At 2:14 on a Tuesday morning, a packaging line in a plant nobody was standing in stopped itself. A vision camera had flagged a run of cartons whose glue seams were drifting out of spec, the line controller throttled the conveyor, a work order went to the on-call tech's phone, and by the time the first shift walked in at six the line was running clean again with a logged note explaining what happened. That is the dream the word "lights-out" sells: a line that runs, watches, and corrects itself while the building sleeps. It is real, and on that one cell it worked. But walk fifty feet down the same plant and you find the truth nobody puts in the keynote. The assembly cell next door still needs two operators per shift because the parts arrive in a tangle no robot can reliably pick, the welding station throttles itself but cannot decide whether a borderline weld ships or scraps, and the final inspection booth still has a human in it because the customer's quality standard says a human signs the record. The autonomous line is not a switch you flip for the whole plant. It is a gradient, cell by cell, where the physics, the economics, and the accountability decide what actually goes dark and what stays staffed. This lesson is about reading that gradient honestly, so when an executive asks "why are we not lights-out yet," you can answer with the loss chart and the constraint, not the hype.

What Lights-Out Actually Means, and Where It Is Real

The phrase "lights-out manufacturing" gets used as if it were a single destination, but on the floor it describes a spectrum. At one end is a fully attended line where every station has an operator. At the other is a cell that runs unattended for a full shift, watching itself, correcting small deviations, and calling a human only by exception. Most plants in 2026 live somewhere in the middle, and the honest framing is not "are we lights-out" but "which cells can run unattended for how long before a human has to decide something."

The cells that genuinely go dark share three traits. The process is stable and well understood, meaning the same inputs reliably produce the same output without a human making judgment calls. The inputs are clean and presented consistently, so a machine is not fighting variation it cannot model. And the decisions the cell makes are bounded, meaning every choice the automation faces has a predefined right answer the engineers already encoded. A CNC machining cell loading from a pallet of identical fixtured blanks fits this profile. A high-volume bottling or capping line fits it. A pick-and-place surface-mount line placing components on a circuit board fits it, because the board, the parts, and the placements are all defined to the micron.

Where it breaks is variation the automation was not designed to handle. The classic example is a part that arrives wet, or with the lighting changed between shifts, or at a camera angle that has drifted two degrees since the cell was commissioned. The cell does not know it is now wrong; it keeps running on assumptions that no longer hold. This is why the same plant can run one cell unattended overnight and need two operators on the cell beside it. The difference is not ambition or budget. It is whether the process and its inputs are stable enough that the bounded decisions stay correct without a human watching.

The market reality anchors this. Adoption of AI in quality reached 47% of manufacturers in 2026, up from 33% the prior year, which tells you the watching part (a camera grading parts, a model flagging a failure) is spreading fast. But adoption of AI in direct control of anything that moves lags far behind, because watching is advisory and controlling is consequential. A model that flags a defect and is wrong costs you a false reject. A model that controls a press and is wrong costs you a crushed fixture or a hurt person. The gradient from advisory to autonomous is the gradient from cheap mistakes to expensive ones, and plants move along it exactly as fast as the cost of being wrong allows.

Lights-out is not a plant-wide switch. It is a per-cell question: can the bounded decisions stay correct without a human watching, for how long, before the cost of being wrong gets too high.

The Three Jobs on a Line, and Which One Goes Dark First

To predict what goes autonomous, separate the work a line does into three jobs: moving the part, judging the part, and deciding what to do about the judgment. These three jobs automate at completely different rates, and conflating them is how a plant buys a "lights-out" promise that never arrives.

Moving the part is the most mature. Conveyors, robots, pick-and-place, automated guided vehicles: this is the oldest automation on the floor, and where the part presentation is consistent, it runs unattended reliably. The limit is not the motion; it is the variation in what is being moved. A robot picking identical fixtured blanks runs lights-out. The same robot picking randomly tangled parts from a bin still fails often enough that a human stands by, because bin-picking of high-variation parts remains genuinely hard. So the rule on motion is: it goes dark when the thing being moved is presented the same way every time.

Judging the part is where AI made its biggest recent gains. A vision model grading a cosmetic surface, a predictive model reading vibration on a bearing, a dimensional check at line speed: these are the 47% adoption number in action. The machine judges faster and more consistently than a tired inspector on a third shift. But judging has a quiet trap that decides whether it can run unattended: the false-reject rate (FRR, the share of good parts the system wrongly rejects). A vision system can run unattended only if its false-reject rate is low enough that nobody needs to stand there overruling it. If the FRR is high, you have not removed the human; you have given the human a new job of babysitting a twitchy alarm, and an operator who has been burned by false alarms will eventually disable the green light, which destroys the value entirely. So judging goes dark when its error rate, in both directions, is low enough that the exceptions are rare.

Deciding what to do about the judgment is where autonomy hits its hardest wall, and where it stays staffed the longest. A camera can judge that a weld is borderline. Deciding whether that borderline weld ships to the customer, gets reworked, or scraps the assembly is a different kind of decision, because it carries accountability. Under the cardinal rule of this whole program, the customer audits the plant, not the vendor, and "the model decided to ship it" is never an acceptable answer in a containment meeting or a customer audit. So the disposition decision (ship, rework, scrap, hold) tends to stay human even on cells where the moving and the judging are fully automated. The human is not there because the machine cannot judge. The human is there because somebody has to own the consequence.

Work the math on a real cell to see how these three jobs interact. Say a final-inspection booth handles 1,000 parts a shift, and the true defect rate is 1%, so 10 real defects and 990 good parts. A vision model judges them at 95% recall and a 3% false-reject rate. It catches roughly 9 to 10 of the real defects, good, but it also wrongly rejects about 30 of the 990 good parts (3% of 990 is 29.7). That is 30 false rejects per shift, 90 per day across three shifts. If a false reject costs five minutes of operator time to clear plus the risk of scrapping a good part worth $40, the false rejects alone burn real money and real attention every single shift. The judging is automated, but the disposition of those 30 borderline calls keeps a human in the loop. The cell is not lights-out; it is human-light, and the limiting factor is the false-reject economics, not the camera.

The Economics That Decide, Not the Technology

The most common mistake leaders make about autonomous lines is treating it as a technology question when it is almost always an economics question. The technology to run many cells unattended exists. Whether it pays to run them that way depends on volume, mix, the cost of being wrong, and the brownfield reality of the plant you actually have.

Volume and mix decide the payback. Automation amortizes over throughput. A cell running one stable high-volume product for years pays back the engineering, fixturing, and integration cost of full autonomy. A job shop running 200 part numbers in lots of 50 cannot, because every changeover reintroduces the variation that autonomy cannot absorb, and the engineering cost to make each of 200 parts run unattended never amortizes. This is why high-volume discrete and process plants go lights-out on specific cells while job shops stay heavily attended. It is not that the job shop is behind. It is that the economics genuinely do not support it for low-volume high-mix work.

The cost of being wrong sets the autonomy ceiling. Recall the program's core fact that a vision system's false-reject rate can quietly cost more than the escapes it catches. The same logic governs autonomy. The more expensive a wrong autonomous decision is, the more a human stays in the loop, regardless of how good the model is. A capping line where a wrong call wastes a $0.04 cap will go autonomous early. A line making a safety-critical aerospace bracket under AS9100 (the aerospace quality management standard) where a wrong call ships a part that could fail in flight keeps a human on the disposition forever, because the cost of being wrong is unbounded. The autonomy ceiling is set by the worst-case cost of an autonomous error, not by the average accuracy of the model.

Brownfield is the default, and it taxes everything. Most readers run a 1990s PLC (programmable logic controller, the industrial computer that actually runs the machine) and a historian (the time-series database logging sensor tags) nobody has queried in years. Greenfield plants deploy AI 40 to 60% faster than brownfield plants for exactly this reason: a new plant can design the data, the network, and the cell layout for autonomy from day one, while a brownfield plant retrofits autonomy onto equipment that was never instrumented for it. When a board asks why the reshored greenfield site is more autonomous than the flagship brownfield plant, this is the honest answer. The greenfield site did not try harder. It started without the 30-year tax.

Put it in a worked number a CFO understands. Suppose making one assembly cell fully unattended costs $480,000 in robotics, vision, fixturing, integration, and the safety system, and it saves two operators per shift across three shifts. At a fully loaded labor cost of roughly $65,000 per operator-year, two operators across three shifts is six positions, around $390,000 a year, but realistically you keep one roving tech to handle exceptions, so the net labor saving is closer to $325,000 a year. The cell pays back in about eighteen months and then runs ahead. Now run the same math on a cell that changes over twice a day and needs an operator to reset fixtures each time: the labor never fully comes out, the payback stretches past five years, and the project dies in the business-case review. Same technology, opposite decision, and the only variable that changed was the mix.

The OT Boundary and the Accountability Wall

There are two hard walls that no amount of model accuracy gets you past, and a leader who pretends otherwise will get a plant hurt. The first is the OT boundary. The second is accountability.

The OT boundary is the line between IT (information technology, the business network of email and ERP) and OT (operational technology, the network of PLCs, SCADA, and the machines that move). On the OT side, a wrong command does not corrupt a spreadsheet; it moves a thousand pounds of steel. The program's anchor fact is blunt: 78% of OT networks lack centralized monitoring. You cannot make a line autonomous on a network you cannot see. If you cannot centrally monitor the OT environment, you cannot detect when an autonomous cell is being fed bad data, drifting out of calibration, or behaving abnormally, and an unmonitored autonomous cell is a more dangerous version of the same cell with a human watching. So the prerequisite for autonomy is not a better model. It is OT visibility, and most plants do not have it yet. The honest sequencing is: get the network visible, then keep AI advisory, then automate the bounded low-consequence decisions, and only then consider unattended operation on the cells where the cost of being wrong is low.

The rule that follows is one this program repeats deliberately: keep AI advisory and out of direct control of anything that moves unless it is properly governed. Advisory means the AI recommends and a human or a hard-coded deterministic interlock acts. A safety-critical control loop is not where AI goes first, and on many cells it is not where AI goes at all. The autonomous packaging line from the opening worked because throttling a conveyor and sending a work order are low-consequence, reversible actions inside a bounded envelope. The same architecture would be reckless on a press that can crush a hand. Autonomy advances on the cells where a wrong move is cheap and reversible, and it stops cold at the cells where a wrong move is expensive or dangerous.

The second wall is accountability, and it does not move with technology at all. The customer audits the plant, not the vendor. When an aerospace customer's auditor walks the floor, or an automotive customer invokes IATF 16949 (the automotive quality management standard), they hold the plant accountable for every quality decision, including the ones an autonomous cell made. "The line decided autonomously" is the manufacturing version of "the model said no," and it fails the audit exactly the same way. This is why the disposition decision and the signed quality record stay human on regulated and customer-audited work even when everything around them is automated. The plant keeps a human in the loop not because the human is faster or more accurate, but because the human is the entity the customer and the auditor can hold responsible. Autonomy can remove the operator from moving and judging. It cannot remove the accountable human from the record, because the record is what the customer audits.

Think about what this means for the talent cliff that drives this whole program. With roughly 2 million manufacturing workers needing reskilling by 2026 against about 500,000 unfilled roles, and 85% of manufacturers saying staffing shortages are hurting product quality, the appeal of autonomy is obvious: if the line runs itself, you need fewer of the people you cannot hire. But the accountability wall flips this. Autonomy does not eliminate the skilled human; it moves the skilled human from doing the work to owning the exceptions and signing the records. The plant that goes autonomous well does not have fewer skilled people. It has skilled people doing higher-judgment work (handling exceptions, verifying autonomous decisions, owning the audit trail) instead of lower-judgment work (standing at a station moving parts). Autonomy is a knowledge multiplier for a thinner crew, not a replacement for the crew.

The Honest Roadmap from Attended to Autonomous

If you accept that autonomy is a per-cell gradient governed by physics, economics, and accountability, then the roadmap to it is not "buy a smart factory." It is a disciplined sequence you can run cell by cell, and it is the same sequence whether you have one plant or a network of them.

Stage one: instrument and make it visible. You cannot automate what you cannot measure, and you cannot run unattended what you cannot monitor. This stage is the unglamorous work of getting sensors on the cell, getting the historian populated, and closing enough of the 78% OT-monitoring gap to actually see the cell when no one is standing there. A plant that skips this stage and jumps to autonomy is flying blind, and the first unmonitored failure teaches the lesson the expensive way.

Stage two: AI watches, humans decide. Deploy the judging models (vision QA, predictive maintenance) in advisory mode. The model flags, the human decides. This is the 47% adoption stage, and it is where most plants should be living in 2026. It captures most of the quality and uptime value while the cost of a model error stays bounded to a false alert. Measure the false-reject rate relentlessly here, because the FRR you accept in advisory mode is the FRR that will determine whether the cell can ever run unattended.

Stage three: automate the bounded, low-consequence decisions. Where the decision has a predefined right answer and the cost of being wrong is low and reversible, let the cell act on its own judgment: throttle the conveyor, divert the suspect part to a hold bin, raise the work order. Keep the high-consequence decisions (ship versus scrap, anything that moves with force near a person) human. This is the packaging-line architecture from the opening, and it is the realistic ceiling for most cells.

Stage four: unattended operation on the qualified cells only. A cell graduates to running unattended for a shift only when the process is stable, the inputs are consistent, the error rates in both directions are low, the cost of any autonomous error is bounded, the OT environment is monitored, and the accountability for the records is clearly assigned to a human who reviews them. Most plants will have a handful of cells that qualify and many that never will, and that is the correct outcome, not a failure. The skill is knowing which cells belong in which stage and being honest about the ones that will stay attended for the foreseeable future.

The number that makes this concrete: a plant that moves three high-volume cells from stage two to stage four, each saving a net $325,000 a year in labor while holding first-pass yield steady and cutting unplanned downtime through the predictive layer, books close to a million dollars a year in defensible savings, and it does so without pretending the job shop down the aisle can do the same. That selective, honest autonomy is what survives a board review. The plant-wide lights-out promise is what gets a leader fired when the reshored line still needs operators a year after the press release.

What This Means for the Leader Who Has to Answer the Question

An executive or board member who toured a competitor's "smart factory" will ask why your plant is not lights-out. The graduate of this program does not answer with enthusiasm or apology. They answer with the gradient. They walk the loss chart cell by cell and say: these three cells qualify for unattended operation because the process is stable, the inputs are consistent, the error cost is low, the OT is monitored, and the accountability is assigned, and here is the roughly $325,000 a year each one saves. These cells stay attended because the mix is too high, the cost of a wrong call is unbounded, or the OT is not yet visible, and here is the cost and the timeline to change that for the ones worth changing. And these cells will stay attended indefinitely because the economics or the accountability will never support autonomy, and that is the right call, not a gap to apologize for.

That answer does three things a hype answer cannot. It tells the truth, which protects the leader's credibility when the autonomous line still needs people next year. It ties every autonomy decision to a dollar figure and a constraint, which is the language the board funds. And it keeps the human accountability the customer audit requires, which protects the plant from shipping an unowned decision into a containment. The leader who can give that answer is not the one who promised lights-out. They are the one who delivered selective, defensible, audited autonomy where it actually paid, and kept the crew safe and the records signed everywhere else.

Key Takeaways

• Lights-out is not a plant-wide switch but a per-cell gradient: a cell goes unattended only when the process is stable, the inputs are consistent, the decisions are bounded, and the cost of an autonomous error is low and reversible.
• Separate the three jobs on a line: moving the part automates first, judging the part is where the 47% AI-quality adoption lives, and deciding what to do about the judgment (the disposition) stays human longest because it carries accountability.
• The false-reject rate is the hidden gate on unattended judging: a 3% FRR on 990 good parts is roughly 30 false rejects a shift, which keeps a human babysitting the cell and can cost more than the escapes the system catches.
• Autonomy is an economics decision, not a technology one: high volume and stable mix pay back full autonomy (a worked cell at about $325,000 net annual labor saving, eighteen-month payback), while high-mix job-shop work never amortizes and should stay attended.
• Two hard walls cap autonomy: you cannot run unattended on an OT network you cannot see (78% lack centralized monitoring), and you cannot remove the accountable human from the signed quality record because the customer audits the plant, not the vendor.
• Keep AI advisory and out of direct control of anything that moves unless properly governed; a safety-critical control loop is not where AI goes first, and autonomy advances only on cells where a wrong move is cheap and reversible.
• Against the talent cliff (2 million needing reskilling, 500,000 roles unfilled, 85% saying shortages hurt quality), autonomy moves skilled people from doing the work to owning the exceptions and signing the records, multiplying a thinner crew rather than replacing it.
• The leader's job is to answer "why are we not lights-out" with the cell-by-cell gradient, a dollar figure, and a constraint for each cell, delivering selective audited autonomy where it pays and honestly defending the cells that will stay attended.