By Teodor Vasile-Grünberg (ERNI Germany)
Physical AI works convincingly in the lab. Getting it to run reliably on a real production line, within tightening EU regulation and under budget pressure, is a different challenge. We look at what separates successful deployments from stalled pilots, and why the model is rarely the deciding factor.
Seven crucial factors for moving physical AI from pilot to production
Physical AI, meaning AI systems that perceive, decide and act in physical environments, has passed its demo phase. Robots that adapt to changing situations, cameras that make quality decisions and machines that react to their surroundings now work convincingly in labs and trade show halls. The crucial question is no longer whether physical AI can work, but whether it can run reliably in a real factory every day, comply with regulation and still pay for itself.
The adoption gap is visible in the figures. According to Bitkom, 90 percent of companies describe AI as decisive for competitiveness. Yet only six percent of German industrial companies already use physical AI, while almost a third are planning concrete applications. That puts most organisations somewhere between ambition and production. Our project experience suggests that this gap is caused less by model performance than by three production readiness issues: safety, compliance and scalability.
In production, the model is rarely the reason physical AI fails. It fails because of missing system validation, regulatory architecture and scalable integration. Companies that secure these three dimensions early shorten the path from pilot to production. Safety and compliance are entry conditions; economy is what turns a working pilot into a scalable one. The seven factors below are grouped accordingly.
Safety: The system has to prove itself
The management question here is simple: can you prove safe behaviour after a sensor or component change? The two factors below set out what that takes.
1. Validate the whole system, not just the model
A physical AI system consists of more than a neural network. It includes sensors, lighting, optics, actuators, firmware, pre-processing and the edge hardware the model runs on. Change any one of these, and behaviour can change, even if the model stays the same.
This is why the robotics industry is rethinking validation. In a 2026 report, the IFR notes that AI-driven autonomy fundamentally changes the safety landscape. Testing, validation and human oversight become more complex, but also more necessary. Teams that treat the full perception and action chain as the unit of validation avoid unpleasant surprises during commissioning.
2. Design human oversight and safe fallback behaviour from the start
Every physical AI system will meet situations it was not trained for. The decisive design question is what happens afterwards. Clear confidence thresholds, defined procedures for escalating to an operator and a safe fallback state belong in the architecture, not in a later safety review. A system that knows when to stop is more valuable in production than one that is slightly more accurate but unpredictable at the edges.
Compliance: Regulation is an architecture decision
The management question here is just as direct: can you produce evidence for every model, dataset and update in use, on demand? The two factors below turn compliance from paperwork into architecture.
3. Treat the regulatory stack as a design input
Regulation is no longer a legal afterthought. It determines architecture. For companies bringing AI-enabled machines to the European market, several regulatory obligations are converging between 2026 and 2028. The Cyber Resilience Act reporting obligations apply from 11 September 2026, the Data Act’s data access by design from 12 September 2026, the new Product Liability Directive must be transposed by 9 December 2026, and the Machinery Regulation applies from 20 January 2027. The Machinery Regulation in particular raises requirements for software-based controls and safety-relevant AI functions. The AI Act’s high-risk obligations for AI embedded in regulated products will apply from 2 August 2028.
Each of these rules affects the architecture: how data is stored and shared, how updates are delivered, how decisions are logged. Addressing them retrospectively is far more expensive than designing for them from the outset.
4. Build traceability into the delivery pipeline
Compliance evidence is best produced as a by-product of good engineering, not assembled before an audit. This means versioned models and datasets, requirements traceable to tests, a software bill of materials, vulnerability management and a controlled update process. Companies with experience in MedTech or pharma know this discipline well. For many machine builders, it is new, and it is quickly becoming a market requirement.
Economy: Value comes not from pilots but from scale
The management question here is: can you scale this use case without rebuilding the data foundation? The three factors below cover what production-grade scaling requires, including the operating costs that ‘economically’ have to cover: lifecycle cost, maintenance, retraining, downtime, edge hardware refresh and who owns the system once it runs.
5. Fix the data foundation before adding autonomy
Analysing Hannover Messe 2026, IoT Analytics identified a central tension. Vendors are monetising autonomous AI decisions on the shop floor before many manufacturers have the trusted, contextualised data foundation needed to execute those decisions safely. Autonomy built on inconsistent sensor data or unclear process context does not scale. It generates exceptions. Investing in clean data flows and context first is usually the faster route to return.
6. Plan for integration into existing environments
Few factories are greenfield. Physical AI has to work alongside existing PLCs, MES, LIMS, robot controllers and safety systems. Bitkom names integration into existing processes and ensuring operational reliability as the central tasks for broader rollout. The report also ties scaling to standardised data interfaces and clear governance models. In practice, integration effort often exceeds model development, and should be budgeted accordingly.
7. Start bounded, test early and measure hard
Budgets are tight. VDMA expects revenue in the German robotics and automation sector to fall by 5 percent to EUR 14.1 billion in 2026. Under this pressure, convincing business cases come from clearly bounded use cases with a measurable effect, such as fewer rejects, less waste or shorter set-up times.
Simulation and realistic test environments reduce risk before anything touches a production line. Major robot manufacturers such as ABB, FANUC, KUKA and YASKAWA are increasingly using simulation and digital twin environments, including NVIDIA Omniverse and Isaac-based workflows, to reduce commissioning risk. At ERNI, we follow the same principle in our Physical AI Labs in Germany and Spain, where camera, model and robot are tested together under real conditions before a solution moves to the customer’s site. The model matters here, but it is only one part of the production system: sensors, integration and process context decide as much as the algorithm does.
From impressive to dependable
What these seven factors have in common is that none of them is primarily about AI. They are about systems engineering, regulatory discipline and operational realism. That is good news for European industry, which has deep expertise in all three.
A pilot proves that physical AI can work. Production proves that it works reliably, every day, under real regulatory and commercial constraints. Closing that gap is what the seven factors above are designed to do.
A practical first step is a production readiness assessment: validate the use case, map the regulatory obligations, inspect the data and integration landscape, and define a bounded prototype with measurable ROI. ERNI supports this process through Physical AI Labs in Germany and Spain, combining feasibility assessment, compliance-oriented architecture and realistic testing before deployment on the customer site.
