Human-Governed Automation: Where AI Should Stop
A practical control model for automation, informed by NIST guidance on human-AI configuration, over-reliance, and risk.
Based on research by
Chloe Autio et al. — NIST
Published -
The useful question is not whether a workflow can be automated. It is where automation should begin, where a person must review, and how the process stops safely when the situation no longer matches the rules.
Good automation removes repetition without removing responsibility.
Automation needs visible boundaries
The NIST Generative AI Profile identifies risks in human-AI configuration, including automation bias and over-reliance. Those are operating-design problems as much as model problems: people need to know what the system did, why it acted, and when their judgment is required.
Three control points
A practical control model defines the conditions that start an automated action, the point at which a person reviews it, and the mechanism that pauses the workflow safely.
Trigger — begin routine work only when defined conditions are present.
Review — require a responsible person to check consequential outputs before action.
Stop — let owners pause, correct, and investigate without waiting for a developer.
Relief without loss of control
The best candidates are repetitive, observable, and reversible tasks. Consequential decisions require stronger review, provenance, and escalation. This boundary lets teams gain speed without pretending that responsibility has moved to the machine.
Source: NIST Generative AI Profile. NIST says unmarked site information may be distributed or copied with appropriate credit requested. This is an AURENDOR interpretation and is not endorsed by NIST. Reuse terms
More Insights
