A system says that ten units are available. The shelf, bin, or warehouse location contains six. That four-unit gap looks simple, but its business effect depends on where it occurs, which item is involved, what demand is expected, and whether the system will use the wrong number to promise, replenish, value, or move stock.
This is why inventory-record inaccuracy should not be reduced to a single accuracy percentage or a blanket estimate of “lost value.” A mismatch can suppress replenishment and create an unknown stockout. It can also hide usable stock, trigger an unnecessary order, occupy scarce capacity, or distort an inventory balance. The sign and location of the error matter as much as its size.
The practical objective is therefore not perfect records at any cost. It is a control system that finds the discrepancies most likely to affect service, working capital, operations, or reporting; corrects them quickly; and removes their causes.
Start with a precise definition
Inventory-record inaccuracy is the difference between a system’s recorded quantity and the quantity physically available at the relevant SKU-location.
Because terminology varies, every operating review should state its sign convention. A useful convention is:
Signed variance = physical quantity − system quantity
Under that definition:
- A negative variance means the system overstates available stock. This is often called ghost inventory. It can suppress replenishment, create false availability, and cause a picker or customer to discover an out-of-stock condition that the system did not know existed.
- A positive variance means the system understates stock. This is often called hidden inventory. It can prompt avoidable replenishment, increase working capital and space requirements, or leave product unavailable for sale even though it is physically present.
- A location variance means the total quantity may be correct at an aggregate level while stock is assigned to the wrong shelf, bin, store, lot, or status. A company-wide total can therefore reconcile while the operationally relevant SKU-location remains wrong.
The distinction is not semantic. A system that overstates stock by four units creates a different decision error from one that understates it by four.
The problem is material, but popular prevalence figures need restraint
The most frequently cited field study remains Nicole DeHoratius and Ananth Raman’s analysis of nearly 370,000 inventory records across 37 stores of one retailer. The researchers found 65% of the records to be inaccurate and observed substantial variation across product categories and stores. They also found that auditing practices mitigated inaccuracy, while store complexity and distribution structure could exacerbate it. The study is rigorous and operationally useful, but it was published in 2008 and examined one retailer. Its 65% figure is evidence that severe inaccuracy can exist, not a universal benchmark for every industry, system, or company in 2026.
Newer evidence reinforces the importance of context. A 2026 grocery-retailing study covered about 24,000 SKUs across 11 stores. Inventory inaccuracy was positively associated with average inventory level, restocking frequency, and perishability, and negatively associated with promotional activity. In a field quasi-experiment, an inventory audit was associated with an 11% store-wide sales lift, concentrated in items for which system inventory exceeded physical inventory; the effect was more pronounced for perishables. These are important findings, but they come from one grocery setting and a quasi-experimental design, so the sales effect should not be treated as a transferable ROI promise.
Together, the studies support a more defensible conclusion: inventory-record inaccuracy can be widespread and economically important, but its frequency and impact vary by category, process, environment, and error type.
How a mismatch becomes an operational or financial problem
A record discrepancy causes damage when another process trusts it.
Availability and fulfillment
When the system shows stock that is not physically available, an order may be accepted, a pick task released, or a replenishment decision deferred. The failure appears later as a short pick, substitution, cancellation, late shipment, or shelf stockout. The cost is not simply the value of the missing unit; it can include extra labor, expedited movement, customer-service effort, lost margin, and reduced service reliability.
Replenishment and working capital
When physical stock exceeds the record, the system may order more. That increases inventory, consumes storage and handling capacity, and can raise markdown, obsolescence, or spoilage risk. For constrained items, an inaccurate location can also send scarce supply to the wrong node while another location faces a shortage.
Warehouse productivity
Discrepancies create searches, recounts, exception handling, and rework. A 2023 system-dynamics study modeled inventory inaccuracy and cycle counting across five warehouse-performance scenarios. Its simulations linked inaccuracy to lower picking productivity, higher lost sales, and inventory build-up. It also found that cycle counting alone was insufficient in some modeled warehouse conditions, where receiving and picking errors also needed attention. Because this was a simulation study rather than a field trial, its value is in demonstrating mechanisms and tradeoffs, not supplying a universal effect size.
Financial reporting
Inventory is both an operating resource and an accounting balance. The US Public Company Accounting Oversight Board’s inventory-audit standard recognizes physical observation and comparison with perpetual records as core procedures. It also states that tests of accounting records alone are insufficient when the auditor has not otherwise obtained evidence about quantities through physical counts and related procedures. The standard does not imply that every operational discrepancy is a material financial misstatement. It does show why system records cannot be treated as self-validating evidence of physical inventory.
Why a single “inventory accuracy” percentage is inadequate
An exact-match rate is useful, but it can conceal the risk that operators need to manage.
Imagine two locations, each reporting 98% record accuracy. At the first, the mismatches are one-unit positive variances on slow-moving, low-value items. At the second, the mismatches are negative variances on a few high-demand components that can stop production or cause customer orders to fail. The headline metric is identical; the operating exposure is not.
A useful measurement stack separates prevalence, direction, magnitude, concentration, and consequence.
| Measure | What it reveals | What it can miss |
|---|---|---|
| Exact-match rate by SKU-location | How often physical and system quantities agree | Size, sign, and business importance of errors |
| Signed unit variance | Whether records tend to overstate or understate stock | Materiality across items with different values or demand |
| Absolute unit and value variance | Total magnitude without positive and negative errors canceling | Availability impact and location criticality |
| Negative-variance rate | Exposure to ghost inventory and unknown stockouts | Hidden stock and excess-replenishment risk |
| Unknown out-of-stock rate | Cases where the system shows stock but none is available | Partial shortages and positive variances |
| Location and status accuracy | Whether stock is in the usable node, bin, lot, or quality state | Total network-level value |
| Count yield | Material discrepancies found per count or labor hour | Errors outside the selected population |
| Recurrence and time to correction | Whether root causes persist and controls respond quickly | The value of errors not yet detected |
These measures should be segmented by item velocity, margin or value, criticality, perishability, lead time, node, and process. Segmentation prevents a large population of harmless exact matches from masking a small population of consequential failures.
Count for the objective, not for tradition
Many organizations use annual physical counts, ABC schedules, or random cycle counts because the cadence is familiar. The evidence suggests that the selection rule should follow the decision objective.
A study using audit data from European retailer dm-drogerie markt compared rule-based and model-based count-prioritization methods. Policies favoring high sales volume, high inventory, and past errors found more than twice as many discrepancies as random selection. A different rule that prioritized low recorded inventory found more than eight times as many unknown out-of-stocks as random selection. The lesson from the 2022 study is not that one policy always wins. It is that a counting program optimized to detect quantity discrepancies may differ from one optimized to find silent stockouts.
Analytics can also improve targeting, but claims need boundaries. A 2024 paper framed large-scale inventory detection as anomaly identification across stores and SKUs. On synthetic data and real data from a consumer-goods retailer, the proposed method produced up to a tenfold cost reduction relative to incumbent anomaly-detection approaches. That result supports the use of cross-sectional signals where longitudinal histories are weak. It remains an algorithm-specific result in the studied data, not proof that any “AI inventory” product will produce the same outcome.
The count policy should therefore begin with a question:
- Are we trying to find unknown stockouts?
- Protect high-value inventory?
- Reduce recurring transaction errors?
- Validate a financial balance?
- Improve pick reliability?
- Detect shrinkage, damage, or status mistakes?
- Learn which process creates the discrepancy?
Without a stated objective, a high count volume can become activity without control.
Cycle counting corrects records; it does not automatically fix the process
A count can reveal and correct a mismatch. It cannot, by itself, prevent the next one.
Common causes include receiving mistakes, unit-of-measure conversions, unrecorded damage or loss, mis-picks, returns posted to the wrong status, production consumption not captured at the right time, transfers completed in one system but not another, and stock placed in the wrong location. Timing differences can also create apparent errors when physical movement and system posting occur at different moments.
Recent analytical work makes the distinction between causes explicit. A 2025 Bayesian inventory model separated unobservable physical loss from transaction errors in the record process. The study found that transaction errors can substantially increase uncertainty about shelf inventory and make optimal replenishment calculations more difficult. Its proposed policy is mathematical rather than a general implementation recipe, but it supports an important operating principle: not all inventory uncertainty has the same source or should receive the same response.
A mature control loop has five stages:
- Detect: Prioritize counts and alerts according to the failure mode being controlled.
- Classify: Record the sign, magnitude, location, status, timing, and probable process source.
- Correct: Update the authorized record, resolve affected orders, and reconcile connected systems.
- Prevent: Change the receiving, picking, transfer, return, production, or master-data process that produced the error.
- Verify: Measure recurrence and the operational outcome after correction.
Reason codes are essential. “Adjusted after count” describes an accounting action, not a cause. A useful taxonomy distinguishes transaction omission, duplicate posting, timing, wrong location, wrong unit of measure, damage, shrinkage, return-status error, production-consumption error, and unresolved cause.
Estimate financial impact without pretending every unit is a loss
The safest approach is to model decision consequences rather than multiply every variance by unit cost.
For each material discrepancy, estimate the applicable pathways:
- margin at risk from an order that could not be fulfilled;
- incremental labor for searching, recounting, re-picking, or rescheduling;
- expedite or transfer cost;
- additional carrying, storage, spoilage, markdown, or obsolescence exposure;
- production downtime or lost throughput for a constrained component;
- write-down or control remediation where the discrepancy affects reporting;
- customer credits, penalties, or service recovery.
Do not add every pathway automatically. A hidden unit that is found and sold next week is not equivalent to a missing unit, and a record error that is corrected before any dependent decision may have little financial effect. Use observed outcomes where available, scenario ranges where they are not, and label assumptions.
Practical recommendations
In the next week
Define one sign convention and one SKU-location grain. Separate negative, positive, location, and status errors. Review a risk-based sample that includes low recorded stock, high-activity items, critical components, perishables, and recently adjusted records. Name an owner for each major source process.
In the next quarter
Replace one undifferentiated accuracy target with a metric stack. Match cycle-count rules to explicit objectives, record root-cause reason codes, and measure count yield and recurrence. Link operational adjustments to finance where quantities or values are material. Review whether corrections propagate consistently across ERP, warehouse, planning, commerce, and reporting systems.
In the next year
Invest in better transaction capture and process controls before buying more detection alone. Add anomaly detection where the data and operating scale justify it, but benchmark it against current count yield and false-positive cost. Build a closed loop from discrepancy to process change to outcome, with auditable evidence of who approved adjustments and why.
What remains uncertain
The evidence base is strongest in retail and often comes from a small number of companies. Definitions, count tolerances, item populations, and operational settings differ, which makes prevalence figures difficult to compare. Newer algorithmic studies show promising targeting methods, but external validation across industries remains limited. Even well-designed field studies cannot guarantee that the same intervention will produce the same sales or productivity effect elsewhere.
There is also no universal optimal accuracy level. The economically rational target for a low-value, slow-moving item may differ from the target for a regulated lot, a perishable SKU, or a component capable of stopping a production line.
Conclusion
Inventory accuracy is not one number, and a mismatch is not one kind of loss. The relevant question is whether the record error will cause a bad promise, replenishment, movement, valuation, or operating decision.
Organizations improve control when they measure discrepancies by sign, location, magnitude, and consequence; target counts to a defined objective; classify causes rather than merely posting adjustments; and verify whether corrective actions reduce recurrence. That approach treats inventory accuracy as an operating control system, not a periodic counting exercise.
Sources
- Inventory Record Inaccuracy: An Empirical Analysis: Management Science, April 1, 2008. Field analysis of nearly 370,000 records across 37 stores of one retailer; supports prevalence, variation, and process-factor claims.
- Inventory Record Inaccuracy in Grocery Retailing: Impact of Promotions and Product Perishability, and Targeted Effect of Audits: Journal of Business Logistics, July 7, 2026. Study of about 24,000 SKUs in 11 grocery stores and a field quasi-experiment; supports context-specific drivers and audit effects.
- Evaluating Count Prioritization Procedures for Improving Inventory Accuracy in Retail Stores: Manufacturing & Service Operations Management, September 8, 2022. Retail audit-data study; supports objective-specific count prioritization.
- Fixing Inventory Inaccuracies at Scale: Manufacturing & Service Operations Management, March 14, 2024. Algorithmic study using synthetic and retailer data; supports cross-sectional anomaly detection, with stated limits.
- Inventory Systems with Record Inaccuracy: Transaction Errors vs. Unobservable Loss: Manufacturing & Service Operations Management, June 4, 2025. Analytical Bayesian model; supports separating transaction errors from physical loss.
- The Impacts of Inventory Record Inaccuracy and Cycle Counting on Distribution Center Performance: Production, April 2023. System-dynamics simulation; supports operational mechanisms and the limits of cycle counting alone.
- AS 2510: Auditing Inventories: Public Company Accounting Oversight Board. Supports the role of physical observation, count testing, and reconciliation in audit evidence.