Why shouldn't an AI model authorize actions on its own?
A language model produces the most plausible response rather than a guaranteed-correct one, and that plausibility is what lets a bad action slip through. A model that's been manipulated through injected content can produce an answer that looks right and triggers a payment or an access change that should never have happened. For regulated, high-stakes operations, the white paper's rule is that policy enforcement sits outside the model and stays independently testable, so an action goes through because it passed a check, not because an answer sounded convincing.
What is least privilege for AI agents?
Least privilege means an AI component reaches only the data and tools a task requires, and nothing else. The OWASP AI agent security guidance, which the paper cites, stresses constrained tool access, least privilege, human oversight for high-impact actions, and protection against prompt injection. OWASP also traces excessive agency to three causes: too much functionality, too many permissions and too much autonomy. A governed workflow narrows all three, because each is set by the process rather than left to the agent.
Where does enterprise AI governance live?
It lives in the steps between the model and the action. A request comes in, the model interprets it, and the workflow decides what's allowed: it applies the rule, records who or what did each step, and sends anything high-impact to an authorized person. The white paper's architectural comparison makes the point that human authorization can be a native gate in a workflow, rather than something bolted onto an agent afterward. Because those controls are explicit and reusable, you apply the same guardrails across many processes and test them like any other business rule. Where the platform relies on certifications, those are ISO 27001 and ISO 27701, held by the RunMyProcess platform.
It helps to see what each of those three causes looks like on the ground. Too much functionality is an agent wired to tools it never needs for the task in front of it. Too many permissions is a tool that can write where it only had to read, or reach a system outside the job. Too much autonomy is a design that lets a response act without a check in between. A governed workflow closes each gap by construction: the process grants a step only the tools and permissions it uses, and any high-impact action waits for a rule or a person.
The same discipline covers untrusted content. Because the white paper treats how an application handles injected content as a core risk, a governed design validates what the model returns before any of it reaches a system, so a manipulated answer has nowhere to go. It also makes governance auditable: when the rule, the identity and the process state are recorded at each step, you can show a reviewer why an action was allowed, instead of pointing at a model response no one can replay.
If you're piloting AI and can't yet point to where authority lives, start by drawing that line. Separate what the model may advise from what the workflow must authorize, and you'll have governance you can test.
Keep reading: Agentic AI token costs · Human in command: three decision levels



