The definitional explainer
"Autonomous" is the word every facility-management vendor reaches for, but it covers everything from autocomplete to a system that runs the operation while you sleep. The useful version is a ladder of five stages — manual, assisted, supervised, bounded-autonomous, and exception-managed — and the honest question is not whether a system is autonomous, but which rung it stands on and where a human stays accountable.
STEADYWRK operates at the bounded-autonomous stage: routine decisions run on their own inside an explicit policy boundary, anything under 70% confidence escalates to a person, and the people who catch those exceptions are nearshore — an economic model the category has not claimed.
The autonomy spectrum
Autonomy in facility management is not on or off. It moves along a spectrum defined by one question: how much does the system decide on its own, and what does it hand back to a person? Each stage is a different answer. "Autonomous" properly begins at stage 3.
The human decides everything
A coordinator reads every work order, picks a contractor from memory, types each quote, and chases status by phone. The software, if any, is a filing cabinet. Throughput is capped by one person’s attention, and the operation does not run while they sleep.
Software suggests, the human commits
Tools surface a recommended contractor, a templated quote, or a priority flag — but a person still approves and executes every step. This is where most "AI-powered" facility software actually sits: helpful autocomplete on a workflow a human still drives end to end.
Software acts, the human watches in real time
The system drafts and routes, and a human reviews each action as it happens before it commits. Latency improves, but a person is still in the path of every decision, so the operation scales with headcount rather than independently of it.
Software acts within a policy boundary
Routine, high-confidence decisions execute on their own inside an explicit policy boundary; anything that crosses the boundary stops. This is where genuine autonomous facility management begins — and where STEADYWRK operates. At STEADYWRK the boundary is an explicit confidence threshold: any decision under 70% confidence is escalated to a person, and every decision is written to an append-only audit log.STEADYWRK
The human manages the exceptions, not the volume
The inverse of manual. The system handles the routine volume autonomously, and human attention is reserved for the residual — the unusual scope, the high-value quote, the safety-sensitive trade, the low-confidence call. The job is no longer to dispatch; it is to manage the exceptions the autonomy hands up.
The spectrum is a way to read the category, not a scoreboard. Most facility software today sits at stage 1; "autonomous" only becomes a fair description at stage 3, when decisions execute inside a boundary rather than waiting for a click.
Human-on-exceptions
The leap from supervised to autonomous is not the removal of the human — it is the change in what the human does. Instead of approving every decision, a person manages only the exceptions the system escalates. Four things make that model safe and affordable to run.
In-the-loop means a person is in the path of every decision — the operation can only move as fast as they can click. On-the-loop means the person supervises the exceptions the system escalates and lets the rest run. The shift from in to on is what separates assistance from autonomy.
Autonomy without a brake is just a black box. STEADYWRK’s brake is an explicit confidence threshold: decisions above it execute, decisions below 70% confidence escalate to a person by design. The 3% human-override rate STEADYWRK reports is the measured residual of that line — the share of agent decisions a human steps in on.
When a decision escalates, who catches it — and at what cost — is the unclaimed part of the category. STEADYWRK’s human layer is nearshore: bilingual English and Arabic operations based in Aqaba, Jordan, working US hours at roughly 60% of US compensation. The economics of human-on-exceptions are what make bounded autonomy affordable to run at volume, instead of a cost that scales with every ticket.
Autonomy you cannot replay is autonomy you have to take on faith. Every decision the control plane makes — autonomous or escalated — is written to an append-only audit log, so a dispatch can be reconstructed after the fact. The receipt is what makes the autonomy defensible for regulated or high-value work.
Where STEADYWRK sits
STEADYWRK runs an agentic control plane at the bounded-autonomous stage. The posture below is qualitative by design — specific work-order volumes are operationally sensitive and reported as posture, not a count. The measured numbers behind it are published, dated, at /proof/autonomous-operations.
The autonomy figures are self-reported and estimated from STEADYWRK’s own production telemetry on a rolling 30-day window — not an independent third-party audit. The defensible claim is that the autonomy is measured and published openly, from a single canonical registry.
Autonomous facility management is facility maintenance where software agents — not people — make and execute the operational decisions: reading a work order, drafting a quote, routing it to a contractor, and tracking it to completion. It is best understood as a spectrum rather than a switch, running from fully manual, through assisted and supervised, to bounded-autonomous and exception-managed. "Autonomous" properly begins at the bounded-autonomous stage, where routine decisions execute on their own inside an explicit policy boundary and anything that crosses the boundary stops for a human.
No, and a system that claimed to be should be treated with suspicion. Responsible autonomous facility management is "human on the loop": the system handles the routine volume on its own, and a person manages the exceptions it escalates — the unusual scope, the high-value quote, the safety-sensitive trade. The human moves from being in the path of every decision to supervising only the ones that need judgment.
Human in the loop means a person is in the path of every decision and must approve each one — the operation moves only as fast as they can act. Human on the loop means the system acts autonomously within a policy boundary and the person supervises the exceptions it escalates. The shift from in to on is the difference between software that assists a human workflow and software that runs the workflow with a human watching the edges.
STEADYWRK operates at the bounded-autonomous stage. The boundary is an explicit confidence threshold: any decision under 70% confidence is escalated to a person, and STEADYWRK reports a 3% human-override rate — the measured share of agent decisions a human steps in on — alongside a 90% completion rate on a rolling 30-day window. These figures are self-reported and estimated from STEADYWRK’s own production telemetry, and they are published on a public, dated basis at /proof/autonomous-operations and /evals.
Two ways. First, the autonomy is bounded and measured rather than asserted — an explicit confidence threshold decides what runs on its own, with every decision written to an append-only audit log. Second, the human-on-exceptions layer is nearshore: bilingual English and Arabic operations based in Aqaba, Jordan, working US hours at roughly 60% of US compensation. That economic model is what makes catching the exceptions affordable to run at volume, instead of a cost that scales with every ticket.
It can be, when autonomy is bounded and accountable rather than open-ended. The controls that make it safe are an explicit policy boundary on every decision, automatic escalation of anything below the confidence threshold to a human, and an append-only audit log that records what the system decided and why. Together they let an operator adopt autonomy for routine volume while keeping a defensible, replayable trail for sensitive or regulated jobs.
Bring a real work order. We will show you which stage your current process is at, what bounded autonomy would change, and exactly where the human-on-exceptions checkpoint sits.