The Missing Safety Case for AI on the Shop Floor

Two different gaps, often confused

A recent Rockwell Automation survey found that industrial security leaders’ stated confidence in their cyber readiness does not match what their own incident and patching data actually show, a mismatch that deserves board-level attention. That finding is about network segmentation, access control, and patch discipline. It says nothing about whether a plant has defined what an AI system is allowed to decide on its own. Conflating the two problems is a mistake compliance leaders cannot afford to make, because the fixes are different and the second gap has no mature cybersecurity playbook to borrow from.

Vision systems, robotics platforms, and agentic tools are moving from surfacing information to making operational calls on the plant floor. Industry coverage of this shift notes that spatial and agentic AI are increasingly taking on decisions that used to require a human in the loop, and training programs for combining machine vision with robotic actuation are expanding accordingly, reflecting how fast this integration is moving from pilot to production. Firewalls and segmented networks do not answer whether a given decision was safe to hand to a machine in the first place.

Where the standard is actually catching up

This is not a governance void manufacturers have to improvise through. ISO/IEC TS 22440 is emerging as the reference point for functional safety specifically where AI systems are embedded in industrial equipment, addressing how machines and their AI components should be validated against safety requirements before they’re trusted with operational authority. That is the enforceable layer the Rockwell survey does not touch, and it is the one compliance teams should be mapping against now, not after an incident forces the question.

Legal counsel advising manufacturers on safety programs are already treating AI this way. Recent guidance on integrating AI into plant safety programs argues that AI-driven decisions need to be folded into documented safety protocols, including training records, incident investigation procedures, and clearly assigned human oversight roles, rather than treated as a bolt-on to existing cyber and OT controls.

The honest caveat

None of this makes cybersecurity posture irrelevant. A segmented network and clean patch record remain necessary conditions for any AI deployment on a plant floor, and the Rockwell findings are a legitimate warning on their own terms. The point is narrower: closing the cyber gap does not close the decision-authority gap, and boards that treat the Rockwell survey as evidence of AI readiness are answering a question nobody asked.

What to ask before the next pilot

Before approving the next vision, robotics, or agentic pilot, ask who signed off on the AI’s decision boundaries, whether that sign-off maps to ISO/IEC TS 22440’s functional safety requirements, and whether it shows up in the same documentation your safety program already audits. If the answer lives only in a vendor’s cybersecurity questionnaire, the pilot is not ready.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The core argument—that cybersecurity readiness and AI decision-authority governance are distinct problems requiring different fixes—is coherent and well-supported, though the claim that ISO/IEC TS 224
Source & Claim Verification	Qwen · local	held. The draft leaves multiple factual claims—such as the Rockwell survey’s specific focus, ISO/IEC TS 22440’s enforceability, and the expansion of training programs—unsupported by the cited sources and do
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately references ISO/IEC TS 22440 for AI functional safety but does not substantively address ISO 42001, the EU AI Act, NIST AI RMF, 21 CFR Part 11, or EU GMP Annex 11/22.
Technical Accuracy	Llama	cleared. The text correctly distinguishes between cybersecurity (network/patching) and functional safety (decision boundaries/ISO/IEC TS 22440), maintaining high technical accuracy regarding industrial automat
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively distinguishes between two distinct safety concerns, directly addresses potential conflation, and avoids vendor hype by focusing on a specific standard and legal guidance rathe
Novelty & Non-Duplication	Grok	cleared. The two-gaps framing (OT cyber readiness vs. AI decision-authority/safety case) plus the concrete ISO/IEC TS 22440 pilot gate is a real synthesis, not a rehash of any single wire story or obvious back
Validation	DeepSeek	cleared. The central claim that industrial AI safety requires distinct validation beyond cybersecurity is well-supported by cited sources on governance standards and legal guidance.

Sources cited: 13. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.