Inventory Optimization: Methods, Formulas and What AI Changes

Last updated: October 5, 2026

Inventory optimization is the practice of deciding how much stock to hold of each item, at each location, so the business meets its customer service targets at the lowest total cost. It sets order quantities, reorder points, and safety stock, and in networks with several tiers it decides where along the chain those buffers should sit.

Several terms get used interchangeably with it. Inventory management is the broader operational job of tracking, receiving, storing, and moving stock. Inventory planning is the recurring function that turns a demand plan into stocking targets. Replenishment is the execution step that places the orders. Multi-echelon inventory optimization (MEIO) is one family of inventory optimization methods, built for networks where a plant feeds distribution centers that feed customers or stores.

Why does inventory optimization matter?

Inventory optimization matters because inventory is the largest pool of working capital most product companies control directly, and errors in either direction are expensive. US manufacturers, wholesalers, and retailers held $2.76 trillion of inventory at the end of July 2026, about 1.3 months of sales (US Census Bureau). IHL Group estimates that out-of-stocks and overstocks together cost retailers $1.73 trillion worldwide in 2025, about 6.5% of retail sales, with $415 billion of that in North America (IHL Group).

Two cases from 2022 show how fast the cost arrives. When shoppers shifted spending away from the home goods and apparel they had bought during the pandemic, Target was left holding excess coffee makers, lamps, home décor, and clothing. It canceled more than $1.5 billion in orders and cut prices to clear the rest, and second-quarter net income fell almost 90%, from $1.82 billion to $183 million (CNBC). Six weeks later Nike reported inventory of $9.7 billion, up 44% from a year earlier, driven largely by goods stuck in transit during supply chain volatility. Clearing the excess through markdowns helped cut its gross margin by 220 basis points (Nike, SEC filing).

At Target the demand mix moved, and at Nike the lead times did. Those are the two inputs the classical formulas below treat as stable.

How does traditional inventory optimization work?

Most inventory optimization in production today rests on a handful of models developed between 1913 and the early 2000s. They run fast, and a competent planner can explain every number they produce, which is a large part of why they have lasted.

Economic order quantity (EOQ)

The economic order quantity, first published by Ford W. Harris in 1913, is the order size that balances the cost of placing orders against the cost of holding stock (Erlenkotter, Operations Research).

EOQ = √(2 × annual demand × cost per order ÷ annual holding cost per unit)

For an illustrative item selling 36,500 units a year, with a $200 cost per order, and a $4 annual holding cost per unit, EOQ is √(2 × 36,500 × 200 ÷ 4), or about 1,911 units. EOQ assumes steady demand, a fixed order cost, and no quantity discounts, so most teams use it as a starting point and round to case packs or supplier minimums.

What is the safety stock formula?

Safety stock is the buffer that absorbs the difference between expected and actual demand during the replenishment lead time. The standard formula, when only demand varies, is:

Safety stock = z × σd × √L

Here z is the service factor for the target cycle service level, σd is the standard deviation of daily demand, and L is the lead time in days. The reorder point, which tells you when to order, equals expected demand over the lead time plus safety stock.

Take an illustrative item with average daily demand of 100 units, a daily standard deviation of 30, a nine-day lead time, and a 95% cycle service target (z ≈ 1.65). Safety stock is 1.65 × 30 × 3, or about 149 units, and the reorder point is 900 + 149 = 1,049 units.

When lead time also varies, the formula widens:

Safety stock = z × √(L × σd² + d² × σL²)

If the same supplier’s lead time has a standard deviation of two days, safety stock rises to roughly 362 units. Lead-time variability more than doubles the buffer in this example, which is why supplier reliability deserves as much attention as forecast accuracy. Both versions assume demand and lead time are normally distributed and independent of each other (King, APICS Magazine).

Service level targets and segmentation

The z value comes from a service target. Cycle service level is the probability of not stocking out during a replenishment cycle. Fill rate is the share of demand filled from stock on hand. The two can diverge widely for the same item. Most companies set targets by segment, using ABC analysis (by revenue) and XYZ analysis (by demand variability), so an “AX” item might get 98% and a “CZ” item 85%.

What is multi-echelon inventory optimization?

Multi-echelon inventory optimization sets buffers across every tier of a network at once, instead of letting a distribution center and the stores it serves each hold protection against the same uncertainty. The theory began with Clark and Scarf’s 1960 work on serial systems (Management Science). The guaranteed-service approach developed by Graves and Willems made it practical to decide where in a network safety stock should sit (Manufacturing & Service Operations Management). MEIO typically lowers total network inventory at the same service level by pooling buffers upstream, where one unit can protect several downstream locations.

How does mathematical optimization fit in?

Mathematical optimization sets inventory as part of one model of the network, where the formulas above work item by item. A linear or mixed-integer program describes the decisions (how much to make, stock, and ship, and where), the constraints that bound them, and an objective such as total cost. A solver then searches for the best plan that satisfies every constraint. Those constraints come in two kinds. Hard constraints, such as plant capacity or a minimum order quantity, can never be broken. Soft constraints, such as a preferred production run length, can be bent at a stated cost. Stochastic and robust versions add uncertainty, either by optimizing across many demand scenarios or by protecting the plan against a defined range of outcomes (Bertsimas and Thiele, Operations Research).

Which metrics show whether inventory is optimized?

Inventory optimization is judged by cash and service together, so no single metric is enough on its own:

• Inventory turns or days of supply, read against fill rate or OTIF, so a stock reduction always carries the service it held.

• Excess and obsolete inventory, read against stockouts and lost sales at SKU and location level, where totals can hide both in the same item family.

• Expedite and freight spend, which rises when buffers sit in the wrong place even if the total is right.

• Forecast error at the level where stock is held, since national accuracy says little about one distribution center.

Inventory optimization methods compared

Where do the classical inventory optimization assumptions break?

The classical methods break where real networks violate their assumptions, and planners compensate by hand. Each method in the comparison above is sound on its own terms, but ask an inventory planner how they set safety stock and you will hear about a formula followed by a list of exceptions. The formula says 149 units, but the planner holds 250 because the supplier in Vietnam missed two of its last five ship dates. The formula says 95% service for a “B” item, but the item is in a retailer’s promotion next month, so it gets treated as an “A.”

Each correction is reasonable. Together they move the effective inventory policy into spreadsheets and planners’ heads. A spreadsheet can hold the exceptions, but it can’t test hundreds of alternative policies, answer questions about the data quickly, or compare how each policy would actually have performed. When the planner who built it leaves, the policy leaves too (Why Spreadsheets Aren’t the Answer).

Four gaps account for most of that hand correction.

Demand rarely looks normal at the level where decisions happen. A normal curve fits a high-volume item at the national level reasonably well. At SKU and location, demand is often intermittent or skewed, so the formula sets buffers too thin for some items and too heavy for others.

History records what you sold, which can differ from what customers wanted. When an item stocks out, the sales record shows the stock you had, and the unmet demand disappears from the data. When an order is re-routed from a constrained plant to another one, history shows a demand shift that was really a capacity decision. A model fitted to that record learns your constraints along with your market, and every stocking decision built on it tightens those constraints again. Growth gets planned out of the business one cycle at a time.

Disruptions arrive together. The wider safety stock formula treats demand variability and lead-time variability as independent. A tariff change or a port closure can move demand and lead times at once, across many items and suppliers. Buffers sized as if those shocks were unrelated run short exactly when they are needed, which is close to what Nike’s in-transit inventory showed in 2022.

The forecast and the stocking decision are optimized for different things. Traditional inventory optimization runs in sequence: a forecasting step minimizes error, then an inventory step takes that forecast as a fixed input. A forecast can improve on its own scoreboard while the inventory built on it gets no better, because it was never told which errors are expensive. Over-forecasting a slow, low-margin item by 20% costs little. Under-forecasting a high-margin item with a twelve-week lead time by the same 20% can cost a quarter of its sales.

The gaps share a cause. The formulas see history. The structure of the network and the knowledge in planners’ heads sit outside them, and that missing context is what planners keep adding back by hand.

How is AI changing the theory of inventory optimization?

AI changes the theory by learning from the business the terms the classic equations ask planners to assume, such as the shape of demand and the cost of each kind of error. Three lines of research show how.

Learning from the cost of the decision. Elmachtoub and Grigas proposed training prediction models on the downstream decision error they cause, and showed this can produce better decisions than models trained purely for accuracy, especially when the prediction model is imperfect (Management Science). Since every real forecast is imperfect, that covers most of practice.

Going from data straight to a decision. Bertsimas and Kallus used machine learning to set inventory decisions directly from data, without first estimating a demand distribution. For a large media distributor, combining sales history with external signals such as search trends closed 88% of the gap between a decision based on sales history alone and one made with perfect foresight (Management Science). Their method also handles censored observations, the stockout problem described in the second gap.

Learning a policy where no formula exists. Gijsbrechts, Boute, Van Mieghem, and Zhang tested deep reinforcement learning on lost-sales, dual-sourcing, and multi-echelon problems, which have no clean closed-form answer, and found the learned policies matched the best known heuristics (Manufacturing & Service Operations Management). That is more modest than much AI marketing implies. Our reading: on a textbook problem with stable inputs, a learned policy reaches parity, and the larger gains come from learning parts of the business a heuristic never sees.

How is AI changing the practice of inventory optimization?

In practice, AI is being used in two very different ways: to make the existing inventory decision faster, or to make a better decision. McKinsey reported that early adopters of AI-enabled supply chain management had improved inventory levels by 35% and logistics costs by 15% compared with slower-moving competitors (McKinsey). That average blends both uses.

Faster versions of the same decision

Much of what is sold as AI for inventory automates the existing workflow: recalculating safety stock nightly, flagging exceptions, drafting purchase orders, answering questions in a chat window. Those planner hours matter, but automation’s value is capped by the size of the planning team, and the policy underneath stays the same. A safety stock formula recalculated every night still assumes normal, independent demand and lead time.

Starbucks shows where task automation stops. In 2025 it rolled out a computer-vision system to automate store inventory counts, with restocking orders meant to follow. About nine months later it dropped the system, which employees found unreliable, and moved to one standardized counting process. It is now working toward replenishing stores within 24 hours, with CEO Brian Niccol saying the goal is that “we’re never out of stock” (Supply Chain Dive). Faster counting would improve one input to that goal. Reaching it depends on deciding how much to send to each store every day, which is the inventory optimization problem.

Better decisions about where cash sits

Optimization reaches every dollar of inventory planners manage, where automation reaches the hours they spend managing it. AI-led inventory optimization looks different in four specific ways.

An economic objective replaces blanket service targets. Instead of assigning 95% to every “B” item, the model is directed at the trade-off the business cares about, such as margin captured against working capital and expedite spend. Service level stays in the plan as a hard constraint the result must hold, and preferences such as longer production runs enter as costs the model can weigh against everything else.

Inventory and production decisions get made together. Many of the largest inventory gains sit at the boundary between functions: which raw material to buy, which plant should make which product, which lane to ship it on, and when to substitute one item for another. Sequential tools hand each decision to a separate module. A model that sees the whole problem can find positions none of them would reach alone.

Planner knowledge becomes an input the model can use. Planners know things no system records: a supplier’s quality problem, a retailer’s plan to reset a category, a competitor pulling out of a region, or a launch slipping by a month. Fildes and colleagues found that large planner revisions to statistical forecasts tend to improve accuracy, while small, frequent ones add little (International Journal of Forecasting). In a spreadsheet, that knowledge becomes a manual buffer on one item. In a planning platform that can reason about it, the knowledge changes the plan for every SKU and location it affects and stays on record for the next cycle.

The team compares many plans before choosing one. In a classic tool, each alternative service policy means another run and a manual comparison. A model of the business can evaluate many inventory positions before the review, so the discussion starts from trade-offs.

What AI does not change

Lead times, minimum order quantities, shelf life, warehouse capacity, and supplier contracts are physical and commercial facts. A good model represents them accurately and finds the best position within them.

AI is also not always worth the effort. A business with a few hundred stable, high-volume SKUs, one warehouse, and reliable suppliers will get most of the available value from well-maintained safety stock formulas and a disciplined review cadence. The case for AI grows with SKU count, location count, demand volatility, and how much of the effective policy currently lives in manual corrections.

Where does MEIO fit when AI enters the picture?

To much of the industry, “inventory optimization” means MEIO, and MEIO is a real advance over setting buffers one location at a time. Its limitation is structural: it fixes the network’s constraints in advance and front-loads the work, so it depends on having guessed those constraints right. When a plant hits capacity, a supplier changes its lead time, a lane is disrupted, or a product is substituted, the configured network no longer matches the real one until someone updates it. The same limit applies to any solver. A solver can prove that a plan is optimal for the model it was given. Whether that model still describes the network is a separate question, and the answer changes every time a plant, supplier, lane, or product does.

An AI approach treats those constraints as part of what the model learns, and keeps learning them. MEIO solves an equation after someone has fixed its terms. The newer approach optimizes the supply chain those terms describe.

What does AI inventory optimization look like in practice?

In practice, AI inventory optimization means one model of the business deciding stock and production together, using what the business has recorded and what its people know. Omnifold builds reasoning AI for supply chain planning. Each customer receives its own planning platform containing its own model: a reasoning model of the supply chain, trained with reinforcement learning to reduce forecast error and directed at an economic objective. In plain terms, the model learns how your network actually behaves and aims every plan at profit and cash.

The model draws on three inputs, in what Omnifold calls the Context Gap: history (what happened), structure (how the network works), and knowledge (what people know is changing). Inventory Planning is one of four co-equal functions in the platform, alongside Demand Planning, Production Planning, and Materials Planning, and all four share the same model of the business.

What would change for the item in the worked example?

Go back to the illustrative item from the safety stock section. The formula says 149 units. The planner holds 250 because the supplier missed two of its last five ship dates. Both numbers are reasonable, and nobody has tested either one.

A model of the business starts from that supplier’s actual delivery record. If its lead time varies by about a day and a half, the lead-time formula puts the buffer near 289 units (illustrative), much closer to the planner’s instinct than to the textbook number. The model then weighs that buffer against what the item earns and what it costs to hold. A high-margin item heading into a retailer promotion may justify more. A low-margin item that customers will readily swap for a sister SKU may need less, because the sister SKU’s stock already covers part of the risk.

The planner’s knowledge changes too. In Omnifold’s platform, what the planner knows about the late supplier becomes an Enhancement: an input the model reasons with, applied to every item that supplier ships. In a spreadsheet, the same knowledge would have padded one row.

What happened at a paper packaging company?

A multi-facility paper packaging company buys paper in rolls of set widths and cuts them to fill customer orders of many sizes, so its inventory question was tangled up with a production question: which widths to stock, and how to cut them. Treated as one problem, the plan cut inventory by 41% and improved margin by 8%, while service held at 99.98%.

Is your inventory optimization hitting its ceiling?

These signals suggest the method, rather than the team, is limiting results:

1. Planners routinely override calculated safety stock. If most of the effective policy lives in manual buffers, the formula is no longer setting your inventory.

2. Forecast accuracy has improved, but inventory and service have not. The forecast is being optimized for a metric the inventory decision does not use.

3. Excess and stockouts appear in the same item family at the same time. Aggregate buffers are masking errors at SKU and location level.

4. Stockout history feeds straight into the next forecast. Lost sales are being recorded as low demand, and buffers shrink where they should grow.

5. Your MEIO configuration is older than your last network change. New plants, lanes, suppliers, or channels mean the optimizer is solving for a network you no longer run.

6. Testing a different service policy takes days, or lives in one person’s spreadsheet. If the team can’t compare several inventory plans before the review, every decision rests on one guess.

If two or more apply, start by measuring how far your planners’ manual buffers sit from the calculated ones, item by item. That gap is the clearest estimate of how much of your inventory policy the current method is no longer making.

Frequently asked questions

What is inventory optimization? Inventory optimization is the practice of deciding how much stock to hold of each item, at each location, so the business meets its customer service targets at the lowest total cost. It covers order quantities, reorder points, safety stock and, in multi-tier networks, where buffers should sit.

What is the safety stock formula? The most common version is safety stock = z × σd × √L, where z is the service factor for the target service level, σd is the standard deviation of demand per period and L is the lead time in the same periods. When lead time also varies, the formula becomes z × √(L × σd² + d² × σL²). Both versions assume demand and lead time are normally distributed and independent.

What is the difference between inventory optimization and inventory management? Inventory management is the operational work of tracking, receiving, storing, and moving stock. Inventory optimization is the decision layer above it, setting how much of each item to hold at each location and when to reorder, so service targets are met at the lowest total cost.

How does AI improve inventory optimization? AI improves inventory optimization mainly by learning how demand and lead times actually behave, and what each kind of error costs, from the business itself, where classical formulas assume them. Research shows that models trained on the cost of the inventory decision can outperform models trained on forecast accuracy alone.

Is multi-echelon inventory optimization (MEIO) the same as AI inventory optimization? No. MEIO is a set of mathematical methods for placing safety stock across a multi-tier network, given a fixed description of that network and its constraints. AI inventory optimization can include multi-echelon decisions, and it learns the network’s behavior and constraints from data and updates them as the business changes.

When is AI not worth it for inventory optimization? A business with a few hundred stable, high-volume items in one warehouse, with reliable suppliers, can usually get most of the available value from classical formulas and a disciplined review process. AI adds the most value with many SKUs, many locations, volatile demand, or frequent disruption.

Related reading

• Every Supply Chain Needs a Different Forecast. Your APS Gives Everyone the Same One.

• Why Predicting Isn’t Understanding

• Why Spreadsheets Aren’t the Answer to Demand Forecasting

Sources

• US Census Bureau, Manufacturing and Trade Inventories and Sales, July 2026

• IHL Group, Retail Inventory Crisis Persists Despite $172 Billion in Improvements (2025)

• CNBC, Target’s profit falls nearly 90% as it unloads unwanted inventory (2022)

• NIKE, Inc., Fiscal 2023 first quarter results, SEC exhibit 99.1 (2022)

• Supply Chain Dive, Starbucks targets 24-hour inventory replenishment (2026)

• Erlenkotter, D. (1990). Ford Whitman Harris and the Economic Order Quantity Model. Operations Research.

• King, P. L. (2011). Understanding safety stock and mastering its equations. APICS Magazine.

• Bertsimas, D. and Thiele, A. (2006). A Robust Optimization Approach to Inventory Theory. Operations Research.

• Clark, A. J. and Scarf, H. (1960). Optimal Policies for a Multi-Echelon Inventory Problem. Management Science.

• Graves, S. C. and Willems, S. P. (2000). Optimizing Strategic Safety Stock Placement in Supply Chains. Manufacturing & Service Operations Management.

• Elmachtoub, A. N. and Grigas, P. (2022). Smart “Predict, then Optimize”. Management Science.

• Bertsimas, D. and Kallus, N. (2020). From Predictive to Prescriptive Analytics. Management Science.

• Gijsbrechts, J., Boute, R. N., Van Mieghem, J. A. and Zhang, D. J. (2022). Can Deep Reinforcement Learning Improve Inventory Management? Manufacturing & Service Operations Management.

• Fildes, R., Goodwin, P., Lawrence, M. and Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments. International Journal of Forecasting.

• McKinsey & Company (2021). Succeeding in the AI supply-chain revolution.