Physical AI faces its real test after the pilot, not during it

Budget is not the obstacle. BCG's 2026 AI Radar found that corporations expect to more than double AI spending this year, from 0.8 percent of revenue to roughly 1.7 percent, the article says. Value has not followed at the same rate: PwC's 2026 AI Performance Study found that 74 percent of AI's economic value is captured by just 20 percent of organizations. On that reading, the binding constraint is neither ambition nor spending but execution.

The problem is sharper for physical AI, which operates where decisions affect safety, uptime, service delivery and cost rather than where they affect the drafting of an email or the summarizing of a document. Moving such systems beyond a pilot requires more than a capable model, the article states. It lists four requirements: choosing a problem worth solving, giving the system the operational context to understand it, designing the technology around how work actually gets done, and linking every insight to an appropriate action.

On the first point, the article argues that much enterprise AI discussion has centered on large language models and office productivity, while some of the most costly business problems sit outside the office, including vehicle collisions, equipment failures, unplanned downtime and disruption to essential services. Physical AI addresses them by drawing on data from vehicles, equipment, cameras and sensors. In fleet safety, for example, tools can combine information from connected vehicles and cameras with weather, road conditions and driving patterns, helping an organization tell an isolated incident from a broader pattern and direct attention where it matters most.

Interest and deployment remain far apart. In robotics, described in the article as one of the most prominent applications of physical AI, Capgemini found that 79 percent of organizations are engaging with the technology while only 27 percent are deploying or scaling it.

The article therefore says the right starting question is not where AI can be used but what operational outcome needs to change, what is causing the problem, whether the necessary data exists and which decision or workflow the technology should improve. A pilot that cannot answer those questions may demonstrate technical capability without proving business value.

Selecting the problem is only the beginning. Knowing a vehicle's registration or model is insufficient, according to the piece; a system may also need its maintenance history, current location, typical route, driver behavior and operating conditions, and must understand how those factors interact and what normal performance looks like in comparable situations. When such information is fragmented across systems, delayed or recorded by hand, the system sees only part of the operation. Digitizing workflows and connecting operational data create the foundation that more advanced systems depend on, and the article stresses that the significance lies not in data volume but in having the data available in a shared operational system through which teams manage vehicles, maintenance and customer commitments.

A pilot can be closely supervised; a scaled deployment cannot depend on constant intervention from a technical team. Because physical operations are distributed across vehicles, worksites, warehouses and frontline teams, the article says technology must be easy to install, reliable in changing conditions and connected to systems employees already use, and it must work for the people expected to act on its output. That makes change management part of the technical strategy: employees need to understand what the technology is identifying, how it supports their work and what action is expected of them, while leaders need to establish ownership, measure whether workflows are changing and feed frontline responses back into the process. Governance should also be built in, with organizations defining which actions systems may take autonomously, when human review is required and how decisions can be understood or challenged, at a level of oversight matched to the potential consequences.

Finally, the article states that an insight is not an outcome and that physical AI creates value when it shortens the distance between identifying a problem and responding to it, a progression it frames as a sense, decide and act model.