12. Operations Analytics, AI and Decision Intelligence

Introduction and Context

Operations and supply chains generate decisions continuously: how much to order, where to position inventory, which supplier to select, when to schedule production, how to allocate scarce capacity, and how to respond to disruption. Traditionally, these decisions depended on managerial experience, periodic reports, spreadsheets, and relatively stable assumptions. Contemporary operations, however, are characterized by volatile demand, geographically dispersed networks, shorter product life cycles, digital transactions, connected assets, and increasing expectations of speed and personalization. Under these conditions, data and computational models have become central to operational competitiveness.

Operations analytics refers to the systematic use of data, quantitative models, algorithms, and managerial judgment to understand and improve operational performance. Artificial intelligence extends this capability through machine learning, natural-language technologies, computer vision, intelligent agents, and generative models. Decision intelligence goes further by connecting data, models, organizational objectives, human judgment, and action. It is concerned not merely with predicting what may happen but with designing and governing decisions that produce desirable outcomes.

The value of analytics is therefore not equivalent to the volume of data or sophistication of an algorithm. A highly accurate forecast may be operationally useless if it is not connected to replenishment rules. An optimization model may produce an elegant solution that cannot be implemented because it ignores labor agreements or supplier reliability. Similarly, an AI system may automate a biased or poorly defined decision. The central managerial question is how analytical capability can be embedded in an operating system while preserving transparency, resilience, accountability, and human oversight.

Learning Objectives

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

1. Distinguish descriptive, diagnostic, predictive, and prescriptive analytics in operations.

2. Explain the role of optimization, simulation, machine learning, and generative AI in operational decision-making.

3. Evaluate data quality, model performance, explainability, bias, cybersecurity, and implementation risk.

4. Design a decision-intelligence process linking forecasts and models to operational action.

5. Assess the managerial and ethical implications of automating supply-chain decisions.

6. Apply analytical reasoning to a complex operations case involving uncertainty and competing objectives.

From Reporting to Decision Intelligence

A useful progression begins with four analytical questions. Descriptive analytics asks, “What happened?” It summarizes orders, service levels, lead times, production losses, transportation costs, and inventory positions. Diagnostic analytics asks, “Why did it happen?” It investigates relationships and causal mechanisms, such as whether stockouts resulted from forecast error, supplier lateness, inaccurate inventory records, or an allocation policy.

Predictive analytics asks, “What is likely to happen?” Forecasting models estimate future demand, lead times, equipment failures, or delivery delays. Prescriptive analytics asks, “What should be done?” It evaluates alternative actions subject to constraints and objectives. For example, a prescriptive model may recommend production quantities and transportation modes that minimize expected cost while maintaining a target service level.

These categories are analytically distinct but operationally connected. A dashboard that reports late deliveries is descriptive. A model identifying weather and carrier congestion as likely causes is diagnostic. A model estimating the probability that a shipment will be late is predictive. A routing and expediting recommendation that responds to those probabilities is prescriptive. Decision intelligence integrates all four levels into a repeatable decision process.

The decision-intelligence perspective emphasizes six elements: a clearly defined decision, relevant data, a model of the operating environment, explicit objectives, feasible actions, and feedback from outcomes. This perspective prevents organizations from treating analytics as an isolated technical project. It also recognizes that decisions are situated in institutions. A purchasing recommendation is shaped by contractual authority, risk appetite, sustainability commitments, and supplier relationships as well as by price and predicted demand.

Core Analytical Methods in Operations

Forecasting and predictive analytics

Forecasting estimates future values using historical observations, explanatory variables, or both. Time-series methods identify patterns such as trend, seasonality, cycles, and autocorrelation. Causal or regression-based models incorporate drivers such as price, promotions, macroeconomic indicators, weather, and product availability. Machine-learning methods, including tree-based algorithms and neural networks, can represent nonlinear relationships and interactions.

Forecast accuracy should be evaluated in relation to the decision it supports. Common measures include mean absolute error, root mean squared error, mean absolute percentage error, and weighted variants that give greater importance to high-volume or strategically important items. No single metric is universally appropriate. Percentage-based measures can be unstable when actual demand is near zero, while squared-error measures penalize large errors disproportionately. Forecast bias—the tendency to overpredict or underpredict systematically—is often more damaging than random error because it creates persistent inventory or capacity consequences.

Forecasting is also a process of uncertainty management. A point forecast should normally be accompanied by prediction intervals or scenarios. Demand may be forecast at multiple hierarchical levels, such as product, region, customer, and total network. Reconciliation methods can ensure that detailed forecasts aggregate consistently with higher-level plans. Human judgment can add value when structural breaks, new products, or unusual events make historical patterns unreliable, but judgment should be documented and evaluated rather than treated as automatically superior to statistical methods.

Optimization

Optimization selects the best feasible decision according to a stated objective. Linear programming is appropriate when relationships are linear and variables may be continuous. Mixed-integer programming incorporates discrete decisions such as opening a facility, selecting a supplier, or assigning a truck. Network optimization addresses flows through facilities and transportation links. Stochastic optimization represents uncertain parameters through scenarios or probability distributions, while robust optimization seeks solutions that perform acceptably across adverse conditions.

A general optimization model contains decision variables, an objective function, constraints, and parameters. Consider a production-planning problem in which the organization chooses production quantities, overtime, inventory, and outsourced volume. The objective may minimize production, holding, overtime, outsourcing, and shortage costs. Constraints may include labor hours, machine capacity, minimum order quantities, storage limits, and customer commitments.

The mathematical optimum is meaningful only if the model represents the organization’s real choices and constraints. In practice, analysts should conduct sensitivity analysis, examine alternative objective weights, test feasibility under disruptions, and communicate trade-offs rather than presenting a single answer as unquestionable. Optimization can expose the cost of a managerial preference. For instance, raising the required service level may reveal the additional inventory and capacity necessary to achieve it.

Simulation

Simulation represents the behavior of a system over time and under uncertainty. Discrete-event simulation is useful for queues, warehouses, hospitals, production lines, and transportation terminals. Monte Carlo simulation repeatedly samples uncertain inputs to estimate distributions of outcomes. Agent-based simulation models interactions among decentralized actors such as customers, vehicles, suppliers, or autonomous equipment.

Simulation is particularly valuable when the system is too complex for an exact analytical solution or when managers want to compare policies without experimenting on the real operation. A warehouse simulation can test alternative picking layouts, staffing levels, batching rules, and automation investments. Its limitations include dependence on assumptions, computational cost, and the risk of producing precise-looking results from an invalid model. Verification asks whether the model has been implemented correctly; validation asks whether it adequately represents the real system.

Machine learning

Machine learning identifies patterns in data and uses them to classify, predict, recommend, or detect anomalies. Supervised learning uses labeled outcomes, such as whether an order was late or whether a machine failed. Unsupervised learning identifies structure without a specified target, as in customer segmentation or unusual transaction detection. Reinforcement learning learns policies through interaction and reward, although its operational use requires careful control because exploratory actions can impose real costs.

Machine learning is valuable where relationships are nonlinear, data are high-dimensional, or manual rules are difficult to maintain. Applications include demand sensing, predictive maintenance, quality inspection through computer vision, fraud detection, ETA prediction, and supplier-risk classification. However, predictive performance in a test dataset does not guarantee operational value. Models may fail under distribution shift, when market conditions, suppliers, products, or customer behavior change. Monitoring must therefore continue after deployment.

Data as an Operational Asset

Analytical capability depends on the quality and governance of data. Important dimensions include accuracy, completeness, consistency, timeliness, uniqueness, validity, and lineage. An inventory record can be accurate at the time of entry but obsolete by the time a replenishment decision is made. A supplier master may contain duplicate entities, inconsistent units of measure, or missing country and certification information. These are not merely information-technology inconveniences; they directly affect working capital, service, compliance, and risk.

Data governance assigns ownership and establishes rules for definitions, access, quality monitoring, retention, and correction. A “customer,” “available inventory,” or “on-time delivery” must have an agreed operational definition. Metadata and lineage allow analysts to determine where a measure originated and how it was transformed. Master-data management is particularly important in multi-site organizations where different facilities use different product, supplier, or location codes.

Data quality should be managed according to decision criticality. A small error in a low-value report may be tolerable, whereas a systematic error in safety-stock parameters can create substantial shortages. Organizations should identify critical data elements, define quality thresholds, establish exception workflows, and measure the cost of poor data rather than treating quality as an abstract technical target.

Explainability, Bias, and Human Oversight

Explainability concerns the ability to understand and communicate how a model reaches a result. Some models are intrinsically interpretable, such as a small decision tree or a linear model. More complex models may require post hoc methods that identify influential variables, local explanations, counterfactual examples, or sensitivity patterns. Explainability is not the same as proving that a model is correct. An explanation can make a flawed model easier to understand without making it valid.

The appropriate level of explanation depends on the decision. A planner may require a concise reason for a forecast change, while a regulator or affected supplier may require documentation of data sources, model purpose, validation procedures, and appeal mechanisms. Explanations should be intelligible to the relevant audience and should not disclose sensitive information unnecessarily.

Bias can enter through historical data, sampling, labels, proxy variables, objectives, or deployment practices. A supplier-risk model trained on historical purchasing decisions may reproduce a preference for incumbent firms, even if the historical preference was unrelated to actual performance. A workforce scheduling system may disadvantage particular groups if availability data reflect unequal access to transportation or care responsibilities. Bias assessment should examine outcomes across relevant groups, error rates, access to opportunities, and the consequences of false positives and false negatives.

Human oversight is not achieved simply by placing a person somewhere in the workflow. Effective oversight requires authority, competence, time, information, and a genuine ability to challenge or override the system. Automation bias may cause employees to accept algorithmic recommendations without scrutiny, while alert fatigue may cause them to ignore legitimate warnings. Organizations should define escalation thresholds, override procedures, audit trails, and accountability for both automated and human decisions.

Generative AI in Operations

Generative AI produces text, code, images, summaries, scenarios, and other content from prompts and contextual data. In operations, it can assist with natural-language querying of planning data, summarizing supplier communications, drafting work instructions, explaining forecast changes, generating simulation scenarios, supporting maintenance technicians, and helping analysts write or review code. It can also serve as an interface to enterprise systems, allowing users to ask operational questions in ordinary language.

Its use requires disciplined boundaries. Generative models may produce fabricated information, misinterpret ambiguous instructions, expose confidential data, or generate code with security vulnerabilities. They are not inherently reliable sources of numerical truth. A generative assistant should retrieve authoritative data from governed systems, show sources where possible, distinguish facts from suggestions, and require approval before changing orders, schedules, or master data. High-impact actions should remain subject to deterministic rules, validation checks, and human authorization.

A useful operating principle is “copilot before autopilot.” Early applications should augment planners and analysts by reducing search, documentation, and routine communication effort. As evidence accumulates, selected low-risk tasks may be automated. The decision should be based on reversibility, materiality, uncertainty, and the availability of effective controls.

A Decision-Intelligence Architecture

A mature architecture usually contains several interconnected layers. The data layer integrates enterprise-resource-planning systems, warehouse systems, transportation platforms, Internet-of-Things devices, supplier portals, market data, and external risk information. The analytical layer contains forecasts, optimization models, simulation environments, machine-learning services, and business rules. The decision layer translates analytical outputs into recommendations, approvals, alerts, and workflows. The execution layer connects decisions to purchasing, production, inventory, routing, maintenance, and customer-service systems. A governance layer surrounds all others with security, access control, model validation, monitoring, documentation, and accountability.

The architecture should support a closed learning loop. Actual outcomes—demand, delivery, failure, cost, quality, and customer response—are compared with predictions and decisions. Errors are investigated, parameters are updated, and policies are revised. This is different from a one-time analytics project. It is an organizational capability that requires process ownership, technical infrastructure, change management, and performance review.

A practical implementation sequence is to begin with a decision inventory. Managers identify recurring decisions, their frequency, economic importance, uncertainty, current process, data requirements, and decision rights. Candidate use cases can then be prioritized according to value, feasibility, risk, and learning potential. A pilot should include a clear baseline, defined success measures, an implementation owner, and a plan for adoption. Scaling requires integration with daily work rather than merely publishing a model or dashboard.

Applications Across the Supply Chain

In procurement, analytics can compare total cost, quality, lead-time reliability, capacity, financial risk, and sustainability indicators. Predictive models can identify suppliers likely to miss delivery commitments, while optimization can allocate demand across qualified suppliers. The organization must avoid allowing a cost-minimization model to select a supplier that violates ethical, regulatory, or resilience requirements.

In inventory management, analytics supports segmentation, safety-stock calculation, multi-echelon positioning, service-level differentiation, and exception management. The relevant decision is not simply how much inventory to hold but where to hold it, for which customer promise, and under which uncertainty assumptions. Machine learning may improve demand sensing, but replenishment policies still require explicit treatment of lead-time variability, order constraints, obsolescence, and substitution.

In production, optimization can support sequencing, lot sizing, workforce allocation, and capacity planning. Predictive maintenance estimates failure risk from sensor and maintenance data. Computer vision can identify defects, although performance must be evaluated across products, lighting conditions, and operating environments. Digital twins combine real-time data with models to test operational changes and anticipate bottlenecks.

In logistics, ETA prediction, route optimization, load consolidation, warehouse slotting, and dynamic dispatch can improve cost and service. These applications often involve multiple objectives: fuel consumption, emissions, delivery promises, driver regulations, and customer priorities. A technically optimal route may be unacceptable if it is unsafe, legally infeasible, or operationally fragile.

Cybersecurity and Resilience

The expansion of connected analytics increases the attack surface. Threats include theft of operational data, manipulation of sensor readings, ransomware, unauthorized model access, prompt injection against generative systems, and disruption of automated controls. An attacker who changes inventory or supplier data may cause harm without directly attacking physical equipment.

Security controls should include identity and access management, least-privilege permissions, encryption, network segmentation, secure application programming interfaces, logging, vulnerability management, backup procedures, and incident-response plans. Models and data pipelines require protection against tampering. Organizations should also test whether decisions remain safe when data are missing, delayed, corrupted, or deliberately manipulated.

Resilience requires graceful degradation. If a forecasting service fails, planners should have a documented fallback method. If a warehouse-control system becomes unavailable, safe manual procedures should exist. Business continuity is therefore part of analytical design, not an afterthought.

Illustrative Case: Predictive Allocation in a Medical-Device Network

Background

The following fictional case illustrates the integration of forecasting, optimization, machine learning, governance, and human judgment. MedAxis produces sterile medical devices through three plants and distributes them to hospitals through regional warehouses. Some products are essential for scheduled procedures, while others have substitutes. Stockouts create financial costs and may create serious service consequences.

Situation and problem

The company experienced rising emergency shipments and uneven service levels. A data-science team proposed a machine-learning demand forecast and an optimization model for weekly allocation. Historical data, however, contained frequent stockouts. Observed sales therefore understated unconstrained demand. Supplier lead times also varied substantially, but the planning system used average values. In addition, one regional warehouse had incomplete inventory records.

Analysis

The team first mapped the decision process and established definitions for demand, available inventory, service level, and emergency shipment. It separated demand loss caused by stockouts from ordinary demand variation and created lead-time distributions rather than single averages. Forecast performance was evaluated by product class and hospital segment, using both accuracy and bias measures.

The optimization model minimized expected operating cost and emergency freight while imposing constraints for plant capacity, shelf life, regulatory qualification, minimum hospital service levels, and allocation of scarce devices. Scenario analysis represented supplier delay, demand surges, and warehouse-record uncertainty. The team discovered that an apparently high-performing predictive model produced poor allocations when its outputs were not adjusted for stockout censoring.

Approach and solution

MedAxis adopted a hybrid process. The machine-learning model generated baseline forecasts, while planners reviewed only exceptions involving new products, unusual clinical events, or substantial forecast changes. The optimizer proposed allocations, but products designated as clinically critical were subject to protected service constraints. A data-quality dashboard flagged inventory records that exceeded defined age or reconciliation thresholds.

A governance committee approved model changes, reviewed subgroup service outcomes, and maintained an override log. The system displayed the principal drivers of each recommendation, the uncertainty range, and the cost of alternative service levels. No automated recommendation could release a purchase order without authorization from the responsible planner.

Outcome

In the pilot, the organization evaluated success against a pre-established baseline rather than relying solely on model accuracy. It observed fewer emergency shipments, more consistent service across hospital segments, and improved visibility of data problems. The most important benefit was not the algorithm alone but the redesign of the decision process: data definitions, exception management, scenario testing, and accountability were made explicit.

Learning points

First, sales data are not always equivalent to demand. Second, optimization is only as credible as its constraints and objective function. Third, data-quality exceptions should be visible within the decision workflow. Fourth, human oversight is most effective when focused on defined exceptions rather than informal approval of every routine calculation. Finally, operational performance and fairness across customer groups should be evaluated together.

Limitations, Failure Modes, and Professional Judgment

Common failures include automating an undefined decision, optimizing a local metric at the expense of system performance, ignoring feedback loops, confusing correlation with causation, and deploying a model without monitoring. A forecast may alter purchasing, which alters availability, which then changes observed demand; retraining on that data can reinforce the original decision. Performance can also deteriorate when products, suppliers, regulations, or customer behavior change.

Another risk is metric gaming. If a warehouse is rewarded only for picking speed, it may increase errors or shift work to downstream inspection. If procurement is rewarded solely for purchase price variance, it may increase total cost through poor quality, late delivery, or excess inventory. Decision systems should therefore use balanced measures covering cost, service, quality, resilience, sustainability, safety, and employee impact.

Managers should distinguish model uncertainty from operational uncertainty. Model uncertainty concerns whether the analytical representation is correct; operational uncertainty concerns variation in the world itself. Both should be communicated. Scenario analysis, stress testing, confidence intervals, and sensitivity analysis are tools for making uncertainty actionable rather than concealing it behind a single number.

Governance, Regulation, and Ethics

Responsible analytical governance requires a documented purpose, defined decision owner, data-protection controls, validation evidence, records of model versions, monitoring thresholds, and procedures for complaints or correction. Privacy principles such as purpose limitation, data minimization, access control, retention limits, and lawful processing are relevant whenever employee, customer, driver, or supplier data are used.

Ethical evaluation should ask who benefits, who bears risk, who may be excluded, and whether affected parties can understand or challenge a decision. In employment contexts, algorithmic scheduling and performance monitoring can affect autonomy and dignity. In supplier management, automated risk scores can restrict commercial opportunity. In logistics, route and workforce systems can influence safety and working conditions.

Boards and senior executives should treat AI and analytics as enterprise-risk issues rather than delegating all responsibility to technical teams. Accountability cannot be transferred to an algorithm. A model may recommend, but the organization remains responsible for the consequences of its decision system.

Emerging Directions and Best Practices

Emerging developments include probabilistic forecasting, causal machine learning, graph analytics for supply networks, foundation models connected to enterprise data, autonomous planning agents, synthetic data for testing, and digital twins that combine simulation with real-time operational data. These developments may improve adaptability, but they also increase the importance of model governance, secure interfaces, and clear limits on autonomous action.

Best practice is to:

• start with a consequential decision rather than a fashionable technology;
• define objectives, constraints, decision rights, and success measures before modeling;
• establish data ownership and quality controls;
• compare advanced models with transparent baselines;
• evaluate accuracy, bias, robustness, interpretability, and operational impact;
• use pilots, shadow mode, and staged deployment for high-risk applications;
• preserve audit trails, override mechanisms, and fallback procedures;
• monitor drift, exceptions, unintended effects, and cybersecurity events;
• train users to challenge recommendations constructively; and
• review the entire decision process, not merely the predictive model.

Chapter Summary

Operations analytics progresses from describing performance to diagnosing causes, predicting future conditions, and prescribing feasible actions. Forecasting, optimization, simulation, and machine learning address different analytical needs and are most powerful when integrated into a decision-intelligence process. Data quality, definitions, uncertainty, and feedback loops are as important as algorithmic sophistication. Generative AI can improve access to operational knowledge and support planners, but it requires governed data, validation, security, and human authorization. Explainability, bias management, cybersecurity, resilience, privacy, and accountability are essential design requirements. Effective organizations treat analytics as a sociotechnical capability: models, technology, processes, expertise, and governance must work together to produce responsible operational decisions.

Key Terms

Descriptive analytics: Analysis that summarizes what has happened or is currently happening.

Diagnostic analytics: Analysis that investigates the causes or contributing factors of an outcome.

Predictive analytics: Use of data and models to estimate future events or conditions.

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

Decision intelligence: An integrated approach linking data, models, decisions, actions, outcomes, and governance.

Forecast bias: Systematic tendency of forecasts to overestimate or underestimate actual outcomes.

Optimization: Selection of the best feasible solution according to a defined objective.

Mixed-integer programming: Optimization involving continuous and discrete decision variables.

Robust optimization: Optimization seeking solutions that perform acceptably under uncertainty.

Simulation: Computational representation of system behavior used to examine possible outcomes.

Machine learning: Methods that learn patterns from data for prediction, classification, or decision support.

Model drift: Decline in model performance as data or operating conditions change.

Explainability: The extent to which a model’s reasoning or influential factors can be understood.

Algorithmic bias: Systematic and potentially unfair differences produced by data, models, or deployment.

Data lineage: Record of a data element’s origin, transformations, and movement through systems.

Digital twin: A dynamic digital representation of a physical asset, process, or network.

Human oversight: Structured human authority to review, challenge, approve, or override automated outputs.

Graceful degradation: Ability of a system to remain safe and functional at reduced capability after failure.

Closed learning loop: Process in which operational outcomes are used to evaluate and improve decisions and models.

Generative AI: AI that produces new content such as text, code, scenarios, or images from learned patterns.

Review Questions

Short Answer Questions

1. Distinguish predictive analytics from prescriptive analytics using an inventory example.

2. Why can stockout data distort a demand forecast?

3. What are verification and validation in simulation?

4. Identify three dimensions of data quality relevant to operations.

5. Why is human oversight more than simply placing a person in an automated workflow?

Long Answer Questions

1. Explain how a decision-intelligence architecture connects data, analytical models, workflows, execution systems, and governance.

2. Compare optimization, simulation, and machine learning in terms of purpose, assumptions, strengths, and limitations.

3. Discuss the role of explainability and bias assessment in operational AI.

4. Evaluate the opportunities and risks of generative AI as a planning copilot.

5. Develop a governance framework for an AI-enabled supply-chain planning system.

Analytical / Application Questions

1. A retailer has accurate sales records but frequent stockouts and changing promotional patterns. Design a forecasting and evaluation approach that addresses censored demand, uncertainty, and forecast bias.

2. A transportation optimizer reduces cost but increases driver overtime and late deliveries for remote customers. Explain how you would revise the objective, constraints, metrics, and governance process.

3. A supplier-risk model disproportionately classifies small regional suppliers as high risk. Propose an investigation and remediation plan covering data, model design, business policy, and affected suppliers.

22. Multiple Choice Questions with Answers

Question 1

Which analytical question is most directly addressed by prescriptive analytics?

A. What was the average delivery time last month?

B. Why did the warehouse experience a labor shortage?

C. What delivery plan best meets service and cost constraints?

D. What is the probability that a shipment will be late?

Correct Answer: C. What delivery plan best meets service and cost constraints?

Question 2

Why may observed sales understate true demand in a stockout period?

A. Sales data always contain excessive seasonality

B. Customers who cannot obtain the product are not recorded as sales

C. Optimization automatically removes lost sales

D. Lead-time distributions eliminate demand variation

Correct Answer: B. Customers who cannot obtain the product are not recorded as sales

Question 3

Which method is most appropriate for testing alternative warehouse layouts when queueing and process interactions are important?

A. Discrete-event simulation

B. Simple descriptive reporting

C. Supplier segmentation only

D. A point forecast without scenarios

Correct Answer: A. Discrete-event simulation

Question 4

What is a principal purpose of robust optimization?

A. To guarantee that all data are complete

B. To select solutions that perform acceptably across adverse conditions

C. To replace all human decisions with machine learning

D. To maximize model complexity

Correct Answer: B. To select solutions that perform acceptably across adverse conditions

Question 5

Which situation best illustrates model drift?

A. A model is documented before deployment

B. A forecast is accompanied by a prediction interval

C. A delivery-time model becomes less accurate after a major network redesign

D. A planner overrides a recommendation using approved authority

Correct Answer: C. A delivery-time model becomes less accurate after a major network redesign

Question 6

Which control is most appropriate when using generative AI to support purchase-order decisions?

A. Allow the model to release all orders without review

B. Use only unverified text generated by the model

C. Require authoritative data retrieval, validation, and approval before execution

D. Remove audit logs to protect confidentiality

Correct Answer: C. Require authoritative data retrieval, validation, and approval before execution

Question 7

A supplier-risk model favors incumbent suppliers because historical purchasing decisions favored them. This is primarily an example of bias arising from:

A. Historical data and labels

B. Encryption standards

C. Simulation verification

D. Prediction intervals

Correct Answer: A. Historical data and labels

Question 8

What distinguishes validation from verification in simulation?

A. Validation checks whether the model represents the real system adequately; verification checks whether it was implemented correctly

B. Validation measures cybersecurity; verification measures privacy

C. Validation selects the objective; verification removes all uncertainty

D. Validation is performed only after deployment; verification is unnecessary

Correct Answer: A. Validation checks whether the model represents the real system adequately; verification checks whether it was implemented correctly

Question 9

Which feature most strongly supports graceful degradation in an analytics-enabled operation?

A. A single centralized model with no fallback

B. Documented manual procedures when an automated service fails

C. Removing access controls from operational systems

D. Using only historical averages in every situation

Correct Answer: B. Documented manual procedures when an automated service fails

Question 10

Why should an operations analytics project begin with a decision inventory?

A. To prioritize consequential decisions, required data, feasibility, and accountability

B. To ensure that the most complex algorithm is selected first

C. To avoid defining success measures until after deployment

D. To eliminate the need for operational managers

Correct Answer: A. To prioritize consequential decisions, required data, feasibility, and accountability