What does the myth say?
That an AI employee that still needs approval is a half-built one, and the goal of any pilot is to get to the point where nobody looks. Under this belief, the approval gate is friction, the autonomy ladder is a delay, and a Digital FTE that is still at Level 0 after a month has failed. Managers who hold it promote classes early to prove progress, and then experience the send-backs that follow.
The examination is about where the value actually sits, and what autonomy is for.
Where does the belief come from?
From the word autonomous as a selling point, from demos that show an agent doing things end to end, and from a reasonable instinct that anything a person still has to touch is not yet automated. It also comes from a real cost: a manager who reviews a hundred items a day without evidence attached is spending hours, and any promise to remove that is welcome. The belief is a rational response to bad approval design.
It becomes a myth when the approval takes seconds, which is what evidence attached to every proposal is for.
What is true in it?
That approval time is real and should fall over time, and that a role whose routine classes never earn autonomy is being managed too tightly. That a promoted class does save the manager's minutes, and for a high-volume class those minutes add up. That some functions, overnight order-status replies for instance, are worth more when they do not wait for a morning approval. Autonomy is worth having, for the right classes, at the right time, and the ladder exists to grant it.
The myth's error is in the order: value first, then autonomy, from the record, not the other way round.
What does the mechanism actually do?
At Level 0 a Digital FTE on ftes.ai does all the work: it gathers evidence from the connected tools, proposes with the evidence attached, drafts through the fact gate, measures, detects changes, and reports in the morning brief. The manager approves in seconds because the evidence is there, or sends back with a reason. The approval record accumulates per class. At a monthly review the manager reads the record and promotes a routine class within bounds if it is clean; the FTE then runs that class without per-item approval, still logged, and steps back if send-backs rise. Nothing about the FTE's output changes with promotion; only the approval step thins.
Cost per outcome, measured against the baseline, is the same at Level 0 and Level 2 for the same outcomes; what changes is the manager's minutes.
- Level 0: all the work, evidenced, approved in seconds
- Promotion: same output, fewer approvals, still logged, reversible
- Value: proposals, drafts, measurement, record, at every level
What does early promotion actually cost?
A class promoted before its record supports it produces the errors the record would have caught: a reply sent that should have been escalated, a match posted that a person would have questioned. The send-backs rise, the class steps back, and the manager, having been burned, trusts the record less and keeps every class gated longer than it deserves. The net effect is a slower ladder, not a faster one. The guide on stepping autonomy back in the library describes the symptoms and the recovery.
The step-back is designed to be cheap. The lost trust is not.
How can you check for yourself?
Time your approvals for a week: total minutes spent in the Approvals tab per day, per role. Compare with the hours the proposals replaced, from the baseline sheet. For most roles the approvals are minutes and the replaced work was hours, which means the value was captured at Level 0 and autonomy is about the minutes. Then read one class's approval record and ask whether a promotion would have caught anything in the last month; if the send-backs were about substance, it is not ready, and promoting it would have cost more than the minutes it saved.
The playbook on earning the first promotion sets out what a ready record looks like.