Ghost Capacity: The Third Shift You Already Own

The Shortage That Wasn't There

A protein co-packer thought it was running out of room. The onsite read said the lines were tight, growth was coming, and the plant would need more cook capacity to absorb it. That is the moment most plants reach for capital. It is also the moment the spreadsheet quietly takes over the decision.

We did not buy the read. We asked for a year of production data instead of a month, specifically to see the variability over time, the seasonality, the weeks that did not look like the average. The data came back clean enough to build on. We built a digital twin from it and ran thousands of simulations until the model landed within 2 percent of what the plant actually did. A model that is 98 percent accurate to real life is not a forecast. It is a mirror.

The mirror said something the onsite read could not. Across four scenarios, from incremental automation to aggressive line consolidation, the plant could roughly triple in size before cook capacity became the constraint. The capacity everyone thought was missing was already sitting on the floor. It was just filed under the wrong line.

Why Capacity Goes Ghost

Ghost Capacity is the gap between the number on the spreadsheet and the throughput the plant can actually access. It shows up in two shapes. The phantom shortage, where the math says you are short and the floor is not. And the phantom abundance, where nameplate says you have the capacity and the line cannot hold it. Both come from the same move: the spreadsheet flattens a varying process into a single average, and the average erases the exact thing that governs real output.

Take the mechanism this plays out in most often. Incoming material variance changes how long product has to sit in the oven. Moisture content drifts batch to batch, and dwell time has to move with it to hit the target. The spreadsheet models the oven at one speed per product. The floor runs it at whatever speed the material demands that shift. The capacity you budgeted against is the average speed. The capacity you get is the slow-material speed, and nobody put the variance in the model.

The same flattening hits the cost side. A molder watching HDPE move 40 cents a pound knows the spreadsheet line item called "resin" is a range, not a number, and the range is wide enough to swallow the savings the project was supposed to deliver. Upstream variance does not stay upstream. It propagates through every downstream step until it lands on the one number the budget cycle is staring at, and by then it looks like noise.

There is a tell that the baseline itself is broken. In that year of co-packer data, seven weeks showed OEE above 100 percent. That is not a great week. That is a structural data error, a sign the denominator is wrong. We did not match the model to those weeks. We matched it to reality and treated the impossible readings as what they were. Most capacity arguments are built on exactly those numbers, never scrubbed, and a shortage calculated off a baseline that reads above 100 percent is a ghost by construction.

Finding the Capacity Before You Buy It

The work is to build the mirror before you trust the read. Pull a full year, not a month, so the variability is in the data and not in someone's memory. Validate the model against actuals until it is within a couple of points, then let it name the binding constraint to a specific line and a specific process step. Until the constraint has a name, every capacity number is a guess wearing a decimal point.

Then test the spreadsheet's shortage against the model. The protein co-packer was about to fund cook capacity it would not touch for years. The model moved the decision from "add capacity" to "sequence and automate against the real constraint," and changed what the next dollar bought. That is the entire return on building the twin: not a prettier chart, a different check getting written.

Watch the seam where ghost capacity hides at the packaging end too. A food co-packer had a box that fit the pallet perfectly on paper. It could not run it. Product came off the auto-bagger holding air that settled out as the bags sat, so a box sized to the paper dimensions was too small to pack off the line. The "right" box on the spreadsheet was the wrong box on the floor, and the only way to know was to pack it live and watch the air leave. Pallet overhang is not a rounding error either; distributors fine for it. The capacity to run that SKU existed only at the dimension the line could actually fill, not the one the drawing optimized.

And know what keeps ghost capacity ghost: time. A natural-foods brand had a real corrugated optimization on the table, sized and credible. The client locked specs and signed long-term supplier contracts before the work could land, and the leverage evaporated. The savings did not get disproven. They got frozen, pushed past year-end, and once the momentum is gone the suppliers have little incentive to reopen it. Capacity and savings you can model but cannot access are still ghosts. The model tells you they are there; the calendar decides whether you ever spend them.

What a Well-Run Capacity Decision Looks Like

The number on the capital request traces to a model validated within 2 percent of a full year of actuals, not to an onsite gut read. The binding constraint is named to one line and one process step, with the material variance that governs it, moisture, viscosity, resin price, modeled as load rather than waved off as noise. The OEE feeding the baseline has been scrubbed, with zero weeks reading above 100 percent. Changeover sequence is in the model, not assumed away. And a 12-month equipment lead time is decided against that model, so the plant is not buying a shift to chase capacity it already owns.

The spreadsheet was not wrong about the arithmetic. It was wrong about which number was the constraint. The capacity was never missing. It was misfiled, and a model accurate to real life is the only thing that tells you which drawer to open.