't ship
Confidence-aware edge AI for buildings too small for a building automation system. We normalize your sensors, route alerts a human will trust, and integrate into how your building actually runs. Local models, local compute, no per-token billing. $99 to $199 per month, hardware under $3,000.
See how it worksSources: Progressive Robot, "Edge AI in Manufacturing: Predictive Maintenance Guide" (2026-06-01); Weekly Robotics #362, "How to build a shitty robot" by Mario Zechner (2026-06-01); NVIDIA blog, National Robotics Week 2026 roundup; field deployments at The Intersecto Watertown sump pump site and Northampton 40-device building.