Every agent platform demo ends the same way: the agent does the thing. Books the meeting, files the ticket, moves the money. The demo question is always “can it act?” — and the answer is always yes. That is precisely why it’s the wrong question.
The right question is “what has it earned?”
We think about agent autonomy as a four-rung ladder. Observe: the system watches and reports; full tracing, no voice, no hands. Advise: it recommends, with citations; humans decide. Act with Approval: it executes behind a human approval queue, with caps on what any single action can do. Act Autonomously: it operates under continuous supervision, circuit breakers, and rollback — because at this rung, review after the fact is the only kind there is.
Gartner’s prediction that 40% of autonomous agents will be demoted or decommissioned by 2027 is, read correctly, a prediction about ladders: those agents were placed on rung four without ever earning rungs one through three.
What “earning” looks like
Promotion up the ladder should be an evidentiary event, not an enthusiasm event. A system moves from Advise to Act-with-Approval when its recommendation acceptance rate, grounding accuracy, and exception behavior have been measured — on real traffic, over a defined period — and someone accountable signs the promotion. The signature matters: autonomy without a named sponsor is autonomy nobody owns.
Autonomy is a privilege earned on the record — never a default, never a demo.
The ladder also gives you the thing every risk officer quietly wants: a dignified direction for demotion. When drift appears, you don’t kill the project and the political capital behind it — you move it down a rung while the numbers recover. Governance that can de-escalate gracefully is governance people will actually use.