Progressive autonomy: earning the right to remove the human.

The most consequential choice in deploying an agentic system is rarely what the system can do. It is how much the system is permitted to do without a person watching. That question — how much authority to hand a machine, and how quickly — is where an ambitious program and a defensible one part ways. The seductive answer is everything, immediately: full automation on day one, presented as conviction. But authority granted before it is earned is not conviction. It is an untested assumption carrying real liability, and it tends to surface at the worst possible moment.

Autonomy must be earned because the evidence required to grant it does not yet exist at the outset. Clearing a pilot proves that a system can succeed on the cases it was shown, with practitioners standing by to catch what it missed. It says little about the long tail it will meet in production — the malformed request, the user acting in bad faith, the period when the source data moves away from what the system was validated on. To automate fully from the start is to extend trust on the strength of the easy cases and assume the hard ones will resemble them. They seldom do.

Progressive autonomy replaces the switch with a dial. The system begins supervised, running alongside the people who perform the process today, its judgments checked against theirs and measured against a defined target rather than a general impression. As a record accumulates, authority is released in graded steps — the routine decisions first, the consequential ones later, each increment granted only when the evidence supports it. Autonomy is scaled at the pace of proof, not the pace of ambition. The human role migrates by design: from performing the process, to reviewing it, to handling only the exceptions the system escalates.

That dial, however, is engineered never to reach zero. A permanent floor — a share of decisions that always route through a person — is designed in, and it is not evidence that the system is unfinished. It is a deliberate reservation of judgment, not a gap waiting to be closed. Production is not stationary. Models drift, upstream data changes, and the world a system was validated against stops matching the world it now runs in. A deployment that removed its last human would also remove its earliest warning: the one vantage point from which a novel failure is noticed before it compounds. The floor is where someone sees the failure the automated checks were never designed to catch.

It is tempting to read a standing human checkpoint as a cost, or as a quiet admission that the automation is incomplete. The reverse holds: it is what makes the autonomy governable. A board, an examiner, or an acquirer running diligence takes no comfort in hearing that a system operates entirely on its own. They ask where the checkpoints sit, what conditions trigger them, and who answers when one fires. A system whose authority can be narrowed as readily as it was widened is worth more than one running wide open — because its behavior can be steered, evidenced, and, when it matters, pulled back.

The calculus a serious operator applies is not whether the system can do this, but what a wrong answer costs here, how often it occurs, and who absorbs it. Authority is granted decision-class by decision-class against that asymmetry. A mature autonomy plan therefore reads less like a launch date than a schedule of thresholds. Where an error is cheap and reversible, autonomy arrives early and broad. Where an error is expensive, irreversible, or falls on a customer, on capital, or on a regulated judgment, the human stays — in some places permanently. The floor is not spread evenly across the process; it is placed precisely where the downside is worst, and left there on purpose.

Full automation on day one optimizes for the launch — the unveiling, and the authority handed over before it has been earned. Earned autonomy with a floor optimizes for the long stretch afterward, when conditions move and the validation that once justified full autonomy no longer describes the inputs the system now meets. The first is a claim about capability; the second is a discipline about consequence. Removing a person from a process is a threshold a system should have to reach — by proving, repeatedly and under real conditions, that it can be trusted with more. Autonomy that was earned, and that a person can still take back, is the only kind worth running in production.