Physical AI: The Operating Reality

In July, we argued that physical AI can be understood as a systems market rather than a model market: that the hardware, data, evaluation, safety and deployment operations around the model matter as much as the model itself.

Since then, we have surveyed 85 robotics buyers ("Buyers") on the operating reality of their implementations. They are plant directors, heads of automation engineering, VPs of operations and COOs, all with decision-making authority and all already running robots. The full sample is described at the end.

Most of what the Buyers report supports our argument that physical AI is a systems market. One finding, however, challenges it.

We gave Buyers eight options and asked them to rank the three hardest. Fitting the robot into the way the plant already runs came first: 69% put it in their top three and 32% called it the single hardest thing, the highest first-place share of any option. Operating it live came second, at 55%. Taken together, 84% named workflow integration, live operations or both, and 41% named both. The AI model tied with safety and security for fifth, at 32%.

Asked about barriers to scaling, Buyers are more divided. Integration is still the barrier named most often, by 51%, but cost is the one Buyers most often put first.

Asked whether they can tell if a change improved real-world performance, only 20% of Buyers say their measurement is reliable enough to trust. For how the system handles edge cases it was not built for, the figure is 19%. About half say they measure each of these, but only roughly.

The metrics Buyers track consistently are about keeping the plant running rather than autonomy. Intervention rate, for instance, is tracked by only 42%, and that barely shifts between piloting and scaling.

76% of Buyers told us human intervention (how often a person has to stop what they're doing and go help the robot) has fallen over the past year. Asked what contributed, 88% of them said they had narrowed the range of tasks or conditions the system attempts. Other contributors were routing harder cases to people before the system tries them (75%), slowing the system down for reliability (71%) and adding supervision (63%).

Each of these changes can lower intervention without necessarily changing the robot's underlying capability.

Buyers are wary of demonstrations: 69% agree they are not a reliable indicator of production readiness. Yet 89% of those who reported a drop in intervention believe it reflects a genuine gain in autonomy rather than a narrower job.

We asked Buyers to split 100 points across the three parts of a deployment. On their own estimates, robot hardware takes 41% and the AI model and autonomy 28%, roughly 70% between them. Everything else (data, evaluation, operations, integration, safety and support) takes 30%, even though it includes much of what Buyers describe as hardest.

This is not consistent with our working hypothesis that physical AI is a systems problem, with two caveats. First, the spread is wide: hardware alone runs from 10% to 80% across individual Buyers, and 36% of Buyers put at least as much on the surrounding work as on hardware. Second, the question asks Buyers to divide their own cash outlay. Where a Buyer's internal engineering team does the integration, that effort may not show up as deployment spend at all, so the surrounding share could be understated.

71% of Buyers take a complete solution from a robotics vendor, and that barely shifts as deployments mature. What seems to shift is what they add around it: 53% of Buyers at scale also buy hardware separately and 56% use a systems integrator, against 21% and 25% of those still piloting. Buyers are not dropping the packaged vendor as they grow; they are adding integration work on top of it.

Among the 43 Buyers who put at least 30 of their 100 points on the surrounding work, 70% direct the largest part of it to live operations or workflow integration.

19% of Buyers say the economics of their deployments work well today, while 56% say they don't work yet but are improving over time.

By deployment maturity, 41% of Buyers who have scaled across multiple sites say their economics work well today, against 7% of those in production and 4% of those still piloting. 13 of the 16 Buyers who say the economics work are running at scale.

That does not necessarily mean scaling fixes the economics. It could also be that the deployments that work are the ones that get scaled, while the ones that don't never leave limited production.

The 48 Buyers who say things are improving credit four conditions: less human intervention (69%), the system handling more of the workflow itself (54%), faster deployments (48%) and a falling support burden (42%). Only 25% credit cheaper hardware. Asked what has actually changed over the past 12 months, more Buyers report movement on what the robot does than on what it takes to stand up a new deployment.

Payback is slow. Only 4% of Buyers expect their money back inside a year, and 55% expect it to take more than two years.

85% of Buyers say human oversight is essential, and 91% stop short of full autonomy. The most common position is that robots act on their own in low-risk situations only (32%).

Overall, 57% are comfortable with autonomous operation inside an explicit boundary with an escalation path, while 34% want a person in the execution loop, either controlling every action or approving before execution.

This matches what Buyers told us about intervention: they narrow what the system attempts, route hard cases to people and add supervision.

Our paper argued that physical AI is a systems market rather than a model market, and on this evidence the claim holds in part. Buyers put integration and live operations ahead of both the model and the hardware, and they say much the same thing when asked what is stopping them from scaling. Their budgets, though, do not follow: most of what they spend still goes to the robot and the AI model. Buyers appear to experience physical AI as a systems problem while continuing to buy it as a hardware-and-model purchase. A single survey cannot tell us whether that gap is a lag or a correction to our argument.

Running underneath all of this is a measurement problem. A falling intervention rate is one of the numbers this category reaches for most often as evidence of progress, yet fewer than half of these Buyers track it consistently, and most of those reporting a drop achieved it partly by narrowing the robot's job. The Buyers themselves believe the gains are real. But intervention rate alone cannot tell us how much of the gain comes from better autonomy and how much from changing the operating envelope. Until that changes, claims about progress in physical AI, including ours, rest on numbers that the people reporting them cannot fully check.