Multi-echelon optimisation treats a network of stocks as a single system rather than a succession of independent points. Instead of sizing each warehouse for its own service, it allocates protection across echelons so as to reach final service at the lowest total stock. Its contribution fits in one sentence: the network optimum is not the sum of local optima.
Why it matters
Common practice sets a service target at each stocking point, then sizes each one to meet it. It is simple, it holds sites accountable, and it is what most systems do. Yet it almost always produces a network holding too much stock for the service it delivers.
The reason is that echelons are not independent. A regional warehouse backed by a well-placed central stock does not need the same buffer as one supplied from the other side of the world. The service the upstream echelon needs is not the service the end customer sees, and setting them at the same level means funding the same protection twice.
The signature illustrates this trade-off. Sliding the allocation to the left means letting each site protect itself alone: pooling is nil and the network funds the same uncertainty several times over. Pushing it right means concentrating everything upstream: capital falls, but sites no longer have enough to hold out while being replenished, and final service degrades.
The optimum lies between the two, and this is the most important point on this page. It is interior, hence inaccessible to any method that handles echelons one at a time: each such method leads mechanically to one end of the curve or the other.
The typical gain is not marginal. On multi-echelon networks, reallocating protection commonly frees a significant share of stock at constant service, or clearly improves service at constant stock. The gain comes almost entirely from moving stock, not removing it: you do not take inventory out, you put it somewhere else.
That is also what makes the approach politically difficult. Moving protection upstream lowers the stock of sites that are themselves measured on their local service rate. Without a change of indicators, local managers will quietly rebuild their buffers, and the benefit will evaporate within months. The condition for success is therefore to steer on final service, not on each node’s service.
The expert lesson concerns prerequisites. The method demands reliable visibility of stock and lead times at every echelon, which is rarely a given. A company that does not know its real internal lead times will get elegant and wrong recommendations from a multi-echelon engine. Better to make that data reliable first, even if it means starting with a coarse two-tier trade-off, than to deploy a sophisticated model on doubtful inputs.
The mechanism
Two opposing forces structure the problem. The first pushes stock upstream: concentrated, it pools the uncertainty of every site and benefits from the offsetting effect, reducing the total buffer required. The second pushes stock downstream: close to the customer, it responds immediately, whereas central stock must first travel the lead time separating it from the field.
Optimisation consists of finding the allocation that balances these two forces for a given final service. It reasons in cumulative lead time rather than local lead time: what matters for a site is not only transport time from the upstream warehouse, but the time it would take to rebuild the chain if upstream were itself exposed.
From this logic comes a central notion, internal service. Each echelon renders a service to the next, and that service has no reason to equal final customer service. A moderately serviced upstream echelon that can replenish quickly may be enough to support a very high-performing downstream, for a total capital lower than two equally demanding echelons.
Finally, it is worth stating what the method is not. It replaces neither buffer calculation nor segmentation: it decides where to place a protection whose principle remains that of the preceding pages. And it does not remove the need to shorten lead times, which remain the most powerful lever, because they set the height of the whole curve rather than the position of its minimum.
The traps
Applying the final customer target to every echelon means funding the same protection several times. An upstream echelon’s internal service can sit well below final service without degrading it, provided replenishment is fast.
If sites remain assessed on their local service rate, they will rebuild their buffers whatever the model recommends. Moving protection holds only if measurement is on final service and on total network stock.
A multi-echelon engine amplifies the quality of its inputs, in both directions. Without measured internal lead times and reliable stock visibility at each node, it produces precise and wrong recommendations, which will discredit the approach for a long time.
The rollout
Measure real internal lead times between echelons and guarantee a reliable view of stock at each node. This is the prerequisite, and it often represents most of the work.
Write explicitly that only final customer service is a commitment, and that each echelon’s service is an adjustment variable serving that commitment.
Handle a central and regional pair on a representative family first, rather than modelling the whole network. Most of the gain appears at this first trade-off.
Change how sites are measured before relocating buffers. Without that prior step, the stock removed will return within months by indirect routes.
Continue reducing internal lead times in parallel, which lowers the whole cost curve instead of moving its minimum. It is the lever multi-echelon never removes the need for.
Neighboring concepts
From knowledge to action
Demanding final customer service at every node funds the same protection several times. Our Inventory & distribution file allocates protection across echelons and aligns the indicators that keep it there.