Your Deployment Is Generating Two Assets. Your Ops Stack Is Only Catching One.
Every Physical AI deployment produces two things: a service and a signal. The service is what you sell. The signal is what tells you which unit is about to fail, which site is hostile, and what to build next. Most companies are instrumented for neither. They learn a unit is failing when the customer calls, and they learn what the field was teaching them only after the conditions that produced it have passed.
When VS Networks put touchscreens into John Deere dealerships, the pitch to customers was a better in-aisle selection experience. That was the service, and it was real. But the data those screens generated, product registrations, customer purchase intent, brand engagement at the point of decision, turned out to be equally valuable. Some of our most important revenue streams came not from the device on the wall but from what the device was learning about the people standing in front of it. We were running a service business and a data business simultaneously. The companies that recognized that early grew differently from the ones that did not.
Physical AI is repeating this pattern at a larger scale and a faster pace, and most of the builders in this space are only managing one of the two assets their deployments are generating.
The Service Is Obvious. The Signal Is Not.
A Physical AI deployment earns its keep by doing something: moving inventory, serving food, navigating a facility, running a diagnostic. That output is easy to price, easy to sell, and easy to measure.
The second asset is quieter. Every unit in the field is also reporting on itself and on the place it lives. Motor draw creeping up over six weeks. Error rates climbing at one site and nowhere else. A battery that behaves one way in a conditioned showroom and another way in an unheated warehouse in February. That stream is the difference between a small central team running a fleet and a company that hears about every failure from an angry phone call.
Most companies are not capturing it. The sensors are already on the device. The thresholds were never set, the alerting was never built, and nobody owns the question of what normal is supposed to look like for a unit at month eighteen.
The Same Signal Builds the Other Asset
Here is what makes the engineering worth doing twice over: the telemetry that tells you a unit is degrading is the same telemetry that makes the product better. It reflects real-world conditions, real customer behavior, and real operating environments that no lab can replicate. It gets more valuable with every additional deployment, and it is almost impossible for a late entrant to reproduce without running the same deployments for the same duration in the same environments.
The companies treating this seriously are building accordingly. Mind Robotics, spun out of Rivian, was founded explicitly to use data from Rivian’s manufacturing operations to train industrial robots. The factory floor is not just the customer. It is the data source that makes the product better. Physical Intelligence has structured its entire business model around the same insight: every robot running their models feeds the next version of those models.
You do not have to be training foundation models for this to pay. The pipeline that would feed a model is the same pipeline that tells your field team which three units to visit this week, and which site to stop selling into until the product changes.
Two Assets, One Deployment
The operational question is not whether your deployment generates valuable data. It does. The question is whether you have built the architecture to capture it, the contracts to own it, and the team to act on it.
Most Physical AI companies have not done that work. The data is there. The pipeline is not.
The fleet you cannot see is the fleet you cannot fix. Every month you run without that pipeline you pay for it twice: once in the truck rolls you could have avoided, and once in a dataset you will never get back, because the conditions that would have produced it have already come and gone.