AI and Advanced Decision Support

Introduction and Context

Warehouse Management Systems (WMS) and Transportation Management Systems (TMS) have evolved from transaction-processing applications into decision platforms. A traditional WMS records receipts, locations, inventory movements and shipments; a TMS plans loads, selects carriers and monitors transport. Advanced systems increasingly combine these records with machine learning, mathematical optimization, real-time sensor data, computer vision and generative artificial intelligence (AI).

The managerial purpose is not to replace operational judgment with an algorithm. It is to improve the speed, consistency and quality of decisions while preserving accountability. AI is valuable when decisions are repetitive, data-rich and measurable. It is less reliable when conditions change radically, historical data are biased, or the cost of an error is high.

This chapter examines AI applications across demand and workload prediction, dynamic slotting, estimated time of arrival (ETA) forecasting, routing, anomaly detection, computer vision and conversational decision support. It also addresses explainability, bias, model validation, governance and human oversight.

Learning Objectives

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

1. Distinguish prediction, optimization, simulation and generative decision support.
2. Select suitable machine-learning methods for warehouse and transport problems.
3. Interpret forecasting and optimization metrics and their limitations.
4. Design applications for slotting, ETA prediction, routing, anomaly detection and computer vision.
5. Evaluate explainability, fairness, robustness, privacy and human-oversight requirements.
6. Propose a controlled implementation and validation plan for an AI-enabled WMS or TMS.

1. A Decision-Centric AI Architecture

An AI application should begin with a decision, not with a fashionable algorithm. The basic sequence is:

data → prediction → decision model → action → feedback

For example, a WMS may predict tomorrow’s order-line volume, optimize labour assignments subject to skill and shift constraints, execute the plan, and compare actual performance with the prediction. A TMS may predict arrival times, generate feasible routes, issue customer updates and learn from completed journeys.

Four analytical functions should be separated:

Prediction estimates uncertain future states. Optimization chooses an action against an objective and constraints. A highly accurate forecast does not automatically produce a good plan: the plan must reflect capacity, labour rules, service commitments and risk.

Data foundations

Useful data include orders, order lines, stock movements, location dimensions, travel paths, labour rosters, equipment status, carrier events, GPS traces, appointment times, weather and calendar variables. Data engineering commonly requires:

• a common event time standard and time zone;
• stable identifiers for orders, pallets, locations, vehicles and workers;
• treatment of missing, duplicated and delayed events;
• versioned master data and historical layouts;
• controls against data leakage, where future information accidentally enters training data;
• documented data ownership, retention and access rights.

The unit of analysis must match the decision. A daily facility workload forecast may be appropriate for labour planning, whereas wave release may require hourly or even minute-level predictions.

2. Demand and Workload Prediction

Demand forecasting estimates future customer demand. Workload forecasting translates demand into operational effort. The distinction is important because 1,000 order lines do not imply the same workload when the lines differ in unit quantity, travel distance, handling requirements or value-added services.

A simple workload model is:

Here, W_t is workload in time period t (for example, labour-minutes), qᵢ,t is the quantity of activity type i, sᵢ is standard time per unit of that activity, mⱼ,t is the number of tasks of type j, and rⱼ is standard time per task. The equation assumes reasonably stable standard times. Travel congestion, fatigue and interruptions may require a learned correction factor.

Useful methods include:

• time-series models, such as exponential smoothing, for stable seasonal patterns;
• regression and gradient-boosting models, when promotions, holidays, weather or price variables matter;
• count models, when the target is a non-negative number of orders or arrivals;
• probabilistic forecasts, which provide prediction intervals rather than a single estimate;
• hierarchical forecasting, which reconciles forecasts across SKU, customer, facility and network levels.

For intermittent demand, ordinary accuracy measures may be misleading because many periods have zero demand. Croston-type methods, intermittent-demand models and classification-plus-quantity approaches can be considered, but performance must be assessed against a simple operational baseline.

Forecast evaluation

Mean absolute error is:

where yₜ is actual demand, ŷₜ is forecast demand and n is the number of evaluation periods. MAE has the same units as the target and is easy to interpret.

Root mean squared error is:

RMSE penalizes large errors more heavily than MAE. Percentage measures such as MAPE can become unstable or undefined when actual demand is zero. Weighted measures should be used carefully: weighting high-volume SKUs may hide poor performance on critical low-volume items.

A forecast should be evaluated using rolling, time-ordered validation. Randomly mixing past and future observations can produce unrealistically optimistic results. Evaluation should also examine bias, peak-period performance, service-level effects and forecast intervals.

3. Dynamic Slotting and Warehouse Optimization

Slotting assigns products to storage locations. Static slotting often uses historical velocity, cube and compatibility rules. Dynamic slotting updates recommendations as demand, product dimensions, order profiles and congestion change.

A practical slotting objective may minimize expected handling effort:

subject to:

where xᵢⱼ equals 1 if item i is assigned to location j, cᵢⱼ is expected cost or travel effort, vᵢ is required volume, and Cⱼ is location capacity. Additional constraints may enforce weight limits, temperature zones, hazardous-material separation, compatibility, replenishment access and ergonomic restrictions. This is a simplified assignment formulation; real slotting may be mixed-integer and dynamic.

Machine learning can estimate cᵢⱼ from observed travel time, congestion, co-pick relationships and replenishment activity. Optimization then uses those estimates to generate feasible assignments. A reinforcement-learning approach may learn policies for sequential placement or replenishment, but it requires a safe simulator or tightly controlled exploration because poor experimentation can disrupt operations.

Dynamic slotting should not move items whenever a small theoretical improvement appears. Relocation has a cost, creates execution work and can introduce inventory-control errors. A useful policy includes a minimum improvement threshold, a relocation budget and an approval process for exceptions.

4. ETA Forecasting and Transport Visibility

ETA forecasting estimates arrival time using planned route, actual vehicle position, stop history, appointment windows, traffic, weather, driver-hours constraints and facility dwell time. The prediction target must be defined precisely: arrival at the gate, arrival at the dock or completion of unloading are different events.

A point forecast alone is insufficient for appointment management. A prediction interval or probability of on-time arrival is more actionable. For example, a TMS might report an ETA of 14:20, a 90% interval of 14:05–14:45, and a 0.72 probability of meeting a 14:30 appointment.

Models may include gradient boosting, recurrent or transformer-based sequence models, survival models for time-to-event prediction and hybrid systems combining routing logic with learned delay estimates. Models should distinguish recurring lane patterns from unusual disruptions. A model trained on normal traffic may fail during a strike, flood or new road restriction.

ETA performance can be assessed with MAE in minutes, but calibration is also essential. If a system labels 80% of cases as having at least 80% on-time probability, approximately 80% should actually be on time over a sufficiently large and representative sample. Calibration should be checked by lane, carrier, region, time of day and disruption condition.

5. Routing, Scheduling and Network Decisions

Routing is usually an optimization problem, not merely a prediction problem. In a capacitated vehicle-routing problem, a solution must assign customers to vehicles and sequence visits while satisfying capacity, time windows, depot rules and sometimes driver regulations.

A generic objective can combine distance, fixed vehicle cost, lateness and emissions:

where D is distance or travel cost, L is lateness, V is vehicle-use cost, E is emissions, and α, β, γ, δ are business weights with compatible monetary or normalized units. The weights must be documented because changing them changes the meaning of “best.”

Exact mixed-integer optimization can provide high-quality solutions for smaller instances but may be computationally expensive. Heuristics and metaheuristics—such as insertion methods, local search, tabu search and adaptive large-neighbourhood search—often produce good solutions quickly. A mature TMS may use optimization for the initial plan and re-optimization when new orders, cancellations or disruptions occur.

Operational feasibility is more important than a mathematically attractive but unexecutable route. The system must represent loading compatibility, vehicle access restrictions, legal driving time, service duration, appointment tolerance and driver or customer preferences. Human planners should be able to lock mandatory stops and inspect trade-offs.

6. Anomaly Detection and Computer Vision

An anomaly is an observation or sequence inconsistent with expected behaviour. Examples include unusual inventory adjustments, repeated short picks, unexpected temperature excursions, excessive dwell, suspicious route deviation or a scan pattern inconsistent with a normal process.

Supervised classification is suitable when labelled examples of fraud, damage or error are available. Unsupervised or semi-supervised methods are useful when incidents are rare. Isolation forests, density methods, autoencoders and change-point detection can identify unusual records, but an anomaly score is not proof of wrongdoing. It is a prioritization signal for investigation.

A simple standardized residual is:

where xₜ is the observed measure, μₜ is the expected value and σₜ is the expected standard deviation. This approach is useful only when the expected distribution and variance are reasonably stable. A threshold such as |z| > 3 is not universally valid, especially with autocorrelation, skewness or changing variance.

Computer vision can support barcode and label reading, dimensioning, pallet-count verification, damage detection, PPE compliance and parcel sortation. The operating environment creates difficult conditions: glare, occlusion, damaged labels, camera movement, diverse packaging and changing illumination. Performance must be tested on the actual site, not only on curated images.

False positives consume supervisor attention; false negatives allow defects to pass. Threshold selection should therefore reflect the relative cost of both errors. High-risk detections should trigger human review rather than automatic disciplinary action.

7. Generative AI as Decision Support

Generative AI can summarize exceptions, translate operational language into system queries, draft customer communications, retrieve procedures and explain the factors behind a recommendation. Retrieval-augmented generation can ground responses in approved policies, inventory records and operating manuals rather than relying solely on model memory.

However, a language model may generate fluent but false statements, confuse units, expose confidential data or call an unsafe system action. It should not be treated as an authoritative database. A controlled architecture should include:

1. identity and role-based access;
2. retrieval from governed sources;
3. citations or links to source records;
4. validation of units, dates and identifiers;
5. tool permissions restricted to approved actions;
6. confirmation before irreversible changes;
7. logging of prompts, retrieved data, responses and user approvals.

A suitable interface might answer: “Which outbound orders are at risk of missing today’s cut-off, and why?” The answer should show order identifiers, predicted completion time, uncertainty, relevant constraints and recommended interventions. It should not merely produce a confident narrative.

8. Explainability, Bias and Human Oversight

Explainability has several meanings. A global explanation describes overall model behaviour; a local explanation describes one prediction; a procedural explanation states how a recommendation was generated; and a counterfactual explanation states what change could alter the result. Feature importance is not necessarily causation. For example, a carrier identifier may correlate with lateness because of lane mix, not because the carrier is intrinsically poor.

Bias may enter through historical decisions, missing data, unequal measurement, proxy variables or unequal exposure to monitoring. A labour model trained on historical productivity may reproduce biased work assignments. A delivery model may systematically underpredict delays in rural areas if data are sparse. Fairness assessment should be relevant to the decision, not reduced to a single universal statistic.

Human oversight should be risk-based:

Oversight is meaningful only when the human has time, information, authority and training to challenge the model. “Human in the loop” is ineffective if operators merely rubber-stamp recommendations.

9. Validation, Monitoring and Model Risk

A production model requires more than a high test score. Validation should cover:

• temporal validity: performance on future periods;
• operational validity: impact on throughput, service, cost and safety;
• stress validity: disruptions, peaks, new products and layout changes;
• subgroup validity: facilities, lanes, carriers, shifts and customer types;
• calibration: whether probabilities correspond to observed frequencies;
• robustness: sensitivity to missing, delayed or manipulated inputs;
• economic value: benefit after implementation and exception-handling cost.

A/B testing is appropriate when interventions can be safely randomized. Otherwise, stepped-wedge rollout, matched comparison sites or interrupted time-series analysis may be used. Before-and-after comparisons alone are weak because demand, staffing and network conditions may also change.

Monitoring should include data drift, concept drift, forecast error, constraint violations, override rates, latency, cost and adverse incidents. Retraining should not be automatic without review: a new model can be statistically better but operationally worse.

10. Illustrative Case: AI-Assisted Fulfilment and Transport Control Tower

Background

A fictional omnichannel distributor operates a 30,000-square-metre facility and ships to retail stores and direct customers. It experiences Monday labour shortages, congested fast-pick locations and unreliable ETA updates during afternoon traffic.

Problem

The existing system forecasts daily orders but does not forecast order-line workload by hour. Slotting is reviewed monthly. ETA is based on planned transit time, and planners manually search for delayed loads.

Analysis

The project team identifies three linked decisions: staffing by hour, relocation of high-impact items, and prioritization of at-risk deliveries. It finds that order volume alone explains only part of labour time; lines per order, travel distance, replenishment tasks and value-added services are also important.

Approach and solution

A time-ordered gradient-boosting model predicts hourly order lines and labour minutes, with prediction intervals. A constrained optimizer recommends slot changes using travel cost, cube, compatibility and a weekly relocation limit. An ETA model combines GPS, stop sequence, historical dwell and appointment data. A control-tower screen ranks exceptions by expected service impact and displays the main contributing factors. Generative AI summarizes approved exceptions, but planners authorize route changes.

The pilot uses two comparable operating zones for eight weeks. The evaluation measures labour-plan error, picks per labour-hour, travel distance, on-time delivery, planner overrides and safety incidents. The system is not judged only by forecast accuracy.

Outcome

In this illustrative case, the pilot reduces peak staffing error and unnecessary slot moves while improving the timeliness of exception escalation. Some ETA benefits disappear during an atypical road closure, demonstrating the need for disruption flags and human intervention. Planners reject several mathematically efficient routes because of customer-specific unloading constraints that were missing from master data.

Learning points

1. Forecasts must be connected to decisions and measurable outcomes.
2. Constraints and master data determine whether optimization is executable.
3. Exception ranking is often more valuable than attempting full automation.
4. Pilot designs must include unusual conditions and operational overrides.
5. User trust improves when recommendations show evidence and uncertainty.

11. Implementation Roadmap and Best Practices

A responsible implementation can proceed through the following stages:

1. Frame the decision: define the owner, action, objective, constraints and acceptable risk.
2. Establish a baseline: compare against current practice and simple statistical rules.
3. Audit data: document quality, lineage, missingness, access and leakage risks.
4. Build a small pilot: select a bounded site, lane, process or decision.
5. Validate prospectively: use time-based testing and operational KPIs.
6. Design the interface: present uncertainty, reasons, alternatives and override controls.
7. Deploy with safeguards: use permissions, logging, rollback and incident response.
8. Monitor continuously: review drift, subgroup performance, overrides and realized value.
9. Revalidate after change: reassess the model when layouts, carriers, products or policies change.

Senior managers should assign clear accountability among operations, IT, data science, security, compliance and frontline users. Procurement contracts should address data ownership, model updates, audit rights, service levels, cybersecurity and exit arrangements.

Challenges and Limitations

AI systems face changing demand, sparse disruption data, inconsistent event capture, process workarounds and incompatibility between legacy systems. Optimization may produce unstable plans when small input changes create large route or slot differences. Forecast accuracy can improve while business performance declines if users ignore the output or if the objective function rewards the wrong behaviour.

Privacy and employment concerns are significant where worker scans, location data or productivity measures are used. Data minimization, purpose limitation, access controls and retention rules are essential. Safety-critical and employment-related decisions require especially strong evidence, transparency and appeal mechanisms. Cybersecurity must protect models, APIs, operational data and connected equipment from manipulation.

Chapter Summary

AI-enabled WMS and TMS capabilities should be designed around operational decisions rather than isolated algorithms. Machine learning can forecast demand, workload and ETA; optimization can convert those forecasts into feasible slotting, labour and routing plans; anomaly detection and computer vision can focus attention on unusual or defective events; and generative AI can improve access to governed information and explanations. Each capability has distinct data, validation and risk requirements. Reliable deployment requires time-ordered testing, calibration, constraint modelling, subgroup assessment, monitoring and meaningful human oversight. The central managerial principle is controlled augmentation: automate routine, low-risk work while preserving accountable human judgment for uncertain, consequential or irreversible decisions.

Key Terms

Anomaly detection: Identification of observations or event sequences that differ materially from expected behaviour.

Calibration: Agreement between predicted probabilities and observed frequencies.

Computer vision: Use of image or video analysis to identify objects, conditions or actions.

Concept drift: Change in the relationship between inputs and the target being predicted.

Data leakage: Accidental use of information unavailable at the time a prediction would have been made.

Dynamic slotting: Repeatedly updating storage assignments in response to changing operating conditions.

ETA forecasting: Estimation of the time at which a vehicle or shipment will reach a defined event point.

Explainability: Ability to communicate how a model or recommendation operates or was produced.

Generative AI: AI that creates text, code, images or other content from learned patterns and prompts.

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

Machine learning: Methods that infer patterns from data to support prediction or classification.

Model drift: Decline in model performance caused by changes in data or operating conditions.

Optimization: Selection of the best feasible action according to a stated objective and constraints.

Prediction interval: Range intended to contain a future outcome with a specified probability.

Prescriptive analytics: Analysis that recommends actions rather than only describing or forecasting events.

Retrieval-augmented generation: Generating responses using information retrieved from approved external sources.

Route optimization: Selection of vehicle assignments and visit sequences subject to operational constraints.

Workload forecast: Estimate of future operational effort, such as labour-minutes or task volume.

Review Questions

Short Answer Questions

1. Distinguish predictive analytics from prescriptive analytics in a WMS.
2. Why can order volume be an inadequate measure of warehouse workload?
3. What is the purpose of time-ordered validation?
4. Explain why an anomaly score is not proof of misconduct.
5. State two safeguards for generative AI used in a TMS.

Long Answer Questions

1. Explain how machine learning and optimization can be integrated for dynamic slotting.
2. Compare MAE, RMSE and percentage-based forecast measures, including their limitations.
3. Discuss the design of an ETA forecasting system for a multi-carrier network.
4. Evaluate the role of explainability and human oversight in AI-enabled warehouse decisions.
5. Develop a governance framework for monitoring model drift, bias and operational risk.

Analytical / Application Questions

1. A facility has accurate daily demand forecasts but repeated hourly labour shortages. Diagnose the likely causes and propose a workload-prediction and validation design.
2. A routing model reduces distance but increases missed appointments. Formulate the managerial and optimization changes required to correct the decision objective.
3. A computer-vision system flags damage more often on one packaging type. Design an investigation to distinguish genuine damage from measurement or sampling bias.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from optimization?

A. Prediction selects the best feasible action, while optimization estimates uncertainty.

B. Prediction estimates a future state, while optimization selects an action subject to objectives and constraints.

C. Prediction is always more accurate than optimization.

D. Optimization can be used only when historical data are unavailable.

Correct Answer: B. Prediction estimates a future state, while optimization selects an action subject to objectives and constraints.

Question 2

Which forecast metric has the same units as the forecast target?

A. Mean absolute error

B. Mean absolute percentage error only

C. A dimensionless calibration ratio

D. A standardized residual

Correct Answer: A. Mean absolute error

Question 3

Why is random train-test splitting risky for time-dependent warehouse data?

A. It always increases computational cost.

B. It prevents the use of categorical variables.

C. It can allow information from future periods to influence training.

D. It makes all forecasts probabilistic.

Correct Answer: C. It can allow information from future periods to influence training.

Question 4

In a slotting assignment, what does a capacity constraint ensure?

A. Every item is assigned to every location.

B. Assigned item volume does not exceed location capacity.

C. All locations have identical travel distance.

D. Demand is always forecast without error.

Correct Answer: B. Assigned item volume does not exceed location capacity.

Question 5

What does calibration of an ETA probability mean?

A. Predicted probabilities correspond reasonably to observed frequencies.

B. Every ETA prediction is exactly correct.

C. The model uses only GPS data.

D. The route has minimum distance.

Correct Answer: A. Predicted probabilities correspond reasonably to observed frequencies.

Question 6

Which is an appropriate use of an anomaly-detection output?

A. Automatic dismissal of an employee without review

B. Proof that a carrier committed fraud

C. Prioritization of records for investigation

D. Replacement of all operational controls

Correct Answer: C. Prioritization of records for investigation

Question 7

Which safeguard is most appropriate for generative AI that can alter transport plans?

A. Allow unrestricted system access to improve creativity.

B. Require approved tools, role permissions and confirmation before irreversible actions.

C. Delete all source records after generating a response.

D. Use generated text as the sole source of truth.

Correct Answer: B. Require approved tools, role permissions and confirmation before irreversible actions.

Question 8

What is concept drift?

A. A change in the relationship between inputs and the prediction target.

B. A permanent improvement in data quality.

C. A method for assigning warehouse locations.

D. A type of barcode failure only.

Correct Answer: A. A change in the relationship between inputs and the prediction target.

Question 9

Why should slotting systems include relocation limits?

A. To prevent all use of optimization.

B. Because every theoretical improvement justifies unlimited movement.

C. Because relocation creates handling cost and operational disruption.

D. To ensure that demand forecasts are always zero.

Correct Answer: C. Because relocation creates handling cost and operational disruption.

Question 10

Which principle best describes meaningful human oversight?

A. The human automatically accepts every recommendation.

B. The human has information, authority and time to review and challenge outputs.

C. The model is hidden from the human to avoid bias.

D. Only the software vendor can override decisions.

Correct Answer: B. The human has information, authority and time to review and challenge outputs.