Abstract
Industrial AI risk is often discussed as though intelligence itself creates danger. A more useful lens is authority. The consequence of an AI error changes dramatically as the system moves from retrieving information to modifying records, releasing work, changing configuration, and altering physical equipment. This paper proposes an authority ladder that ties increasing AI permissions to stronger evidence, verification, auditability, and reversibility.
Authority is the consequential variable
A weak model that can only search manuals is annoying. A stronger model that can change work-order status is consequential. A model connected to physical equipment can be consequential even when its reasoning is excellent. The relevant design question is not “How smart is the AI?” but “What may this system cause to happen?”
Current enterprise platforms already show the progression. Oracle’s Maintenance Work Order Builder can create and update records. Maximo 9.2 exposes MCP tools for actions such as changing work-order status, creating follow-up work orders, recording meter readings, and invoking custom tools. Anthropic’s Model Hardware Standard research preview extends the idea to physical devices.
The authority ladder
A practical authority ladder can be defined as follows.
Level 0, Observe: retrieve documents, histories, sensor values, and records.
Level 1, Interpret: identify patterns, contradictions, anomalies, and missing evidence.
Level 2, Recommend: propose verification steps, work scope, priority, or maintenance action.
Level 3, Prepare: draft official records, work packages, schedules, or configuration changes without committing them.
Level 4, Transact: create or modify official enterprise records and workflow states.
Level 5, Actuate: alter physical equipment state or control parameters within bounded permissions.
Level 6, Closed-loop autonomy: observe physical outcomes and independently choose subsequent actions.
The ladder matters because products are already being sold across several rungs at once without saying where the line sits. At IMTS in September 2026, deviceWISE demonstrated agentic fault recovery on live robotic cells, and its own release says the system "recommends or executes" recovery steps. That one phrase spans Level 2 to Level 5. Nothing public says which faults it may recover on its own, which need an operator to confirm, or what forces it to stop and escalate. That is not a criticism of the demonstration. It is the unwritten boundary this paper argues has to be set per action, in writing, before the system is connected to anything that moves.
Evidence burden should rise with authority
The required strength of evidence should increase as the action becomes more consequential, less reversible, or more difficult to detect. A manual-retrieval error can be corrected by a technician. A false work-order closure can pollute records and hide work. An inappropriate physical command can create immediate operational consequences.
ADMITBench offers a useful parallel by evaluating industrial advisories against authority, procedure, consequence, and reversibility. The implication is clear: permissions should not be granted simply because a model performs well on average. Authority should be conditional on the action class and the verification environment around it.
A practical deployment path
Organizations should climb the ladder deliberately. Start with read-only retrieval. Add structured review and exception detection. Allow drafting only after outputs can be checked reliably. Permit transactions when role permissions, audit logs, approval workflows, and rollback mechanisms are mature. Physical actuation should remain tightly bounded by engineered limits and independent controls.
This is slower than marketing language about autonomous operations, but faster than recovering from an agent that was given enterprise-wide write permissions because somebody liked the demo.
Conclusion
Industrial AI becomes consequential when its output acquires authority. A mature governance model should therefore tie permissions to evidence quality, verification strength, consequence, reversibility, and human accountability. The authority ladder provides a simple way to talk about autonomy without pretending every AI capability belongs in the same risk category.
References
- IBM Docs, “Creating custom AI tools in Maximo by using object structures.” https://www.ibm.com/docs/en/masv-and-l/cd?topic=tools-by-using-object-structures
- Oracle, “AI Agent: Maintenance Work Order Builder.” https://docs.oracle.com/en/cloud/saas/readiness/scm/25d/maint25d/25D-maintenance-wn-f40777.htm
- Anthropic, “Previewing the Model Hardware Standard,” 2026. https://www.anthropic.com/news/model-hardware-standard-research-preview
- Misra et al., “ADMITBench,” 2026. https://arxiv.org/abs/2608.03866
- deviceWISE (Telit Cinterion), “deviceWISE to Demonstrate Agentic AI and Automated Fault Detection & Recovery on Live Robotic Lines at IMTS 2026,” PR Newswire, September 2026. https://www.prnewswire.com/news-releases/devicewise-a-telit-cinterion-company-to-demonstrate-agentic-ai-and-automated-fault-detection--recovery-on-live-robotic-lines-at-imts-2026-302875313.html