Physical AI in Manufacturing: Why Executives Must Act Now

Artificial intelligence has become a boardroom priority, but most of the discussion still centers on knowledge work: automating reports, generating software, helping customer service teams. While it is true that those applications create real value, they cover only part of what AI can do economically. For industrial companies, the bigger opening is in the physical world.

Physical AI applies artificial intelligence to machines, production systems, and industrial operations. Equipment can perceive its environment, make informed decisions, and coordinate complex activities with little human intervention. Over the next decade that capability will change who competes well in manufacturing and who will be left behind.

Adoption is still in its early stages, which strengthens the argument for putting it on the agenda now rather than later. Deloitte's 2026 manufacturing outlook, citing a Manufacturing Leadership Council survey, puts current use of physical AI at 9 percent of manufacturers, with 22 percent expecting to deploy it within two years¹. That is a technology moving from pilot to rollout inside a single capital planning cycle. The installed base it will run on is already substantial: 542,000 industrial robots were installed worldwide in 2024 alone, taking the operational fleet to roughly 4.66 million units, more than double the number a decade earlier². Most of those machines execute fixed programs today. The question for the next ten years is how many of them will start making decisions.

The pressure is already there

Manufacturers are under pressure from several directions at once. Skilled trades are short of people. Deloitte and The Manufacturing Institute estimate that US manufacturing could need 3.8 million additional workers between 2024 and 2033, and that as many as 1.9 million of those roles may go unfilled if the skills gap is not closed. Sixty-five percent of the manufacturers they surveyed named attracting and retaining talent as their top business challenge³.

At the same time, product portfolios keep expanding while batch sizes shrink. Supply chains stay unpredictable, sustainability targets demand better energy efficiency and less waste, and customers want faster responses without compromising quality or cost. Conventional automation was built for stable, predictable environments, so it can only take you part of the way.

What Physical AI adds is adaptability. Rather than asking engineers to anticipate every operating condition in advance (which is borderline impossible), intelligent systems evaluate production status as it changes and either recommend appropriate action to an operator or carry it out themselves. The engineering work shifts from writing exhaustive rules toward defining objectives, constraints, and safe boundaries.

Where the value is generated

Robotic systems handle product variation without extensive reprogramming, which makes production more flexible. Orchestration software balances workloads across machines, so assets get used better. Inspection, leveraging AI-powered machine vision, gets smarter and defects issues earlier, before more value is added to a part that will end up as scrap. Maintenance moves from reactive repair toward predictive intervention, and energy use can be tuned continuously instead of periodically.

Two of those are worth spelling out. Handling variation without reprogramming is what makes small batches pay, which matters if batch sizes keep falling. A cell that can recognize a part it has not seen before, adjust its grip and keep running reduces the changeover penalty that pushes manufacturers toward longer production runs than their order book actually calls for. Predictive maintenance changes what a maintenance budget buys. Instead of servicing on a calendar and still being surprised, teams intervene when a particular machine shows signs of a particular failure mode.

Sites that have pushed hardest on this give some sense of the range that can be expected. The World Economic Forum's Global Lighthouse Network, now 201 factories, reports average improvements of 40 percent in labor productivity, 48 percent in lead time and 28 percent in energy consumption across its newest cohort, with AI enabling up to half of the top use cases implemented⁴. These are flagship sites, so the figures are better read as the upper end of what disciplined execution produces than as a forecast for any particular site. They do show that the gains are operational rather than theoretical.

The gain that may matter most is resilience. When something goes wrong, an intelligent system can find alternative production paths, reallocate resources, and get back to running faster than conventional automation architectures allow. Siemens estimates that unplanned downtime costs the world's 500 largest companies around $1.4 trillion a year, about 11 percent of revenue, up from $864 billion five years earlier. In automotive, an hour of stopped production runs to $2.3 million. Incidents actually became less frequent over that period, but the cost of each one climbed as material values rose and spare capacity thinned out⁵. In a volatile market, recovering faster is worth something on its own.

Building a strong foundation

The mistake would be to treat Physical AI as one more standalone technology purchase. Its value rests on the digital foundation underneath it: connected equipment, standardized industrial data, interoperable software platforms, digital twins, secure infrastructure, and software-defined automation architectures. Without those, companies tend to get isolated AI wins but fall short of holistic successes on an enterprise scale.

McKinsey's 2025 global AI survey found 88 percent of organizations using AI in at least one business function, about two thirds of them yet to start scaling it across the enterprise, and only 39 percent able to point to any effect on operating earnings⁶. In industrial settings the constraint is rarely the model. It is that the data needed to train and run it is siloed away in controllers, historians, and line-side systems that were never designed to talk to each other. The practical symptom is familiar: a use case works on the line where it was built, then takes another six months of integration work at the second site, and the business case dies somewhere in the queue for the third. Anything that has to be rebuilt per machine will not reach a whole plant, let alone a network of them.

That points to four investments worth making now:

Modernize industrial software architectures so equipment and business systems can actually interoperate. This is where separating control logic from specific hardware starts to pay, because it lets you change behavior without requalifying the machine underneath it.
Invest in semantic industrial data platforms that support AI training and ongoing operational intelligence. Semantic here means governed, contextualized, and consistent enough that two plants describe the same event the same way.
Put money into workforce development, because engineers, operators, and maintenance teams will spend more of their time supervising intelligent systems and less of it controlling every detail by hand. The World Economic Forum expects around 40 percent of the skills workers need to shift by 2030, and 63 percent of employers already name skills gaps as their main barrier to transformation⁷. Retraining takes longer than installing software, so it needs to start before the technology lands rather than after.
Start with high-value use cases that produce measurable business results before pushing toward broader autonomous operations. A single line with a quantified outcome and an architecture you can reuse is worth more than ten pilots that never connect to each other.

There is a governance question sitting alongside all of these. A machine that decides for itself needs boundaries someone can inspect, an audit trail of what it did and why, and a safety case that holds up when its behavior is no longer fully deterministic. Certification and liability practice for learned behavior in safety-relevant functions is not settled, and how it gets handled varies by industry and by region. Companies that take up that question early usually move faster later, since they are not retrofitting explainability and access control onto a system already running in production.

People stay in the loop

None of this is about replacing people. The organizations that get the most out of it will pair human judgment with machine intelligence. AI is good at working through large volumes of operational data and spotting where things could run better. People supply context, experience, ethical judgment, and strategic decisions. Deloitte estimates that more than 81 percent of task hours in manufacturing stay human driven even as AI use expands (cf. ¹). The combination produces manufacturing systems that are both more productive and more resilient.

What does change is what a plant needs from its people. Supervising a system that makes its own decisions calls for someone who can tell a sensible decision from a bad one and who knows when to step in. That is a higher bar than operating a machine, and it is the part of the transition that budget alone will not speed up.

Executives tend to get the timing wrong in both directions. Moving before the foundation exists produces expensive demonstrations that cannot be extended anywhere and don't scale. Waiting for the technology to settle means buying capability at a point when competitors have already worked out how to use it. The slow work is not the AI. Most companies have already bought it. It is connecting equipment, cleaning up data models and retraining teams that takes years, and none of it depends on which vendor eventually wins.

Every major industrial transformation so far has rewarded the companies that built foundational capabilities before the technology matured. Physical AI looks like the next one. For manufacturing leaders the open question is no longer whether intelligent industrial systems become commonplace. It is whether their own organization leads that shift or spends the next decade catching up.

Why Physical AI Should Be on Every Manufacturing Executive's Strategic Agenda

The pressure is already there

Where the value is generated

Building a strong foundation

People stay in the loop

Sources

This belongs on the agenda now