What does the myth say?
That an AI agent is an autonomous actor: switch it on, give it tools, and it does things in your systems without anyone looking. The fear that follows is reasonable: an agent with access to a ledger, a helpdesk, or a publishing route, acting unsupervised, is a liability waiting for its first bad day. Leaders who hold this belief refuse pilots, and they are right to refuse the version they are imagining.
The examination here is about which agents the belief describes, and what a Digital FTE does instead.
Where does the belief come from?
From agent frameworks and demos where autonomy is the point: a loop that plans, acts, observes, and acts again, with tools wired directly and a human watching a log if at all. Those exist, are useful for some tasks, and are what most people have seen. It also comes from the word autonomous being used as a selling point, and from incidents where an agent did something nobody approved. The belief is an accurate description of a real category.
It becomes a myth when applied to every product that calls itself an AI employee, including ones designed on the opposite principle.
What is true in it?
That an agent with direct tool access and no gate will act on its own judgement, and that its judgement will sometimes be wrong. That a single trust setting, on or off, is not supervision. And that on ftes.ai, a routine class that has earned autonomy does run within its bounds without a person approving each item, which is a real form of unattended action, deliberately granted, bounded, logged, and reversible. Anyone who says a Digital FTE never acts without approval is overstating it; the honest statement is that it never acts without approval until a specific class has earned it.
The difference between the two is the whole design.
What does the mechanism actually do?
Every task class a Digital FTE handles starts at Level 0 on the autonomy ladder: the FTE proposes with evidence and a person approves, sends back with a reason, or escalates. Promotion is per class, decided by the manager at a monthly review from the approval record, within bounds the manager sets, and it steps back if send-backs rise. The always-escalate list, complaints, money, decisions about people, anything sensitive, holds at every level and grows after incidents; the never-automated list cannot be changed by anyone. Every proposal, approval, send-back, and autonomous action is in the audit trail with a name and a timestamp.
The manager's morning brief lists what is waiting; the Approvals tab is the gate; the Setup tab shows each class's level.
- Level 0 for every class at hiring
- Promotion per class, bounded, from the record, reversible
- Always-escalate and never-automated lists at every level
- Every action in the trail
Why not a single trust setting?
Because trust is not one thing. A manager can reasonably trust a Finance Controller FTE to post exact matches under a threshold and not trust it to post anything else; can trust a Support FTE with password resets and not with refunds. A single setting forces the choice between gating everything, which wastes the record, and freeing everything, which is the unsupervised agent the myth describes. Per-class autonomy lets the record decide, one class at a time, and lets a promotion be undone without undoing the others.
It also produces a defensible answer to the question every auditor asks: what can this thing do on its own, and who decided?
How can you check for yourself?
Open a new role's Setup tab and read the autonomy ladder: every class should show Level 0 and the escalate list should be visible and fixed. Approve one proposal in the Approvals tab, then open the audit trail and find it with your name and the time. Attempt to change the never-automated list and confirm you cannot. Ask in writing what happens when the manager is away: the answer should be that proposals wait and nothing happens without a decision. Ten minutes, and you have verified supervision rather than been assured of it.
If a vendor cannot show you a per-class ladder, a fixed escalate list, and a trail entry with a name, you are looking at the agent the myth is about.