The Lights-Out Factory Is a Labor Theory in a SKU Complexity World

A CEO at Hannover Messe in April 2026 walked past a humanoid-robot booth, watched the demo for about ninety seconds, and said to the colleague next to him, "this is the cheapest decade of labor we will ever see."

He had the math backwards.

Direct labor is roughly ten to fifteen percent of cost of goods sold for most discrete manufacturers in North America. Materials and SKU-driven overhead are sixty to seventy percent. The lights-out factory pitch is a promise to take a wedge out of a fifteen percent line. Meanwhile changeover time, short-run yield, engineering change order propagation, and SKU-level forecast error are eating two to three times more margin. The reshoring wave that is supposed to make all of this matter is structurally high-mix. AI in manufacturing is not undersold in 2026. It is misaimed.

The Labor-Cost Theory the Pitch Depends On

Walk the floor at Hannover this year and the narrative is consistent across booths. AI plus humanoids plus simulation equals the lights-out factory. The implicit financial case is that headcount is the dominant operating cost in a modern plant, and removing it is the dominant lever.

That case stopped being true about a decade ago for most North American discrete manufacturers. Industry benchmarks from APQC, the 2024 MPI Manufacturing Study, and standard cost-of-quality references put direct labor in the ten to fifteen percent of COGS range for the majority of producers in automotive, aerospace, electronics, and industrial equipment. Materials are typically forty-five to fifty-five percent. Overhead, much of it driven by complexity rather than direct production, is another twenty to thirty percent. COGS itself has climbed from sixty percent of revenue in 2020 to sixty-seven percent in 2024 across the MPI sample, and almost all of that increase is materials and complexity-related overhead, not direct labor.

Take the most aggressive lights-out claim a vendor is willing to make on the record. Sixty percent headcount reduction in a specific production area. Apply it to a twelve percent direct labor line. The arithmetic gets you a 7.2 point reduction in COGS in the affected area, before you net out the capital cost of the robots, the integration work, the new maintenance contract, the ongoing software licensing, and the engineering team you have to hire to keep the system running. A conservative net is three to four points of COGS, in the area where you deployed.

That is a real number. It is also a number that is materially smaller than what a competent SMED program, a short-run yield project, or a serious SKU rationalization will deliver in the same plant in the same year. The lights-out factory does not lose to those projects because the technology is bad. It loses because it is aimed at the wrong cost line.

What Is Actually Binding in 2026

Here are the four lines on the P&L that, for most discrete manufacturers I see, are eating real margin right now.

Changeover time. A typical mid-mix plant runs somewhere between fifteen and forty changeovers per week, with average duration between forty-five minutes and three hours. Plants on a manual SMED program are leaving forty to sixty percent of that time on the table, according to published OEE benchmarks. For a plant running thirty changeovers a week at an average of ninety minutes, an honest SMED program with sensor and computer-vision instrumentation can recover ten to fifteen hours per week of productive capacity. That is not a wedge of the labor line. That is throughput, against a fixed cost base that includes materials and overhead.

Short-run yield. The yield gap between long-run and short-run production on the same line is consistently in the five to fifteen percentage-point range in industries I have looked at. On a line producing high-margin parts, three points of recovered yield on short runs can move the gross margin of those runs from break-even to acceptable. The capital required is sensor instrumentation and a small AI model trained on a few hundred runs. The headcount affected is zero.

ECO propagation lag. An engineering change order in a serious manufacturer triggers cascading updates across bills of material, routings, purchase orders, work instructions, supplier qualifications, and quality records. In aerospace and medical, the same ECO can take four to twelve weeks to fully propagate through a tier-two supply chain, and the cost of that lag shows up as expedite charges, scrap from old-revision parts, and field rework. Even a thirty percent reduction in propagation time, achievable with an AI agent that routes the ECO and pre-validates downstream documents, produces a measurable margin improvement that no lights-out demo can match.

SKU-level forecast error. As SKU counts grow, traditional forecasting methods perform worse on the long tail. Published research from ToolsGroup and others puts forecast value-added for top sellers at roughly four times the value-added for long-tail items, which is a polite way of saying long-tail forecasts are barely better than guessing. The carrying cost of that error shows up as obsolete inventory, expedite charges on misforecast components, and lost margin on stock-outs. For a plant with five thousand active SKUs, taking five points of inventory cost off the long tail beats the entire lights-out pitch.

None of these four require fewer humans on the floor. All four are AI problems that vendors are under-selling because they do not film as well as a humanoid pouring a beer.

The Reshoring Wave Is Structurally High-Mix

This is the part that makes the misaiming structural, not anecdotal.

The Reshoring Initiative's 2024 annual report counted 244,000 manufacturing jobs announced in the United States via reshoring and foreign direct investment. Semiconductors accounted for thirty-five percent of those announcements. Electrical equipment, dominated by EV batteries and solar, accounted for thirty-one percent. Two sectors, two-thirds of the wave. Add transportation equipment, defense, and biotech and you have over eighty percent of the announced capacity.

Look at what those sectors actually produce.

TSMC's Arizona Fab 21 is running 4nm at full volume, with 3nm equipment install starting summer 2026 and 2nm in Phase 3. A modern advanced node involves more than a thousand process steps, and a fab routinely runs multiple nodes, multiple device families, and multiple customer-specific masks in the same facility. The defining operational problem is not headcount. It is yield against process drift across a wildly varied mix.

Ford and SK On split BlueOval SK in 2026, and both halves of the divorce produce batteries in at least two chemistries (NCM and LFP) across at least two form factors (pouch and prismatic) for at least four vehicle programs. A modern cell has on the order of 3,600 control points. The economics of that plant turn on yield-per-variant and changeover between chemistries, not on how many people are on the line.

GE Aerospace's Greenville plant ships about a thousand high-pressure turbine blades a day across multiple CFM56 configurations including Tech Insertion and PIP. Defense work adds further variants under different traceability regimes. The plant has been on a lean program for years. The remaining margin sits in ECO propagation, short-run yield, and supplier qualification, not in removing the humans who already barely touch the blade.

Eli Lilly's Indiana expansions are batch biologics. Tier-one EV battery customization is per-vehicle. Defense components are per-program with security overlays. Custom industrial equipment is per-customer by definition.

This is what the largest greenfield manufacturing wave in forty years actually looks like. High mix. Tight tolerance. Heavy regulation. Long bills of material with frequent revisions. Selling a lights-out factory into that environment is selling a labor theory into a complexity world. The marketing video does not match the P&L.

Where the Leverage Actually Is

If you are a CFO writing the AI manufacturing budget for the next eighteen months, here are the four surfaces that beat lights-out economics in the kind of plant the reshoring wave is actually building.

Changeover optimization. Sensor and vision-instrumented SMED that converts internal setup steps to external in real time, learns from prior changeovers, and flags drift on critical fixtures. Capex is modest. The payback is in throughput on existing equipment with existing crews. The vendor pitch is unsexy. The margin is real.

Short-run zero-shot defect detection. Vision models that detect defects on new part variants without per-variant retraining, using foundation-model embeddings plus a small labeled set. Closes the long-run versus short-run yield gap that is structurally present in any high-mix plant. Beats traditional machine vision by an order of magnitude on time-to-deploy for a new SKU.

ECO propagation routing. Agentic workflow that takes an engineering change order, identifies every affected BOM, routing, work instruction, supplier qualification, and audit record, drafts the downstream changes, routes them to the named human for approval, and posts under separation-of-duties. Eats the propagation lag that costs aerospace and medical manufacturers tens of millions a year. Pairs naturally with the agent control plane work I have written about elsewhere on this site.

SKU-level production scheduling. Mixed-integer optimization combined with intermittent-demand forecasting on the long tail. Cuts inventory and scrap on the SKUs traditional planning systems are worst at. Frees the planners who are currently spending eighty percent of their time on the twenty percent of SKUs that matter least.

Each of these has the property that the lights-out pitch lacks. The benefit hits a sixty percent line on the P&L, not a twelve percent line. The capital required is software and sensors, not capital robots. The integration is into the systems of record the manufacturer already has, not a new mechanical layer in front of them. None of these require a headcount story to clear the investment committee, which is part of why they do not get pitched.

Why Vendors Sell Lights-Out Anyway

A one-pager that says "remove sixty percent of the labor in this cell" fits on a slide. A CFO can model it. A board can approve it. The ROI calculation is a multiplication.

A one-pager that says "reduce SKU forecast error on the long tail by twenty-two percent and recover four points of working capital across the affected categories" requires the CFO to actually understand the working-capital impact of the long tail of SKUs, which is something most CFOs at the discrete-manufacturer scale have never built a model for. The ROI calculation has eight variables and a confidence interval. The board glazes over.

The vendor sells what the buyer can model. The buyer models what the vendor can pitch. Both walk out of the meeting feeling smart. Both are wrong about where the margin is.

The robotics-vendor incentive is also worth naming. A humanoid demo is the most expensive piece of capital equipment in the building. The deal size is large. The discount conversation is easy. The implementation services attach is enormous. There is no equivalent business model around an ECO-routing agent that costs a tenth as much and delivers more margin. The vendor ecosystem points its best salespeople at the wrong target by economic gravity.

A Buyer's Framework

Four questions, in order, before signing the capital request for the lights-out scenario.

What percentage of your cost of goods sold is direct labor today, by plant and by line? If the number is below twenty percent for any plant where the lights-out pitch is being made, the project economics have a ceiling that almost certainly fails to clear a serious capital threshold. If you do not know the number, the project is not ready to be evaluated.

How many changeovers per week, at what average duration, in the target plant? Multiply duration by frequency. If the answer is more than ten hours per week, your dominant operational loss is changeover, not headcount, and the AI project that pays back is not the lights-out one.

What is the delta between your short-run yield and your long-run yield on the same line? If the gap is more than five percentage points, your dominant quality loss is short-run, not long-run, and your highest-leverage AI investment is short-run defect detection, not lights-out.

What is the five-year trend on your active SKU count? If it is growing more than five percent a year, your structural problem is complexity, not headcount, and any AI investment that does not directly attack complexity is solving a problem you do not have.

If three of the four answers point at mix and complexity, the lights-out pitch is the wrong purchase. That does not mean robotics is wrong. It means robotics is a 2030s conversation about reliability and unit economics on long-run cells, not a 2026 conversation about reshoring discrete manufacturing.

Close

This is the largest greenfield manufacturing build-out in North America in forty years. The decade ahead will be defined by what gets installed in the next three to five years. If the operating system of the next manufacturing decade is a labor-substitution theory imported from the simulation videos, the wave will leave a generation of high-mix plants under-instrumented at the layer where their real margin lives. The wrong theory now compounds for a decade.

AI in manufacturing is not undersold in 2026. It is misaimed. The vendors selling lights-out are selling the right technology to the wrong cost line. The CFOs buying it are approving the easy ROI instead of the real one. The reshored plants being built today are structurally complexity-bound, not labor-bound, and the AI surfaces that match that reality are sitting unloved because they do not film well.

If you are sketching the AI budget for a reshored or expanded plant and want a second pair of eyes on whether the spend is aimed at the right cost line before it goes to your audit committee, reach out. The next twenty-four months are the budget window for the decade.

Shubhendu Tripathi is an AI and ERP strategy consultant based in Toronto, advising organizations on digital transformation, enterprise AI adoption, and technology leadership. Connect on LinkedIn or reach out at tripathis@qubittron.com.