The Variability Tax a 30-Day Model Cannot See

The model that flipped when we fed it a year

A packaged-meat producer had four sausage lines in one hall. Same packaging process, different vintages of equipment: two older machines and two newer ones built for quick changeovers. We went in with 30 days of clean production data, built a digital twin, and ran the simulation. The answer looked reasonable. Keep three of the four lines, shuffle a little production between them, and live with it.

Then we pulled the full 12 months. The recommendation inverted. With a year of real data in the model, the right move was to cut from four lines to two, drop the two worst-performing older machines, move from one shift to two, and optimize scheduling onto the quick-change equipment. That combination returned about a 12 percent throughput improvement, and we took roughly six people off each line through a mix of front-end automation and simply modeling utilization at the worker level. Nothing about the machines changed between the 30-day run and the 12-month run. What changed was that the year exposed the variability the month had hidden: a lot of SKU changes inside the sausage category, and the allergen changeovers that ride along with them.

Changeover is a graph, not a line item

Most plants carry changeover as a single average in their heads. Forty minutes, say. That average is the tax rate. It is not the tax. The tax is set by how many distinct transitions the schedule has to make, and that number does not grow with the SKU count. It grows with the square of it. Two SKUs have one changeover between them. Ten SKUs have forty-five possible transitions. Thirty SKUs have four hundred thirty-five. Add allergen classes and each one drops a flush-or-sequence constraint on top, carving dead zones into the schedule where a feasible run order simply disappears. The floor does not experience an average. It experiences the specific sequence it got dealt that week, and the bad sequences are the ones a short window rarely samples.

That is why a 30-day snapshot lies with a straight face. Thirty days sees a slice of the transition space, usually a calm one, and reports back that the schedule is tractable. The season is where the SKU proliferation and the allergen mix actually collide. A veteran I talked with recently, fifty years in meat processing, put the same point from the other side. Small and mid-size companies, he said, have far more variability in their everyday production than the long-run plants everyone benchmarks against. Some of his lines run one product all day. Others never do. The interesting optimization, he noted, is not "run it faster." It is "do not go from product A to B to C; go A to C to B, and here is why." He was describing a changeover graph without calling it one.

Model the season, price the sequence

If you are about to consolidate lines, buy a piece of automation, or defer either, do three things before the capital decision, not after.

First, feed the model a full year, not a convenient month. The window has to be long enough to contain your seasonal peak and your messiest SKU mix, because those are the weeks that set your real capacity, and they are exactly the weeks a 30-day pull skips. In the sausage hall, the 30-day version and the 12-month version did not disagree at the margin. They gave opposite answers.

Second, measure changeover per transition, not as an average. Build the SKU-to-SKU matrix and mark the allergen boundaries. The average tells you the tax rate; the matrix tells you which sequences are cheap and which are ruinous, and it turns "sequencing" from a scheduler's art into a number you can optimize against.

Third, remember that automation is a sequencing question wearing a capital costume. The same veteran named the trap precisely: automation lends itself to product A and fights product B, so if product B still needs twelve people on the line, the payback on cutting product A's crew evaporates. The line-consolidation decision above only worked because the model showed which products belonged on the quick-change machines and which did not. That is also why you catch it in simulation first. Labor is the one lever you cannot pull twice; once a plant loses its labor base it is very hard to get back, so the move has to be right before anyone touches the floor.

What a well-run sequencing model reads

The capacity model runs on 12 months of production data, and it is re-run before any line-consolidation or automation-capital decision, not to justify one already made. Changeover time lives in a SKU-to-SKU matrix with the allergen boundaries marked, not as a single floor-wide average. Someone owns the sequence, not just the schedule, and can tell you why the run order is A to C to B this week. The plant knows which of its lines are quick-change and routes the high-mix SKUs there on purpose. When a marginal line comes up for a keep-or-cut call, the answer comes from the annual model, not the last calm month.

Closing

The old lines were never the constraint. The sequence was. Nobody had priced the sequence, so the plant kept paying the tax the only way it knew how, in labor thrown at every slowdown, and a 30-day model was happy to tell them the lines were fine.