We measured sustained autonomous resolution rate — the share of incidents closed end to end with no human action — across production deployments in six industries, after each had passed 90 days post-observe-mode. The range is narrower than most people expect, and the reasons for the variance are almost entirely organisational.
The numbers
A 7-point spread across six industries is remarkably tight. It tells you that the underlying failure physics of enterprise infrastructure does not vary much by sector — a full filesystem behaves identically whether it is under a core banking system or a warehouse management system.
Banking: lowest rate, strictest gates
Banking's 68% is not a technical ceiling. In the same estates, the share of incidents that were *technically* autonomously resolvable measured 74%. The gap is governance: a larger proportion of policies are deliberately held at approve-then-act because the change advisory process requires a named human on every production change touching systems in regulatory scope.
This is a legitimate choice, not a failure. What autonomy buys these estates is not the removal of the human, but the removal of the diagnosis and preparation work — the engineer approves a fully-formed action with evidence attached, rather than starting from a red alert at 2am.
Healthcare: high rate, narrow windows
Healthcare reaches 72% despite equally strict compliance, largely because clinical systems have unusually well-documented runbooks — regulatory pressure has already forced procedural rigour that translates directly into policy definitions. The constraint here is timing rather than approval: many actions are prohibited during clinical hours and queue for defined windows.
Retail: the counter-intuitive peak
Retail hits 75% during high-traffic events — Diwali, Black Friday, end-of-quarter sales — which surprises people who assume peak load means more chaos. In practice, peak-event incidents are *more* repetitive: capacity exhaustion, connection pool saturation, cache pressure, and queue backlog, over and over across a horizontally scaled fleet. Repetition is exactly what autonomy is good at.
Off-peak retail rates settle around 66%, where the incident mix is more varied and includes more one-off integration failures.
What actually predicts a high rate
Regressing autonomy rate against estate characteristics, three factors dominate:
- Standardisation. Estates with a small number of OS builds, consistent agent deployment, and uniform naming conventions score 8 to 11 points higher than heterogeneous estates of the same size.
- Runbook maturity. Teams that already maintain written, tested runbooks convert to policies in days. Teams relying on tribal knowledge spend the first month writing down what they know.
- Monitoring hygiene. Estates with high false-positive rates depress the measured autonomy rate because noise alerts are correctly escalated rather than acted upon.
Measuring it honestly
Beware vendors quoting autonomy rates against alert volume rather than incident volume. Deduplicating a storm of 400 alerts into one incident and resolving it counts as one resolution, not four hundred. Every number in this article is measured against deduplicated incidents, and we recommend you hold any vendor to the same definition.
The second honest measure is sustained rate after 90 days, not first-month rate. Early rates are flattered by the fact that the easiest, most repetitive signatures are automated first.
“If your autonomy rate is climbing every month with no new policies written, you are not measuring autonomy. You are measuring alert drift.”
Arjun Kumar, Chief AI Officer