What has to be true before a digital employee works alone
Autonomy is not a setting you turn on. It is the result of four decisions. What the employee may touch. What it may do there. Whether an action can be undone. And whether you can read back what it did. Get those right and approval becomes the exception.
What does autonomy actually mean here?
Not that the digital employee decides what to work on. That it finishes work without waiting for you.
The useful unit is the action, not the employee. Reading an inbox, drafting a reply and sending a reply are three actions with three different risk levels. Treating them as one thing is why so many pilots end up with a human clicking approve on everything.
So the question is never "do we trust it". It is "which of these actions still needs a pair of eyes".
Which actions should always wait for a person?
Anything that leaves the building, anything that moves money, and anything you cannot take back.
Sending to a customer, issuing a credit, changing a price, deleting a record. These are worth keeping on approval permanently, not just during the first weeks. The cost of the wait is small and the cost of being wrong is not.
Everything inside the building can usually run free. Reading, sorting, drafting, preparing, flagging. That is where most of the hours are anyway.
How do you know what it did?
Because every action is written down in a form a person can read, not a log file.
A readable record means four things. What it did, when, which data it used, and why it proposed this. If you cannot answer those four for any single action, you do not have oversight, you have a feeling.
This is also what makes the first weeks work. You start by reading everything, notice that a category is always right, and stop reading that category. The record is what lets you narrow your attention honestly.
What happens when it is wrong?
You want to know before the customer does. And you want the fix to be one change, not a conversation.
Wrong output is usually not random. It comes from a rule that was too broad, a source that was stale, or a case nobody foresaw. All three are fixable at the source, which is why the record matters so much.
Plan for it explicitly. A digital employee that has never been wrong has either not run long enough or is not being checked.
How do you widen the leash over time?
One category at a time, on evidence rather than on comfort.
The pattern that works has three steps. Run everything on approval for two weeks. Count how often you changed something before approving. Release the categories where you changed nothing. Repeat monthly.
That gives you a number to argue with instead of an opinion. It also means autonomy grows fastest where it is safest. That is the opposite of what happens when you decide up front.
Written from the approval and control layer AImplement builds into every implementation.