Why This Topic Lacks Reliable Statistics
Manufacturing organizations have adopted AI agents for tasks ranging from production scheduling assistance to quality data analysis, but no independent, methodologically transparent survey currently quantifies how many of these deployments have direct write or control access to operational technology (OT) systems. Vendor-commissioned surveys and marketing materials frequently cite adoption percentages or incident counts without disclosing sample size, methodology, or verification, which makes them unsuitable as a basis for governance decisions. This is a data availability gap, not a definitive marker of a small or large-scale problem. Sound governance planning should be built on documented structural gaps rather than unverifiable figures.
The Framework Gap Between OT Security and Agentic AI
IEC 62443 defines security requirements for industrial automation and control systems, including network segmentation, zone and conduit models, and role-based access. The NIST AI RMF provides a general risk management structure for AI systems, covering governance, mapping, measuring, and managing risk. Neither framework was written with autonomous agent tool-call behavior in mind. IEC 62443 assumes human operators or deterministic software interacting with defined interfaces; it does not address an agent dynamically selecting which system to query or which action to take. The NIST AI RMF addresses AI risk broadly but does not specify manufacturing OT controls, such as how to scope an agent's access to a programmable logic controller or manufacturing execution system. This leaves a real gap between two mature but non-overlapping frameworks, which manufacturers must fill with agent-specific controls.
What the Gap Means in Practice
In practice, this means that a manufacturer following both IEC 62443 and the NIST AI RMF to the letter can still have no formal control over what an AI agent is permitted to do once it has network access to an OT-adjacent system. Compliance with existing frameworks does not equate to agent governance.
Why Tool-Call Access Is a Distinct Risk Category
Traditional ICS access control is built around static roles: a given user or service account is granted a fixed set of permissions, typically reviewed periodically. AI agents behave differently. An agent's specific tool calls in a given session are determined by a model's interpretation of a prompt and available context, which means the exact sequence of system interactions is not fully predictable in advance. This does not mean agents are inherently unsafe, but it does mean that permission models designed for static roles do not naturally extend to dynamic, per-call decision-making. Distinguishing between read-only telemetry access and write or control actions on production and safety systems becomes essential, since the consequences of an unauthorized read differ substantially from those of an unauthorized write.
Operational and Implementation Considerations
Manufacturers evaluating agent governance should treat it as an extension of existing OT security practice rather than a parallel program. Practical considerations include:
- Maintaining a single source of truth for network segmentation, rather than layering a separate access model on top of IEC 62443 zones and conduits.
- Requiring audit logs for agent tool calls that meet the same evidentiary standard used for human operator actions in regulated environments.
- Scoping agent permissions at the level of individual actions, not just system-level access, so read and write operations can be governed independently.
- Requiring runtime enforcement, meaning anomalous behavior is blocked as it occurs, rather than relying solely on retrospective log review.
- Evaluating vendor claims against independently verifiable deployment evidence rather than adoption or incident statistics that cannot be traced to a disclosed methodology.
How Trussed AI Fits This Problem Space
Trussed AI positions its runtime governance layer as a complement to existing OT security investment, not a replacement for it. The premise is that manufacturers already have identity, network segmentation, and access control programs in place; what is missing is a mechanism purpose-built to log, scope, and constrain agent tool-call behavior at runtime, in a way that produces audit-ready evidence and integrates with existing IEC 62443-aligned infrastructure.