Physical AI changes the consequence of an AI error. A poor text answer can waste time. A poor physical decision can damage equipment, interrupt a production line or place a person at risk. That difference should change how founders design the product, how buyers evaluate it and how investors define progress.
My thesis is that the strongest physical-AI ventures will not be those promising the widest autonomy first. They will be those that can prove a useful operating envelope: the task the machine may perform, the conditions it can handle, the signals that stop it, the person who can intervene and the evidence retained after every exception.
Market momentum is not operating proof
The market signal is real, but still early. Reuters reported in September 2026 that the International Federation of Robotics estimated about 7,000 humanoid robots were sold worldwide in 2025 for industrial and professional service uses. The same report contrasted that figure with roughly 542,000 conventional industrial robots installed in 2024. Many humanoids were bought for research and training-data collection rather than productive work.[8]
A J.P. Morgan forecast reported by Reuters on 12 November 2026 projected US humanoid production of 300,000 to 400,000 units annually by 2030. That is a forecast, not an achieved deployment result. For a founder, the useful conclusion is not that demand is guaranteed. It is that buyers will increasingly need credible ways to distinguish production evidence from impressive movement.[9]
Physical autonomy needs coverage, not only accuracy
NIST describes autonomous-systems assurance as a coverage problem. A system may face lighting changes, weather, people, animals, markings and unexpected objects that conventional testing cannot exhaust. The commercial implication is straightforward: a single success rate cannot explain where a robot is safe to operate. The buyer needs to know which conditions were tested, which remain outside scope and how the machine behaves at the boundary.[2]
Connectivity adds another layer. NIST's work on uncrewed systems connects increasing automation with changing cybersecurity and AI risks, especially in public safety, inspection, hazardous-materials handling, transport and delivery. A robot can fail because perception is wrong, because a command channel is compromised, because software is unavailable or because an authorized operator cannot recover control. These are one operating system, not separate procurement checklists.[3][1]
Define the operating envelope before the demonstration
I would require a physical-AI proposal to define six boundaries before discussing scale: the specific task; the permitted environment; the people and assets that may enter it; the maximum tolerated force, speed or deviation; the fallback state; and the authority to resume operation. Those boundaries turn a demonstration into an acceptance plan.
The ISO 10218 standards, published in 2025, separate the industrial robot from the integrated application and cell. The first part addresses the robot as partly completed machinery. The second part covers integration, commissioning, operation, maintenance and decommissioning. The standards also list important exclusions, including airborne, medical, service and consumer robots. They are therefore useful industrial references, not blanket certification for every physical-AI product.[4][5]
That distinction matters in Saudi Arabia. A warehouse manipulator, a hospital service robot, an inspection drone and a school robot may share software components, but they do not share the same hazards, regulator, user population or evidence threshold. A reusable autonomy stack can reduce engineering cost. It cannot erase the need for application-specific safety and authorization.
Saudi governance should become product architecture
SDAIA's AI Ethics Principles emphasize privacy and security, reliability and safety, transparency and interpretability, and identifiable accountability. A physical-AI company can translate those principles into architecture: authenticated commands, bounded permissions, event logs, clear human override, documented model and software versions, and named responsibility for release decisions.[6]
For robots connected to industrial environments, the National Cybersecurity Authority's OT controls add another practical layer. They set minimum requirements for protecting operational-technology and industrial-control environments and extend the Essential Cybersecurity Controls. The founder should expect asset inventory, vulnerability management, network boundaries, controlled access, monitoring, recovery and evidence to be part of the sale.[7]
This is not compliance theatre. In physical systems, a security update can affect timing, sensors or integration. A vulnerability scan can interrupt production if handled carelessly. Governance must therefore connect engineering, cybersecurity, operations and customer acceptance in one change process.
The entrepreneurial mechanism is verified deployment
A defensible venture can make verified deployment its product. It can package site assessment, task design, simulation, controlled trials, safety and cyber controls, operator training, maintenance, incident review and performance reporting. The robot may be visible, but the recurring business value comes from keeping the approved task reliable.
The metrics should reflect accepted work rather than theatrical capability. Track accepted tasks per operating hour, interventions per hundred tasks, unsafe-state triggers, recovery time, unresolved software or sensor defects, energy use, maintenance hours, false stops and the cost per accepted outcome. Separate laboratory, supervised-pilot and unsupervised-production results.
For a Robotics.sa intervention experiment, a credible protocol could compare manual work, assisted robotics and bounded autonomous operation on one defined task. Measure completion quality, cycle time, interventions, near misses, operator workload and recovery. That is a proposed research design, not a claim that such an experiment or partnership has been completed.
For a BLACKBAT-style drone feasibility study, the same discipline applies. Define the permitted mission, weather and communications envelope; geofence and abort rules; data retention; pilot or operator authority; maintenance condition; and post-mission evidence. A mission-planning interface is not flight approval, safe operation or commercial acceptance.
The counterargument: general capability will improve quickly
The strongest counterargument is that foundation models, simulation and large-scale robot data will make task-specific controls obsolete. If one model can perceive, reason and act across many machines, narrow engineering may look temporary. Broad capability could also reduce integration cost and accelerate learning across sites.
I expect capability to improve, but I do not think it removes the operating envelope. A more capable system can encounter more situations and create more valuable options, yet the buyer still needs defined authority, safe fallback and evidence. General intelligence expands what might be attempted. It does not decide what an organization has accepted.
Saudi, GCC and MENA paths will differ
Saudi Arabia offers large industrial, logistics, infrastructure, healthcare and education opportunities, together with national AI and cybersecurity governance. The first commercial wedge should still be a bounded task with an accountable buyer and measurable cost. Localization should include Arabic operations, climate and site conditions, maintenance capacity, data handling and the real approval chain.
The GCC should not be treated as one deployment market. Procurement practices, aviation and health rules, industrial ownership, data requirements and local-service expectations differ. A venture can reuse technology and evidence templates, but it should revalidate the operating envelope and commercial owner in every country.
Across wider MENA markets, access to capital, replacement parts, connectivity, insurance and trained maintenance can vary sharply. A robot that depends on continuous cloud access or imported specialist support may have poor unit economics outside a controlled flagship site. Resilience and serviceability are therefore product features, not after-sales details.
The decisions I would make now
Choose one task where failure modes can be observed. Name the human authority and safe state. Create a test matrix that includes normal work, degraded sensors, lost connectivity, unexpected people and failed updates. Price deployment and maintenance separately from hardware. Keep evidence by software, model, site and operator version. Do not claim autonomy beyond the accepted envelope.
Physical AI will create durable value when the machine can do useful work repeatedly, stop safely, recover predictably and show what happened. The next advantage is not a more dramatic demonstration. It is an operating system that customers can trust, maintain and buy again.
Key takeaways
- Define the task, environment, limits, fallback and human authority before the demonstration.
- Treat safety, cybersecurity, operations and maintenance as one deployment system.
- Measure accepted work, interventions, recovery and cost per outcome separately across lab, pilot and production.
- Revalidate the operating envelope and accountable buyer in every Saudi, GCC and MENA market.