Agentic AI in OT: New Capabilities, Emerging Risks, and Governance Challenges

As industrial systems become more interconnected, organizations are discovering that visibility alone is inadequate. Dashboards can show what is happening, but without context and coordination, they rarely drive effective action. Advanced software platforms can move beyond passive monitoring toward agentic control, where systems interpret conditions, apply rules, and even initiate corrective actions autonomously within complex industrial environments. Reaching this level of maturity will rely on information technology/operational technology (IT/OT) convergence and open-architecture software capable of closing the “context gap” that often renders raw data operationally useless.

The Limits of Dashboard Visibility

Industrial systems generate continuous data streams that teams often struggle to interpret meaningfully in real time. A context gap prevents humans from turning raw, fragmented data into actionable insights. For example, ambient temperature fluctuations in a facility are typically managed as a comfort issue, isolated from the process control layer. Yet those same temperature shifts can introduce measurement inaccuracies that affect product quality, with no mechanism in place to surface that relationship to the operators monitoring output.

The correlation problem contributes directly to the human context gap. Humans don’t easily correlate data streams from different systems and areas within a plant. Missing these correlations means failures remain unseen or unexplained for too long. A welder change on a production line could introduce electrical noise that causes a separate machine to fault intermittently. An operator examining the machine may eventually correlate the two events, but it would likely take considerable time and manual effort. The underlying cause was always present in the data; identifying it was the obstacle. As IT/OT integration pushes more operational data into shared environments, the volume of uncorrelated signals grows. Artificial intelligence (AI) systems also suffer from a context gap when they lack sufficient historical data to guide operations. Without better tools, these context gaps widen, leaving operations data-rich but insight-poor.

The IT/OT Infrastructure Agentic Software Requires

Running agentic software requires an open infrastructure that delivers continuous, interoperable access to data from across the entire operational environment. Open architecture is foundational to that capability. When industrial software platforms restrict data access through proprietary formats or closed application programming interfaces (APIs), they limit what any analytical or agentic system can see and act on. Platforms built on open, industry-standard protocols enable data to move more freely among equipment, control systems, and higher-level analytics layers. This interoperability is a prerequisite for the cross-system correlation that makes agentic AI viable in industrial settings.

The digital thread generated extends that foundation across time. By maintaining a continuous historical record of data as a product or process moves through its lifecycle, the digital thread provides agentic systems with the contextual depth needed to identify patterns that point-in-time dashboards cannot surface. Early-stage process variations that predict downstream quality failures, for example, only become visible when data from across the production lifecycle can be queried together. For AI agents, that accumulated, curated history becomes valuable training material, provided organizations carefully vet the data they feed into models, rather than indiscriminately dumping every signal into them. This curated approach closes the context gap for the AI agents, which need access to the right data and sufficient exposure over time to learn the same kinds of relationships that experienced operators carry as tribal knowledge.

As IT/OT convergence extends connectivity deeper into the control layer, those gaps become new attack surfaces. Legacy OT hardware was designed for isolated environments, with limited authentication, no encryption, and communication protocols that were never intended to defend against external threats. Mitigation typically involves a layered approach, mapping all active ports and protocols, segmenting networks into defined security zones, and routing legacy device communication through hardened gateways rather than exposing the devices directly.

Beyond connectivity and security lies the question of where to run an agent. Deploying agents close to the operational process reduces latency, which is crucial for time-sensitive use cases such as image-based quality inspection, whereas historical analysis and pattern correlation can tolerate higher latency. Local hosting can increase cost and operational complexity, but it keeps sensitive operational data within the plant’s security boundary. Cloud environments offer elastic compute that is attractive for training and experimentation, yet raise additional security and compliance concerns when production data leaves the organization’s control.

The Agentic Controls Maturity Model

The path from passive dashboard to autonomous control is a slow progression, and most industrial deployments are still in the earliest phase.

Current State: Monitoring and Notification

Today’s agentic deployments represent early investigations into what is possible, even testing the architecture where agents are hosted and how they interact with local controllers. In practice, this means agents continuously monitor conditions across systems, analyzing performance against thresholds and expected patterns, and flagging anomalies for human review. A typical early-phase use case could be an agent tracking equipment performance, noticing that failure rates correlate with operational cycles rather than elapsed time, and recommending shifting from a time-based to a cycle-based preventive maintenance schedule. An operator reviews the pattern evidence highlighted by the agent and makes a decision. That division of responsibility is deliberate. Trust in agentic performance will be develop incrementally before organizations test the feasibility of autonomous action.

The Trajectory: Recommendation and Autonomy

The next phase includes building systems, such as virtual commissioning and digital twins, to serve as an active testing environment. Using these environments, humans can test agent recommendations as a pathway to building trust and allowing incremental agent autonomy. These environments also allow agents to autonomously refine possible control logic and generate a log and report of their decision-making process, before recommending a specific resolution. Human engineers can review the logs before introducing anything into the real-world system. Initial deployments will target narrow, well-defined processes where models are reliable, and the consequences of error are limited.

Full autonomy, in which an agent updates the control logic in real time without human intervention, remains the longer-term trajectory. Even once truly autonomous agents are deployed, human oversight is expected to persist in high-risk environments regardless of how well the agent’s decision-making is validated. Mathematical models that perform reliably within normal operating ranges can break down at physical extremes. In hazardous environments, the consequences of an edge-case failure are too significant to absorb, making the level of human oversight a risk-based decision rather than a purely technical one.

Governing Agentic Control Before It’s Fully Visible

The shift toward agentic control in industrial environments is underway. It may not happen from the top down, but it is already occurring on the floor. Engineers and operators are independently experimenting with AI tools, often without clear organizational frameworks to guide how those tools interact with critical systems. The risks are more contained when agents are in the early phases, with humans making the decisions and taking action. Yet the trajectory points toward agentic control systems that recommend, simulate, and eventually act.

The infrastructure being built today, driving IT/OT convergence, is the same infrastructure that will carry these agents’ autonomous decisions into the OT control layer. It’s imperative for organizations to answer key questions before technological capability outpaces governance, including: What defines the boundary between agent recommendation and autonomous action, and who owns it? As IT/OT convergence extends connectivity deeper into legacy systems, how are the emerging security gaps being addressed? Technology is maturing faster than the governance frameworks designed to manage it. The time to build those frameworks is now.

About the Authors

Roy Krans is a Software Development Manager and Dan Idzikowski is an Instrumentation & Controls Engineer for ACS. ACS engineers, integrates, and builds technically complex equipment, controls, and facilities for industry-leading companies in markets including automotive, aerospace, energy, chemical, manufacturing, and more. ACS specializes in control systems, custom machines, testing solutions, automation, and production systems, as well as the design and construction of integrated facilities. For more information, please call (608) 663-1590 or visit http://www.acscm.com.

Disclaimer: The authors are completely responsible for the content of this article. The opinions expressed are their own and do not represent IEEE’s position nor that of the Computer Society nor its Leadership.