The question we are asked most often is some version of “can the agent just do it?” The honest answer is: for some steps, yes; for others, not yet; and for a few, never. The useful work is deciding which is which, before anything is built.
We draw the line with three questions. Is the action reversible? Is the evidence the agent used visible to the person accountable for the outcome? And does the organisation already have a rule for who may take this action? Where all three answers are yes, automation is usually safe. Where any answer is no, the agent proposes and a person approves.
Separate suggestion from action
In an AKE deployment, drafts and approved actions are different objects with different permissions. A drafted reply is not a sent reply. A proposed journal entry is not a posted one. This sounds obvious, yet many AI tools blur the two, and the blur is where most incidents come from.
Approval points are also where the evaluation lives. Each decision a person makes on an agent’s proposal is a labelled example. Over time those decisions tell you where the agent is reliable enough to be trusted with more, and where it is not.
Make the stop visible
A good approval point shows the person three things: what the agent is proposing, what evidence it used, and what will happen if they approve. It should take seconds to review and one action to approve, return or escalate. If reviewing takes longer than doing the task by hand, the workflow design is wrong, not the reviewer.
Governance designed this way is not a brake on transformation. It is what lets an organisation extend automation step by step with its eyes open.