Supply Chain Analytics and Decision Support

Introduction and Context

Supply chain management (SCM) software has evolved from transaction processing toward integrated decision support. Enterprise resource planning (ERP), warehouse management, transportation management, procurement, customer relationship, and planning systems now generate large volumes of operational data. The managerial challenge is not simply to collect more data, but to convert data into timely, explainable, and economically sound decisions.

Supply chain analytics links data, models, visualization, and managerial action. It supports decisions such as how much inventory to hold, which supplier to prioritize, whether to expedite an order, how to allocate scarce capacity, and how to respond to a disruption. A well-designed analytical system distinguishes measurement from explanation, prediction from prescription, and statistical confidence from managerial judgment.

A useful decision-support architecture contains five connected layers:

1. Data layer: transactions, master data, external signals, event logs, and sensor data.
2. Semantic layer: common definitions for products, locations, customers, suppliers, time periods, currencies, and measures.
3. Analytical layer: descriptive, diagnostic, predictive, and prescriptive models.
4. Presentation layer: dashboards, scorecards, alerts, reports, and control-tower views.
5. Action and learning layer: workflows, approvals, execution, feedback, and model monitoring.

The final layer is essential. An analytics system that does not change decisions, actions, or learning is reporting infrastructure rather than effective decision support.

Learning Objectives

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

• distinguish descriptive, diagnostic, predictive, and prescriptive analytics;
• design supply chain KPIs with clear definitions, owners, targets, and action thresholds;
• evaluate dashboards, alerts, scorecards, and control-tower designs;
• calculate and interpret important service, inventory, cost, and forecast measures;
• apply scenario analysis, simulation, and optimization to supply chain decisions;
• assess decision quality, model risk, data quality, ethics, and governance;
• recommend an analytics operating model for a complex supply chain.

The Four Forms of Supply Chain Analytics

Descriptive analytics: what happened?

Descriptive analytics summarizes historical or current performance. Typical outputs include order fill rate, inventory by location, supplier lead-time averages, transportation cost by lane, and forecast accuracy by product family. Its value depends on consistent definitions and appropriate aggregation. A global average can conceal serious problems in a particular customer segment, region, or stock-keeping unit (SKU).

Diagnostic analytics: why did it happen?

Diagnostic analytics investigates causes, relationships, and deviations. Techniques include drill-down, variance analysis, segmentation, correlation analysis, process mining, and root-cause analysis. For example, a decline in on-time delivery may result from supplier lateness, carrier capacity, customs delays, inaccurate promised dates, or poor warehouse release discipline. Diagnostic analysis should test plausible explanations rather than assume that the most visible correlation is causal.

Predictive analytics: what is likely to happen?

Predictive analytics estimates future demand, arrival times, stockout risk, supplier failure, or transportation delays. Forecasts are conditional statements based on available information and assumptions; they are not guarantees. A predictive model should be evaluated using data that was not used to train it and should be monitored after deployment for drift and bias.

Prescriptive analytics: what should be done?

Prescriptive analytics recommends actions subject to objectives and constraints. Examples include inventory reorder decisions, allocation of constrained supply, production sequencing, transport-mode selection, and workforce scheduling. Prescriptive models may use optimization, simulation, decision rules, or combinations of these methods. A recommendation is useful only when its assumptions, constraints, costs, and implementation requirements are credible.

These forms are complementary. A control tower may show that service fell (descriptive), identify a capacity bottleneck (diagnostic), estimate future shortage exposure (predictive), and recommend reallocating inventory (prescriptive).

Designing Supply Chain Measures

A KPI is not merely a number on a screen. It is a defined measurement linked to an objective, decision, owner, time horizon, and action. A strong KPI specification includes:

• name and business purpose;
• numerator, denominator, and exclusions;
• grain, such as order line, order, shipment, SKU-location, or customer;
• time window and refresh frequency;
• data source and data owner;
• target, tolerance, and escalation threshold;
• segmentation requirements;
• known limitations and intended decisions.

Important KPI formulas

On-time delivery (OTD) measures the proportion of eligible deliveries received by the agreed date or date window.

The denominator must exclude formally cancelled orders and define how partial deliveries, grace periods, and customer-requested date changes are treated. OTD is not identical to complete and on-time delivery.

On-time-in-full (OTIF) requires both timing and quantity compliance.

If 920 of 1,000 eligible deliveries meet both criteria, OTIF is 92%. A firm should not compare this value with another firm unless definitions and eligibility rules are comparable.

Inventory turnover relates annualized cost of goods sold to average inventory measured at the same valuation basis.

A common average for a stable period is opening inventory plus closing inventory divided by two. Monthly or daily averages are preferable when inventory is seasonal. Turnover has units of cycles per year when annual cost and average inventory are used. High turnover is not automatically desirable if it causes lost sales or excessive transport cost.

Days of supply (DOS) expresses inventory in time terms.

If inventory is 4,800 units and average daily demand is 160 units, DOS is 30 days. The measure assumes that the demand basis is appropriate and nonzero. Forecast demand may be more relevant than historical demand when conditions are changing.

Forecast error for period t can be written as:

where A_t is actual demand and F_t is the forecast. A positive error indicates under-forecasting under this convention. Mean absolute error is:

where n is the number of periods. MAE is expressed in demand units and is easy to interpret, but it cannot compare items of very different scale. Mean absolute percentage error can be unstable or undefined when actual demand is zero; therefore, weighted or scaled measures may be preferable for intermittent demand.

Process cycle efficiency and lead-time measures must specify start and end events. For example, order-to-cash lead time may run from order acceptance to customer payment, whereas order fulfillment lead time may end at shipment or receipt. Ambiguous event definitions create apparently precise but operationally weak KPIs.

Leading and lagging indicators

Lagging indicators describe outcomes, such as total logistics cost or service level. Leading indicators signal future outcomes, such as overdue purchase orders, supplier capacity utilization, forecast bias, inventory record accuracy, or unresolved exception age. A balanced dashboard combines both. Managing only lagging indicators encourages late reaction; managing only leading indicators may reward activity without results.

KPI conflicts and gaming

SCM objectives are interdependent. Reducing inventory can increase stockouts; maximizing vehicle utilization can reduce responsiveness; minimizing purchase price can increase quality and disruption costs. KPIs can also be gamed when targets are narrow. For example, a warehouse may improve measured pick productivity by prioritizing easy orders while difficult orders accumulate. Governance should therefore monitor complementary measures and use balanced objectives rather than isolated targets.

Dashboards, Scorecards, Alerts, and Control Towers

Dashboard design

A dashboard supports repeated monitoring and interpretation. It should answer three questions quickly: What is the current state? What has changed? What decision is required? Effective design generally follows a hierarchy:

• executive view: strategic outcomes and material risks;
• functional view: operational drivers and exceptions;
• analyst view: transactions, dimensions, and diagnostic evidence.

A dashboard should avoid excessive visual density. Every visual should have a decision purpose, a defined refresh time, a visible period comparison, and a path to underlying detail. Color should communicate status consistently; red should not be used merely for visual emphasis. Accessibility requires adequate contrast and alternatives to color-only signals.

Scorecards

A scorecard compares performance with targets across dimensions such as service, cost, cash, resilience, sustainability, quality, and people. Supplier scorecards commonly combine delivery, quality, responsiveness, innovation, compliance, and commercial measures. Weighting should reflect strategy and risk. A single composite score may hide a critical failure, so threshold rules and dimension-level results should remain visible.

Alerts and exception management

An alert is valuable when it identifies a meaningful deviation and enables an action. An alert specification should state:

• event or threshold;
• severity;
• recipient and owner;
• expected response time;
• recommended or permitted action;
• suppression and escalation rules;
• closure evidence.

Too many alerts produce alert fatigue. Static thresholds are often inadequate because demand, lead time, and seasonality vary. A statistically informed alert may compare a measure with an expected range, but the range must account for forecast uncertainty and data quality. Alerts should be prioritized by risk, urgency, value at stake, and reversibility.

Control-tower views

A control tower is an integrated view of end-to-end flows, exceptions, risks, and response workflows. It is not simply a large dashboard. A credible control tower requires:

• multi-enterprise visibility across relevant partners;
• common identifiers and event definitions;
• near-real-time or appropriately timed data;
• exception prioritization;
• scenario and impact analysis;
• collaboration and case management;
• audit trails for decisions and interventions.

A control tower should distinguish observed facts, model estimates, and assumptions. For example, “shipment departed at 08:30” is an observed event; “arrival probability before Friday is 65%” is a model estimate; “carrier capacity is available” may be an unverified assumption.

Figure: Analytics decision support connects data and definitions to models, action, and organizational learning.

Scenario Analysis, Simulation, and Optimization

Scenario analysis

Scenario analysis asks how outcomes change when assumptions or decisions change. A scenario should specify a baseline, changed inputs, constraints, time horizon, and outcome measures. Examples include a 20% demand increase, a port closure, a supplier qualification delay, a revised service target, or a change in transport mode.

Scenarios should not be presented as forecasts unless probabilities are assigned and justified. They are structured “what-if” explorations. Sensitivity analysis varies one or more assumptions to identify drivers of outcome. A tornado chart, for example, can rank the effect of demand, lead time, capacity, and price assumptions on total cost.

Simulation

Simulation represents uncertainty and system behavior by repeatedly generating possible outcomes. Monte Carlo simulation is useful when demand, lead time, failures, or recovery times are uncertain and analytical formulas are insufficient. A basic procedure is:

1. define the system, decisions, inputs, and outputs;
2. select probability distributions or empirical samples using evidence;
3. generate random input scenarios;
4. run the operational logic for each scenario;
5. summarize cost, service, utilization, and risk distributions;
6. validate the model and test sensitivity.

Simulation produces a distribution of outcomes rather than one supposedly exact answer. The number of replications should be sufficient for stable estimates, and random seeds should support reproducibility. The model must be validated against known historical or operational behavior.

Optimization

Optimization identifies the best feasible decision under a specified objective and constraints. A generic linear optimization model is:

Here x_j is a decision variable, c_j is its objective coefficient, a_ij is the use of resource i by decision j, and b_i is the available amount or required level. Linear programming assumes linear relationships and continuous variables. Integer or mixed-integer programming is needed for indivisible choices such as opening a facility, assigning a truck, or selecting a supplier. Nonlinear, stochastic, or dynamic formulations may be more appropriate when relationships or uncertainty require them.

An inventory allocation model might maximize contribution margin while ensuring that allocated units do not exceed available supply and that customer or contractual minimums are respected. The solution is only as meaningful as the objective function, constraints, and data. An optimizer can return an elegant answer to the wrong problem.

Combining simulation and optimization

Optimization often uses expected values and may understate tail risk. Simulation can evaluate candidate policies under uncertainty. A practical workflow is to optimize a policy, simulate its performance, examine service and risk distributions, revise constraints or penalties, and repeat. Robust optimization and stochastic programming incorporate uncertainty more directly, but they require stronger assumptions and greater computational and modeling expertise.

Worked Application: Reorder Point and Service Risk

Consider a SKU with average daily demand of 100 units, demand standard deviation of 20 units per day, and a constant replenishment lead time of 9 days. Assume independent daily demand, normally distributed aggregate demand, no planned lead-time uncertainty, and a cycle-service target of 95%. For a one-sided normal target, the associated z-value is approximately 1.645.

Mean demand during lead time is:

Under the independence assumption, lead-time demand standard deviation is:

The reorder point with safety stock is:

Thus, the system would trigger replenishment at approximately 999 units of inventory position. This calculation does not include lead-time variability, order quantity, review-period demand, forecast bias, minimum order quantities, or correlated demand. If lead time is variable and independent of daily demand, a commonly used approximation is:

where σ_L is the standard deviation of lead time in days. The formula assumes independent demand across days and independence between demand and lead time. It should not be used uncritically when demand and lead time are correlated or distributions are strongly non-normal.

The example illustrates an important decision-quality principle: a target service level is a managerial choice, while the safety-stock calculation is a model translation of that choice under assumptions. The dashboard should display the assumptions and monitor whether they remain credible.

Decision Quality and Governance

Decision quality concerns both the quality of the decision process and the quality of the outcome. A good decision can produce a poor outcome because of random events; a poor decision can occasionally produce a good outcome. Evaluation should therefore examine the information available at decision time, the alternatives considered, assumptions, constraints, uncertainty, and consistency with the stated objective.

A practical decision-quality checklist asks:

• Is the decision problem and time horizon clearly framed?
• Who has authority and accountability?
• Are objectives explicit and prioritized?
• Are feasible alternatives represented?
• Which inputs are observed, estimated, or assumed?
• What uncertainty and downside risks exist?
• Are constraints real, current, and correctly coded?
• Can the recommendation be explained and challenged?
• What is the value of additional information?
• How will the decision and outcome be reviewed?

Data and model governance

Governance should assign ownership for data definitions, quality rules, access, lineage, retention, and correction. Important controls include validation of units and currencies, duplicate detection, reconciliation to source systems, time-stamped snapshots, and monitoring of missing or late data. Master-data errors can propagate through every downstream KPI and model.

Model governance includes version control, documentation, validation, approval, performance monitoring, change management, and retirement. A model should not silently change when its data pipeline or feature definitions change. Analysts should record training periods, exclusion rules, evaluation metrics, and known failure modes.

Ethics, privacy, and regulation

Supply chain analytics can affect workers, suppliers, customers, and communities. Location tracking, workforce productivity monitoring, customer profiling, and supplier risk scoring may involve personal or commercially sensitive information. Organizations should apply purpose limitation, data minimization, access control, retention limits, and applicable privacy law. Individuals should not be subjected to consequential automated decisions without appropriate oversight, explanation, and appeal mechanisms.

Fairness is context-dependent. A supplier model may systematically penalize small suppliers because they have less historical data, not because they are less reliable. An AI-generated recommendation may reproduce historical allocation bias or optimize cost at the expense of labor, safety, or environmental objectives. Human oversight must be substantive rather than a nominal approval step.

Sustainability metrics also require careful boundaries. Carbon estimates should state whether they cover transport only, full life-cycle emissions, market-based or location-based electricity, and which allocation method is used. A lower reported value is not necessarily a lower real-world impact.

Artificial Intelligence and Contemporary Developments

Machine learning can improve demand sensing, ETA prediction, anomaly detection, document extraction, and predictive maintenance. Generative AI can support natural-language querying, explanation, workflow drafting, and interaction with control-tower data. These uses are valuable when connected to governed data and bounded actions.

Risks include data leakage, hallucinated explanations, concept drift, adversarial manipulation, automation bias, and unclear accountability. A generative interface should not be treated as a system of record. Critical actions should be grounded in retrieved enterprise data, constrained by authorization rules, logged, and subject to human review. Performance should be evaluated using operational outcomes, not only language quality.

Digital twins can represent facilities, networks, inventory, and flows for scenario analysis. Event-driven architectures can ingest updates from partners and IoT devices. However, greater speed does not eliminate the need for semantic consistency, process ownership, and disciplined decision rights. Technology amplifies both good and poor management practices.

Illustrative Case: Northstar Consumer Products

Background

Northstar Consumer Products operates three distribution centers and sells seasonal household goods through retailers and e-commerce channels. It has an ERP system, a transportation management system, supplier portals, and separate spreadsheets used by regional planners. Service performance appears acceptable at the national level, but major retailers report frequent partial deliveries.

Problem

The executive dashboard reports 96% OTD, while the commercial team reports customer dissatisfaction. Investigation finds that OTD excludes partial deliveries and uses the requested date after it has been changed by the customer-service team. Inventory turnover has improved, but stockouts have increased for high-margin SKUs. Alerts are sent for every late purchase order, causing planners to ignore them.

Analysis

A cross-functional team first created a KPI dictionary and reconciled order, shipment, and receipt events. It separated customer-requested date changes from supplier-caused delays and introduced OTIF at order-line level. Diagnostic analysis segmented performance by retailer, SKU velocity, distribution center, and carrier. The team found that a small group of seasonal SKUs caused a disproportionate share of lost margin. Forecast error was biased downward before promotional events, while safety stock settings were based on stable historical demand.

Approach and solution

Northstar implemented a control-tower workflow with four layers:

1. an executive view of OTIF, lost-margin exposure, inventory value, and high-severity risks;
2. a planner view of SKU-location exceptions, root-cause categories, and recommended actions;
3. a supplier and carrier scorecard using common event definitions;
4. a case-management process with owners, due dates, and closure evidence.

The team used scenario analysis for three alternatives: expedite selected inbound shipments, reallocate inventory between distribution centers, or accept lower service on low-margin items. A mixed-integer allocation model respected transportation capacity, retailer commitments, and minimum presentation quantities. Candidate policies were then evaluated through simulation using historical demand variability and lead-time distributions.

Alerts were redesigned around impact and urgency. A late order generated an alert only when it threatened a committed customer order, exceeded a material value threshold, or lacked a recovery plan. Dashboard users could inspect the data lineage and assumptions behind each recommendation.

Outcome

In the first quarter after implementation, Northstar reported better agreement between executive and operational measures, fewer unresolved high-severity exceptions, and more disciplined allocation during promotions. The company did not treat the initial results as proof that the model was permanently correct. It established monthly model reviews, quarterly KPI governance, and post-event analysis after each major promotion.

Learning points

The case demonstrates that analytics value depends on event definitions, segmentation, workflow, and decision rights—not only on visualization technology. It also shows why improving one metric, such as turnover or OTD, without considering OTIF, margin, and customer commitments can produce misleading conclusions.

Best Practices for Implementation

A graduate-level implementation plan should proceed incrementally:

1. identify decisions with measurable economic or service value;
2. define process events, entities, measures, owners, and grain;
3. assess data quality and reconcile key totals to source systems;
4. establish a KPI dictionary and governance forum;
5. design role-specific dashboards before adding advanced models;
6. pilot alerts with explicit owners and response workflows;
7. validate predictive models using temporal holdout data;
8. test recommendations through scenarios, simulation, or controlled trials;
9. document assumptions, limitations, and escalation rules;
10. monitor adoption, outcomes, drift, unintended effects, and equity;
11. retire measures and alerts that no longer support decisions.

The appropriate level of sophistication depends on decision value, uncertainty, data maturity, and organizational capability. A transparent rule with reliable data may outperform an opaque algorithm that users do not trust or cannot operationalize.

Challenges and Limitations

Common limitations include inconsistent master data, delayed partner events, missing timestamps, biased historical records, incompatible KPI definitions, model drift, and organizational resistance. Forecasts may fail during structural change because the future is unlike the training data. Optimization may be brittle when constraints are incomplete. Simulation may create false confidence when probability distributions are invented or poorly fitted. Dashboards may encourage local optimization if they do not reveal end-to-end effects.

There are also trade-offs between speed and control. Near-real-time information can improve responsiveness but may expose provisional events that later change. More granular measures can improve diagnosis but increase complexity and privacy risk. Automation can reduce workload but may weaken human expertise if users stop questioning recommendations. These limitations should be made visible rather than hidden behind polished interfaces.

Chapter Summary

Supply chain analytics converts governed data into insight and action. Descriptive analytics measures what happened; diagnostic analytics investigates why; predictive analytics estimates what is likely; and prescriptive analytics recommends what should be done. Effective KPIs require precise definitions, appropriate denominators, owners, targets, and action rules. Dashboards support monitoring, scorecards support structured comparison, alerts support exception response, and control towers integrate visibility with workflow and scenario analysis.

Scenario analysis explores alternative assumptions, simulation represents uncertainty, and optimization selects feasible actions under explicit objectives and constraints. Their value depends on valid data, credible assumptions, and careful interpretation. Decision quality requires separating facts from estimates, considering alternatives and uncertainty, and reviewing decisions rather than judging them solely by outcomes. Governance, privacy, fairness, explainability, and human accountability are essential, particularly when AI influences consequential supply chain decisions. The central managerial lesson is that analytics is an operating capability: technology creates value only when it improves accountable decisions and organizational learning.

Key Terms

KPI: A defined measure linked to an objective, decision, owner, and target.

Descriptive analytics: Analysis that summarizes historical or current conditions.

Diagnostic analytics: Analysis that investigates causes and drivers of observed outcomes.

Predictive analytics: Analysis that estimates future outcomes or probabilities.

Prescriptive analytics: Analysis that recommends actions under objectives and constraints.

OTD: On-time delivery, the percentage of eligible deliveries meeting the agreed timing criterion.

OTIF: On-time-in-full performance, requiring both timing and quantity compliance.

Inventory turnover: Annual cost of goods sold divided by average inventory value.

Days of supply: Inventory quantity divided by average daily demand.

Forecast error: The difference between actual and forecast demand under a stated sign convention.

Dashboard: A visual interface for monitoring measures and supporting interpretation.

Scorecard: A structured comparison of performance across targets or dimensions.

Alert: A prioritized notification of a condition requiring investigation or action.

Control tower: An integrated visibility, analytics, exception, and response capability.

Scenario analysis: Systematic examination of outcomes under alternative assumptions or decisions.

Simulation: Repeated representation of uncertain system behavior to estimate outcome distributions.

Optimization: Selection of the best feasible decision under a specified objective and constraints.

Decision quality: The quality of framing, alternatives, evidence, assumptions, and reasoning available when a decision is made.

Model governance: Controls for model documentation, validation, monitoring, approval, and change.

Data lineage: Traceability from a reported measure or model input to its source and transformations.

Review Questions

Short Answer Questions

1. Distinguish descriptive analytics from diagnostic analytics in a supply chain context.
2. Why must an OTIF KPI define its denominator and treatment of partial deliveries?
3. What makes an alert actionable rather than merely informative?
4. State two assumptions in the worked reorder-point example.
5. Explain why a control tower is more than a dashboard.

Long Answer Questions

1. Explain how descriptive, diagnostic, predictive, and prescriptive analytics can be integrated into one supply chain decision process.
2. Design a supplier scorecard for a strategic supplier, including measures, weights, thresholds, and governance safeguards.
3. Compare scenario analysis, simulation, and optimization, including the situations in which each is most appropriate.
4. Discuss the relationship between KPI design, local optimization, and end-to-end supply chain performance.
5. Evaluate the governance requirements for using artificial intelligence in supply chain control-tower decisions.

Analytical / Application Questions

1. A firm reports 98% OTD but 89% OTIF. Diagnose at least three possible causes and propose the data fields needed to test them.
2. Using the worked application assumptions, explain how the reorder point would change if lead-time variability were material. State why the original formula would no longer be sufficient by itself.
3. Design a control-tower response for a supplier delay that threatens two customers with different margins and service commitments. Specify the data, alternatives, objective, constraints, and escalation logic.

Multiple Choice Questions with Answers

Question 1

Which type of analytics recommends an allocation of scarce inventory subject to customer and capacity constraints?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: D. Prescriptive analytics

Question 2

A firm has 920 eligible deliveries, of which 874 meet both the agreed date and quantity criteria. What is its OTIF?

A. 92.0%

B. 94.0%

C. 95.7%

D. 98.0%

Correct Answer: B. 94.0%

Question 3

Which feature most directly distinguishes an actionable alert from a simple notification?

A. It uses bright colors

B. It has an owner, response time, and permitted action

C. It is sent to every employee

D. It contains the largest possible amount of data

Correct Answer: B. It has an owner, response time, and permitted action

Question 4

If annual cost of goods sold is $12 million and average inventory value is $3 million, inventory turnover is:

A. 0.25 cycles per year

B. 3 cycles per year

C. 4 cycles per year

D. 9 cycles per year

Correct Answer: C. 4 cycles per year

Question 5

In the worked reorder-point example, mean demand during a nine-day lead time is:

A. 180 units

B. 900 units

C. 999 units

D. 1,645 units

Correct Answer: B. 900 units

Question 6

Which statement about simulation is correct?

A. It always identifies one mathematically optimal decision

B. It eliminates the need to validate input distributions

C. It estimates a distribution of outcomes by representing uncertainty repeatedly

D. It is applicable only when all inputs are deterministic

Correct Answer: C. It estimates a distribution of outcomes by representing uncertainty repeatedly

Question 7

In a linear optimization model, x_j most commonly represents:

A. A decision variable

B. A data-quality exception

C. A forecast error percentage

D. A dashboard color code

Correct Answer: A. A decision variable

Question 8

Why can mean absolute percentage error be problematic for intermittent demand?

A. It has units of currency only

B. It can be unstable or undefined when actual demand is zero

C. It cannot measure forecast error

D. It always rewards under-forecasting

Correct Answer: B. It can be unstable or undefined when actual demand is zero

Question 9

Which practice best supports model governance?

A. Changing model inputs without recording the change

B. Removing assumptions from user-facing documentation

C. Maintaining version control, validation records, and performance monitoring

D. Allowing any user to alter production rules

Correct Answer: C. Maintaining version control, validation records, and performance monitoring

Question 10

A control tower should distinguish observed facts from model estimates because this distinction:

A. Prevents all supply chain disruptions

B. Clarifies certainty, assumptions, and accountability

C. Makes data governance unnecessary

D. Ensures every prediction is correct

Correct Answer: B. Clarifies certainty, assumptions, and accountability