Eppen's 1979 pooling result is the strongest argument for centralising, and it is an argument about where stock sits rather than about where planners sit. Those two decisions get bundled into one programme and only one of them has a theorem behind it. The test that sorts the rest comes from Jensen and Meckling's 1992 split between general knowledge, which transfers cheaply, and specific knowledge, which does not, so for every planning decision ask whether the information behind the last three overrides came out of a report or out of a phone call. Model selection, reconciliation, parameter setting and master data stewardship carry to a hub. Promotion timing, distributor credit and anything needing a person in the room stay in the market. Price the coordination in cycle days before moving anyone, because a hub that adds a day and a half of hand-off latency to a five day cycle has spent thirty percent of the cycle on the org chart.
The business case for a planning hub usually has three lines in it. Headcount comes down. Everyone works off one version of the numbers. Planners get a proper career path instead of being the only analyst in a market of forty people.
Only the first line has a number attached, so the first line is the one that gets tracked afterwards. Eighteen months later somebody asks whether the hub worked, and the available evidence is a headcount figure and a set of opinions. The markets say the hub does not understand their customers. The hub says the markets will not follow the process. Both are usually right, and neither observation tells you whether the structure was wrong or whether the wrong things were moved.
The useful version of this question gets settled one decision at a time, and the org chart falls out of the answers rather than driving them.
Pooling stock and pooling people are different results
There is a real theorem behind centralisation and it is worth stating precisely, because it is routinely used to justify something it says nothing about.
Eppen showed in 1979 that consolidating independent, normally distributed demand from several locations into a single stocking point reduces expected holding and shortage cost, and that under identical demand at each location the safety stock requirement falls by a factor of the square root of the number of locations. Nine markets consolidated to one need roughly a third of the buffer, holding service constant. That is a large number and it is genuinely available.
Two conditions sit under it. The demands have to be close to independent, and the single stocking point has to actually serve all the demand.
The independence condition is the one that bites. Markets under a common macroeconomic cycle, a shared promotional calendar or a single category trend are correlated, and correlation eats the benefit quickly. With nine locations of equal variability and a pairwise correlation of 0.5, the pooled standard deviation works out at about 75 percent of the sum of the individual ones rather than 33 percent. The saving drops from two thirds to a quarter. Still worth having, and a quarter is a different business case from two thirds.
The part that matters for organisation design is that none of this is about staffing. Eppen's result concerns the physical or logical location of inventory. You can hold stock in one place and have nine planners plan it. You can hold stock in nine places and have one planner plan it. These are separable decisions, and they get bundled into the same programme so often that the inventory benefit ends up quoted as the justification for a headcount move that has no equivalent result behind it.
Before going further, separate the two on your own case. If the stock consolidation is where the money is, do the stock consolidation. It does not require anybody to change desks.
The test that sorts the decisions
Jensen and Meckling, writing in 1992 on specific and general knowledge in organisational structure, give the cleanest available test. Some knowledge is cheap to transfer: a demand history, a lead time, a service target, a cost. Other knowledge is expensive or impossible to transfer, because it is tacit, perishable, or exists only as a relationship. Decision rights want to sit where the specific knowledge already is, because moving the decision to the knowledge is cheaper than moving the knowledge to the decision.
That gives a procedure rather than a principle. Take a quarter of forecast overrides, replenishment parameter changes and expedite approvals. For each one, write down where the information behind it came from. A report, a system field, a supplier portal, a data feed. Or a conversation, a store visit, a customer's tone on a call, something the local sales director happened to mention.
The first pile can move to a hub, because the information moves with it at almost no cost. The second pile tells you which seats have to stay where the conversation happens.
Jensen and Meckling add a second half that people skip. Wherever you place a decision right, you have to pay for it. Put the decision next to the local knowledge and you owe a measurement and reward system, because you have given away control. Put the decision in a hub and you owe an information transfer mechanism, because you have separated it from what it needs to know. Every structure has a bill. The failure mode is choosing a structure and not paying either bill.
What carries to a hub and what stays put
Working through a normal planning job, the split lands in a fairly consistent place.
Carries well:
- Model selection and re-fit cadence. Choosing between a smoothing model, an intermittent demand method and a regression with causal factors is general knowledge applied to transferable data. It also benefits from concentration, because one person doing it across nine markets gets nine times the practice.
- Hierarchy reconciliation. Making the market forecasts add to the regional one is a mathematical operation on a full data set, and it has to be done somewhere that can see all the levels at once.
- Safety stock and reorder parameterisation. The inputs are lead times, variability and cost. All of them exist in systems.
- Master data stewardship. Unit of measure, lead time fields, product hierarchies. Concentration here is the whole point, since the errors are exactly the ones that come from nine people applying nine conventions.
- Exception rule design and threshold tuning. Setting the thresholds that decide what counts as an exception is a statistical question about distributions, best answered once across the whole catalogue. Working the resulting queue is a different job and sits further down this list.
Stays in the market:
- Promotion timing and the trade calendar. The relevant fact is often that the customer moved their event by a week and told the account manager over lunch.
- Customer and distributor order behaviour that has no field in the system. Why one wholesaler buys in a lump every eleven weeks, or which distributor is close enough to their credit limit that next month's order will not arrive.
- Supply commitments where the local commercial relationship is the collateral. Somebody has to be able to make a promise and be personally exposed to it.
- Anything requiring a negotiation with a person who is in the room. Allocation between two country managers is a negotiation, and negotiations conducted over email between time zones take four times as long and produce worse outcomes.
The list is not a template. Two businesses in the same category will split it differently depending on how much of their commercial reality has made it into a system, which is precisely the variable the test measures.
Coordination cost, priced in cycle days
Galbraith's 1974 framing of organisation design as information processing explains where the hidden cost goes. An organisation facing uncertainty either reduces its need to process information, by holding slack or by making tasks self-contained, or increases its capacity to process it, through vertical information systems and lateral relations.
A regional planning team is a self-contained task. Centralising removes that, so the bill arrives as either slack, which means inventory, or lateral relations, which means meetings, liaison roles and a queue.
Price the queue in cycle days, because that is the unit the planning calendar runs on. Suppose a five working day cycle. Before the hub, a planner notices a distributor's order pattern shift on Tuesday and has adjusted by Wednesday. After the hub, the observation has to be written up, reach the hub, sit in a queue, get worked and come back. A median round trip of a day and a half consumes thirty percent of the cycle, every cycle, on decisions that used to take an afternoon.
Then price the volume. Nine markets, twelve exceptions a cycle each that need local context to resolve, gives about 108 round trips a week landing on the hub. If each one costs the hub twenty minutes and the market fifteen, that is 63 hours a week, call it 1.6 full time equivalents of pure coordination. Against a claimed saving of sixteen heads, coordination has quietly taken ten percent of it back, and it has taken it from people whose time nobody is counting because it is spread in fifteen minute pieces across a sales organisation.
Neither figure appears in a business case, and both are measurable before you commit. Count your context-dependent exceptions for one cycle and time three of them end to end.
The career ladder the hub deletes
The third line in the business case, the one about career paths, is the one most likely to invert.
The role a hub usually removes first is the local junior analyst, because it is the cheapest per head and the most obviously duplicated across markets. That role was also the entry point to the function. Five years on, the market has nobody who came up through planning, and the only people who understand both the numbers and the customers sit in sales.
The hub does create a career path, and it is a different one. Hub careers run up a technical ladder: analyst, senior analyst, methods lead, planning manager. Local careers ran up a commercial one, ending in a supply chain seat next to a country manager. Businesses that centralise hard often discover four years later that they cannot fill that commercial seat internally, because the ladder that used to produce candidates for it has been taken out.
This is checkable rather than theoretical. Track the ratio of internal to external hires into market-facing planning manager roles for the three years before and after. If it inverts, the career path line in the business case was wrong, and it was wrong in a way that takes several years to repair. The retention consequences that follow have their own arithmetic and their own post.
Two adjustments make the difference. Keep at least one market-facing planning seat in every material market, even where the workload does not justify a full role, and staff it as a rotation. And rotate hub analysts into markets for a full cycle rather than a visit, because a planner who has sat in a distributor meeting works exceptions differently for years afterwards.
Where this stops
The honest limits here are about evidence rather than about method.
There is no clean experiment. You cannot run both structures over the same weeks on the same business, so every assessment is a before-and-after read that confounds the reorganisation with the demand environment, the assortment changes and whoever left during it. The nearest thing to a controlled comparison is a partial rollout, moving three markets and holding six, and running both for two quarters on the same calendar. It is slower and it is the only version that produces an answer somebody outside the programme will believe.
Every reorganisation also costs two to three planning cycles of degraded output in either direction, because relationships and undocumented routines have to be rebuilt. The first read after go-live is always bad and means nothing. Plan the assessment for month six at the earliest, and say so at the start, before anyone has an incentive to read the month two numbers as a verdict.
The deeper limit is that the org chart is a weaker instrument than most reorganisations assume. Where decision rights are written down, who holds them, what their limits are and what happens when they disagree, both structures work acceptably. Where they are not written down, neither structure works, and moving the boxes changes which meeting the confusion happens in. A regional model with clear decision rights beats a hub without them, and the reverse is equally true. Some of the businesses that got the largest gain from a centralisation programme got it from writing the decision rights down, which was a prerequisite for the move rather than a benefit of it, and which they could have done without moving anybody.
One more caution on cost. Where the hub is justified by wage arbitrage, the arbitrage is real and it decays. Hub locations chosen for cost become expensive as they become popular, and the retained coordination layer grows rather than shrinks as scope widens. Model the cost line over five years with a wage inflation assumption specific to the hub city, and see whether the case still closes in year four.
Pull last quarter's overrides and write next to each one where the information behind it came from; the ones sourced from a report can move to a hub, and the ones sourced from a conversation tell you which seats have to stay in the market.
More on this in organisation on the DemandLabs platform.
