Buy the Model Before You Buy the Line

The producer who wanted to buy a line he already had

A packaged-meats producer came in convinced he was short on capacity. Rate was missing, the floor felt maxed, and the obvious answer was sitting in the capital budget: add line, add automation, add shift. The number was large and the equipment lead time was over a year, so the decision had real weight.

We did not start with a quote. We started with a year of his own production data and built a digital twin of the plant. After thousands of scenario runs the model landed within about 2 percent of real operations, accurate enough to trust against the actual floor. Then it answered the question he came in with. He could grow roughly threefold before the cook, the thermal stage, became the binding constraint. The capacity he was about to buy was not the capacity he was short on, because on most of those lines he was not short at all.

Why the spreadsheet keeps lying about capacity

The reason this happens is mechanical, and it repeats across plants. A spreadsheet sizes capacity on an average rate per format. It treats the line as one number. But throughput is not set by an average; it is set by how the stages interact. The cook, the bagger, the labeler, and the case packer each have their own rate and their own variability, and the slowest interaction under real conditions sets the pace. Push more units at the front and the wall just moves a few feet. Add three heads to a line whose constraint is a thermal stage and the oven still runs at yesterday's speed.

Capital Confidence is the difference between committing money because a flattened number said "short" and committing it because you modeled the interaction and saw exactly which stage binds and when. The spreadsheet answer feels safe because it is conservative on rate. It is not safe. It is precise about the wrong stage. Confidence built on an average is just a well-formatted guess, and at 12-month lead times a guess is expensive to be wrong about.

The same pattern showed up at a precooked-protein processor we scoped a digital-twin pilot with. Strict packaging test requirements were quietly cutting plant flexibility, and the team was leaning on a single higher spec to cover it. That masks root cause the same way an average rate does. You end up paying, in flexibility or in capital, for a problem you never actually located.

Model first, then decide invest or defer

The move is to build the model before you build the line. It runs in phases, and the phases are cheap relative to the capital they govern.

Phase one is replication. Pull real throughput data, the kind that already exists in the control system: line rates, fill rates per minute, case and pallet level timestamps. Rebuild the plant's current OEE and its variability until the model reproduces the real floor within a couple of points. Watch the data while you do it. One line in the producer's set reported OEE above 100 percent for seven straight weeks, a structural data error. If you size capital on numbers like that, you buy the wrong thing with confidence. Reconciling that gap is not a detour; it is the work.

Phase two is experiments. Once the twin is accurate, run the scenarios you would otherwise run with real money: incremental automation at one end, aggressive line consolidation at the other. Each scenario tells you the new binding stage and the headroom before it binds. That turns a single "buy or do not buy" into a sequence. You learn which spend actually moves output, which spend lands upstream of the wall, and how far the existing asset stretches before any of it is required.

Phase three is the decision itself, and now it is honest. Invest or defer is no longer a feel. The producer above did not need the multi-line request this year. He needed a phased plan that bought capacity in the order the constraints actually bound, sequenced against the 12-month lead times so the equipment arrived the quarter it became the wall, not two years early as idle depreciation.

What a well-modeled capital decision looks like

The model reproduces a full year of OEE within 2 percent against real production data, not a vendor curve. The binding stage is named, not assumed, and for most products on the line it is one specific stage. The case states how many units of growth the line absorbs before that stage binds, expressed as a multiple of today's volume. Capital is sequenced against equipment lead time, so the spend lands the quarter the constraint binds rather than the quarter rate got missed. And the data feeding all of it has been reconciled, so no number above 100 percent OEE is quietly steering a seven-figure decision.

Closing

The spreadsheet said buy capacity. The model said he already had it for the next tripling. The cheapest line on the floor is the one he did not buy, because he finally modeled the cook instead of the average.