All resources
Agents5 May 2026·6 min read

The approval-gate ladder: a practical path from human-only to autonomous

Autonomy isn't a switch you flip. It's a ladder you climb one rung at a time — and each rung is what gets a nervous stakeholder to say yes.

The fastest way to lose an agent project's political support is to propose full autonomy on day one. Legal, compliance, and operations teams have good reason to be skeptical of a system that acts without a human checking — their skepticism isn't an obstacle to route around, it's a signal to design for. The approval-gate ladder gives every stakeholder a rung they're comfortable with today, and a visible, evidence-based path to the next one.

The five rungs

  • Human does everything, agent drafts — the agent proposes an action or response; a person reviews and executes every single instance. This is where trust-building starts.
  • Human approves each action — the agent takes the first step (queries a system, prepares a change) but a person clicks approve before anything executes.
  • Human approves batches — once error rates are known and low, approvals move from per-action to per-batch, reviewed on a schedule rather than in real time.
  • Agent acts, human reviews a sample — the agent executes autonomously; a person audits a statistically meaningful sample after the fact, not before.
  • Full autonomy with monitoring — the agent runs unattended; humans are alerted only on anomalies or threshold breaches, with the full audit trail available on demand.

What moves you up a rung

Not time — evidence. A documented error rate below an agreed threshold, over an agreed volume of real cases, reviewed by the stakeholder who was skeptical in the first place. Promotion criteria should be written down before the project starts, so 'are we ready for more autonomy' is a data question, not a debate.

The ladder isn't a compromise on ambition — it's how ambitious autonomy actually gets approved. Nobody signs off on full autonomy from a standing start. Plenty of skeptics sign off on 'draft only' today and 'batch approval' next quarter, if the evidence keeps arriving.

Where teams go wrong

Skipping rungs to hit a launch date, or never defining what evidence would justify climbing at all — leaving the agent parked at 'human approves each action' indefinitely, which is expensive to run and rarely the autonomy anyone actually wanted. The ladder only works if someone owns moving up it.

Related capability

This is what our Autonomous AI Agents practice is built around.

See how it's scoped