Every Supply Chain Needs a Different Forecast. Your APS Gives Everyone the Same One.
A fast-growing beverage brand and a paper box manufacturer buy the same advanced planning system (APS). It gives the same setup and instruction to both: let's understand your SKUs and your locations, then we'll create a demand forecast. For the beverage brand, it's roughly the right question. For the box maker, it's the wrong question entirely, and no amount of model tuning will fix it.
Every demand planning conversation runs on the same premise: get the model right and accuracy follows. The real work is two steps upstream from picking a model. First you have to forecast the right thing, and then you have to build a system that can actually forecast it. Most APSs get neither step right. They point every customer at the same question, and even when that question roughly fits, their models can't answer it well enough to maximize growth. Getting both halves right is what changes the number.
The right thing to forecast is different for make-to-stock, make-to-order, and engineer-to-order supply chains. Here's what that looks like across three real customers.
Make-to-stock: the standard target
When you make-to-stock, you hold finished product. Your cash sits in inventory on the balance sheet. Your costs are warehousing and holding. The question that matters is how much of each finished SKU each distributor needs, and when they need it. You want to balance service levels against inventory levels.
What a generic APS does: forecasts at SKU and location. That's roughly the right target, and this is the case where the target choice isn't the story. Every planning vendor points at SKU and location for make-to-stock. Where they lose is on the second half: they can't actually forecast well against that target.
What doing it right looks like: we work with one of the fastest-growing US beverage brands, a make-to-stock supply chain. The prior system was forecasting at SKU level and struggled at the distributor level. We kept the target, then built the system to hit it: a purpose-built AI trained on more than sales history, including demographic data, brand-growth signals, and third-party data for new launches where there is no history to lean on. Same target as the incumbent. Roughly double the accuracy. Seven figure P&L impact from carrying 20% less inventory.
The make-to-stock lesson is the one people don't want to hear: even when the incumbent picks the right target, its forecast usually falls apart at the SKU/Location level. Push to SKU/Location/Customer and it gets worse.
Make-to-order: a different target
When you make-to-order, you don't hold finished product at all. You hold raw materials. Your exposure is supplier lead times. A material that takes two months to arrive has to be forecast two months out, based on what your plants will actually consume. You want raw materials on hand when your plants need them, without paying to sit on months of supply.
What a generic APS does: forecasts finished SKUs at the plant level, then explodes the bill of materials to size the raw materials underneath. When SKUs share common materials, that indirection compounds error at every layer. The forecast may be accurate, but not at an operationally useful level.
What doing it right looks like: we replaced a top-five planning system for a manufacturer whose raw materials had long lead times. We stopped forecasting finished SKUs and forecast the raw materials directly, tied to each material's lead time and the plants that consume it. Same customer, same demand history, same commercial reality. Repositioning-driven transportation costs dropped 32%.
The make-to-order lesson is that a forecast can be technically accurate and operationally useless. The incumbent's aggregate SKU forecast was accurate. But the SKU/Location forecast was not, and the materials plan built off it was wrong.
Engineer-to-order: the target the APS may not see at all
When you engineer to order, there may be no fixed bill of materials. One of our customers makes paper boxes. A forty-inch box can be cut from any paper forty inches or wider. One finished product maps to many possible raw materials, and the choice of which one changes the economics. You want inventory that can become anything a customer orders, without stockpiling material that never moves.
What a generic APS does: asks the box maker to forecast SKUs at locations. There is no useful answer to that question, because the box forecast isn't the constraint. The paper is, and which paper to build with is a decision, not a demand signal. So the APS does the wrong math. There is no point predicting demand for raw materials when the choice of raw materials in the history was sub-optimal. The past poisons your future choices. You get an "accurate" demand forecast that perpetuates the inventory position you already have.
What doing it right looks like: we optimized the bill of materials order by order. The system reads current inventory alongside short and long-term demand, then allocates raw material to each box — the "which paper" decision. Every order gets built from the paper that leaves the best inventory position behind it, rather than the paper history would have picked. Margin improved 8%. Inventory dropped 41%. Service levels held at 99.98%.
The engineer-to-order lesson is that accuracy was never the failure. The math is aimed at the wrong decision, and no amount of forecasting skill recovers from that.
Three supply chains, one principle
The beverage story is the make-to-stock case: the incumbent picked the right target and simply wasn’t accurate enough. The manufacturer is the make-to-order case: the incumbent executed well on a target that was technically fine and operationally wrong. The box maker is the engineer-to-order case: no APS in the market will do the right math. Across all three, forecasting the right thing and actually forecasting it well produced double-digit gains on the metric that mattered.
What we built
Purpose-built means forecasting the right target at the right level of granularity. For every supply chain we work with, Omnifold:
Learns the target that runs your business. Finished goods, raw materials, or the material choice itself. Often several at once.
Builds the system that can hit it. Not a menu of generic time-series models, but a system designed for the specific mathematical structure of your supply chain.
Adapts as the business moves. New SKUs, new channels, new constraints. The system updates itself, so the forecast keeps pointing at the right thing.
If your forecast keeps missing and your inventory stays elevated, and no one can explain why, you are probably debugging the wrong layer. Ask two questions instead. What is the system actually aimed at? And can it hit even that?