Autonomy Readiness: Measuring When an AI Workforce Has Earned Independence

Autonomy readiness is an evidence-based measure of whether an AI system has earned the right to act without human review on a defined type of work. Readiness is never a default setting, never a feature flag, and never a function of how long a contract has been active. It is a calculation performed against the system's own record of real decisions inside a real business.

Four inputs feed the measure. Volume of proven decisions establishes that the pattern has been tested enough times to mean something. Correction rate shows how often a human had to intervene and how recently. Outcome quality asks whether the completed work produced the result the business wanted, which is a separate question from whether the work looked right at the moment of approval. Elapsed operating time places the whole record against the calendar.

Time carries a hard floor. A task type can produce two hundred clean decisions in nine days and still fail the readiness test, because nine days has not exposed the system to a month-end close, a holiday week, a price adjustment, or the customer who calls at 6pm on a Sunday. Independence that arrives faster than the business cycle is a guess dressed as a metric.

Why It Matters.

What goes wrong when independence is granted by calendar or by vendor promise.

Software vendors sell autonomy as a launch date. Go live, flip the switch, and the system starts running the work. Owners of service businesses learn the flaw in that promise from a customer complaint: an appointment booked into a slot that never existed, a quote sent at last year's pricing, a review request fired at the one client whose job went sideways.

The opposite failure is quieter and more common. Nervous about the first category, an owner reviews everything forever. Twelve months in, the software has produced a great deal of activity and zero relief. Sitting in every approval is the same bottleneck described in The Founder Bottleneck Nobody Talks About.

A readiness measure gives the owner a defensible middle. Independence expands where the evidence supports it and stays closed where it does not, and the reasoning is visible in both directions. Explaining to a service manager why dispatch messages run on their own while refund decisions still route to a human becomes a matter of showing the record rather than defending a hunch. The per-task numbers underneath that record are covered in Confidence Scores.

How It Works in AA19.

The gate between the three modes, and what opens it.

AA19 runs every task type through the ladder described in Approval, Hybrid, Autonomous. Readiness sits at each gate. Approval mode routes everything to a human and builds the initial record. Hybrid mode opens once the system has proven the common patterns, letting familiar work execute while novel work waits. Autonomous mode opens only after sustained performance across a defined operating window, and only inside a stated boundary.

Evidence behind each gate comes from the approval queue. Every yes, every edit, and every rejection is captured with context and preserved as Decision History. Outcome checks come from the layer described in Verification, which distinguishes work that was completed from work that was correct. Corrections turn into persistent instruction through AI Agent Guidance, so the same failure does not have to be caught twice.

Readiness produces eligibility, not permission. The Brain surfaces a graduation recommendation with the full evidence attached: decision count, correction history, outcome record, operating window, and the boundary being proposed. A human authorizes the change. That separation is the practical form of the argument in Trust in Autonomous Business Systems: authority gets handed over deliberately, by a person, against evidence.

Demotion runs automatically. A correction inside an autonomous lane, a cluster of weak outcomes, or a change to underlying policy pulls readiness down and returns that lane to Hybrid. Nothing accumulated is discarded, so the second ascent runs on a stronger base than the first. Across dozens of task types, that ratchet is what turns a supervised workforce into an independent one, and the wider system view lives on the Organizational Intelligence page.