Analytics, Optimisation and AI-Based Decision-Making

Introduction

Warehouses and distribution networks generate large volumes of operational data: orders, receipts, stock balances, locations, equipment movements, labour activities, transport events, customer priorities and exception records. The managerial challenge is not simply to collect data, but to convert it into better decisions. Analytics provides the methods for describing what has happened, predicting what is likely to happen, and prescribing what should be done.

In warehousing and distribution, analytics supports demand forecasting, inventory positioning, slotting, labour planning, replenishment, capacity management, vehicle routing and service-level control. Optimisation extends analytics by identifying the best feasible decision under specified objectives and constraints. Artificial intelligence (AI), machine learning, digital twins and autonomous systems can improve responsiveness, but they do not eliminate the need for sound process design, human judgement, data governance and accountability.

This chapter develops an integrated decision framework. It distinguishes descriptive, diagnostic, predictive and prescriptive analytics; presents quantitative models used in warehouse and distribution decisions; examines optimisation and simulation; and critically evaluates AI-enabled operations.

Learning objectives

After studying this chapter, the learner should be able to:

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics.
2. Apply forecasting, inventory, slotting, labour, capacity and routing models.
3. Formulate operational decisions as optimisation problems with objectives and constraints.
4. Interpret predictive-model accuracy and recognise data and model limitations.
5. Explain the role of digital twins, AI and autonomous systems in logistics.
6. Evaluate analytics initiatives using operational, financial, ethical and governance criteria.

1. The analytics hierarchy in logistics

Analytics is most useful when connected to a decision and a measurable outcome. A dashboard showing yesterday’s order lines is descriptive; a model explaining why productivity fell is diagnostic; a forecast of tomorrow’s workload is predictive; and a recommendation for staffing, wave timing or carrier allocation is prescriptive.

The hierarchy is cumulative. Prescriptive analytics depends on reliable descriptive data and an adequate predictive model. A sophisticated algorithm cannot compensate for inaccurate stock records, inconsistent item identifiers or poorly defined service objectives.

A useful decision statement has five elements: decision owner, decision variable, objective, constraints and time horizon. For example: “The fulfilment manager must assign 42 available workers to picking and packing for the next shift to minimise late orders while respecting skill, safety and labour-hour constraints.” This is more actionable than “use AI to improve productivity.”

2. Data foundations and performance measures

Warehouse analytics commonly integrates a warehouse management system (WMS), enterprise resource planning system, transport management system (TMS), labour-management system, order-management system, telematics and automated equipment. Master data should include SKU dimensions, weight, handling class, storage requirements, supplier and customer attributes, unit conversions and location capacities.

Data quality has several dimensions:

• Accuracy: the recorded value agrees with reality.
• Completeness: required fields and events are present.
• Timeliness: data arrives quickly enough for the decision.
• Consistency: definitions agree across systems.
• Validity: values conform to permitted units, ranges and codes.
• Traceability: a result can be linked to its source data and transformation.

Important measures include order-cycle time, dock-to-stock time, pick rate, lines per labour hour, inventory accuracy, order accuracy, fill rate, on-time-in-full delivery, travel distance, space utilisation and cost per order. Measures should be interpreted together. For example, a high pick rate may reflect easier orders, reduced quality checking or the postponement of difficult work. Local optimisation can damage total system performance.

For a service measure, fill rate is often defined as the proportion of demand supplied immediately from available stock. If 9,700 of 10,000 requested units are supplied immediately, the unit fill rate is:

The measure is meaningful only when the numerator and denominator use the same unit, period and order-status rules. Order fill rate, line fill rate and unit fill rate are not interchangeable.

3. Forecasting warehouse demand and workload

Forecasting supports replenishment, labour scheduling, dock planning, vehicle capacity and inventory positioning. The forecast target must be specified. Demand for a SKU, order lines by hour, cartons by wave and inbound pallet arrivals are different time series.

A simple moving average for a level series is:

where F_t+1 is the forecast for the next period, D_t is observed demand in period t, and n is the number of observations. It assumes recent observations are representative and gives equal weight to them. It is unsuitable when strong trend or seasonality is present unless the series is adjusted.

Exponential smoothing gives greater weight to recent observations:

Here α is the smoothing parameter, D_t is actual demand and F_t is the previous forecast. A larger α responds more quickly to changes but may react excessively to random variation. This level-only model does not explicitly represent trend or seasonality.

Forecast errors should be monitored, not hidden. Common measures include:

MAE is expressed in the original demand units. RMSE is also in those units but penalises large errors more heavily. Percentage measures can be misleading when actual demand is zero or very small. Forecast accuracy must be evaluated using a time-based holdout or rolling-origin procedure; evaluating on the same data used to fit a model produces over-optimistic results.

In warehouses, the operational forecast may be derived from demand forecasts. For example, cartons per order, lines per order and units per carton convert sales demand into picking and packing workload. These conversion factors should be forecast or monitored because product mix changes can make a unit forecast a poor labour forecast.

4. Inventory and capacity decisions

A basic economic order quantity (EOQ) model illustrates the trade-off between ordering and holding costs:

where Q^* is the economic order quantity in units per order, D is annual demand in units, S is ordering or setup cost per order, and H is annual holding cost per unit. The model assumes relatively stable demand, constant lead time, instantaneous replenishment, no quantity discounts and no stockouts. It is a benchmark, not an automatic purchasing instruction.

For uncertain demand during a replenishment lead time, a reorder point can be represented as:

where μ_L is expected demand during lead time, σ_L is the standard deviation of lead-time demand, and z is the standard-normal safety factor associated with the desired cycle-service probability. If daily demand is independent with mean μ_d and standard deviation σ_d, and lead time is a fixed L days, then:

These expressions assume independent daily demand and a fixed lead time. If lead time is variable, demand and lead-time uncertainty should be modelled jointly; simply adding unrelated safety factors can be incorrect.

Capacity analytics compares required workload with available capacity. If a process has m identical workers, each available for h hours, and effective productivity is p units per labour hour, then nominal capacity is:

Effective capacity should adjust for breaks, meetings, absenteeism, congestion, equipment downtime and learning. A capacity utilisation ratio is:

Utilisation above 1 indicates infeasibility for the period. Sustained utilisation close to 1 may also produce queues and poor resilience because variability has no buffer.

5. Slotting and warehouse layout analytics

Slotting assigns SKUs to storage locations to balance travel, access, replenishment effort, space and safety. A common starting principle is to place high-frequency or high-volume items near appropriate picking and dispatch areas, but frequency alone is insufficient. Dimensions, weight, cube, compatibility, ergonomics, security, temperature and co-pick relationships must also be considered.

A simple travel-priority score can combine normalised measures:

where F_i is pick frequency for SKU i, V_i is volume or workload, C_i is co-pick benefit, R_i is replenishment or handling risk, and the w terms are managerial weights. Because the variables must be normalised before combination, the score is a prioritisation heuristic rather than a universal law.

Slotting can be formulated as an assignment problem. Let x_ij=1 if SKU i is assigned to location j, and zero otherwise. A simplified objective is:

subject to each SKU receiving an allowable location, each location respecting capacity, and incompatible assignments being prohibited. Here c_ij estimates travel, handling and replenishment cost. Real implementations may require multiple periods, replenishment frequency, family constraints and re-slotting cost. The best solution is therefore often a mixed-integer optimisation or a heuristic supported by simulation.

6. Labour planning and routing

Labour planning begins with a workload forecast by process and time interval. The manager then decides staffing, shift timing, cross-training, overtime and task allocation. A sound model includes skill restrictions and legal or contractual constraints, not merely headcount.

For distribution routing, the vehicle-routing problem (VRP) assigns customer visits to vehicles and sequences stops. A simplified objective minimises distance or travel cost while satisfying vehicle capacity, customer service windows, driver regulations and depot requirements. Exact optimisation becomes computationally difficult as the number of customers and constraints increases; practical systems therefore use mixed-integer programming for smaller instances and heuristics or metaheuristics for larger ones.

Routing models should include realistic travel times, loading constraints, delivery priorities, break rules, time windows and failed-delivery probability. A mathematically shorter route may be operationally inferior if it uses congested roads, creates difficult loading sequences or increases customer waiting time. Scenario testing is essential when travel times are uncertain.

7. Prescriptive analytics and optimisation practice

An optimisation model has three core components:

1. Decision variables: what the organisation controls, such as quantities, assignments, sequences or schedules.
2. Objective function: what is minimised or maximised, such as cost, lateness, distance, emissions or service.
3. Constraints: physical, commercial, regulatory and policy limits.

Multi-objective decisions require explicit priorities. A weighted objective might be:

Weights have meaning only after component measures are scaled comparably. Otherwise a variable measured in large numerical units can dominate the objective. An alternative is lexicographic optimisation, in which service feasibility is enforced first and cost is minimised second.

Optimisation outputs should be treated as recommendations subject to validation. Planners should inspect infeasible constraints, shadow prices or marginal values where available, sensitivity to assumptions, and the operational consequences of exceptional cases. A model that is optimal under inaccurate travel times or obsolete capacities is not operationally optimal.

Simulation complements optimisation. Discrete-event simulation represents arrivals, queues, resources and process variability. It is useful for testing layout, staffing and automation without disrupting live operations. Simulation does not find the best decision automatically; it evaluates specified policies under specified scenarios. Combining optimisation with simulation can reveal whether an apparently optimal plan remains effective under variability.

8. Artificial intelligence and machine learning

Machine learning can identify patterns in demand, travel time, equipment failure, image quality, labour productivity and order risk. Supervised learning uses labelled outcomes, such as whether an order was late. Unsupervised learning identifies structure without a target, such as SKU or customer clusters. Reinforcement learning learns policies through interaction with an environment, but its use in safety-critical warehouse control requires strong safeguards and carefully designed reward functions.

Useful applications include:

• probabilistic demand and workload forecasting;
• anomaly detection for inventory discrepancies or equipment behaviour;
• computer vision for damage, counting and safety monitoring;
• predictive maintenance for conveyors, sorters and vehicles;
• dynamic ETA and route prediction;
• intelligent exception prioritisation;
• natural-language interfaces for operational queries.

Prediction is not prescription. A model may predict that an order is likely to be late without identifying an action that is feasible, cost-effective or fair. AI should therefore be connected to a decision workflow with constraints, approval rights, monitoring and fallback procedures.

Model performance should be evaluated using a baseline, out-of-sample data and operational metrics. Classification models require attention to precision, recall and cost-weighted errors. A false negative for a safety hazard may be much more serious than a false positive. Data drift occurs when demand, product mix, processes or sensor behaviour change. Models need monitoring, retraining criteria, version control and documented ownership.

Explainability is especially important when AI affects worker evaluation, service allocation, safety interventions or customer priority. Explanations need not reveal proprietary code, but affected decision-makers should understand relevant inputs, limitations and appeal routes.

9. Digital twins and autonomous systems

A digital twin is a continuously updated digital representation of a physical asset, process or network, linked through data and used for monitoring, prediction or experimentation. It is more than a static simulation model: its value depends on a meaningful connection between the physical and digital systems.

A warehouse digital twin may represent inventory, locations, conveyors, robots, workers, orders, queues and environmental conditions. It can test alternative layouts, wave policies, robot fleets and staffing patterns. Its credibility depends on calibration against observed cycle times, travel times, failure rates and queue behaviour. A visually impressive model with poor calibration is not a reliable decision tool.

Autonomous systems include automated storage and retrieval systems, autonomous mobile robots, robotic picking, automated sortation and dynamic control software. Their benefits may include improved consistency, reduced exposure to hazardous work, extended operating hours and higher information visibility. Risks include capital exposure, cyber vulnerability, maintenance dependence, reduced flexibility, workforce displacement and unsafe interaction between people and machines.

A responsible deployment uses staged pilots, geofencing, emergency stops, human override, fail-safe states, maintenance procedures and clear accountability. Automation should be evaluated on total cost of ownership, including integration, training, downtime, software licences, energy, cybersecurity and end-of-life costs—not only purchase price.

10. Illustrative case: Northstar Fulfilment Centre

Background. Northstar, an illustrative omnichannel distributor, operates a 20,000-square-metre fulfilment centre. It experiences seasonal order peaks, high travel distance and frequent late cut-offs. Management proposes an AI-enabled transformation.

Problem. A preliminary dashboard reports 92% on-time dispatch, but managers distrust the result because order-status definitions differ between the WMS and transport system. Pickers also report that fast-moving items are frequently stored in inconvenient locations.

Analysis. The project team first establishes a common event dictionary and reconciles inventory and order timestamps. A rolling forecast is created for hourly order lines and cartons, with separate models for promotional and ordinary periods. Forecast performance is evaluated against a seasonal-naïve baseline using rolling holdouts. A slotting model uses pick frequency, cube, weight, replenishment frequency and co-pick relationships. Labour planning converts forecast cartons into process workloads and includes absenteeism and effective productivity.

A routing model is tested in a calibrated simulation. It considers vehicle capacity, time windows and realistic travel-time distributions rather than average distance alone. An anomaly model flags likely inventory discrepancies, but no stock adjustment is posted automatically; a cycle-count task is created for supervisor approval.

Approach and solution. Northstar pilots the changes in one zone. High-frequency, light items are relocated, while heavy and awkward products remain within ergonomic limits. The labour recommendation is displayed with the assumptions behind it, and supervisors can override it with a reason code. A digital twin tests alternative wave schedules before implementation. The AI system predicts late orders and ranks interventions, while the optimisation layer checks dock, labour and carrier constraints.

Outcome. In the pilot, travel distance per order falls, late dispatch decreases and inventory exceptions are resolved earlier. However, the improvement is not attributed solely to AI: data cleansing, slotting discipline, supervisor involvement and process standardisation contribute materially. The pilot also identifies an unexpected risk: a promotion changes product mix rapidly, causing the original slotting scores to become obsolete. Northstar therefore introduces weekly monitoring and trigger-based re-slotting.

Learning points. Analytics programmes should begin with decision definitions and data governance. Predictive accuracy does not guarantee operational value. Human review is appropriate where consequences are material, data is incomplete or the model is unfamiliar with exceptional conditions. Benefits should be assessed against a controlled baseline and total implementation cost.

11. Governance, ethics and implementation

Analytics may affect worker productivity scores, schedules, job security, customer priority and access to service. Governance should address privacy, proportionality, fairness, security, transparency and accountability. Data collected for safety or process improvement should not automatically be repurposed for punitive surveillance without notice and legitimate justification.

Implementation is best managed as a cycle:

1. Define the business decision and success measures.
2. Map data sources, ownership, quality and access rights.
3. Establish a baseline and test assumptions.
4. Develop the simplest adequate model.
5. Validate technically and operationally.
6. Pilot with human oversight.
7. Monitor performance, drift, exceptions and unintended effects.
8. Scale only when benefits and controls are demonstrated.

Cybersecurity is part of operational continuity. Connected conveyors, robots, sensors and cloud systems expand the attack surface. Access control, network segmentation, secure updates, backups and incident-response plans are necessary. Contractual arrangements should define data ownership, model responsibilities, service levels and what happens when a vendor system fails.

Chapter Summary

Analytics converts warehouse and distribution data into decisions. Descriptive and diagnostic analytics establish what happened and why; predictive analytics estimates future demand, workload, risk or travel time; prescriptive analytics recommends actions under objectives and constraints. Forecasts must be evaluated out of sample, and operational measures must use consistent definitions.

EOQ, reorder points, capacity relationships, slotting scores, assignment models and routing models provide useful foundations, but each depends on assumptions. Optimisation requires explicit variables, objectives and constraints; simulation tests policies under variability. AI can enhance forecasting, anomaly detection, vision, maintenance and exception management, but prediction is not prescription. Digital twins and autonomous systems are valuable only when connected to reliable data, calibrated models and safe operating procedures.

Successful deployment combines technical validity with process redesign, human judgement, governance, cybersecurity and continuous monitoring. The central managerial principle is to optimise the complete service system rather than a single local metric.

Key Terms

Analytics: Systematic use of data, models and evidence to support decisions.

Descriptive analytics: Analysis that reports what has happened.

Diagnostic analytics: Analysis that investigates causes and contributing factors.

Predictive analytics: Estimation of future outcomes or probabilities.

Prescriptive analytics: Recommendation of actions subject to objectives and constraints.

Forecast error: Difference between observed and forecast demand.

Exponential smoothing: Forecasting method that weights recent observations more heavily.

EOQ: Benchmark order quantity balancing ordering and holding costs.

Reorder point: Inventory position at which replenishment is initiated.

Safety stock: Inventory held to protect against uncertainty.

Slotting: Assignment of SKUs to storage or picking locations.

Optimisation: Selection of the best feasible decision under a defined model.

Simulation: Representation of system behaviour to evaluate policies or scenarios.

Machine learning: Methods that learn patterns or relationships from data.

Data drift: Change in the data-generating environment that reduces model relevance.

Digital twin: A connected digital representation of a physical asset or process.

Autonomous system: Equipment or software capable of acting with limited direct intervention.

Model governance: Controls for ownership, validation, monitoring, use and accountability.

Review Questions

Short Answer Questions

1. Distinguish descriptive, predictive and prescriptive analytics with one warehouse example for each.
2. State the assumptions of the basic EOQ model.
3. Why must forecast accuracy be evaluated using out-of-sample data?
4. Identify four dimensions of data quality relevant to warehouse analytics.
5. Explain why prediction is not the same as prescription.

Long Answer Questions

1. Explain how analytics can support inventory, labour, capacity and routing decisions across a distribution centre.
2. Discuss the formulation of an optimisation problem, including decision variables, objectives and constraints.
3. Critically evaluate the use of AI and machine learning in warehouse operations.
4. Compare simulation and optimisation as tools for distribution-system design.
5. Explain the governance, ethical and cybersecurity requirements for autonomous warehouse systems.

Analytical / Application Questions

1. A warehouse forecasts 1,200 cartons for a shift. Effective packing capacity is 48 cartons per labour hour, and each worker is available for 6 productive hours. Determine the minimum number of workers required and discuss practical reasons why the result may need adjustment.
2. A manager proposes placing every high-frequency SKU nearest to dispatch. Analyse the proposal using frequency, cube, weight, replenishment, ergonomics and co-pick relationships.
3. Design a pilot for an AI model that predicts late orders. Specify the baseline, data, evaluation measures, human controls, drift-monitoring process and criteria for scaling.

Multiple Choice Questions with Answers

Question 1

Which analytics type recommends a staffing plan subject to labour and service constraints?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: D. Prescriptive analytics

Question 2

In the EOQ model, what does H represent?

A. Annual demand

B. Annual holding cost per unit

C. Ordering cost per order

D. Reorder-point service factor

Correct Answer: B. Annual holding cost per unit

Question 3

A moving average forecast gives which weighting to the observations included in its basic form?

A. Equal weighting

B. Weighting based only on item value

C. Increasing weights for older observations

D. Random weighting

Correct Answer: A. Equal weighting

Question 4

If required workload is 960 cartons, productivity is 40 cartons per labour hour, and each worker provides 6 productive hours, how many workers are required?

A. 2

B. 3

C. 4

D. 6

Correct Answer: C. 4

Question 5

Which statement best describes a digital twin?

A. A static warehouse layout drawing

B. A connected digital representation used to monitor or test a physical system

C. A spreadsheet containing only historical orders

D. A replacement for all physical safety controls

Correct Answer: B. A connected digital representation used to monitor or test a physical system

Question 6

Why can high warehouse utilisation create operational risk?

A. It eliminates all inventory records

B. It leaves little buffer to absorb variability and disruption

C. It always reduces order accuracy

D. It prevents the use of forecasting

Correct Answer: B. It leaves little buffer to absorb variability and disruption

Question 7

Which measure penalises large forecast errors more strongly than MAE?

A. RMSE

B. Fill rate

C. Utilisation

D. Pick rate

Correct Answer: A. RMSE

Question 8

What is a key reason to include human approval for an AI-generated inventory adjustment?

A. AI models never use data

B. The adjustment may have material consequences when data or model confidence is inadequate

C. Optimisation cannot use constraints

D. Human review guarantees perfect accuracy

Correct Answer: B. The adjustment may have material consequences when data or model confidence is inadequate

Question 9

In a routing problem, which item is a constraint rather than normally a decision variable?

A. Vehicle route sequence

B. Customer visit assignment

C. Vehicle capacity limit

D. Delivery departure time

Correct Answer: C. Vehicle capacity limit

Question 10

Why should slotting variables be normalised before combining them in a weighted score?

A. To make all SKUs identical

B. To prevent numerical scales from unfairly dominating the score

C. To remove all operational constraints

D. To guarantee the global optimum

Correct Answer: B. To prevent numerical scales from unfairly dominating the score