AI and Data-Driven Decision Support

Introduction

Reverse logistics is characterised by uncertain timing, heterogeneous product condition, incomplete information and competing economic, environmental and service objectives. These features make conventional, rule-based decision making difficult. Artificial intelligence (AI), machine learning (ML), computer vision, optimisation and real-time data platforms can improve decisions concerning return volumes, inspection, grading, disposition, routing, fraud prevention and asset maintenance.

AI does not replace operational judgement. It provides predictions, classifications, recommendations or automated actions that must be embedded in a controlled decision process. In reverse logistics, the quality of a decision depends not only on predictive accuracy but also on data lineage, cost-sensitive evaluation, explainability, fairness, cybersecurity and the organisation’s ability to execute the recommendation.

A useful distinction is between prediction and decision support. Prediction estimates an uncertain outcome, such as whether a returned appliance is repairable. Decision support combines that estimate with costs, capacity, environmental impacts, policies and constraints to select an action. A highly accurate prediction may still produce poor decisions if it is connected to an inappropriate objective function.

Learning Objectives

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

1. Explain the role of AI and data analytics in reverse logistics.
2. Select suitable ML approaches for forecasting, inspection, grading, fraud detection and maintenance.
3. Formulate data-driven disposition and routing decisions.
4. Evaluate models using operationally meaningful metrics.
5. Identify problems of data quality, explainability, bias, privacy and cybersecurity.
6. Design a responsible AI implementation for a reverse supply chain.

1. The Data and Decision Landscape

A reverse-logistics system generates data at several stages:

• Initiation: return reason, order history, customer information, authorisation records and promised return date.
• Transportation: carrier, route, scan events, location, weight, package condition and transit time.
• Receipt: arrival time, warehouse, serial number, packaging condition and receiving photographs.
• Inspection: test results, defect codes, technician observations, images and sensor readings.
• Disposition: resale, repair, refurbishment, remanufacture, recycling, donation, vendor return or disposal.
• After disposition: resale outcome, warranty claim, recovery value, environmental performance and customer feedback.

The analytical challenge is that these data are not automatically decision-ready. Events may be recorded at different times, identifiers may be inconsistent, outcomes may be missing, and labels may reflect historical decisions rather than objective product condition. For example, a model trained on past disposition decisions may learn that products from a particular region are more likely to be scrapped, even when the true condition is equivalent.

A data-driven architecture normally contains five layers:

1. Data capture: enterprise systems, warehouse management systems, transport systems, IoT devices, cameras and mobile applications.
2. Data management: master-data management, event storage, validation, access control and lineage.
3. Analytical models: forecasting, classification, regression, computer vision, anomaly detection and optimisation.
4. Decision and workflow: recommended action, human approval, exception handling and execution.
5. Monitoring and governance: accuracy, drift, fairness, security, auditability and realised business outcomes.

The unit of analysis must be defined carefully. A forecasting model may operate at product–warehouse–week level, whereas an inspection model operates at item–image level and a routing model operates at shipment–vehicle level. Combining records without preserving these levels can create misleading results.

2. Core Analytical Methods

2.1 Supervised, unsupervised and reinforcement learning

Supervised learning uses labelled examples. Inputs such as product age, return reason and test readings are mapped to an observed output, such as repair success. Classification predicts categories; regression predicts numerical values.

Unsupervised learning identifies structure without a labelled outcome. Clustering may reveal groups of return reasons or customer behaviours. Anomaly detection may identify unusual return patterns, although an anomaly is not necessarily fraud.

Reinforcement learning learns actions through interaction with an environment and feedback. It may be relevant to dynamic inventory allocation or routing, but it requires careful control because experimentation in live reverse networks can create service, safety or financial risks. In many organisations, constrained optimisation with forecasts is more transparent and easier to govern.

2.2 Forecasting returns

Return forecasting estimates the quantity and timing of future returns. A basic decomposition of observed returns is:

where R_t is the expected number of returns in period t, B_t is the relevant base of sales or installed units, and p_t is the probability of return during that period. This relationship is a conceptual expectation, not a complete forecasting model. It assumes that the base and probability are appropriately defined and that aggregation does not conceal important product, channel or cohort differences.

For a time-series model, a simple autoregressive form is:

Here, c is an intercept, φ is the effect of the previous period’s returns, X_j,t are explanatory variables such as promotions or product age, β_j are coefficients, and ε_t is an error term. The model assumes a stable enough relationship for estimation and requires attention to seasonality, changing policies and autocorrelation.

Forecast accuracy should be assessed with operationally appropriate measures. Mean absolute error is:

where R_t is actual returns, F_t is forecast returns and n is the number of evaluation periods. MAE is expressed in return units and treats overprediction and underprediction symmetrically. In reverse logistics, that symmetry may be inappropriate: underpredicting returns can cause congestion and missed recovery opportunities, while overpredicting can create idle capacity. Therefore, service-level and cost-weighted measures should supplement MAE.

Forecasts should be generated at the level at which decisions are made. A national forecast may be accurate overall but useless for allocating inspection labour among facilities. Hierarchical forecasting can reconcile product, site and regional forecasts so that detailed forecasts remain consistent with aggregate totals.

2.3 Automated inspection and image-based grading

Computer vision can detect missing components, scratches, cracks, stains, dents, seal damage and packaging deformation. A typical pipeline includes image capture, quality checking, object localisation, feature extraction, classification or segmentation, confidence estimation and human review for uncertain cases.

A classification model can be evaluated using a confusion matrix. For a positive class such as “repairable,” precision is:

and recall is:

where TP denotes true positives, FP false positives and FN false negatives. Precision answers: among items predicted as repairable, how many were actually repairable? Recall answers: among all truly repairable items, how many were identified? The appropriate priority depends on cost. If incorrectly approving unsafe equipment is severe, precision and safety controls may dominate. If missing repairable products causes unnecessary disposal, recall becomes important.

A useful combined measure is the F1 score:

It is appropriate when precision and recall both matter and their relative costs are not sufficiently different. It should not be used mechanically when the consequences of errors are highly asymmetric.

Image-based grading requires a controlled grading ontology. Terms such as “like new,” “good,” “fair” and “poor” must be linked to observable criteria. Lighting, camera angle, background, occlusion, dirt and packaging variation can materially affect performance. Images should therefore be captured using standardised procedures, and performance should be evaluated across product models, sites, lighting conditions and demographic or geographic segments where relevant.

2.4 Disposition selection

Disposition is a constrained economic and environmental decision. Let alternatives be indexed by d, such as resale, repair, refurbishment, recycling and disposal. A simplified expected contribution from disposition d is:

where P_d is expected revenue or recovery value, C_d is processing cost, T_d is transport and handling cost, E_d is the monetised environmental cost if included, and S_d represents avoided costs or strategic benefits. The units must be consistent, usually currency per item. The expression is useful only when uncertainty, capacity and regulatory restrictions are either incorporated or separately controlled.

If outcome probabilities are available, expected value can be written as:

where o indexes possible outcomes, Pr(o|x,d) is the probability of outcome o given item information x and disposition d, and V_d,o is the value if that outcome occurs. A decision rule selects the feasible alternative with the highest expected value, subject to safety, legal, capacity and service constraints.

For multiple items and scarce capacity, a binary optimisation model may be appropriate:

subject to:

Here x_i,d equals 1 if item i is assigned to disposition d, a_i,d is the capacity consumed and K_d is available capacity. The first constraint assigns each item once; the second respects capacity. Additional constraints may address hazardous materials, minimum recycling rates, promised customer dates and facility eligibility.

2.5 Routing and network decisions

AI can estimate travel times, consolidation opportunities, damage risk and facility processing times. Optimisation then uses these estimates to assign returns to collection points, consolidation centres, repair facilities or recyclers. A routing model should not optimise distance alone. Relevant objectives may include transport cost, carbon emissions, time, service reliability, risk and processing capacity.

For a route with segments j, total distance is:

where d_j is segment distance in kilometres. If emissions are estimated using an emissions factor e in kilograms of CO₂-equivalent per kilometre, then a simplified estimate is:

This assumes the factor is appropriate for the vehicle, load and operating conditions. More detailed models account for payload, fuel consumption, empty running and mode changes. AI can improve parameter estimates, but the optimisation model must still represent legal and operational constraints.

2.6 Fraud detection

Fraud patterns may include repeated “item not received” claims, serial-number reuse, manipulated photographs, mismatched weights, unusual geographic movements or collusion between accounts. Supervised models require confirmed fraud labels, which are often incomplete because non-detection does not mean legitimate behaviour. Consequently, a hybrid approach is preferable: rules for known controls, anomaly detection for novel patterns, and human investigation before adverse action.

Fraud scores should trigger review rather than automatically deny a customer unless the evidence, policy and legal basis are robust. False positives can damage customer relationships and disproportionately affect certain groups. Features should be reviewed for privacy, necessity and potential proxy discrimination.

2.7 Predictive maintenance

Reverse facilities depend on conveyors, scanners, compactors, testing equipment, vehicles and refrigeration or battery systems. Predictive maintenance uses vibration, temperature, current, error codes, operating hours and maintenance history to estimate failure risk or remaining useful life.

A hazard model may express the instantaneous failure rate as:

where h(t|x) is the hazard rate at time t given covariates x, h₀(t) is a baseline hazard and β is a coefficient vector. The model assumes that the proportional effect represented by exp(β^T x) is suitable for the data. Maintenance decisions should consider the cost of failure, planned downtime, spare parts and the risk of unnecessary intervention. A prediction is valuable only if it changes maintenance action in a beneficial way.

3. Model Evaluation and Decision Quality

A model’s statistical performance and operational value are different constructs. Evaluation should include:

Random train–test splits can produce leakage in time-dependent reverse data. A return from the same product, customer or incident may appear in both training and test sets. Time-based validation, group-based splitting and out-of-period testing are safer. Model performance should be monitored after deployment because product mix, return policy, camera hardware, fraud behaviour and market prices change.

Threshold selection should reflect consequences. Suppose a model predicts repairability probability. A threshold of 0.5 is not automatically correct. If repair capacity is scarce, the organisation may accept only high-confidence items. If disposal is environmentally harmful and repair capacity is expandable, a lower threshold may be justified. Thresholds should be documented, periodically reviewed and approved by accountable owners.

4. Explainability, Bias and Human Oversight

Explainability is needed for at least four reasons: operational troubleshooting, regulatory accountability, employee acceptance and contestability. Different forms serve different purposes:

• Global explanations describe overall model behaviour, such as feature importance.
• Local explanations describe why a particular item received a score.
• Example-based explanations show similar historical cases.
• Interpretable models use transparent structures, such as decision trees or additive scoring rules.

Feature importance is not automatically causation. A variable associated with scrapping may be a proxy for site, supplier or socioeconomic status. Explanations should therefore be combined with process knowledge and sensitivity analysis.

Bias can enter through unrepresentative samples, historical labels, measurement differences, unequal access to return services or feedback loops. For example, if only items sent to a premium facility receive detailed inspection, a model may confuse inspection intensity with product quality. Mitigation includes representative sampling, label audits, subgroup evaluation, reweighting, revised procedures, human review and periodic impact assessment.

Human oversight must be substantive rather than ceremonial. Reviewers need authority to override a recommendation, a reason code for the override, adequate time and training, and a mechanism for escalating systematic errors. Automation should be graduated: low-risk, high-confidence cases may be automated; ambiguous, high-value, safety-critical or legally sensitive cases should receive human review.

5. Data Quality and Cybersecurity

The principle “garbage in, garbage out” is especially important in reverse logistics. Data-quality controls should address:

• completeness: required fields are present;
• validity: values follow permitted formats and ranges;
• consistency: identifiers and units agree across systems;
• timeliness: events are recorded promptly;
• uniqueness: duplicate returns and serial numbers are detected;
• accuracy: measurements and labels reflect reality;
• lineage: the origin and transformations of data are traceable.

Data contracts can specify the meaning, owner, format, frequency and quality threshold for each critical field. A product serial number, for example, should have a defined structure, uniqueness rule, validation process and ownership across sales, service and returns systems.

AI expands the attack surface. Threats include stolen credentials, manipulated images, poisoned training data, adversarial inputs, model extraction, ransomware, insecure APIs and unauthorised access to customer or employee data. Controls include least-privilege access, encryption in transit and at rest, network segmentation, multifactor authentication, secure software development, immutable audit logs, backup recovery, image provenance checks and incident-response exercises.

Cybersecurity must cover the entire model lifecycle. A protected production endpoint is insufficient if training data can be altered without detection. Versioned datasets, signed model artefacts, approval gates and reproducible deployment help establish accountability.

Privacy principles require purpose limitation, data minimisation, retention controls, access rights and appropriate legal bases. Facial images, precise location data, customer identifiers and employee monitoring data require particular care. Organisations should avoid retaining sensitive data merely because future analytical uses might be possible.

6. Implementation Framework

A disciplined implementation can proceed through the following stages:

1. Define the decision: Specify who decides what, at which time, with which constraints and what success means.
2. Map the process: Document data capture, handoffs, failure points, exceptions and existing human judgement.
3. Establish a baseline: Compare against current rules or expert performance before building a complex model.
4. Audit data and labels: Examine missingness, leakage, sampling, bias, duplicates and label reliability.
5. Develop and validate: Use appropriate temporal or group-based validation and cost-sensitive metrics.
6. Pilot with safeguards: Run in shadow mode or limited scope; preserve human approval and rollback capability.
7. Integrate into workflow: Present recommendations with confidence, explanation, evidence and next action.
8. Monitor and improve: Track drift, overrides, subgroup outcomes, realised value, incidents and user feedback.
9. Govern retirement: Decommission models that no longer meet performance, safety or legal requirements.

The business case should include implementation cost, data preparation, integration, training, ongoing monitoring and failure costs. A model that saves processing cost but increases warranty failures may destroy value. Similarly, a routing model that reduces kilometres but increases unsafe handling is unacceptable.

7. Illustrative Case: Home Appliance Returns

Background

A fictional national appliance retailer receives approximately 18,000 returned appliances each month through stores, parcel carriers and direct collection. Returned items include refrigerators, washing machines, microwaves and small appliances. Historically, employees used inconsistent inspection checklists, and approximately 14% of items were sent to recycling without a documented repairability assessment.

Problem

The retailer faced four problems: uncertain weekly volumes, long inspection queues, inconsistent grading and poor allocation of items to repair centres. It also suspected serial-number abuse and wanted to reduce unnecessary transport and disposal.

Analysis

The project team first mapped the process and found that return reasons were coded differently across channels. Photographs varied widely in lighting and distance. Historical “scrap” labels were not reliable ground truth because some facilities had limited repair capacity. The team therefore created a new sample of 4,000 items assessed independently by two trained technicians, with disagreement resolved by a senior assessor.

A return forecast was developed at product-family–facility–week level using sales, warranty age, promotion calendars and historical returns. Performance was assessed using time-based validation and MAE, while under-forecasting was assigned a higher operational cost because it created queue congestion.

A vision model classified visible damage into standard categories, but it did not make safety decisions autonomously. Items with low image quality, possible electrical hazards or low confidence were routed to manual inspection. A separate model estimated the probability of successful repair using test readings, age, model family and defect category.

For disposition, the retailer estimated expected value for resale, repair, parts harvesting and recycling. Capacity constraints were included because repair centres could not accept unlimited units. A rules layer excluded unsafe, recalled or legally restricted items from automated recommendations.

Fraud analytics combined serial-number validation, account history, package weight and return timing. A high score generated an investigation case rather than an automatic rejection. Access to customer information was restricted to authorised investigators.

Approach and solution

The system displayed each item’s recommended disposition, estimated value, confidence, principal evidence and permitted alternatives. Employees could override the recommendation by selecting a reason. Weekly governance meetings reviewed overrides, false positives, model drift and differences among facilities.

Outcome

After a six-month pilot, the retailer reported a 9% reduction in average inspection queue time, a 6% increase in the proportion of items routed to economically viable repair, and a reduction in unnecessary inter-facility transfers. These figures are illustrative rather than universal results. The project also revealed that performance was substantially worse for one product family because its camera station used a different lighting configuration. Standardising the station improved performance more than changing the model architecture.

Learning points

The case demonstrates that value came from process redesign, reliable labels, capacity-aware decisions and human escalation—not from a model considered in isolation. It also shows why model governance must include operational and equity checks, not merely an accuracy score.

8. Contemporary Developments and Strategic Implications

Foundation models and multimodal AI can combine text, images, sensor readings and service histories. They may assist technicians by summarising evidence or retrieving relevant repair instructions. However, generated text can be inaccurate, and outputs require grounding in authorised knowledge sources and human verification.

Edge AI enables inspection or anomaly detection near the camera or machine, reducing latency and dependence on continuous cloud connectivity. Digital twins can represent return flows, facility capacity and asset condition for scenario analysis. Federated learning may allow organisations to learn from distributed data without centralising all raw records, although it does not remove privacy or security risks.

The strategic question is not whether an organisation uses the most advanced algorithm. It is whether data and decision capabilities create a defensible advantage: faster recovery, better customer service, lower environmental impact, improved resilience or more reliable compliance. Firms should prioritise high-value decisions with measurable outcomes and avoid “AI washing,” in which automation is advertised without meaningful analytical capability.

Challenges and Limitations

AI systems face persistent limitations:

• rare failures and fraud cases create severely imbalanced data;
• historical decisions may encode inefficient or biased practices;
• changing products and policies cause concept drift;
• labels may be subjective or expensive to obtain;
• predictions do not establish causation;
• optimisation can amplify errors in forecasts;
• high accuracy may not translate into financial or environmental value;
• automation can deskill employees if expertise is not retained;
• data sharing across suppliers raises contractual and privacy issues;
• complex models can be difficult to audit or contest.

Accordingly, AI should be treated as a socio-technical capability. Process design, incentives, skills, controls and organisational learning are as important as algorithms.

Best-Practice Checklist

A responsible reverse-logistics AI programme should:

1. Start with a defined operational decision and measurable outcome.
2. Use reliable, representative and traceable data.
3. Separate training, validation and future test periods appropriately.
4. Evaluate costs of different errors, not accuracy alone.
5. Include capacity, safety, regulatory and environmental constraints.
6. Present confidence and evidence with each recommendation.
7. Keep humans involved in high-risk or ambiguous decisions.
8. Test performance across facilities, products and relevant groups.
9. Monitor drift, overrides, incidents and realised benefits.
10. Secure data, interfaces, models and operational devices throughout the lifecycle.

Chapter Summary

AI and data-driven decision support can improve reverse-logistics forecasting, inspection, grading, disposition, routing, fraud detection and predictive maintenance. The central distinction is between predicting an outcome and selecting an action under constraints. Forecasts should be evaluated with measures such as MAE and operational service costs; classifiers require attention to precision, recall, calibration and asymmetric error consequences; disposition models should combine expected value with capacity, safety and environmental constraints.

Successful implementation depends on trustworthy data, reliable labels, appropriate validation, explainability, bias assessment, cybersecurity and meaningful human oversight. Computer vision and predictive models are valuable only when embedded in operational workflows. Contemporary approaches such as multimodal AI, edge analytics, digital twins and federated learning expand possibilities but also introduce new governance requirements. The most effective strategy is therefore responsible, measurable and process-centred adoption rather than indiscriminate automation.

Key Terms

Artificial intelligence: Computational systems that perform tasks associated with human reasoning, perception or decision making.

Machine learning: Methods that learn patterns from data to produce predictions or decisions.

Supervised learning: Learning from examples with known outcome labels.

Unsupervised learning: Learning patterns or groups without specified outcome labels.

Return forecasting: Estimating the quantity and timing of future product returns.

Computer vision: AI methods for interpreting images or video.

Image-based grading: Classifying product condition from visual evidence.

Confusion matrix: A table of correct and incorrect classification outcomes.

Precision: The proportion of predicted positives that are actually positive.

Recall: The proportion of actual positives correctly identified.

Calibration: Agreement between predicted probabilities and observed frequencies.

Disposition: The selected recovery pathway for a returned item.

Expected value: Probability-weighted value of possible outcomes.

Predictive maintenance: Using data to anticipate equipment failure or maintenance need.

Data drift: Change in input-data characteristics over time.

Concept drift: Change in the relationship between inputs and outcomes.

Explainability: The ability to communicate how or why a model produced an output.

Algorithmic bias: Systematic, often unfair differences in model outcomes.

Data lineage: A traceable record of data origin, transformations and use.

Model governance: Policies and controls for developing, deploying, monitoring and retiring models.

Review Questions

Short Answer Questions

1. Distinguish prediction from decision support in reverse logistics.
2. Define precision and recall and explain why both matter in image-based grading.
3. What is data drift, and why can it reduce model performance?
4. Identify four important dimensions of data quality.
5. Why should a high fraud score normally trigger investigation rather than automatic rejection?

Long Answer Questions

1. Explain how AI can support the complete reverse-logistics process from return initiation to final disposition.
2. Discuss the design of a capacity-aware disposition model, including expected value and operational constraints.
3. Compare supervised learning, unsupervised learning and reinforcement learning in reverse-logistics applications.
4. Analyse the requirements for responsible use of computer vision in automated inspection.
5. Explain how explainability, bias assessment and human oversight should be incorporated into AI governance.

Analytical / Application Questions

1. A repairability classifier identifies 90 truly repairable items, misses 10 repairable items and incorrectly labels 20 non-repairable items as repairable. Calculate precision and recall, and discuss which metric may be more important if unnecessary repair processing is very costly.
2. A facility must choose between repair and recycling. Repair produces expected revenue of £120, repair cost of £55 and transport cost of £15. Recycling produces recovery value of £35, processing cost of £8 and transport cost of £5. Ignoring environmental monetisation and capacity constraints, calculate the contribution of each option and recommend a disposition.
3. Design a governance plan for deploying an image-based grading model across three warehouses with different camera systems. Specify validation, monitoring, human review, cybersecurity and model-retirement controls.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from decision support in reverse logistics?

A. Prediction selects an action, while decision support only stores data.

B. Prediction estimates an uncertain outcome, while decision support combines estimates with objectives and constraints to choose an action.

C. Prediction is always manual, while decision support is always automated.

D. Prediction applies only to transport, while decision support applies only to inspection.

Correct Answer: B. Prediction estimates an uncertain outcome, while decision support combines estimates with objectives and constraints to choose an action.

Question 2

A classifier produces 80 true positives and 20 false positives. What is its precision?

A. 20%

B. 25%

C. 75%

D. 80%

Correct Answer: D. 80%

Question 3

Which validation approach is most appropriate for a return-forecasting model whose data are time-dependent?

A. Randomly split all records without regard to date.

B. Use only the most recent observations for training and older observations for testing.

C. Use time-based validation in which future periods are held out from training.

D. Test the model on the same records used for training.

Correct Answer: C. Use time-based validation in which future periods are held out from training.

Question 4

Which factor is most likely to create label leakage in a disposition model?

A. Using a product’s authorised return date as an input.

B. Using a final disposition decision as an input when predicting that same decision.

C. Using product weight measured at receipt.

D. Using the warehouse identifier as a monitored feature.

Correct Answer: B. Using a final disposition decision as an input when predicting that same decision.

Question 5

If false negatives are especially dangerous in safety inspection, which metric should receive particular attention?

A. Recall

B. Mean absolute error

C. Transport distance

D. Specificity of the product catalogue

Correct Answer: A. Recall

Question 6

In an expected-value disposition model, why are probabilities assigned to possible outcomes?

A. To remove all operational constraints.

B. To calculate a probability-weighted estimate of value under uncertainty.

C. To guarantee that the highest-revenue option is selected.

D. To replace data-quality controls.

Correct Answer: B. To calculate a probability-weighted estimate of value under uncertainty.

Question 7

Which is an example of concept drift?

A. A database contains duplicate serial numbers.

B. A camera stores images at a higher resolution.

C. The relationship between return reasons and repair success changes after a product redesign.

D. A required field is missing from a transaction.

Correct Answer: C. The relationship between return reasons and repair success changes after a product redesign.

Question 8

What is the most appropriate response to a low-confidence image classification for a potentially hazardous product?

A. Automatically recycle it without inspection.

B. Automatically resell it at a discount.

C. Route it to a defined human or specialist safety review.

D. Delete the image and retain only the predicted class.

Correct Answer: C. Route it to a defined human or specialist safety review.

Question 9

Which control most directly reduces the risk of unauthorised access to AI production systems?

A. Removing all model monitoring.

B. Multifactor authentication and least-privilege access.

C. Increasing the classification threshold.

D. Using a larger training dataset.

Correct Answer: B. Multifactor authentication and least-privilege access.

Question 10

Why can a model with high predictive accuracy still produce poor reverse-logistics decisions?

A. Accuracy always measures transport emissions incorrectly.

B. The model may be disconnected from costs, capacity, safety or environmental constraints.

C. High accuracy means that no human oversight is possible.

D. Accurate models cannot be used with optimisation.

Correct Answer: B. The model may be disconnected from costs, capacity, safety or environmental constraints.