Workflow-scoped AI and open-ended agents

Two operating logics

In the workflow-scoped pattern, AI functions sit inside an existing procedure: convert an approved document into a structured work instruction, flag what changed in a revised procedure, compare captured evidence against a reference image. The scope comes from the surrounding process — the released source content, the role and qualification of the user, the current workflow state, the record that has to be produced. In the open-ended pattern, an AI agent receives a goal, plans its own steps and discovers its own tools and data sources. Nothing in the task itself limits it, so limits must come from policy, identity, network scope and supervision.

What makes the scoped variant reviewable

Four couplings do the work. Released knowledge: only content that has been approved may be used as a source, and its version release status is visible in the output. Roles: who may see or generate what follows the existing permission model rather than a separate AI model. Workflow state: a suggestion appears at a defined step, not in a vacuum. Evidence: an output that claims a check was passed points to the image, measurement or log that supports it. None of this makes the output correct — it makes it reviewable, which is the precondition for validation and for meaningful human oversight.

What goes into the documentation

Workflow-scoped AI is described as part of the technical documentation of the process it supports: for each step, the permitted sources, the input data, the output class (draft, recommendation, check note, released instruction), the role that may act on it, and the responsibility assignment for review, approval and execution. Add version control for prompts, models and templates, so a later reader can tell which configuration produced a given artifact. Vendor statements about tenant isolation, permission inheritance or human accountability are starting points for an acceptance test in your own environment, not conclusions.