Production planning and control (PPC) has traditionally coordinated demand, materials, capacity, labour, inventory and shop-floor execution. Digitalisation has expanded this task. Artificial intelligence (AI), industrial Internet of Things (IIoT) sensors, cloud platforms, digital twins and advanced analytics can now support decisions across the planning horizon—from demand forecasting and aggregate planning to sequencing and real-time control.

However, technological capability does not remove managerial responsibility. AI-generated plans can amplify poor data, embed historical discrimination, expose confidential information or recommend actions that are operationally efficient but environmentally or socially harmful. Production systems must therefore be designed for resilience, sustainability and accountable governance. The central managerial question is not simply “Can an algorithm optimise the schedule?” but “What objectives should be optimised, who is affected, what evidence supports the recommendation, and who remains responsible for the decision?”

This chapter examines AI-assisted PPC as a socio-technical capability. It integrates operational performance with continuity, environmental stewardship, human judgement, regulatory compliance and Industry 5.0 principles.

Learning Objectives

1. AI in the PPC Decision Architecture

AI is an umbrella term for computational methods that perform tasks associated with learning, prediction, classification, optimisation or decision support. In PPC, AI should be understood as a component within a decision architecture rather than as a replacement for the planning system.

The layers are interdependent. A sophisticated model cannot compensate for inaccurate inventory records or inconsistent product definitions. Likewise, a mathematically optimal schedule is unusable if it violates labour agreements or lacks materials.

1.1 AI-assisted demand forecasting

Forecasting methods range from seasonal time-series models to machine-learning models using promotions, weather, macroeconomic indicators, web traffic and customer behaviour. AI is particularly valuable when many explanatory variables interact or when demand patterns change rapidly.

A basic seasonal-naïve forecast may use demand from the same period in the previous cycle. A moving average for a window of n periods is:

where Fₜ₊₁ is the forecast for the next period, Dₜ is observed demand in period t, and n is the number of observations. The method assumes that recent observations are reasonably representative and that no strong trend or structural change invalidates the average.

Forecast accuracy should not be assessed by one measure alone. Common measures include:

Here, Dₜ is actual demand, Fₜ is forecast demand and n is the number of periods. MAE is expressed in demand units; RMSE is also expressed in demand units but penalises large errors more strongly. MAPE is a percentage but is unsuitable when actual demand is zero or close to zero. Forecast evaluation must use a time-ordered holdout period: randomly mixing future observations into training data creates data leakage.

AI forecasts should be compared with a transparent baseline. A complex model is justified only if it produces materially better decisions, not merely a lower statistical error. Forecast bias is also important:

A persistently positive bias indicates under-forecasting, while a negative bias indicates over-forecasting, assuming the sign convention shown.

1.2 Optimisation and scheduling

Optimisation selects values for decision variables while satisfying constraints and pursuing objectives. In PPC, variables may represent production quantities, machine assignments, overtime, inventory levels or shipment quantities. Constraints may include capacity, material availability, due dates, minimum batch sizes, labour rules and maintenance windows.

Minimise Z = w₁C_cost + w₂C_lateness + w₃C_inventory + w₄C_carbon + w₅C_risk

where each C is a comparable cost or penalty measure and each w is a managerial weight. The terms must be normalised or converted into meaningful common units; otherwise a large numerical scale can dominate the result arbitrarily. The weights express priorities, not objective truth. Different stakeholders may reasonably choose different weights.

where pᵢⱼ is the processing time required on resource i for one unit of product j, xⱼ is the quantity of product j, and Hᵢ is available capacity on resource i. The units on both sides are time. The constraint applies only when processing times and available capacity are defined over the same planning period.

Autonomous scheduling uses algorithms to generate or revise schedules with limited human intervention. It may combine constraint programming, mixed-integer optimisation, simulation, reinforcement learning or heuristic search. It is most appropriate when the decision environment is well instrumented, constraints are explicit, the cost of delay is significant and safe override mechanisms exist. It is inappropriate to grant full autonomy where consequences are irreversible, data are unreliable or legal and safety obligations require human judgement.

1.3 Predictive analytics

Predictive analytics estimates the probability or expected timing of an event. In maintenance, a model may estimate the probability of failure within a defined horizon. In quality management, it may identify process conditions associated with defects. In supply management, it may estimate late-delivery risk.

The operational value of a prediction depends on the decision threshold. A maintenance alert has four possible outcomes: true positive, false positive, true negative and false negative. The choice of threshold should reflect the relative costs of these outcomes. For example, a false negative on a safety-critical component may be unacceptable even if frequent false positives increase inspection cost.

Model discrimination and operational usefulness are different. A model may rank risky machines accurately but still fail to improve performance if planners cannot obtain spares or schedule maintenance. Consequently, predictive analytics must be connected to feasible actions and monitored after deployment.

2. Resilient Production Planning

Resilience is the capability of a production system to prepare for disruption, absorb its effects, recover performance and adapt to changed conditions. It differs from efficiency. Efficiency seeks to reduce slack; resilience deliberately preserves selected forms of flexibility and redundancy.

2.1 Sources and dimensions of disruption

Disruptions may arise from supplier insolvency, geopolitical restrictions, transport interruption, cyberattack, equipment failure, labour shortage, extreme weather, quality incidents, demand shocks or public-health emergencies. Their consequences propagate through material, information and financial flows.

2.2 Planning mechanisms for resilience

A resilient plan normally combines several mechanisms rather than relying on one large inventory buffer:

Buffers should be targeted. Holding excess inventory everywhere can hide poor planning, increase obsolescence and consume working capital. A risk-based approach concentrates protection where the product of likelihood and consequence is high.

where Pₖ is the probability of disruption scenario k and Iₖ is its financial or operational impact. This is not a complete risk model: correlations, non-financial consequences and uncertainty in the estimates must also be considered.

Scenario planning asks how the PPC system performs under plausible futures. Scenarios should vary demand, supply, capacity, lead time, energy availability and recovery duration. The preferred plan may not be the cheapest in the normal scenario; it may be the plan with acceptable performance across a broad range of scenarios.

2.3 Resilience metrics

Relevant measures include recovery time, recovery point, fill rate during disruption, time to detect, supplier concentration, percentage of critical parts with approved alternatives, schedule stability and the proportion of capacity that is flexible. Metrics should be balanced. A plan that achieves high utilisation but cannot recover from a small failure is not resilient.

3. Sustainability in PPC

Sustainable PPC integrates environmental, economic and social performance over the product and process life cycle. The objective is not simply to reduce energy per unit while increasing total output and total emissions. Both intensity and absolute impact matter.

3.1 Environmental planning variables

Important variables include energy consumption, greenhouse-gas emissions, material yield, scrap, water use, hazardous waste, packaging, transport distance and product end-of-life. A production plan can influence these variables through batch sizes, sequencing, machine selection, overtime, shipment consolidation and renewable-energy timing.

If electricity consumption is E in kilowatt-hours and the relevant emission factor is g kilograms of CO₂-equivalent per kilowatt-hour, emissions are:

The units are kg CO₂e. The emission factor must be appropriate to the location, time period and electricity contract. Scope boundaries and data quality should be documented.

A carbon-aware schedule may shift discretionary production to periods with lower grid carbon intensity, provided due dates, labour conditions, product quality and total cost remain acceptable. Such decisions require life-cycle thinking: a local energy saving may be offset by additional transport, spoilage or inefficient machine start-ups.

3.2 Circularity and material efficiency

PPC supports circularity by planning remanufacturing, repair, reuse, take-back, recycled input and segregated material flows. Bills of material should identify recycled content, recoverable components and material restrictions. Reverse logistics must be planned alongside forward production.

The numerator and denominator must use compatible units and system boundaries. Improving yield reduces material demand and waste, but quality and safety must not be compromised.

3.3 Social sustainability

Socially responsible plans consider safe workloads, fatigue, fair scheduling, skills development, employment quality and impacts on communities. An algorithm that systematically assigns undesirable shifts to a particular group may be operationally efficient but socially unacceptable and potentially unlawful. Sustainability therefore includes human rights and decent work, not only carbon accounting.

4. Governance of AI-Assisted PPC

Governance is the system through which authority, responsibility, controls and review are established. It converts AI from an experiment into a controlled organisational capability.

4.1 Accountability and decision rights

Every AI use case should have a named business owner, technical owner, data steward and operational approver. The organisation should record:

Responsibility cannot be transferred to “the algorithm.” If a planner accepts an AI recommendation, the organisation remains accountable for the design, deployment and supervision of the system.

4.2 Human oversight

Risk level	Typical use	Appropriate control
Low	Draft forecast or replenishment suggestion	Planner review and exception monitoring
Medium	Capacity allocation or maintenance prioritisation	Approval gates, explainability and audit trail
High	Safety-related sequencing or workforce-impacting decisions	Mandatory competent human decision, independent validation and documented rationale

Human oversight must be meaningful. A person who can only click “approve” under time pressure is not exercising effective control. Planners need training, access to evidence, authority to override and protection from unreasonable automation targets.

4.3 Bias and fairness

Bias may enter through incomplete samples, historical decisions, proxy variables, measurement errors or feedback loops. For example, a supplier-risk model trained on historical purchasing decisions may label small suppliers as risky because they previously received fewer orders, not because they perform poorly.

Fairness assessment should examine performance across relevant groups, sites, suppliers or product categories. Useful questions include: Does error rate differ materially between groups? Are protected attributes being used directly or indirectly? Does the model create unequal access to overtime, training or production opportunities? Is there a legitimate operational reason for the difference?

Removing sensitive variables does not guarantee fairness because location, job title or purchasing history may act as proxies. Bias mitigation may require better sampling, revised labels, constraints, human review and continuous monitoring.

4.4 Privacy and cybersecurity

PPC systems may process personal data such as attendance, productivity, skills, location and health-related restrictions. Data collection should follow purpose limitation, data minimisation, access control, retention limits and applicable privacy law. Employees should receive clear information about what is collected, why it is used and how decisions can be challenged.

Cybersecurity is a production-continuity issue. Controls should include identity management, least-privilege access, network segmentation, secure software updates, immutable backups, incident response exercises and supplier-security assessment. Connecting an AI model directly to actuators without safeguards can turn a data or model compromise into a physical event.

4.5 Explainability, validation and model drift

Explainability should match the decision. A planner may need to know that a schedule changed because a critical machine became unavailable, a forecast rose because of a promotion and a supplier score fell because of repeated late deliveries. Explanations should not imply causal certainty when the model has only identified statistical association.

Validation must occur before and after deployment. It should test accuracy, constraint compliance, robustness to missing data, edge cases, cybersecurity, fairness and operational outcomes. Model drift occurs when demand, suppliers, processes or behaviour changes. Monitoring should trigger investigation when error, override frequency, data distributions or constraint violations exceed thresholds.

5. A Governance Framework for Implementation

A model should not move directly from a successful laboratory test to autonomous shop-floor control. Production trials reveal practical problems such as missing master data, planner distrust, unstable interfaces and unanticipated workarounds.

6. Industry 5.0 and the Future of PPC

Industry 4.0 emphasised connectivity, automation and data-driven optimisation. Industry 5.0 extends this emphasis by placing human-centricity, sustainability and resilience at the centre of industrial development. The objective is not to reverse automation but to use technology to augment human capability and create systems that are robust, responsible and adaptable.

6.1 Human–machine collaboration

Collaborative robots, augmented-reality instructions, natural-language interfaces and decision-support systems can reduce physical strain and improve access to expertise. Yet workers should participate in process design. Their tacit knowledge often identifies quality risks, unsafe conditions and practical constraints that are absent from databases.

6.2 Digital twins and adaptive planning

A digital twin links a physical asset or process to a continuously updated digital representation. In PPC, it can test alternative schedules, simulate bottlenecks, estimate energy use and evaluate disruption scenarios. Its credibility depends on synchronised data, validated process logic and clearly stated assumptions. A visually impressive twin that does not represent actual changeover times or quality losses can mislead decision-makers.

6.3 Responsible autonomy

Future systems will increasingly recommend or execute local decisions. Responsible autonomy requires bounded authority: the system may optimise within approved limits but must stop, explain or escalate when conditions leave those limits. Examples include a schedule agent that may resequence non-critical jobs but cannot override a safety interlock or labour-rest rule.

7. Illustrative Case: Resilient and Sustainable Scheduling at Meridian Components

Background. Meridian Components, a fictional manufacturer of precision housings, serves automotive and medical-equipment customers. It operates three machining centres and purchases a specialised alloy from a single overseas supplier. Demand is volatile, electricity prices vary hourly and the company has experienced late deliveries and unplanned machine downtime.

Problem. The existing planning process used a spreadsheet forecast and a cost-minimising schedule. The schedule achieved high utilisation but created frequent overtime, excessive changeovers and vulnerability to alloy disruption. Management proposed an AI scheduling system that would use demand history, machine data, supplier performance and electricity information.

Analysis. The project team first defined non-negotiable constraints: medical orders, quality approvals, maximum working hours, preventive-maintenance windows and material traceability. It established a seasonal-naïve forecast as a baseline and tested a machine-learning model using a chronological holdout. The advanced model reduced MAE by 12% overall but performed poorly for low-volume medical products; therefore, the system used the advanced model only where performance was demonstrably superior.

A scenario analysis considered normal supply, a four-week alloy delay, a machine failure and a 25% demand surge. The team found that a small reserve of alloy, an approved second supplier and cross-training produced greater recovery improvement than simply increasing average inventory. The optimisation objective included production cost, lateness, changeovers, estimated carbon emissions and a disruption penalty. The schedule could shift flexible jobs to lower-carbon electricity periods, but medical orders and due dates remained dominant constraints.

Approach and solution. The AI system initially operated in shadow mode. Planners compared its recommendations with their own schedules and recorded reasons for overrides. A governance committee reviewed privacy because operator performance data were collected. Individual productivity scores were excluded from disciplinary decisions, and data were aggregated for workforce planning. Every schedule change generated an explanation and audit record.

The final system used human approval for weekly scheduling and allowed limited automatic resequencing during the shift. It could not alter safety, quality or labour constraints. A dashboard displayed forecast error, schedule stability, emissions per unit, overtime, material coverage, override frequency and recovery time under scenarios.

Outcome. In the illustrative six-month pilot, schedule changeovers fell by 9%, overtime by 7% and forecast MAE by 12% for eligible product families. The organisation also improved disruption recovery because the second supplier and substitution rules had been approved in advance. The system did not eliminate judgement: planners continued to review new products, abnormal demand and model alerts.

Learning points. First, the business case depended on process redesign and data governance, not the algorithm alone. Second, resilience required targeted flexibility rather than indiscriminate inventory. Third, sustainability objectives had to be integrated into the optimisation model and measured after implementation. Finally, a staged autonomy model produced more trust and safer adoption than immediate full automation.

8. Challenges and Limitations

AI-assisted PPC has significant limitations. Historical data may reflect obsolete products, exceptional disruptions or biased managerial decisions. Forecasts cannot reliably predict unprecedented events. Optimisation can produce false precision when costs and probabilities are uncertain. Multi-objective models may conceal value judgements inside weights. Real-time systems can increase instability if they respond to every small fluctuation; planners may need frozen horizons or change thresholds.

There are also organisational risks. Automation may deskill planners, reduce learning from experience or create excessive dependence on vendors. Employees may resist surveillance or fear job displacement. Small organisations may lack specialist data and cybersecurity capabilities. Finally, sustainability metrics can be manipulated through narrow system boundaries or offset claims that do not reduce actual operational impacts.

Best practice is therefore to combine automation with transparent baselines, robust process controls, human expertise, scenario testing and independent review. The standard of success is improved organisational performance under real operating conditions—not model sophistication in isolation.

Chapter Summary

AI can strengthen PPC through demand forecasting, predictive maintenance, anomaly detection, optimisation and bounded autonomous scheduling. Its value depends on data quality, feasible constraints, decision relevance and continuous monitoring. Forecast accuracy measures such as MAE, RMSE, MAPE and bias must be selected with awareness of their assumptions and limitations.

Resilient production planning anticipates disruption, absorbs shocks, recovers quickly and adapts through learning. It uses targeted buffers, supplier alternatives, flexible capacity, scenario planning and recovery priorities. Sustainable PPC incorporates energy, emissions, materials, waste, circularity and social conditions into planning decisions rather than treating them as afterthoughts.

AI governance requires named accountability, meaningful human oversight, fairness assessment, privacy protection, cybersecurity, explainability, audit trails and controls for model drift. Industry 5.0 broadens the purpose of digital production by combining technology with human-centricity, sustainability and resilience. The future PPC professional will therefore need analytical competence as well as ethical judgement, systems thinking and the ability to govern socio-technical change.

Key Terms

Artificial intelligence: Computational methods that perform learning, prediction, optimisation or decision tasks.

Autonomous scheduling: Automatic generation or revision of production schedules within defined authority and constraints.

Bias: Systematic error or unequal treatment arising from data, models or decision processes.

Carbon intensity: Emissions associated with a unit of energy, output or activity.

Circularity: Designing production and consumption systems to reduce waste through reuse, repair, remanufacture and recycling.

Digital twin: A continuously updated digital representation of a physical asset, process or system.

Explainability: The ability to communicate why a model produced a result in an understandable way.

Forecast bias: Persistent tendency of forecasts to be above or below actual demand.

Human oversight: Competent human review, intervention and accountability over an automated system.

Industry 5.0: A human-centric, sustainable and resilient development of advanced industrial systems.

Model drift: Decline in model relevance as data, processes or operating conditions change.

Multi-objective optimisation: Optimisation involving several objectives that may conflict.

Predictive analytics: Analysis used to estimate future events, risks or outcomes.

Resilience: The ability to prepare for, absorb, recover from and adapt to disruption.

Scenario planning: Evaluation of decisions under several plausible future conditions.

Sustainability: Managing environmental, economic and social performance over time.

Time-series leakage: Use of future information in model training or evaluation, causing falsely optimistic results.

Traceability: Ability to identify the origin, transformation and history of materials or decisions.

Review Questions

Short Answer Questions

Long Answer Questions

Analytical / Application Questions

Multiple Choice Questions with Answers

Question 1

Which measure is expressed in the same demand units as the forecast error and does not square the error?

Question 2

What is the main purpose of a chronological holdout period in forecasting evaluation?

Correct Answer: B. To prevent future information from contaminating model evaluation

Question 3

Which action most directly improves resilience to a single-source component disruption?

Question 4

Question 5

Why should objectives in a multi-objective optimisation model be normalised or placed on comparable scales?

Correct Answer: A. To ensure the largest numerical scale does not dominate arbitrarily

Question 6

Which situation is the clearest example of a meaningful human-oversight control?

C. A trained planner can review evidence, override the system and record the rationale

Correct Answer: C. A trained planner can review evidence, override the system and record the rationale

Question 7

Correct Answer: B. A decline in model relevance as operating conditions change

Question 8

Which Industry 5.0 principle most clearly distinguishes it from a narrow automation strategy?

Correct Answer: A. Human-centricity combined with sustainability and resilience

Question 9

A production system has a disruption probability of 0.2 and an estimated impact of $50,000. Using expected loss P × I, what is the result for this scenario?

Question 10

Correct Answer: C. Applying purpose limitation, access control and retention limits

Chapter 15 AI, Resilience, Sustainability and PPC Governance

AI, Resilience, Sustainability and PPC Governance

Introduction and Context