Published on :
May 22, 2026
by
Anisha Bhattacharjee
Views...
What does human-in-the-loop (HITL) actually look like when autonomous maintenance is operating across thousands of assets?
Across a one-year period covering approximately 25,000 assets on Xempla's autonomous maintenance model, around 14% of fault decisions required human review, while approximately 86% progressed autonomously within defined operating boundaries.
The important point isn't that 86% of maintenance happened without people. Physical maintenance still requires people. The figure shows that most fault decisions did not require a person to intervene in the decision-making itself. Human involvement was concentrated where a case needed review, rather than being built into every decision.
A 14% human-review rate gives a practical view of how a HITL model can operate at portfolio scale. For a broader explanation of what human-in-the-loop means in FM and how it fits into governed maintenance workflows, see What Is Human-in-the-Loop in FM AI Systems? In this dataset, most fault decisions progressed without review, while cases that crossed defined boundaries reached a person.
The 14% figure isn't a fixed benchmark for HITL. It reflects how the autonomous maintenance model was operating across this portfolio: fault decisions could progress independently within defined boundaries, while cases requiring human review were routed accordingly.
The balance can look different across deployments. What matters is not achieving a particular percentage, but deciding where human involvement belongs in the operating model.
At portfolio scale, this has a practical implication for FM teams. Instead of spending time reviewing every incoming signal, people can focus their attention on exceptions, specialist decisions, approvals and the physical work that still requires a person.
Three things stand out.
Autonomy does not mean removing people from maintenance. AI can handle part of the decision-making process independently while people retain responsibility for decisions and work that require them.
The percentage matters less than what sits behind it. A 14% review rate is useful because it shows how the workflow allocated human attention across the portfolio. It is not a target that every FM organisation should try to reproduce.
Good HITL design is about allocation, not elimination. The question is not how quickly the human-review rate can be pushed towards zero. It is whether the workflow sends the right cases to the right people at the right point.
That's what human-in-the-loop looks like when autonomous maintenance operates at scale: not a human checking every AI decision, but a defined operating model for deciding when human involvement is needed.
The data shows how human involvement was distributed across a portfolio. What happens to an individual fault from the first deviation through investigation, decision, and verification is what we've walked through in From Fault to Outcome: How the Autonomous Maintenance Workflow Runs.
Across one year of autonomous maintenance data covering approximately 25,000 assets, around 14% of fault decisions required human review, while approximately 86% progressed autonomously within defined operating boundaries.
Human-in-the-loop allows an autonomous maintenance system to progress decisions within defined boundaries while routing cases that require human review, authority, expertise or physical intervention to a person.
No. Requiring a person to review every decision would limit the value of autonomous decision-making. A HITL workflow defines when human involvement is required and allows other decisions to progress independently.
Not necessarily. The review rate depends on how the workflow is designed, the boundaries established for autonomous decisions, and the types of cases being handled. The objective is to involve people where their input matters.
Not necessarily. Some decisions will continue to require human authority, specialist expertise or physical intervention even as autonomous decision-making expands.
© Xempla. All rights reserved.