Physical AI: Decide at the Edge, Learn in the Cloud, Grow Autonomy in Stages

On 12 November 2026, IIoT World published a summary of a panel at the Industrial AI Summit 2026 on how to implement, govern and secure so-called physical AI. The speakers represented Galyos, AWS, InfluxData and Grayhawk Consulting Group. One caveat: the piece is marked as sponsored by InfluxData and discloses that AI tools were used for the summary. It should therefore be read as vendor-influenced opinion, not independent research.

Where to run the AI

The rule the panel proposes can be summed up in its own formula: AI "acts locally and learns globally". The allocation criterion is as follows.

• Edge: safety-critical decisions, quality defect detection and anything that requires sub-second responses (roughly under 200 ms).
• Cloud: cross-plant pattern recognition, training models on aggregated failure data, multi-site energy analysis.
• Hybrid layer: an agentic workflow detects an anomaly at the edge, queries maintenance records, manuals and spare-parts inventory in the cloud, creates the work order and sends the action back to the edge.

Siemens Energy is cited as an example: condition monitoring and energy optimization run locally in each plant with AWS IoT SiteWise Edge, which brings 10 protocol types into the edge layer, including Modbus, OPC UA and the legacy Siemens S7. Failure data from multiple assets is then aggregated to train predictive maintenance models. According to the article, a 200 ms delay can be the difference between a safe shutdown and an accident, a defect or a line stoppage.

Cost and resilience

The panel argues that keeping raw telemetry locally for 5 days to a month, and sending only the relevant subsets to the cloud, can cut storage costs to about a tenth of full ingestion. This is the panel's claim and has not been independently verified. The resilience principle is more solid: edge devices must keep working when the network goes down.

Reliable data before autonomy

Before any autonomous action, the article lists four checks: fresh, trustworthy data; operating conditions within the model's training range; a defined perimeter of authorized actions; and a planned path for exceptions.

For data trust, it points to a unified namespace with a tag ontology, useful for catching naming drift, clock skew and schema changes (for example, an integer that becomes a decimal). Back-testing new sources against historical patterns is a second check. Silent schema drift, the piece warns, degrades models without anyone noticing.

Four stages of autonomy

• Observation.
• Recommendation with human approval.
• Pattern-based action, with proven and reversible outcomes.
• Full autonomy, earned only after consistent performance at each previous stage.

Every decision must be explainable and traceable, with escalation to a person beyond a risk threshold. A model trained at one site should not be applied to another with different equipment or processes.

The security issue

A robot that talks to both a WMS and a MES crosses boundaries that the ISA-95 model meant to keep separate. A compromised edge model could also influence work orders and maintenance across multiple sites, which is why, according to the article, IT and OT security disciplines must converge. There are physical limits too: edge hardware has limited memory and heat dissipation, and in summer inference time increases by about 30 ms because of thermal throttling.

What to do with it in practice

The article does not address the Italian context or specific sectors, but the method is transferable. Start with narrowly scoped projects with clear scope, success criteria and authorized actions, then reuse the architecture, governance and security controls at other plants.

Machine connectivity and clean, contextualized data (MES, OPC UA, unified namespace) come before any autonomy.

For plant managers and IT/OT leads, the takeaway is a checklist for a first pilot: decide what requires local latency, ensure data quality, set the starting autonomy stage, and define escalation thresholds and security controls up front.