Service Levels Are Not Control Knobs

A planner asks for a 95% service level. The system computes safety stock. The inventory goes up. Everyone feels like the business made a serious decision.

Most of the time, it did not.

A service level is often treated as the knob that controls inventory. Turn it up for important products. Turn it down for slow movers. Give executives a clean table of targets. Build the replenishment system around those targets. Then spend the next few years arguing about why inventory is too high, availability is still bad, and nobody can explain why one SKU deserves 95% while another deserves 92%.

The problem is not that service metrics are useless. Fill rate, availability, and stockout frequency are important diagnostics. The problem is treating a diagnostic target as the actual decision objective.

The business does not make money by achieving 95%. It makes money by placing inventory where the expected economic value exceeds the cost and risk of carrying it.

The hidden decision

A target service level hides the real tradeoff.

For an inventory decision, the controllable action is not “achieve 95%.” The action is buying, producing, transferring, reserving, allocating, expediting, or withholding inventory. The tradeoff is economic: margin gained from additional availability versus holding cost, purchase cost, obsolescence, capacity usage, cash, markdown risk, and operational complexity.

A service target compresses all of that into one number. That can be useful for communication, but it is dangerous as the primary control mechanism.

Two SKUs can both have a 95% service target and completely different economics. One may have high margin, low holding cost, stable demand, and long-term assortment value. Another may have low margin, bulky storage, volatile demand, and high markdown risk. The same target does not mean the same decision.

The right question is not, “what service level do we want?” The right question is, “what is the marginal value of one more unit of inventory at this location and time?”

Service levels can reward the wrong behavior

Service metrics can hide stockouts, especially when demand is censored. If sales disappear during an outage, the measured demand may shrink. The model then appears to maintain service against a corrupted denominator.

Service metrics can also overweight cheap units and underweight expensive misses. Achieving high unit fill rate on low-margin, easy-to-serve products does not compensate for missing high-margin products that customers actually came to buy.

Averages make this worse. A category can report acceptable service while a few critical SKUs repeatedly fail. Or a national service metric can look fine while one region suffers because inventory was allocated to the wrong place.

The metric may look professional. The decision may still be bad.

Economic inventory logic

A better inventory policy starts with marginal economics.

For each candidate unit, case, pack, or order increment, estimate the expected incremental value:

[ \Delta V = \mathbb{E}[\text{margin captured}] - \mathbb{E}[\text{holding cost}] - \mathbb{E}[\text{markdown or disposal cost}] - \text{purchase and logistics cost} - \text{cash cost} ]

If the unit has positive risk-adjusted value and fits the operational constraints, buy it. If not, do not buy it. This is the core logic, whether the implementation is a MILP, a stochastic dynamic policy, a greedy marginal allocator, or a simulator-guided search.

This framing naturally produces different availability outcomes for different products. High-margin products with frequent demand and low carrying cost will tend to receive more inventory. Low-margin products with intermittent demand and high obsolescence risk will receive less. The implied service level is an output of the economics, not the input that pretends to be economics.

Decision variables

Suppose the decision is how much to replenish for item (i) at location (l) for arrival period (t):

[ x_{i,l,t} = \text{units ordered or transferred} ]

The objective should evaluate inventory across uncertain demand scenarios:

[ \max \mathbb{E}[\text{gross margin} - \text{holding} - \text{stockout penalty} - \text{markdown} - \text{logistics} - \text{cash}] ]

The stockout penalty should not be an arbitrary punishment selected to back into a desired service level. It should represent lost margin, lost customer value, substitution effects, contractual penalties, or strategic cost where those costs can be defended.

If a service commitment is contractual, then it can appear as a constraint:

[ P(\text{service}_{i,l} \geq s) \geq \alpha ]

But many targets are not contractual. They are habits. If a target is really a business preference, the economics should reveal the cost of enforcing it.

Constraints still matter

Economic logic does not mean ignoring operational constraints.

Real inventory decisions face MOQs, pack sizes, vendor calendars, lead times, warehouse capacity, shelf capacity, truck capacity, budget, expiration dates, product substitutions, receiving labor, and assortment rules. These constraints often matter more than the service-level target itself.

For example, a product may deserve more inventory economically, but the vendor MOQ forces a large lump. The question becomes whether opening that MOQ creates enough value across all affected SKUs. A service target does not answer that. A marginal economic model can.

Similarly, two products may both have high expected stockout cost, but only one can fit in a constrained cold-storage location. The decision is allocation under scarcity, not independent safety-stock calculation.

Uncertainty is the input

Service-level formulas often assume clean demand distributions, stable lead times, and simple loss functions. Real demand is lumpy, intermittent, seasonal, correlated, promotion-sensitive, and frequently censored by availability. Lead times are not constants. Vendor reliability changes. Product substitution matters.

A useful policy should consume uncertainty directly. Demand should be represented as a distribution or scenario path. Lead times should be uncertain where they are uncertain. Lost sales should be estimated where demand is censored. The model should evaluate the economic consequences of different futures.

A target service level takes uncertainty and crushes it into a quantile. That may be fine for a teaching example. It is often too crude for a real planning system.

Metrics to report

Do not throw service metrics away. Demote them to diagnostics.

Report expected profit, inventory investment, lost-margin exposure, expected stockout units, fill rate, availability, tail outcomes, aged inventory risk, markdown exposure, working capital, constraint shadow prices, and recommendation stability. Show the implied service level that results from the economic decision, but do not pretend the implied service level is the reason the decision is good.

Also report marginal value curves. Planners should see whether the next case is worth buying. Executives should see where the inventory budget runs out of positive-return opportunities.

This changes the conversation. Instead of arguing whether 95% is the right target, the team can discuss whether the assumed lost-margin cost, holding cost, substitution behavior, and risk tolerance are credible.

Failure modes

The first failure mode is assigning service levels by product importance without costing the decision. “A items get 98%” sounds disciplined until you realize some A items are expensive, bulky, volatile, and easily substituted.

The second failure mode is using one service target across a category. Categories contain products with different economics.

The third failure mode is treating service level as independent across SKUs. Vendor MOQs, shared capacity, substitutions, and allocation rules create coupled decisions.

The fourth failure mode is tuning service levels until the inventory number looks acceptable. That is not optimization. That is spreadsheet steering.

The fifth failure mode is hiding the cost of service. If leadership wants a higher availability promise, show the required inventory, cash, capacity, and markdown risk.

What to do in practice

Keep service levels as outcome metrics. Stop using them as the primary control knob.

Build the decision around marginal economic value. Make uncertainty explicit. Estimate lost sales where demand is censored. Include real constraints. Show the implied service outcomes and the cost of changing them.

A good planning system should not ask the business to worship 95%. It should show where the next dollar of inventory creates value, where it destroys value, and which constraints make the answer non-obvious.

That is the decision. The service level is just one way to measure what happened after the decision was made.