As the industrial sector accelerates its transition toward autonomous operations, the promise of "Data Streaming to Autonomous Action" has moved from a theoretical aspiration to an immediate, complex reality. Following the Industrial AI Summit 2026, where industry leaders gathered to dissect the future of factory floors, a series of critical questions remained regarding the practical, ethical, and technical implementation of AI agents.
This article, the final installment in a three-part series, addresses the "what happens next" phase: the point at which AI transitions from a passive analytical tool to an active participant in plant operations. Peter Sorowka, CEO of Cybus, and David Puron, CEO of Barbara, provide a roadmap for navigating the volatile intersection of legacy Operational Technology (OT) and cutting-edge Artificial Intelligence.
1. Architectural Strategy: The "Waterfall" vs. The "Use Case"
A common point of contention for CIOs and plant managers is whether to adopt a top-down, waterfall-style digital transformation or a bottom-up, agile approach.
Peter Sorowka argues for a hybrid model that balances velocity with stability. "Both are necessary, but at different speeds," he explains. The strategy should involve running short-cycle use cases—validating them within two to four weeks using real production data—while simultaneously treating the "data foundation" as permanent, evolving infrastructure rather than a project-based deliverable.
The litmus test for this approach is the marginal cost of expansion. If each new use case requires building a separate, siloed pipeline, the foundation is failing. In a robust architecture, each new application should integrate into a unified namespace, driving the marginal cost of additional use cases toward zero.
2. The Economics of Intelligence: Managing Tokenization Costs
As organizations shift from on-premise legacy systems to AI-augmented workflows, the "consumption-based" pricing models of cloud-native LLMs pose a significant barrier to adoption.
Sorowka outlines three levers to maintain cost control:
- Retrieval Optimization: Agents should be trained to query specific data points rather than scanning an entire namespace. Reducing a data set from 550 points to 20 can shift the financial burden by an order of magnitude.
- Model Tiering: Avoid using large, expensive language models for simple tasks. Use small, specialized models for classification and anomaly detection, reserving "expensive" large models exclusively for synthesis and high-level reasoning.
- Policy-Based Guardrails: Implement rate limits and scoping at the source. By bounding costs through policy, an organization prevents "runaway prompts" from inflating the operational budget.
Furthermore, Sorowka notes that production budgets in manufacturing are traditionally built on depreciation cycles. Consequently, the unpredictable nature of cloud consumption pricing remains a primary inhibitor for steady-state industrial loops. The strategic recommendation is to prioritize on-premise infrastructure for critical, constant loops and leverage the cloud for exploratory, intermittent tasks.
3. The "Invasive AI" Paradox and Human Trust
The deployment of AI agents often triggers a form of "Invasive AI" anxiety among plant operators. This manifests in two ways: technical interference and human resistance.
From a technical perspective, an autonomous agent might inadvertently stress an OT network by pulling excessive data, either through a coding bug or a security compromise. This is mitigated through rigid read-scopes and "default-deny" network policies.
More importantly, however, is the psychological dimension. "If operators believe the system is measuring them for surveillance, they will stop feeding it," warns Sorowka. The success of an AI foundation relies on the data provided by the humans on the floor. Transparency is the antidote to this friction: by clearly defining what is being collected—and, crucially, what is not—leadership can maintain the trust necessary to keep the data foundation accurate and reliable.
4. Sovereignty: The Bedrock of Industrial Resilience
Sovereignty in the industrial context is often mistaken for a mere geopolitical buzzword. However, for plant operators, it is a matter of operational continuity.
David Puron of Barbara defines sovereignty through the lens of vendor independence: "If your vendor, your connectivity, or your cloud contract changes tomorrow, does the process keep running?"
Sovereignty requires three pillars of control:
- Data Layer Independence: The data architecture must remain vendor-neutral and reside on-premises to prevent external lock-in.
- Model Interchangeability: Organizations must avoid dependency on a single AI provider, ensuring that models can be swapped without re-engineering the entire pipeline.
- Decoupling from External Volatility: A cloud outage or a change in enterprise software licensing should never reach the shop floor.
As regulatory frameworks like the EU’s NIS2 directive become mandatory, sovereignty is no longer just a strategic preference; it is a compliance requirement.
5. The Human-Control Boundary: Where Autonomy Ends
Defining the limits of AI agency is the most significant hurdle in transitioning to autonomous manufacturing. Sorowka proposes a three-tier model for interaction:
- Observe: The AI accesses data within a strict read-scope. This is generally "free" and low-risk.
- Provision: The AI suggests changes to configuration. This requires a "four-eyes" protocol—a technical enforcement where a human must review and approve the change before it enters production.
- Actuate: The AI writes directly to a machine. This is strictly gated by pre-defined interlocks. If a command violates these hard-coded physical limits, the system rejects it automatically.
Crucially, human oversight should be positioned at the "edge of the mandate" rather than forcing a human to approve every trivial, repetitive action. Auditability is the cornerstone here: every value must be traceable to its machine origin, with logs documenting every read, write, and blocked attempt.
6. Cybersecurity in the Era of Machine-Speed Attacks
The integration of AI into industrial environments changes the threat profile of the manufacturing sector. As noted by David Puron, the fundamentals of cybersecurity have not changed, but the lifecycle of an attack has been compressed from days to mere minutes.
The Essential Security Stack
Sorowka provides a baseline of non-negotiable security controls for AI in manufacturing:
- Identity and Access: Each agent must possess a unique, short-lived identity. No shared passwords in configuration files.
- Least Privilege: Read-only access by default, scoped strictly to the asset level.
- Network Integrity: Use of mTLS (mutual TLS) on every connection and mandatory schema validation to ensure "garbage data" cannot be injected into the namespace.
- Immutable Auditing: A tamper-proof log that records not just what the AI did, but what it tried to do and was blocked from doing.
The Human-Gated Defense
David Puron emphasizes that while AI provides the speed necessary for modern defense, it cannot replace the human element in critical decision-making. "I call it ‘human-gated decision-making’ for AI-enabled tools," Puron states.
In the event of a security alarm, the AI can detect a threat in seconds, but the decision to halt a multimillion-dollar production line must remain with a human. Furthermore, the supply chain is the final, critical vulnerability. Autonomous systems are only as secure as the software dependencies beneath them. Organizations must prioritize Supply Chain Security and Software Bill of Materials (SBOM) management to ensure that the AI foundation remains resilient against evolving threats.
Conclusion: The Path Forward
The journey from data streaming to autonomous action is not merely a technical migration; it is a fundamental shift in how manufacturing organizations manage risk, talent, and control.
By grounding AI in sovereign architecture, enforcing rigorous human-in-the-loop protocols, and treating cybersecurity as a machine-speed necessity, manufacturers can harness the power of AI without sacrificing the stability of the plant floor. As the industry moves toward 2026 and beyond, the winners will be those who successfully balance the breakneck speed of AI innovation with the unwavering, traditional requirements of industrial reliability and safety.
This article is based on the session "From Data Streaming to Autonomous Action" at the Industrial AI Summit 2026. Insights were provided by Peter Sorowka (Cybus) and David Puron (Barbara). The session was supported by HiveMQ, Albemarle, and Hamish Mackenzie.
