Essay 02 · Network
Planning orthodoxy has a clear hierarchy. "How much do we need" is the hard question, the one that gets the forecasting engine, the statistical safety stock and the monthly executive review. "Where should it sit" is a downstream allocation problem, usually handled by a proportional rule that splits the buy across sites by historical share.
That hierarchy made sense when freight was cheap. It doesn't anymore, and the interesting part is that most companies haven't re-litigated it — because of who reports to whom.
Take a business with eight distribution centers. You buy exactly the right total quantity of an item: demand for the season comes in at precisely what you forecast, company-wide. Congratulations on the forecast.
Now distribute it wrong. Sixty percent of the units land in a region that consumes forty percent of them. What you have is a stockout in one half of the country and a markdown in the other — simultaneously, on the same item, in the same quarter, from a perfect forecast. Your service metric is bad. Your inventory metric is bad. Your forecast accuracy is excellent, and it's the number that will get reported.
Then comes the correction, which is where the money actually goes. Someone notices, and units get moved. Not on a planned full truckload built into a scheduled lane, because nobody planned for it — as an expedited LTL transfer, at rates that can consume a meaningful share of the item's margin. Or the item gets bought again for the short region while the long region marks its copies down. Both corrections are expensive, and neither appears in any planning metric. They land in freight and in gross margin, owned by different people, explained separately.
The structural reason is the org chart. In most companies, three different executives own the three pieces of this decision. Planning and procurement decide how much to buy. Transportation owns what it costs to move. DC operations owns where it physically fits and what it costs to hold. A decision that spans all three has no single owner, so it defaults to a rule — and rules don't get revisited, because revisiting them isn't anyone's job.
I'll say this from the inside: I've run all three under one roof, and it changes what you can see. The person deciding how much to buy also owning what it costs to move it and where it has to sit is unusual, and it's exactly why placement decisions get made in that configuration and go unmade in most others. If you want to know whether your company has a placement problem, don't audit the process — look at whether one person could fix it if they wanted to.
The second reason is measurement. Companies measure service and inventory at a company level, where placement errors cancel out beautifully. Aggregate fill rate looks fine. Aggregate inventory looks fine. Both are the average of a stockout and a pile. This is the same pathology as measuring forecast error above the level where it happens — the level of detail is doing the hiding.
The first and cheapest question, asked before any purchase order: do we already own this, somewhere else, in a building with too much of it? In a multi-node network a meaningful share of every buying cycle is items the company already has. Surfacing that in the buyer's workbench — as a suggested transfer, priced against the buy — changes behavior immediately, because it's a decision the buyer can make without convening anyone.
Not everything should be stocked everywhere. Slow, bulky, low-margin items with unpredictable demand are candidates for a single node with a defined service commitment and a freight recovery mechanism, rather than eight partial piles that each stock out independently. The analysis is a viability segmentation: for each item, is the carrying and handling cost of local stocking justified by the freight and service benefit? At one company that exercise covered 13,000+ special-order SKUs and replaced item-by-item judgment calls with a defensible quadrant.
What does a line actually cost to fulfill from each node — labor, handling, outbound freight to the customers that node serves? Most companies genuinely do not know this at item-location level, and it's the input that makes every other placement question answerable. It's also the input that tells you when a node shouldn't exist.
How full the truck is when it moves, and how much life is left on the pallet when it lands, are the two constraints that most often turn a mathematically optimal transfer plan into a bad one. Any deployment model that ignores them will produce recommendations that operations correctly refuses to execute — and once that happens twice, nobody opens the model again.
Everything above can be improved with better rules and better visibility. But some of these questions are genuinely too large to reason through, and pretending otherwise is how companies end up defending a proportional split nobody believes. Which carrier should get which lane across a bid. Which node should own which item, given every item's demand and every lane's cost. Which sites should exist at all and which customers each should serve.
Those are mixed-integer programs — assignment, facility location, network flow with side constraints — and a solver answers them properly. The reason that matters isn't precision for its own sake. It's that a solved model hands you two things a heuristic never can: a bound, so you know whether your current answer is 2% off or 20% off and can stop worrying about it if it's the former, and shadow prices, which tell you what each binding constraint is actually costing. "The binding constraint is dock capacity at this site, and relieving it is worth this much" is a sentence that ends an argument. Three of these are written up here.
If you have one distribution center, none of this applies and you should stop reading. If you have two or three in a compact geography with cheap lanes between them, the prize is real but modest — worth a quarter of someone's attention, not a program.
The threshold where this becomes the highest-return work I can point at is roughly: four or more nodes, meaningful geographic spread, seasonal or regional demand variation, and freight that has materially outpaced your product margin over the last several years. That describes a very large number of mid-market distributors right now, most of whom are still running a proportional allocation rule written when diesel was half the price.
There's also a fair objection that this is what a network optimization study is for. Partly — but the classic study is a one-time engagement that produces a strategic recommendation about facility location and then goes stale, because the model leaves with the consultant. I build the model and hand it over, so "what if this lane's rate moves" or "what if that site's capacity drops" is a re-solve rather than a new project. And the strategic answer still doesn't address the operational one I'm describing here: where the units go this week, which has to be answered by the system, continuously, or it won't be answered at all.
Next step
If the answer is "an allocation rule nobody has looked at in three years," that's worth an hour.