Inventory can be reduced safely only where the protection it provides is not needed. The task is to find those places precisely, rather than cutting a percentage across the board and discovering which customers notice.
Where cuts are usually safe
Short, reliable replenishment. A SKU that can be replenished within one review period from a supplier whose observed lead times and fill are tight needs little protection beyond the demand variation of that one period.
Genuinely independent alternatives. A second approved supplier on a different route, or a branch with excess and a fast transfer lane, reduces the protection the SKU needs. The word genuine matters: shared ports, carriers or plants do not diversify.
Approved substitutes on the shelf. Where a customer has approved a substitute, its stock counts toward the requested SKU's protection at the conversion ratio.
Low-priority demand. Protection sized for a 90% spot customer is smaller than protection sized for a 99% contract. Holding the higher level for everyone is the most common source of excess.
Where cuts are not safe
Long import lanes with a lead-time tail; suppliers with a record of partial deliveries; SKUs where every source shares a node; strategic customers with contractual service; and any SKU whose demand history is too short or has just shifted to say what normal looks like. A recent level change should widen the uncertainty, not narrow the buffer.
The method
Replay the candidate policy on the operator's own past year under many plausible futures and under deliberate stress (a supplier at 70% fill, three weeks late, for two months). Accept a leaner plan only if it meets each customer's target on the fitted futures and does not do worse than the current policy under stress. Measure service by the metric the customer uses: unit fill, line fill, order fill or on-time-in-full are different numbers.
Two things a leaner plan must still do
It must be able to protect again: when a supplier's observed tail widens or a route signal appears, the plan should rebuild protection on that SKU alone, without waiting for a stockout. And it must classify any service miss honestly: unavoidable (no feasible action could have met the target), avoidable (a feasible action was not taken), an economic trade-off the operator explicitly accepted, a model that was wrong, or data that was insufficient. A plan that only reports "service breach" is hiding which of those it was.