The lights-out factory is a distraction. Bounded autonomy wins.
Every public lights-out case study has 200–400% maintenance technician attrition or a hidden human layer just off-camera. The constrained-agency factory floor is the actual scalable pattern.
The lights-out factory is one of the most-photographed concepts in manufacturing journalism. Robots assembling cars in eerie blue light. Empty shop floors with a single human supervisor in a glass control room. The headline reads: this is the future. The footnote, if you can find it, reads: maintenance technician attrition rate at this site is 280%, the lights-out shift produces 30% of design throughput, and the parts of the line that are not photographed contain six people doing the work the robots could not.
The lights-out story is a marketing artifact, not an operating-model template. The factories that actually scale agentic systems run on bounded autonomy: high agent density on well-bounded tasks, named human accountability per workflow, factory-pattern deployment, and constraint engineering that survives shift handoff. The lights-out narrative misdirects operators into chasing a configuration almost no one operates profitably, while ignoring the configuration most operators can ship in 18 months.
Why bounded autonomy wins.
Lights-out is a marketing optimization, not an operating optimization.
The lights-out installation produces a photograph that funds the next round of capital. It rarely produces the unit economics the photograph implies. The operating optimization — the configuration that maximizes EBITDA per square meter of plant floor — is bounded autonomy: more agents per workflow, tighter constraint design, smaller human teams handling exceptions and escalation. The photo is worse. The economics are better.
Maintenance is the variable lights-out economics fail to model.
Lights-out lines treat maintenance as a fixed cost. In practice, maintenance is the binding constraint. Highly autonomous lines have higher mean-time-to-repair because the technicians who can fix them are scarce, expensive, and often poached by the next site over. Bounded autonomy intentionally keeps the line repairable by the workforce the operator can actually retain — a real-world constraint the lights-out narrative repeatedly underestimates.
The bounded-autonomy floor is what the agentic stack was built for.
Every primitive in the constraint stack — Microsoft Entra Agent ID, Copilot Studio scoping, Purview lineage — is designed for bounded delegation with named accountability. Lights-out is the configuration least supported by the platform, because the platform assumes a human-in-the-loop architecture. The operator who tries to ship lights-out on Copilot Studio is fighting the platform; the operator who ships bounded autonomy is riding the platform.
The strongest argument against this position.
The strongest counter is that some narrow categories of manufacturing — semiconductor fabrication, certain pharmaceutical purity processes, deep-sea or hazardous-environment production — genuinely require near-lights-out configurations because human presence is impossible or contraindicated. This is right. The argument here is not that lights-out is never the right answer; it is that it is rarely the right answer, and the analyst narrative misallocates attention by treating it as the headline configuration. For 90% of manufacturing operations, bounded autonomy is the right answer; for the 10% where lights-out applies, lights-out is the right answer because of the specific physical or safety constraints — not because it produces a better photograph.
Three things to do this quarter.
01 · Stop benchmarking against lights-out lines. They are not your peer set unless you are in semi or pharma. Benchmark against the bounded-autonomy operators in your own segment.
02 · Calculate maintenance-availability cost into every autonomy investment. Use a 5-year horizon. If the math depends on retaining maintenance technicians at competitive rates, name how. If it doesn’t, the math is wrong.
03 · Set bounded-autonomy targets per shift, not lights-out targets per facility. The operator metric is “agents per shift, named owner per workflow, constraint engineering coverage” — not “hours of unattended operation per quarter.”