When AI gets a body: reflections on the legal framework and strategic guardrails for deploying physical AI in the EU

Artificial intelligence is reshaping almost every sector, from manufacturing and logistics to healthcare and mobility. Much has been written about its opportunities and risks, and the topic features heavily in our own daily work. One development, however, deserves a closer look, because it shows the overlap of innovation, safety and liability particularly clearly: physical AI. By this we mean AI systems that perceive their environment through sensors, make decisions in real time and act in the physical world through actuators—collaborative and autonomous mobile robots, humanoid robots, autonomous vehicles, drones and AI-controlled machinery.

For manufacturers, integrators and deployers, this brings risks as well as opportunities for a new kind of intelligent industry transformation. Once AI controls physical systems, an algorithmic error can cause physical harm, sensor data can raise questions of workplace surveillance, and a cybersecurity incident can quickly become a safety incident.

Innovation, AI regulation and product safety

EU Regulatory Framework

The European regulatory landscape for AI and products has become considerably more concrete in recent months. The EU AI Act creates a horizontal framework for AI systems,Regulation (EU) 2024/1689 (Artificial Intelligence Act), OJ L, 2024/1689, 12.7.2024; on the staggered application, see Art. 113 AI Act.Show Footnote the new Machinery Regulation addresses self-evolving, machine-learning-based safety functions for the first time and will apply from 20 January 2027,Regulation (EU) 2023/1230 on machinery, OJ L 165, 29.6.2023, p. 1; see in particular Annex I Part A and Annex III.Show Footnote the revised Product Liability Directive expressly treats software and AI as products and must be transposed by 9 December 2026,Directive (EU) 2024/2853 on liability for defective products, OJ L, 2024/2853, 18.11.2024.Show Footnote and the Cyber Resilience Act brings cybersecurity requirements for connected products, with its reporting obligations applying since September 2026. Many public bodies, public enterprises and companies, including ours, are currently working through what this means in concrete terms.

Physical AI shows well how these areas overlap. Before deployment, the risk classification under the AI Act is a key step—and in practice rarely a simple one. AI that serves as a safety component of a product covered by EU harmonisation legislation, such as machinery, will often be classified as high-risk, with the corresponding obligations applying from August 2027. Roles along the value chain are also frequently unclear. A deployer that integrates a third-party model into a robot, fine-tunes it or changes its intended purpose may itself become a provider under the AI Act and, through a substantial modification, a manufacturer under the Machinery Regulation. In our advisory work, these distinctions regularly prove demanding.

Physical AI in public infrastructure modernisation

The public sector is also becoming an important user of physical AI. Given the considerable investment backlog in bridges, roads, rail, sewer networks and public buildings, as well as the shortage of skilled workers, AI-supported inspection drones, sewer and pipe robots, autonomous construction machinery and robotic maintenance systems offer real potential: damage can be detected earlier, maintenance planned on the basis of condition data and work in hazardous environments made safer. Large-scale public investment programmes in infrastructure make these questions all the more pressing.

Public bodies, however, are subject to additional requirements. As deployers of high-risk AI systems, they must meet the deployer obligations under the AI Act and, as bodies governed by public law, generally carry out a fundamental rights impact assessment before first use (Art. 27 AI Act). Procurement law determines how such systems can be acquired: alongside conventional tenders, the innovation partnership and pre-commercial procurement allow solutions that are not yet available on the market to be developed together with suppliers. Contracting authorities should define requirements on transparency, documentation, data access, update obligations and exit options in the tender documents; the EU model contractual clauses for the procurement of AI (MCC-AI) offer a useful starting point.

Special case: physical AI in critical infrastructure

The tension becomes particularly acute where physical AI is used in critical infrastructure—for example inspection robots and drones in energy grids, autonomous systems in water supply or AI-controlled handling equipment in ports and logistics hubs. Operators here want the efficiency and resilience gains that AI promises, but they also bear special responsibility for security of supply. Under the AI Act, AI systems used as safety components in the management and operation of critical digital infrastructure, road traffic or the supply of water, gas, heating and electricity are classified as high-risk (Annex III, point 2), independently of the product-related classification described above.

Sector-specific security law applies on top of this. The NIS2 Directive and its German implementation require affected entities to take appropriate cybersecurity risk management measures, to report significant incidents within short deadlines and hold management personally accountable for implementation. The CER Directive and the German KRITIS umbrella legislation add requirements for the physical resilience of critical entities. For connected, AI-controlled machines, this means that cybersecurity, functional safety and resilience can no longer be considered separately: a compromised robot is a security incident, a safety risk and potentially a supply disruption at the same time. 

The problem: when the system is never ‘finished’

Product safety law has traditionally followed a fairly predictable logic: a product is designed for a defined purpose, tested, assessed for conformity and placed on the market. Its properties then remain largely stable.

Physical AI does not fit this logic. Vision-language-action models translate camera images and spoken instructions directly into movements; they are powerful, but their behaviour is probabilistic and hard to trace. Over-the-air updates, new model versions and fleet learning keep changing the system over its lifecycle. General-purpose and humanoid robots make it harder to distinguish a defined intended purpose from open-ended use. A system that was compliant on delivery may no longer be so after the next update.

The legislator has responded: under the new product liability regime, manufacturers remain liable for defects caused by updates or by missing security updates, and where technical complexity makes proof excessively difficult, presumptions of defectiveness and causation help injured parties.Cf. Arts. 7, 9 and 10 Directive (EU) 2024/2853.Show Footnote The ‘black box’ problem is therefore increasingly the manufacturer’s problem. Meanwhile, the harmonised standards meant to specify the AI Act requirements are delayed, and in its Digital Omnibus the European Commission has proposed linking the application of the high-risk obligations to their availability.European Commission, Digital Omnibus proposals of 19 November 2025. Readers should verify the current status of the legislative procedure.Show Footnote In our view, a possible postponement is no reason to relax: the substantive requirements remain, timeline is still tight and liability law applies in any case.

What does this mean for public authorities, public enterprises and private companies developing, procuring and deploying physical AI?

Physical AI is more than a passing trend; it reflects a structural shift. In our view, this leads to several areas for action that public authorities, public enterprises and private-sector organisations alike, including our clients, are working on:

1. Clarify classification and roles
   
   Before any deployment, it is worth examining carefully which regulatory regimes apply and who assumes which role along the value chain. Clear internal criteria for when changes to software, models or operating parameters amount to a substantial modification help to avoid unexpected provider or manufacturer obligations. At the same time, the deliberate definition of the intended purpose and operating environment offers genuine scope for design.
2. Consider regulatory requirements together
   
   The AI Act, Machinery Regulation, product liability, Cyber Resilience Act, data protection and occupational safety are not isolated regimes; they form a network of interlinked requirements. In practice, it helps to combine risk management, technical documentation, post-market monitoring and reporting obligations in a single governance and safety management system—for example based on ISO/IEC 42001 and relevant robotics standards such as the revised ISO 10218 series.
3. Build in safety and human oversight
   
   A proven design principle is to separate the AI-based ‘intelligence layer’ from a deterministic, verifiable safety layer that the AI cannot override—speed and force limits, protective stops and monitored zones. Just as important is a realistic concept of human oversight: who monitors how many systems, how quickly can they intervene, and are they aware of the risk of automation bias? Clear-cut answers are rare here, and trade-offs are unavoidable.
4. Involve people early and allocate risks by contract
   
   Involving the works council early, communicating openly with employees and designing sensors with privacy in mind—for instance through edge processing and abstraction of image data—are legal requirements, but they are also key to acceptance. Along the value chain, contracts should clearly set out roles, update and support obligations, access to logs and documentation, as well as liability and recourse, and should be aligned with insurance cover.

Outlook

Physical AI will change how we work, produce and move. For public administrations, public enterprises and private companies, for example in manufacturing, logistics, mobility, healthcare and public infrastructure, as well as for us as advisers, this means engaging with the topic early and being prepared to keep learning. The key regulatory instruments, from the AI Act to the Machinery Regulation and the new product liability regime, are in place, and the coming months will be decisive. In our view, the aim should not be perfect solutions but robust governance structures, built deliberately and developed continuously, so that physical AI drives safe innovation rather than creating new liability risks.