The integration of Physical AI, Internet of Things (IoT) devices, and edge platforms is a key direction for creating systems capable of real-time interaction with the physical world. This year and in the coming ones, we are observing these technologies converge, shaping new approaches to automation and operational process optimization, particularly in the manufacturing sector. However, the successful implementation of such systems requires not only technological readiness but also careful data and risk management during the preparation phase.
Data as the Foundation for Physical AI: Why Fragmentation Hinders Innovation
Data is the bedrock of any AI system. For Physical AI, which interacts with the physical world through sensors and actuators, data quality, completeness, and accessibility become paramount. In practice, many manufacturing enterprises face issues with data fragmentation, inconsistency, and gaps. Information about equipment operation, production processes, product quality, and logistics is often scattered across different systems, stored in incompatible formats, or missing entirely. This prevents the effective use of data for training AI models, thereby slowing down the adoption of predictive analytics, automated quality control, and production optimization.
Without a unified, clean, and structured database, Physical AI remains merely a concept. Models trained on poor-quality data will yield inaccurate predictions, leading to erroneous decisions and potential disruptions in physical processes. This is particularly relevant given that by 2028, over half of enterprise GenAI models will be domain-specific, requiring high-quality, specialized data for specific industries and tasks, according to Gartner’s ‘Top Strategic Technology Trends for 2026’ report.
Manufacturing: Where OT and IT Meet, But Data Doesn’t Speak the Same Language
Manufacturing traditionally comprises two worlds: Operational Technology (OT) and Information Technology (IT). OT systems, such as SCADA (Supervisory Control and Data Acquisition), PLCs (Programmable Logic Controllers), and MES (Manufacturing Execution Systems), are responsible for direct control of physical processes, collecting data from sensors, and monitoring equipment. IT systems, including ERP (Enterprise Resource Planning), CRM, and other business applications, manage planning, finance, logistics, and customer interactions.
The challenge lies in the fact that these systems have historically evolved independently, using different protocols, data formats, and architectures. Sensor data from SCADA might be aggregated differently than production batch data in MES, and order information in ERP may not have a direct link to the actual status of the production line. This disconnect creates significant obstacles to building a holistic view of operational activities and leveraging AI for end-to-end optimization.
A Common Mistake: Trying to Clean Everything at Once Instead of Prioritizing
When faced with the problem of poor-quality data, many enterprises attempt to simultaneously clean and integrate all available data from all systems. This approach is overly ambitious, resource-intensive, and often leads to failure due to complexity, duration, and high cost. A more effective strategy is an iterative approach based on prioritization. It is necessary to first identify the most critical business tasks that can be addressed with Physical AI (e.g., predictive maintenance, energy consumption optimization, quality control) and then focus on preparing the data specifically required for these tasks. This allows for quick wins, demonstrates the value of AI, and enables gradual expansion of its application scope.
Architectural Example: Integrating SCADA, MES, and ERP for Predictive Maintenance
Consider a typical scenario for predictive equipment maintenance in manufacturing. For a Physical AI model that forecasts failures to function effectively, data from various sources is required:
- SCADA/PLC: Real-time sensor data (temperature, vibration, pressure, current) reflecting the current equipment status.
- MES: History of production batches, equipment operating modes, production output, downtime information, and previous repair records.
- ERP: Data on maintenance orders, spare parts availability, scheduled maintenance timelines, and maintenance costs.
Integrating these systems allows for the creation of a unified data stream, which is fed to an edge platform for initial processing and then to the AI model. The model analyzes a combination of parameters, detects anomalies, and predicts potential failures. For instance, if engine vibration (SCADA) begins to increase during the production of a critical batch (MES), and repair parts are available in stock (ERP), the system can automatically generate a recommendation for scheduled maintenance before a breakdown occurs. This minimizes downtime and optimizes costs.
Edge AI and IoT: How Sensors, Robots, and Platforms Create an Intelligent Environment
At the core of Physical AI lies the synergy between IoT devices, edge computing, and AI models. IoT sensors collect raw data from the physical world: temperature, humidity, vibration, pressure, images, sound. This data is often large in volume and requires low-latency processing. This is where edge platforms come into play.
Edge computing enables data processing closer to the source, on the device itself or a local gateway, rather than sending it to a centralized cloud. This is critical for scenarios where latency matters, such as controlling robotic arms, autonomous vehicles, or security systems. For monitoring equipment on a factory floor or managing autonomous robots moving through a production hall, the use of 5G networks can provide the necessary bandwidth and low latency. According to the Ericsson Mobility Report November 2025, 5G will become the dominant mobile access technology by subscriptions by the end of 2027, highlighting the growing role of these networks in supporting complex AI and IoT systems.
Lightweight AI models can run on edge platforms for initial data filtering, aggregation, and analysis. This reduces the amount of information transmitted to the cloud, lowers network load, and enhances privacy. More complex models can run in the cloud, but edge platforms ensure immediate responses to local events. For example, the AZIOT Platform (an IoT platform for physical environment management from AZIOT) allows for the management of IoT devices and data processing at the edge, ensuring integration between the physical world and intelligent systems.
Risk Management for Physical AI: From NIST to ISA/IEC 62443
The implementation of Physical AI brings not only benefits but also significant risks related to security, privacy, and ethics. As AI systems directly interact with physical equipment, errors or malicious attacks can lead to severe consequences, from production stoppages to physical damage.
Recognized standards exist for managing these risks. NIST AI RMF 1.0 (Artificial Intelligence Risk Management Framework) structures AI risk management around four functions: Govern, Map, Measure, and Manage. This framework helps organizations identify, assess, and mitigate risks associated with the development and use of AI systems.
Beyond general AI risks, cybersecurity is critical for industrial systems. The ISA/IEC 62443 series of standards (ISA/IEC 62443 Series of Standards) outlines security requirements for industrial automation and control systems. It covers aspects from secure system design to patch management and incident response. Integrating Physical AI into production processes requires adherence to these standards to protect against cyberattacks, ensure data integrity, and guarantee reliable equipment operation.
Risk management for Physical AI is a continuous process that demands constant monitoring, model updates, and adaptation to new threats. It is not just a technical task but also a matter of organizational culture and responsibility.
Data Readiness Checklist for Physical AI
- Have priority business tasks for AI analytics (e.g., predictive maintenance, quality control) been defined before commencing data work?
- Has an assessment of data quality from key OT/IT systems (SCADA, MES, ERP) been conducted for completeness, consistency, and relevance?
- Do unified master data management (MDM) repositories exist and are they applied for key entities (equipment, materials, production batches)?
- Has a mechanism for real-time data quality monitoring been implemented to detect anomalies and drift?
- Have AI-specific metrics for data quality assessment been defined (e.g., time-series completeness for model training)?
- Does a data governance policy exist and is it enforced, covering the data lifecycle for AI systems?
- Has a risk assessment related to the use of AI on physical equipment been conducted in accordance with the NIST AI RMF 1.0 framework?
Successful implementation of Physical AI and IoT in manufacturing depends on data readiness and a systematic approach to risk management. Only with high-quality, integrated data and clear security policies can the potential of systems that react to the physical world in real-time be fully unlocked.