AI in OT: A Standards Stack, from IEC 62443 to ISO/IEC 42001

On 23 October 2026, IIoT World published a Q&A drawn from the session "It Can't Happen to Me – Until It Does" at the Industrial AI Summit 2026. The speakers are Gary Tillery (CEO of Skkynet) and Matt Morris (VP Security and Resilience at EverLine). The article is labeled "Sponsored by Skkynet | Editorially Independent", so it should be read as content with some vendor influence. In addition, the summary below is based only on the first part of the page.

No single standard is enough: you need a stack

Tillery's message is that there is no single reference standard for AI in manufacturing. At the base sit interoperability and OT security; functional safety is a hard boundary the model must not cross; at the top is an AI management system for trustworthiness. The biggest gap concerns agentic and generative AI, because current standards assume a bounded model, not a planner that chooses which tools to use. Tillery himself notes that he looked up the standards after the panel and that the list is not complete.

The standards cited

The main layers of the stack, according to the article:

• OPC UA (IEC 62541): the practical foundation, with secure communication and information models, so the AI reads machine state rather than raw tags. Method calls should stay inside the information model.
• Asset Administration Shell (IEC 63278) and RAMI 4.0, which can be paired with ECLASS to anchor information retrieval to asset data.
• IEC 62443 and NIST SP 800-82 Rev. 3: models, brokers and tool gateways placed in zones and conduits, with least privilege and authentication; IEC PAS 62443-1-6 addresses IIoT.
• Functional safety: IEC 61508, 61511, 62061 / ISO 13849. No model or remote tool should write through that path.
• IEC 62264 / ISA-95 for the enterprise–control boundary, ISO 23247 for digital twins, MQTT with Sparkplug B for lightweight telemetry (without replacing OPC UA semantics for control-relevant data).
• ISO/IEC 27001 for enterprise IT, which does not replace 62443 on the plant network.
• ISO/IEC 42001:2023, a certifiable AI management system, with related standards on risk (23894), lifecycle (5338), data quality (5259) and vocabulary (22989).
• NIST AI RMF (AI 100-1) and the Generative AI Profile (AI 600-1), the voluntary counterpart built on Govern, Map, Measure, Manage.
• EU AI Act: where it applies, it treats safety components and certain industrial uses as high-risk; it does not replace IEC 61508 or 62443.

Architecture: AI as an advisor

The guidance is to treat AI as an advisor. The model should hold no standing credentials to PLCs or historians. Tool calls go through a broker outside the prompt, which uses short-lived, task-scoped credentials and keeps an audit trail of who delegated each action.

The IT/OT boundary according to Morris

Morris proposes designing on the assumption that IT is already compromised. He cites Colonial Pipeline: its OT was not compromised, but the company shut down because without IT it could not bill or track product.

The controls listed:

• a separate OT identity, with no domain trust or shared accounts;
• an IT/OT DMZ where data is pushed outward and nothing initiates inbound connections into OT;
• jump hosts with phishing-resistant MFA, approval, recording and time-limited access;
• application allowlisting on HMIs and engineering workstations;
• OT-aware monitoring at the conduits.

To verify them, he suggests assume-breach tests from a compromised IT workstation (in a replica lab or during maintenance windows), attack-path analysis on the as-built configuration, and island-mode exercises that separate IT from OT. The proposed metrics: time to isolate, percentage of conduits validated, and undocumented dependencies found.

Why it matters and what to do with it

More than an incident or product story, this is a map. Its value for anyone connecting AI to plants is as a checklist: keep AI out of safety paths, give it only mediated and traceable access, and measure isolation. Since it is a list proposed by a single vendor executive, it should be used as a starting point to check against the original standards, not as an exhaustive guide.