▸ proof · autonomous-operations · published May 30, 2026 · v2.0.0
Most "autonomous operations" claims are adjectives. STEADYWRK’s is a number: a 3% human-override rate — the share of agent decisions a person has to step in on — on a rolling 30-day window. Underneath it: a 90% completion rate, 340ms median routing latency, and ±9% variance between the agent’s quote and the settled invoice.
These figures are self-reported and estimated from STEADYWRK’s own production telemetry — not an independent third-party audit. The defensible claim is the sharper one: the autonomy is measured and published openly, on a public, versioned, dated basis, read from the same canonical registry that serves the live /evals endpoint. The prose and the JSON read one source, so they cannot disagree.
curl https://steadywrk.app/api/dispatch/analytics/evals?period=rolling_30d▸ the stat block
Eight operational figures, each defined, sourced, and current as of May 30, 2026. Quote any single line — every statement stands on its own with its ground truth attached. Autonomy leads, because that is the question rivals answer with adjectives.
Human override rate
3%
A human operator escalates or reverses 3% of agent decisions. Everything below 70% confidence is routed to a person by design — so the autonomy is a measured residual, not an assumption. This is the single number that answers "how much runs without a human."
Ground truth: operator audit log.
Completion rate
90%
STEADYWRK closes 90% of dispatched work orders successfully on a rolling 30-day window. Autonomy is only worth anything if the work still gets done — this is the did-the-job-close check.
Ground truth: contractor outcome + payment disposition.
Dispatch latency (p50)
340ms
Median server-side latency from work-order accept to contractor notification is 340ms. The autonomous routing decision itself — the part no human is in the loop for.
Ground truth: server-side request traces.
Dispatch latency (p95)
890ms
Tail (p95) latency for routing plus contractor outreach is 890ms. Reported alongside the median because an autonomous system is judged on its worst common case, not its best.
Ground truth: server-side request traces.
NTE variance
±9%
Final invoices settle within ±9% of the not-to-exceed (NTE) figure the agent quoted at intake — the measure of whether an autonomously-quoted price holds through to the settled invoice.
Ground truth: quoted NTE vs. settled invoice.
Quote turnaround
<2hr
A not-to-exceed quote is returned in under 2hr from work-order intake — a target, not a binding SLA. The autonomous latency-to-first-useful-output for the dispatch category.
Ground truth: intake timestamp vs. NTE-returned timestamp.
Policy-violation catch rate
Tracking
In tracking. Every decision is gated by Zod schemas and a policy layer before it acts. STEADYWRK withholds a headline number until the sample is large enough to publish honestly — naming the control without faking a value is the point.
Ground truth: pending sufficient sample.
Cost per decision
Private
Kept private to protect contractor-margin confidentiality. Operational quality is published openly; unit economics that would leak partner pricing are withheld with a stated boundary.
Ground truth: internal only.
▸ what "autonomous" means here
"Autonomous" is only meaningful if you can say what decides, what bounds it, and what gets recorded. STEADYWRK runs an agentic control plane — the layer that decides, not the workforce that executes.
The control plane is the layer that decides — not the workforce that executes. Agents handle intake, authorization, routing, and escalation; humans stay in the loop where confidence is low.
Every decision is bounded by policy. A policy gate runs on each action, and anything under 70% confidence escalates to human review automatically — which is exactly what the human-override rate measures.
Every decision is written to an append-only audit log — a 7-year, queryable record. Autonomy you can replay after the fact, not autonomy you have to take on faith.
The numbers above live behind a no-auth JSON endpoint and carry a schema version (v2.0.0). Each reading is a rolling 30-day window with a publication date — so a citation can pin exactly what it cited.
▸ operating posture
Work-order volumes are operationally sensitive and reported qualitatively — STEADYWRK publishes posture, not a count that competitors can reverse-engineer.
▸ the timed run
A start-to-finish, timed autonomous run — every work order the control plane took in over a fixed window, what it decided, and where a human had to step in — is being prepared for publication here, with the same per-figure ground truth as the readout above. Until those figures are verified and committed to the canonical registry, this page reports only the rolling-window metrics shown above; no run statistics are published yet.
▸ methodology
Source.Every figure is computed from STEADYWRK’s own production telemetry over a rolling 30-day window and read from a single canonical metrics registry (v2.0.0). The same registry feeds this page, the live evals dashboard, and the public JSON endpoint — so the prose and the data cannot disagree.
Estimated and self-reported.These figures are estimates from STEADYWRK’s production telemetry, reported by STEADYWRK. They are not an independently audited third-party benchmark, and the autonomy figures are not a leaderboard against named competitors. The honest version of the claim is that STEADYWRK measures its autonomy and publishes the measurement openly.
Dated. This reading is current as of May 30, 2026 on a rolling 30-day basis. Operational quality drifts; an honest figure is dated, not eternal. When the definition changes, the registry version (v2.0.0) changes with it, so a citation can pin exactly what it cited.
Withheld on purpose. Specific work-order volumes, cost per decision, and contractor-network counts are not public — they are operationally sensitive or would leak partner pricing. Naming a metric and stating that its value is withheld is part of the discipline; faking one is not.
companion proof surfaces
▸ questions, answered
STEADYWRK reports a 3% human-override rate on a rolling 30-day window — the share of agent decisions a human operator escalates or reverses. Everything below 70% confidence is routed to a person by design, so the autonomy is a measured residual rather than an assumption. The figure is self-reported and estimated from STEADYWRK’s own production telemetry.
It means an agentic control plane: the layer that decides, not the workforce that executes. AI agents handle work-order intake, authorization, contractor routing, and escalation; humans stay in the loop where confidence is low. Every decision is bounded by policy and written to an append-only, queryable audit log.
No. These are STEADYWRK’s own operational figures, self-reported and estimated from its production telemetry. The defensible claim is that STEADYWRK publishes its autonomy metrics openly — on a public, versioned, dated basis, read from the same canonical registry that serves the live /evals endpoint — not that an external body audited them.
Median (p50) dispatch latency from work-order accept to contractor notification is 340ms; the p95 tail latency for routing plus contractor outreach is 890ms. These are the autonomous routing decision itself, measured server-side.
Final invoices settle within ±9% of the not-to-exceed figure the agent quoted at intake, and a quote is returned in under 2hr from work-order intake (a target, not a binding SLA). On average, dispatched work closes in 48 hours and contractors are paid on a 48-hour basis.
Every figure on this page reads the same canonical metrics registry that powers the live evals dashboard at /evals and the public, no-auth JSON endpoint at /api/dispatch/analytics/evals. The page prose and the JSON read one source, so they cannot disagree.