The most important question about an industrial agent is not whether it can produce an answer. It is what authority it has, which evidence it must cite, how uncertainty is handled and who remains accountable when the operating situation changes.
Use an explicit authority ladder
An agent that summarizes a shift report has a different risk profile from one that changes a work-order state or writes a process value. Treating both as the same AI feature hides the engineering controls that matter.
Delfers uses a five-level ladder: observe, explain, recommend, prepare and execute. Each step requires stronger evidence, testing, authorization and monitoring.
- Observe: retrieve governed facts without interpretation
- Explain: summarize conditions and cite the evidence used
- Recommend: propose an action while a person decides
- Prepare: draft a controlled change for authorized approval
- Execute: perform a bounded action through approved interfaces
Citations are part of the control surface
A plant user should be able to inspect the alarms, production events, procedures, model outputs or maintenance history behind a recommendation. Sentrix can organize that reasoning, Delfi can make it accessible conversationally, and Fabrix can preserve the execution context and approval event.
If the evidence is stale, conflicting or incomplete, the system should say so. An explicit limit is safer and more useful than synthetic certainty.
Define stop conditions before success criteria
Teams commonly define accuracy or response-time targets but leave shutdown logic vague. Stop conditions should cover missing data, model drift, unexpected asset state, revoked authorization, network loss, failed write confirmation and any safety interlock.
The fallback must be operationally legible: return control to the established human or automation workflow, preserve the evidence and notify the named owner.
Earn authority through reviewed operation
Promotion to a higher authority level should rely on a documented operating history: accepted and rejected recommendations, overrides, failure modes, incident reviews and measured consequences. The review should include operations, engineering, security and the process owner—not only the AI team.
Autonomy is therefore not a launch setting. It is a privilege granted to a narrow workflow after evidence shows that the controls are appropriate.
Standards and primary sources used for context.
These links provide technical context. Their inclusion does not claim product certification or compliance unless explicitly stated elsewhere.
- Artificial Intelligence Risk Management Framework (AI RMF 1.0)National Institute of Standards and TechnologyA voluntary framework for incorporating trustworthiness considerations into the design, development, use and evaluation of AI systems.
- Cybersecurity Framework (CSF) 2.0National Institute of Standards and TechnologyOutcome-based guidance for understanding, assessing, prioritizing and communicating cybersecurity risk.
- ISA/IEC 62443 seriesInternational Society of AutomationCybersecurity requirements and processes for industrial automation and control systems.
