Case Study · Consumer Products

The $4 Million Retrofit
That Never Had to Be Built

A soap manufacturer was consolidating several plants into one. Over 300 products, new equipment, a clean sheet. One of the machine configurations under consideration would have worked on paper and then, quietly, produced a quality problem expensive enough to need a $4 million retrofit. The model found it before the concrete was poured.

The decision

The company needed to collapse manufacturing from multiple sites into a single optimised facility. The existing sites between them made over 300 different products, and the consolidated plant had to absorb all of it.

That is not an equipment-sizing problem. Whether the plan works depends on things that only exist over time:

The logistics manager chose simulation for a reason the study records plainly: it was the one tool that captured real-world variability and the interaction between subsystems. A spreadsheet can size a machine. It cannot tell you what happens when a seasonal peak, a changeover sequence and an unreliable upstream stage line up in the same week.

The plant model on screen: raw materials feeding two making lines, a bank of surge tanks, and three packing lines running to a warehouse. Each line shows its current product and system status, and a legend maps colours to blocked, satisfied, starved and rate-zero flow states plus running, emptying, changing and ready control states.
The model of the whole plant, from the original study. Raw materials feed two making lines; a bank of surge tanks sits between making and the three packing lines; finished cases run to the warehouse. Every line reports its current product and status, and the legend is the vocabulary this whole site is about — blocked, satisfied, starved, rate = 0, and control states for running, emptying and changing over. Packing 3 is mid-changeover in this frame.

What the model changed

Two things came out of it. The first was ordinary and valuable: the model characterised the balance of the line and predicted the optimal buffer size between processing and packaging — the number that decides whether the two halves of a plant can run independently.

The second was the one that paid for the study. For a given set of products, the envisioned machine assignments produced excess time in system. Material sat longer than it should, and for this product that is not an efficiency problem, it is a quality problem. Left in place it would have surfaced months into production, in a plant already built around those assignments.

One configuration predicted severe long-term quality problems that would have required a $4 million retrofit. Alternative product routings and machine assignments were tested in the model instead, and the problem was avoided completely.

Note what kind of finding that is. Nothing was broken. Every machine met its specification. The failure was in how the parts combined over time, which is exactly the class of problem that survives every static review and then shows up in production.

A nested sub-model of one packing line in detail: product in, conversion, containers, filler, sealer, conveyor and case packer, with cases out.
One packing line opened up — a sub-model nested inside the block on the plant view above. Containers, filler, sealer, conveyor, case packer. The plant view answers where the constraint is; this view answers why.

What happened afterwards

The most telling outcome is not the $4 million. It is that the company changed its process: it now requires a simulation study before committing to any major capital expenditure. The logistics manager on this project went on to extend the same approach to supply and distribution policy across the division.

The method

The model was built by the team using the discrete-rate technology for high-volume processes that ReliaSim now packages, along with an embedded database driving the dynamic factors and multi-stage sequential schedule definition. This work predates ReliaSim as a product; the approach is the same one, and it is the approach being described here rather than the current software.

The client is not named in the source material and is not named here.

Related: How big should a buffer be? · The Hidden Bottleneck · All case studies