Why a chemical-sector AI governance benchmark does not yet exist
Chemical and process manufacturing enterprises are integrating AI agents into engineering, production, and OT-adjacent workflows faster than measurement infrastructure has developed to track that integration. No publicly confirmed 2026 study establishes what percentage of chemical manufacturers have deployed AI agents into production or engineering systems, and no documented incident database captures unauthorized tool-call or access events specific to this sector. This is a measurement gap, not evidence of low risk. Governance leaders should treat the absence of published statistics as a reason for internal data collection, not a reason for inaction. The frameworks that do exist, NIST's AI Risk Management Framework and its Generative AI Profile, were published in 2023 and 2024 respectively, ahead of widespread agentic AI deployment in industrial settings. IEC 62443 predates agentic AI entirely and was written for deterministic OT/ICS security, not for probabilistic systems making tool calls into production environments.
The frameworks currently governing AI risk in process environments
Two distinct bodies of guidance apply to chemical manufacturing today, and they were not designed to work together. NIST AI RMF 1.0 structures AI risk management around four functions: Govern, Map, Measure, and Manage. It is IT-oriented and sector-agnostic. IEC 62443 is OT-oriented, built around zone and conduit segmentation models that define trust boundaries between industrial control systems. Neither framework, based on confirmed public material, contains provisions specific to agentic AI operating inside OT-adjacent environments. NIST AI 600-1, the Generative AI Profile released in 2024, acknowledges risks tied to autonomous and tool-using AI systems, but it does not address process safety validation or the deterministic control logic that chemical manufacturing environments depend on. CISA has flagged AI as an emerging risk factor for critical infrastructure sectors including chemical manufacturing, and recommends aligning AI governance with existing risk-based OT security approaches, but it stops short of prescribing sector-specific controls. The practical result is that governance leaders in this sector are currently expected to synthesize two frameworks that were not built for each other.
Structural gaps facing AI governance in chemical manufacturing
- Deterministic control logic meets probabilistic AI: Safety-critical process environments require auditable, repeatable control behavior. AI agents introduce probabilistic outputs that have not yet been addressed by a confirmed sector-specific validation standard.
- No confirmed incident disclosure norm: No documented AI governance failures or unauthorized tool-call events specific to chemical manufacturing were identified in available public sources, which limits sector-wide risk quantification.
- IT and OT frameworks remain separate: NIST AI RMF addresses AI risk generally, while IEC 62443 addresses OT security specifically. Neither fully covers agentic AI operating across the boundary between the two.
- Vendor-reported statistics are unverifiable: Adoption or maturity figures circulated by vendors should be treated as unverified until sourced independently, given the absence of confirmed primary research for this sector.
- No cross-sector maturity benchmark: No confirmed comparison exists between chemical manufacturing and other industrial sectors, such as discrete manufacturing or energy, on AI governance maturity.
What governance leaders can measure internally in the absence of external data
In the absence of a confirmed sector benchmark, internal measurement becomes the primary evidence base for governance decisions. This starts with an inventory of where AI agents interact with engineering, production, or safety systems, an exercise consistent with the Map function in NIST AI RMF. From there, the Measure function can be applied to track tool-call frequency, access scope, and deviation from expected agent behavior over time. Audit trails covering AI agent activity in safety-critical workflows serve as a direct substitute for the external incident data that does not yet exist publicly. Mapping this activity against existing IEC 62443 zone and conduit architecture allows governance and OT security teams to evaluate AI agent access using the same trust-boundary logic already applied to other industrial systems. This approach does not require new sector-specific regulation to be actionable, since AI RMF and IEC 62443 are already recognized frameworks that organizations can adapt now.