The four things that define one
When you create an AI employee you are really defining four constraints. Get these right and the output is predictable; leave them vague and you get confident nonsense.
- Responsibilities — the outcomes it owns, written as work, not as prompts.
- Tools — the specific actions it can take, such as drafting outreach or updating a task.
- Knowledge scope — which parts of your Business Brain and documents it can read.
- Authority — the value or risk level at which it must stop and request approval.
What it does on its own
Inside its scope, an AI employee reads the relevant context, proposes a plan, and produces work: a draft, a scored list, an analysis, a suggested next action. That work is recorded so you can see what was produced and on what basis.
What it never does
It does not invent authority. It does not act outside its tool list. It does not read knowledge outside its scope. And it does not run on a schedule in the background unless you deliberately start work — which also means it is not quietly spending money while you sleep.
Where humans stay in the loop
Anything that touches a customer, moves money, or changes a policy should sit behind approval while you are building trust. Start with everything requiring approval, then relax the limits per employee once you have seen a few weeks of its output.