What has to be true before an agent touches a real process.

The interesting question about agents is not what they can do. It is what happens when they are wrong, and who finds out.

Three conditions have to hold before autonomy is safe in an enterprise process: the boundary of the action is explicit, the action is reversible or reviewable, and accountability sits with a named human role rather than with the system.

Where those conditions are absent, an agent does not remove work — it relocates it into exception handling, usually into a team that was never consulted.

The organisations getting value are unglamorous about it. They start where the process is well understood, the data is already governed, and the cost of an error is knowable.