Audit Sampling and Inventory Testing

Introduction

Inventory is simultaneously a physical resource, a financial statement balance, a service-level enabler, and a major exposure to theft, damage, obsolescence, and recording error. Consequently, a stock audit cannot normally examine every unit in a large network. The auditor must obtain persuasive evidence from a carefully designed subset of locations, stock keeping units (SKUs), transactions, and monetary values. Audit sampling is therefore not merely a cost-saving device: it is a method for making defensible inferences about an inventory population.

This chapter develops sampling and testing methods for graduate supply-chain professionals. It links statistical sampling, professional judgment, test counts, confidence, tolerable error, stratification, monetary-unit sampling (MUS), and exception evaluation. The emphasis is on designing evidence that is representative, traceable, and useful for controlling shrinkage rather than simply producing a pass-or-fail count result.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Distinguish a population, sampling unit, sample, exception, and projected error.
2. Select probability, systematic, stratified, judgmental, or monetary-unit sampling appropriately.
3. Explain how confidence, tolerable error, expected error, variability, and risk affect sample size.
4. Design and document inventory test counts and record-to-floor and floor-to-record tests.
5. Evaluate exceptions without confusing isolated errors with evidence of control failure.
6. Apply sampling results to operational, financial, and shrinkage-control decisions.
7. Recognize the limitations, ethical requirements, and technology implications of inventory sampling.

1. The Audit Population and the Evidence Problem

An inventory population is the complete set from which a sample is selected. It must be defined by scope, date, location, ownership, and measurement basis. Examples include all pallet positions in a distribution centre at the count date, all SKU-location records in an enterprise resource planning (ERP) system, or all inventory movements during a period.

A population is not necessarily a list of physical units. It may be:

• Unit-based: individual cases, pallets, bins, or serialized items.
• Record-based: SKU-location records, adjustment documents, receiving reports, or dispatch transactions.
• Value-based: inventory lines or monetary units represented by recorded currency amounts.
• Location-based: warehouses, stores, yards, quarantine cages, and third-party logistics sites.

The sampling frame is the operational list used to select items. A frame can be incomplete, duplicated, stale, or systematically biased. For example, selecting only active ERP locations may exclude unrecorded stock in staging areas. Before calculating sample size, the auditor should reconcile the frame to the inventory ledger and investigate missing or unusual categories.

The audit objective determines what constitutes an exception. In a quantity-existence test, the auditor begins with a recorded item and verifies that it physically exists. In a completeness test, the auditor begins with a physical item and verifies that it is recorded. Other assertions include condition, ownership, valuation, cut-off, location, and authorization of adjustments. A sample suitable for existence may be unsuitable for completeness.

Important terminology

2. Sampling Approaches

2.1 Statistical and non-statistical sampling

Statistical sampling uses random selection and probability theory to quantify sampling risk. Every population unit has a known, non-zero probability of selection, and the design supports estimation of precision or confidence. Randomness does not mean that professional judgment disappears; judgment is required to define the population, select the risk model, classify exceptions, and interpret results.

Non-statistical, or judgmental, sampling uses professional judgment rather than a formally quantified probability model. It can be appropriate when the population is small, highly heterogeneous, exploratory, or dominated by individually important items. It is also useful for selecting unusual locations, high-risk SKUs, recent adjustments, and suspected fraud indicators. However, a judgmental sample cannot normally support a claim such as “there is 95% confidence that the error rate is below 2%” unless the design also satisfies probability-sampling requirements.

A strong audit frequently combines methods: 100% testing of individually significant or high-risk items, stratified random sampling of the remainder, and targeted judgmental testing of unusual transactions.

2.2 Random, systematic, and haphazard selection

In simple random sampling, each unit has the same selection probability. A random-number generator should select identifiers from a complete, frozen frame. In systematic sampling, the auditor selects every kth unit after a random start. If the frame contains periodic patterns—such as replenishment cycles arranged in repeating warehouse zones—systematic selection can be biased unless periodicity is assessed.

Haphazard selection attempts to avoid conscious bias but does not establish known probabilities. It is weaker than random selection and should not be described as random. Selection should be reproducible: another auditor should be able to identify the frame, seed or start point, interval, exclusions, and replacements.

2.3 Stratification

Stratification divides a population into relatively homogeneous groups and samples each group separately. Common strata include monetary value, demand velocity, temperature class, theft risk, product age, location, and condition. A useful design often includes:

• a 100% stratum for individually material, sensitive, or high-value items;
• a high-risk stratum with a larger sampling fraction; and
• a general stratum selected randomly or systematically.

Stratification improves efficiency because it reduces variation within strata and ensures that important categories are not overwhelmed by numerous low-value items. Results should be evaluated by stratum and, where appropriate, combined using a documented weighting or projection method. An error in a high-value stratum should not be diluted by a large number of low-value, low-risk units.

3. Sample-Size Logic

Sample size is a design decision, not a universal constant. Relevant inputs include population size, confidence, tolerable error, expected error, variability, risk of material misstatement or control failure, and the cost and quality of testing.

3.1 Estimating a proportion

When the audit estimates a population proportion—such as the percentage of bins containing a count discrepancy—the approximate sample size for a large population is:

Here, n₀ is the initial sample size, Z is the standard-normal critical value for the selected two-sided confidence level, p is the anticipated population proportion, and e is the desired absolute margin of error. The formula assumes a probability sample, independent or approximately independent observations, and a proportion estimate rather than a monetary-error estimate.

If the population size N is not large relative to the initial sample, the finite-population correction is:

The correction matters when the sampling fraction is substantial. For an unknown or conservative expected proportion, p = 0.50 maximizes p(1 − p) and therefore produces the largest sample for a specified confidence and precision. This is not a statement that half the inventory is expected to be wrong; it is a conservative planning assumption.

For a 95% confidence level, Z is approximately 1.96. Suppose a population contains 10,000 bin records, the anticipated discrepancy proportion is 0.10, and the desired margin of error is 0.05. Then:

Thus, the initial sample is approximately 139 records. Applying the finite-population correction:

The auditor would ordinarily round upward to 137 records, subject to risk-based additions and operational feasibility. If expected error rises, the sample should generally increase. If the tolerable margin of error becomes smaller or confidence increases, the sample also increases.

3.2 Attribute sampling and control testing

Attribute sampling asks whether a control characteristic is present—for example, whether every tested inventory adjustment has an approved reason code. The planned sample size should reflect the tolerable deviation rate and expected deviation rate. A sample with zero observed deviations does not prove zero deviations in the population. It only provides evidence bounded by the selected confidence and model assumptions.

A common approximation for a zero-exception upper bound on a population proportion is:

where pᵤ is the approximate one-sided upper bound, α is the significance level, and n is the number of independently selected units with zero exceptions. For example, with 60 independent tests and α = 0.05:

Thus, zero observed exceptions is compatible with an underlying exception rate of approximately 4.9% at this approximate 95% one-sided bound. The approximation is most useful for explaining why “no exceptions found” is not equivalent to “no exceptions exist.” Exact binomial methods may be preferable, especially for small samples or formal reporting.

3.3 Monetary and variable sampling

Variable sampling estimates a numerical quantity such as average inventory error, mean unit-cost error, or total shrinkage. It requires attention to the distribution of values and often benefits from stratification because inventory values are typically skewed. A small number of high-value items may dominate total exposure.

For a simple random sample estimating a mean, an approximate sample-size relationship is:

where s is the estimated standard deviation of the measured variable and E is the desired absolute precision in the mean, in the same monetary or physical units as the variable. This formula assumes a reasonably stable variance estimate and a probability sample. In practice, a pilot sample, historical data, or a conservative variance estimate is required; the formula should not be used with an invented standard deviation.

4. Confidence, Tolerable Error, and Exception Evaluation

Confidence describes the performance of a repeated sampling procedure, not the probability assigned to one already-completed interval under a strict frequentist interpretation. In professional communication, “95% confidence” should be accompanied by the population, parameter, method, and assumptions.

Tolerable error is linked to the audit objective. For a financial statement assertion it may be a monetary amount; for a cycle-count control it may be a deviation rate; for an operational objective it may be a unit or service-impact threshold. Tolerable error must be set before results are known, otherwise the auditor risks moving the goalposts.

Expected error influences design but does not replace tolerable error. If expected error is close to tolerable error, a small sample is unlikely to discriminate reliably between an acceptable and unacceptable population. If observed error exceeds tolerable error, the auditor should not simply increase the sample until a favorable result appears. The appropriate response is to investigate the cause, expand testing according to the audit methodology, correct the population, or conclude that the criterion was not met.

Exceptions should be classified by type and cause:

1. Counting error: wrong quantity, unit of measure, or pack conversion.
2. Identification error: incorrect SKU, lot, serial number, or location.
3. Timing error: receipt or dispatch recorded in the wrong period.
4. Condition error: damaged, expired, obsolete, or quarantined stock treated as saleable.
5. Ownership error: consignment or customer-owned stock included as company inventory.
6. Control error: unauthorized adjustment, missing approval, or unsealed access.
7. System error: interface, master-data, or configuration defect.

An exception record should show the sampled unit, book quantity and value, observed quantity and value, evidence obtained, monetary difference, root cause, responsible process, and disposition. Gross and net errors must be distinguished. Offsetting errors may reduce net financial misstatement while still demonstrating serious control weakness.

Projection of sample errors

For equal-probability unit sampling, a simple ratio projection can estimate total monetary error:

This method assumes the sample is representative and that the ratio of error to recorded value is reasonably stable. It is not appropriate to apply mechanically when the sample is stratified, includes a 100% stratum, contains negative or zero book values, or has highly unusual errors. Each stratum should be projected separately, and individually tested items should be treated directly rather than projected.

For example, a random stratum has recorded value $2,000,000. A sample representing $100,000 contains net overstatement of $3,000. The ratio projection is:

If tolerable error for that stratum is $50,000, the projected error exceeds tolerance. The auditor should consider sampling risk, confidence bounds, error characteristics, and whether correction or expanded testing is required. The $60,000 estimate is not an exact loss measurement.

5. Monetary-Unit Sampling

Monetary-unit sampling treats each currency unit as a sampling unit. A $100,000 inventory line has 100 times as many currency units—and therefore a greater chance of selection—as a $1,000 line. MUS is efficient for detecting overstatement because high-value recorded items receive greater selection probability. It is less effective for completeness, unrecorded inventory, or populations containing many zero and negative values.

A simplified systematic MUS design uses an interval:

where I is the sampling interval, T is total recorded monetary value, and n is the planned number of selections. A random start r is selected between 1 and I; selection points are r, r + I, r + 2I, and so forth. Each point is mapped to the cumulative monetary listing. A line crossed by several points may be selected more than once, signaling a larger projected influence or requiring the methodology’s prescribed treatment.

MUS commonly uses a confidence factor and an allowance for expected error to determine sample size. Exact formulas differ by audit methodology and assumptions about error patterns, so practitioners should use approved tables or software and document the selected reliability factor. MUS should not be presented as universally superior: it may underrepresent low-value items with high error frequency and can be distorted by negative balances, credit notes, and incomplete listings.

6. Inventory Test Counts and Physical Observation

A physical inventory test has two complementary directions:

• Book-to-floor: select recorded items and locate, identify, and count them. This primarily supports existence and condition.
• Floor-to-book: select items encountered physically and trace them to records. This primarily supports completeness and unrecorded stock detection.

Before counting, the audit team should understand warehouse layout, unit-of-measure conventions, blocked or quarantined areas, third-party stock, count instructions, and movement controls. The count date and cut-off sequence should be documented. Receiving and dispatch should be paused or controlled; otherwise, movements can create false discrepancies.

A test-count sheet should include location, SKU, lot or serial number, book quantity, auditor count, unit of measure, condition, ownership, counter and reviewer, time, and resolution. Independent recounts are appropriate for material differences. The auditor should avoid replacing an inaccessible or missing item with a convenient item unless the replacement rule was predefined and recorded; otherwise, selection bias is introduced.

Blind counts, in which the counter does not see the book quantity, can reduce confirmation bias. When operationally impossible, the auditor should record the limitation and use compensating procedures. Photographs, barcode scans, RFID reads, weigh-scale data, and geolocation can strengthen evidence, but technology does not eliminate the need to assess data completeness and reliability.

7. A Risk-Based Sampling Framework

A defensible inventory audit can follow this sequence:

1. Define the objective and assertion. State whether the test addresses existence, completeness, valuation, condition, cut-off, ownership, or control operation.
2. Define the population and frame. Reconcile the listing to the ledger, identify exclusions, and freeze the version used.
3. Assess risk and heterogeneity. Consider shrinkage history, access, product value, demand velocity, obsolescence, temperature sensitivity, and management override.
4. Select strata. Separate individually material, high-risk, and general populations.
5. Choose the sampling method. Use random or systematic selection for measurable inference, judgmental selection for targeted risk, and MUS for suitable monetary populations.
6. Determine sample size. Document confidence, tolerable error, expected error, variability, population size, and resource constraints.
7. Perform the test consistently. Use standardized instructions and preserve an audit trail.
8. Evaluate exceptions. Investigate, classify, quantify, project where appropriate, and assess root cause.
9. Conclude and communicate. Compare results with predetermined criteria and report limitations, unresolved items, and corrective actions.
10. Follow up. Verify corrections, repeat counts where necessary, and update risk assessments.

8. Applied Case: NorthStar Fulfilment Network

The following is an illustrative case, constructed for learning purposes.

Background

NorthStar operates three fulfilment centres and reports $48 million of inventory. The prior year showed low net shrinkage, but repeated cycle-count adjustments occurred in fast-moving consumer electronics. Management claims that the new barcode system has eliminated material discrepancies.

Problem

The audit objective is to assess existence and completeness at year-end and determine whether adjustment controls operate effectively. The frame contains 36,000 SKU-location records. High-value serialized devices represent $12 million, general inventory $30 million, and damaged or returned stock $6 million.

Analysis and approach

The audit team creates three strata. Serialized devices above a predetermined value threshold are tested 100%. Returned and damaged stock is sampled separately because condition and ownership risks differ from ordinary saleable stock. The general stratum is selected using a random seed from the frozen frame. A complementary floor-to-book sample is selected from warehouse zones, including staging and quarantine areas.

For the control test, 80 randomly selected inventory adjustments are examined for approved reason codes, independent review, and supporting evidence. Three lack documented review. The team does not conclude immediately that the control failure rate is 3.75%; it investigates whether the three exceptions are genuine, whether the sample was correctly selected, and whether the control criterion was met.

During book-to-floor testing, two serialized devices cannot be located. One is later found in a locked service cage and was incorrectly mapped to another location; the second was dispatched before the count but remained in the system because of an interface failure. During floor-to-book testing, five returned units are physically present but absent from the saleable inventory listing. They are traced to a returns-management queue and reclassified after investigation.

Evaluation

The first exception is a location and master-data error, not necessarily a loss. The second is a cut-off and interface error that may create an existence overstatement at the reporting date. The unrecorded returned units indicate a completeness and process-ownership problem. Although the net monetary difference may be below materiality, the concentration of exceptions in returns and serialized-device processes indicates elevated control risk.

The audit team expands testing around the interface, reviews dispatch cut-off documents, tests additional returns across all sites, and recommends daily exception reports, ownership of the returns queue, and independent review of location-master changes. Management corrects the year-end records and implements a post-count reconciliation.

Outcome and learning points

NorthStar’s barcode technology improved traceability but did not eliminate errors. Sampling identified not only quantity differences but also process failures, data-interface defects, and classification issues. The case demonstrates that a low net shrinkage number can conceal offsetting errors and that sample design must reflect the assertion and risk pattern. It also demonstrates why exceptions should be evaluated by cause, location, value, and recurrence rather than aggregated into one percentage.

9. Technology, AI, and Contemporary Practice

Warehouse management systems increasingly provide perpetual inventory data, cycle-count histories, scan events, camera analytics, RFID reads, and anomaly alerts. These data can support risk-based sample selection: for example, prioritize bins with repeated adjustments, unusual negative stock, frequent overrides, or long gaps between scans.

AI and machine-learning models can rank locations or SKUs by predicted shrinkage risk, but a prediction score is not a probability sample. If the audit conclusion requires statistically defensible population inference, the selected sample must retain known selection probabilities or use a documented probability-based design. Model-driven targeting is best treated as a complementary risk stratum unless its inferential properties are established.

Data governance is essential. Auditors should validate completeness, reconcile interfaces, preserve timestamps, control access to count results, and protect personal or commercially sensitive data. Automated image or scan recognition can misidentify packaging, mixed pallets, or damaged goods; exception queues require human review. A technology-rich audit can still fail through weak master data or uncontrolled overrides.

10. Limitations, Ethics, and Governance

Sampling cannot compensate for a defective population frame, uncontrolled inventory movement, ambiguous unit measures, or biased selection. Statistical confidence also does not measure non-sampling risk. Poor counting instructions, transcription errors, premature conclusions, and management pressure can invalidate otherwise elegant calculations.

Ethically, auditors must maintain independence, avoid suppressing inconvenient exceptions, disclose limitations, and protect confidential inventory information. Management should not direct auditors to exclude high-risk locations merely because results may be unfavorable. Conversely, auditors should not inflate exceptions by treating known duplicates or corrected errors as separate losses.

Governance should assign responsibility for corrections and root-cause remediation. A sound report distinguishes: observed errors, projected errors, upper confidence bounds where used, confirmed losses, unresolved differences, and control deficiencies. This distinction supports fair accountability and better operational decisions.

Best-Practice Checklist

• Define the assertion before selecting the sample.
• Reconcile and freeze the sampling frame.
• Separate individually significant and high-risk items.
• Use probability selection when making quantitative population claims.
• Document confidence, tolerable error, expected error, and assumptions.
• Use both book-to-floor and floor-to-book procedures where existence and completeness both matter.
• Blind counts when practical and control inventory movement.
• Record every exception with evidence, value, cause, and disposition.
• Project errors by stratum only when the method’s assumptions are defensible.
• Treat technology as evidence requiring validation, not as a substitute for audit judgment.
• Escalate patterns and root causes even when net value is small.

Chapter Summary

Inventory sampling converts limited audit work into evidence about a defined population, but its quality depends on the population frame, assertion, selection method, and evaluation criteria. Statistical sampling supports quantified sampling risk when selection probabilities are known; judgmental sampling is valuable for targeted, high-risk, or exploratory work but should not be represented as statistically representative. Stratification and 100% testing of individually important items improve efficiency and protect against dilution of material risks.

Sample size is influenced by confidence, tolerable error, expected error, variability, population size, and risk. Proportion, attribute, variable, and monetary-unit methods address different evidence needs and have different assumptions. Physical inventory testing requires both book-to-floor and floor-to-book directions, controlled movements, reliable unit measures, and documented test counts. Exceptions must be investigated, classified, quantified, and evaluated for recurrence and root cause. Finally, barcode, RFID, analytics, and AI can improve risk targeting and evidence quality, but they do not replace a valid sampling design, professional skepticism, or ethical governance.

Key Terms

Population: The complete set about which an audit conclusion is required.

Sampling frame: The list or data structure from which sample units are selected.

Sampling unit: The individual item, record, location, or monetary unit selected for testing.

Statistical sampling: Probability-based sampling that permits quantified sampling-risk evaluation.

Judgmental sampling: Selection based on professional judgment without formal probability inference.

Stratification: Division of a population into more homogeneous subpopulations.

Confidence level: The long-run coverage rate associated with a repeated sampling procedure.

Tolerable error: The maximum error acceptable for the audit objective.

Expected error: The error anticipated during sample planning.

Sampling risk: Risk that the sample-based conclusion differs from a whole-population conclusion.

Non-sampling risk: Risk arising from execution, interpretation, or data-quality failures.

Attribute sampling: Sampling used to assess whether a control characteristic or condition exists.

Variable sampling: Sampling used to estimate a numerical amount or measurement.

Monetary-unit sampling: Value-weighted sampling in which currency units are the sampling units.

Book-to-floor test: Tracing recorded inventory to physical stock.

Floor-to-book test: Tracing physical stock to inventory records.

Exception: A deviation from the specified audit criterion.

Projection: An estimate of population error derived from sample results.

Sampling frame bias: Distortion caused when the frame excludes or overrepresents parts of the population.

Shrinkage: Inventory reduction arising from loss, damage, theft, error, or unrecorded movement.

Review Questions

Short Answer Questions

1. What is the difference between a population and a sampling frame?
2. Why is a book-to-floor test primarily associated with existence?
3. Define tolerable error and expected error.
4. State one advantage and one limitation of monetary-unit sampling.
5. Why should inventory exceptions be classified by cause?

Long Answer Questions

1. Compare statistical and judgmental sampling in an inventory audit. Explain when a combined design is appropriate.
2. Explain how confidence, tolerable error, expected error, and variability influence sample size.
3. Describe a complete physical inventory test-count procedure, including cut-off and movement controls.
4. Explain stratification and develop a suitable stratification plan for a multi-site distribution network.
5. Discuss how AI, RFID, and warehouse data can improve sampling while preserving audit validity and governance.

Analytical / Application Questions

1. A population contains 10,000 records. Using 95% confidence, p = 0.10, and a 5% margin of error, calculate the initial sample size and apply the finite-population correction. Explain why the result should be rounded upward.
2. A sample from a random stratum has book value $100,000 and identified overstatement of $3,000. The stratum’s total book value is $2,000,000 and tolerable error is $50,000. Project the error and recommend an audit response.
3. An audit finds no exceptions in 60 adjustment-control tests. Using the approximate one-sided upper-bound relationship in this chapter, estimate the possible exception-rate bound at α = 0.05 and explain why management cannot conclude that the true rate is zero.

Multiple Choice Questions with Answers

Question 1

Which sampling method gives each population unit a known, non-zero probability of selection and supports quantified sampling risk?

A. Statistical sampling

B. Haphazard sampling

C. Convenience sampling

D. Unstructured observation

Correct Answer: A. Statistical sampling

Question 2

A book-to-floor inventory test primarily provides evidence about which assertion?

A. Completeness

B. Existence

C. Future demand

D. Supplier solvency

Correct Answer: B. Existence

Question 3

What is the main purpose of stratification?

A. To eliminate the need for documentation

B. To make every item have identical monetary value

C. To create more homogeneous groups and ensure important categories receive attention

D. To replace all random selection with judgment

Correct Answer: C. To create more homogeneous groups and ensure important categories receive attention

Question 4

Using n₀ = (Z²p(1 − p))/e², what is the approximate initial sample size when Z = 1.96, p = 0.10, and e = 0.05?

A. 39

B. 90

C. 139

D. 390

Correct Answer: C. 139

Question 5

Which situation is a limitation of monetary-unit sampling?

A. High-value recorded items receive greater selection probability

B. It is generally useful for overstatement testing

C. It may underrepresent numerous low-value items with frequent errors

D. It can use a cumulative monetary listing

Correct Answer: C. It may underrepresent numerous low-value items with frequent errors

Question 6

If a sample has book value $100,000 and error $3,000, and the population book value is $2,000,000, what is the simple ratio projection?

A. $6,000

B. $30,000

C. $60,000

D. $600,000

Correct Answer: C. $60,000

Question 7

Which statement about zero exceptions in a sample is correct?

A. It proves the population has no exceptions

B. It provides evidence bounded by the sampling method and assumptions

C. It means the control was tested in every instance

D. It removes non-sampling risk

Correct Answer: B. It provides evidence bounded by the sampling method and assumptions

Question 8

Which test begins with physically observed stock and traces it into the records?

A. Floor-to-book testing

B. Book-to-floor testing

C. MUS interval testing

D. Ledger-to-ledger testing

Correct Answer: A. Floor-to-book testing

Question 9

For a 95% confidence level in a normal approximation, the commonly used critical value Z is approximately:

A. 0.50

B. 1.00

C. 1.96

D. 3. Char

Correct Answer: C. 1.96

Question 10

Why should an AI-generated risk score not automatically be treated as a statistical sample?

A. AI cannot process inventory data

B. A risk score does not necessarily provide known selection probabilities

C. AI always selects only low-value items

D. Risk scores eliminate the need for audit evidence

Correct Answer: B. A risk score does not necessarily provide known selection probabilities