As the dust settles on the Industrial AI Summit 2026, the industry is left with a singular, pressing question: How do we move beyond experimental data streaming into the realm of truly autonomous, reliable action? Following two previous sessions that established the foundational architecture and context for Industrial AI, the concluding installment of this series tackles the "hard" questions—the operational realities of putting agents in the driver’s seat.
In this final deep dive, we explore insights from Peter Sorowka, CEO of Cybus, and David Puron, CEO of Barbara, as they address the critical tensions between innovation, cybersecurity, sovereignty, and human control.
The Strategic Framework: Foundational Philosophies
The transition to autonomous industrial action is not merely a software upgrade; it is a shift in the philosophy of the factory floor. The discourse at the Summit centered on a central conflict: the tension between agile, use-case-driven development and the necessity of a robust, long-term technical infrastructure.
The Hybrid Approach: Use Cases vs. Waterfall
One of the most persistent questions facing CTOs is whether to adopt a "waterfall" plan for AI transformation or to iterate via individual use cases. Peter Sorowka argues that successful industrial organizations must embrace both, but at different speeds.
"Use cases are for short-term validation," Sorowka explains. "Pick one, validate it against real production data within two to four weeks, and be prepared to kill it if the ROI isn’t there." However, he cautions that this must not be confused with infrastructure building. The foundation should be treated as a permanent asset with a lifecycle, not a one-off project deliverable. The litmus test for success is scalability: if each new use case requires a parallel pipeline, the foundation is failing. A mature architecture should allow new use cases to plug into an existing namespace, driving the marginal cost of subsequent implementations toward zero.
The Economics of Autonomy: Tokenization and Cloud Costs
As AI agents begin to query data at scale, tokenization costs can quickly spiral. For organizations accustomed to the predictable depreciation cycles of traditional CAPEX-heavy industrial equipment, consumption-based cloud pricing represents a significant barrier to entry.
Sorowka suggests three levers for managing these costs:
- Refined Retrieval: Move away from scanning broad namespaces. By narrowing agent queries to specific, high-value data points (e.g., 20 points instead of 550), companies can reduce token costs by an order of magnitude.
- Right-Sizing Models: Not every task requires a Large Language Model (LLM). Small, purpose-built models are more than sufficient for classification and anomaly detection, with massive synthesis models reserved for language or complex strategic reasoning.
- Policy-Based Guardrails: Implement rate limits and scoping at the source. By bounding costs through policy rather than relying on the unpredictability of user prompts, organizations can ensure that their AI budget remains within predefined limits.
Sovereignty: Maintaining Control in a Connected World
Sovereignty in the industrial context is often framed through a geopolitical or regulatory lens, but its practical application is far more granular. For plant managers, sovereignty is the answer to the question: “If my cloud contract, connectivity, or vendor relationship changes tomorrow, does the plant keep running?”
The Three Pillars of Sovereignty
According to David Puron, sovereignty is about ensuring that the core production process remains insulated from external shocks.
- Architectural Independence: The data layer should reside on-premises and remain vendor-neutral.
- Model Interchangeability: Organizations must avoid vendor lock-in by ensuring that AI models can be swapped out without re-engineering the entire pipeline.
- Data Flow Control: No single external provider should act as the gatekeeper of shop-floor data.
This is becoming a matter of compliance, particularly in the European Union, where the NIS2 directive is forcing industrial firms to take a hard look at their internal security architectures and dependencies.
The Human-Control Boundary: When to Intervene
A common fear regarding autonomous AI is the loss of human agency. To mitigate this, practitioners advocate for a tiered "impact level" system:
- Observe: Agents have read-only access within a defined scope. This is generally free from oversight.
- Provision: When an agent attempts to change a system configuration, it must be declared and reviewed. Crucially, the "four-eyes" principle—requiring two human approvals—should be technically enforced before any configuration change reaches production.
- Actuate: For direct machine intervention, the system must run against hard-coded interlocks and safety limits. If an action violates these boundaries, the system must trigger an automatic "refuse" command.
"Human-in-the-loop belongs at the edge of the mandate," says Sorowka. "We do not want a human reviewing every single action, but we do want humans to set the boundaries and review the exceptions."
Cybersecurity in the Era of Machine Speed
Perhaps the most critical takeaway from the Summit was the realization that while AI hasn’t changed the fundamentals of cybersecurity, it has drastically compressed the lifecycle of an attack.
The Attack Velocity Problem
David Puron, referencing data from the European Union Agency for Cybersecurity (ENISA), noted that in the past, a zero-day vulnerability might provide a window of days for a security team to respond. In the AI era, that window has shrunk to minutes.
To counter this, industrial firms need machine-speed detection and response. However, Puron warns that automation should not mean removing the human from the final decision-making process. "For critical decisions—like shutting down a production line due to a suspected breach—we still need human-gated decision-making," he asserts.
Essential Security Controls
To operate autonomously, industrial systems must adhere to a "minimum viable security" set. Peter Sorowka outlined the non-negotiable requirements for any autonomous industrial agent:
- Identity Management: Each agent must have a distinct, non-shared identity with credentials that expire by default.
- Least Privilege: Access must be restricted to specific asset subtrees, with read-only as the default state.
- Schema Validation: Invalid or malformed data should be blocked at the point of publication to prevent the injection of corrupted data into the namespace.
- Immutable Audit Logs: A comprehensive, version-controlled record of every action, including blocked attempts, is essential for forensic accountability.
- Supply Chain Assurance: Because systems are only as secure as their dependencies, the use of Software Bill of Materials (SBOMs) is now a requirement for any enterprise-grade deployment.
Implications: The Path Forward
The transition from data streaming to autonomous action represents a paradigm shift in how we manage industrial output. The consensus among the experts at the Industrial AI Summit 2026 is that the technology is ready, but the governance must be matured.
The New Industrial Mandate
The path forward is defined by three converging trends:
- Regulatory Compliance: Frameworks like IEC 62443 and the NIS2 directive are no longer optional "nice-to-haves." They are the baseline for operational resilience.
- Human-Centric Design: If operators feel that AI is a tool for surveillance or a replacement for their expertise, they will disengage. The data foundation will only thrive if it is seen as a tool to empower, not replace, the workforce.
- Resilience through Modularity: By keeping the data layer independent and the models interchangeable, manufacturers can capture the benefits of AI without sacrificing the sovereignty of their production processes.
As the industry continues to integrate AI into the core of manufacturing, the focus must remain on creating "defensible" autonomy—systems that are as secure and predictable as the mechanical processes they are designed to optimize.
Acknowledgments
This article summarizes the final session of the Industrial AI Summit 2026. The discussion featured insights from Peter Sorowka (Cybus) and David Puron (Barbara). Additional contributors to the broader session included Kudzai Manditereza (HiveMQ), Jonathan Alexander (Albemarle), and Hamish Mackenzie. This session was proudly sponsored by Cybus, Barbara, and HiveMQ.
