AI, Robotics and the Future of Warehousing

Introduction

Warehouses are changing from passive storage locations into digitally coordinated, semi-autonomous systems. Sensors, warehouse-management systems (WMS), robotics, artificial intelligence (AI), computer vision, cloud platforms and real-time analytics are increasingly combined to improve inventory accuracy, storage utilisation, order fulfilment and worker safety.

The purpose of these technologies is not simply to replace people. Their primary managerial purpose is to improve decisions and coordinate physical work. A modern warehouse may use AI to forecast demand, a slotting engine to assign storage locations, cameras to detect damaged cartons, predictive-maintenance software to identify failing equipment, and autonomous mobile robots (AMRs) to transport goods. These capabilities can create major benefits, but only when supported by reliable data, suitable processes, trained employees and responsible governance.

This chapter examines the principal technologies shaping intelligent warehousing. It also evaluates their limitations, ethical implications, workforce effects and implementation requirements. The central argument is that a smart warehouse is not defined by the quantity of automation installed; it is defined by the quality of decisions, integration and control achieved.

Learning Objectives

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

1. Explain how AI supports warehouse forecasting, slotting and operational decision-making.
2. Describe applications of computer vision, predictive maintenance, digital twins and AMRs.
3. Distinguish automation, autonomy, analytics and AI in warehouse contexts.
4. Assess data-quality, cybersecurity, privacy, safety and ethical risks.
5. Evaluate workforce and managerial implications of intelligent warehousing.
6. Develop a practical framework for responsible technology adoption.

1. From Automated Warehouses to Smart Warehouses

Traditional mechanisation uses equipment to perform predefined physical actions. Examples include conveyors, pallet trucks, automated storage and retrieval systems (AS/RS), sorters and fixed robotic arms. Automation generally follows explicit rules: a sensor detects a carton, a controller starts a conveyor, and a programmed route moves the carton to a destination.

AI adds the ability to identify patterns, estimate uncertain conditions, classify objects, recommend actions or adapt decisions from data. However, AI does not automatically imply independence. A forecasting model may recommend replenishment while a manager retains approval authority. Similarly, an AMR may navigate independently within defined safety rules but remain subject to human supervision.

A useful distinction is:

A smart warehouse normally contains five interacting layers:

1. Physical layer: racks, goods, conveyors, robots, scanners and workers.
2. Sensing layer: barcode readers, RFID, cameras, weight sensors, temperature sensors and equipment telemetry.
3. Data layer: master data, transaction histories, event streams and external information.
4. Decision layer: forecasting, optimisation, exception management and prioritisation.
5. Execution and governance layer: WMS, warehouse-control systems, policies, access controls, audit trails and human oversight.

Failure at any layer reduces performance. A sophisticated algorithm cannot compensate for incorrectly labelled locations or missing stock transactions.

2. AI-Supported Forecasting and Replenishment

Forecasting estimates future demand using historical observations, business knowledge and relevant external variables. In warehousing, forecasts support purchasing, replenishment, labour planning, space allocation and transport scheduling.

A simple time-series model can be written as:

Here, F_t+1 is the forecast for the next period, D_t is actual demand in the current period, F_t is the previous forecast, and α is the smoothing constant, with 0 ≤ α ≤ 1. The method assumes that recent demand contains useful information about near-future demand and is most appropriate when demand has no strong trend or seasonal pattern.

AI-supported forecasting may use additional variables such as promotions, weather, price, product substitutions, holidays, customer segments and supplier lead-time data. Machine-learning methods can model nonlinear relationships, but they require sufficient, representative and accurately labelled data.

Forecast accuracy should be evaluated rather than assumed. Common measures include mean absolute error (MAE):

where n is the number of observations, D_t is actual demand and F_t is forecast demand. MAE is expressed in the same units as demand, such as units or cartons. It treats all errors linearly and is easier to interpret than squared-error measures.

Mean absolute percentage error is:

It is applicable only when actual demand values are non-zero and is problematic when demand is intermittent or close to zero. For slow-moving warehouse items, MAE, weighted measures, service-level outcomes and inventory-cost measures may be more informative.

Forecasting does not directly determine an order quantity. A basic reorder-point model is:

where ROP is the reorder point in units, μ_L is expected demand during lead time, and SS is safety stock. Under a simplified independent-demand assumption, safety stock may be estimated as:

where z is the standard-normal service-factor corresponding to the desired cycle-service probability and σ_L is the standard deviation of demand during lead time. This formula is not universally valid: lead-time variability, demand intermittence, seasonality, dependence between demand and lead time, and service-cost trade-offs may require more advanced models.

AI can improve forecasting by identifying demand changes earlier, but it can also reproduce historical bias. If past stockouts caused recorded sales to be artificially low, a model trained on sales alone may underforecast true demand. Human review and causal interpretation therefore remain important.

3. Intelligent Slotting and Order-Picking Decisions

Slotting is the assignment of products to storage locations. Its objectives may include reducing travel distance, improving cube utilisation, balancing workload, protecting products and meeting ergonomic or safety requirements.

A conventional slotting policy often places fast-moving items near dispatch or at convenient picking heights. AI-supported slotting can consider order correlations, time-of-day demand, product dimensions, replenishment frequency, congestion, worker travel, temperature zones and changing customer priorities.

A simplified decision objective can be represented as:

where Z is the total evaluated cost, D is travel distance, H is handling effort, R is replenishment effort, S is a safety or service penalty, and c_d, c_h, c_r, and c_s are management-chosen weighting coefficients. This is a conceptual multi-objective model, not a universal formula. The units of the terms must be made comparable, usually by converting them into time, cost or weighted penalty points.

A practical slotting engine should respect constraints such as:

• weight limits and rack capacity;
• compatible storage conditions;
• hazardous-material segregation;
• product dimensions and packaging stability;
• expiry-date and batch-control requirements;
• accessibility for equipment and workers;
• fire, emergency and occupational-safety rules;
• minimum presentation stock and replenishment limits.

A mathematically efficient location may be operationally unsafe. For example, placing a heavy item in a high-level location may reduce travel but increase injury risk. Therefore, safety constraints must be hard constraints, not optional preferences.

Dynamic slotting is especially useful in e-commerce and seasonal operations. Nevertheless, frequent relocation can generate confusion, replenishment work and transaction errors. The cost of changing a location should be included in the evaluation.

4. Computer Vision in Warehousing

Computer vision uses cameras and image-processing or machine-learning models to interpret visual information. Applications include:

• barcode and label reading;
• pallet and carton dimension measurement;
• detection of damaged packaging;
• verification of quantity and product identity;
• monitoring of blocked aisles or unsafe conditions;
• checking personal protective equipment;
• counting inventory and identifying empty locations;
• analysing vehicle–pedestrian interactions.

Vision systems do not observe reality perfectly. Performance can decline because of poor lighting, glare, occlusion, dust, unusual packaging, camera movement or changes in labels. A model trained on one product range may perform poorly on another. Performance should therefore be tested by relevant operating condition, not only by an overall accuracy percentage.

Important evaluation measures include precision, recall and false-alarm rates. In safety applications, a false negative—failing to detect a dangerous condition—may be more serious than a false positive. Thresholds should reflect risk, and uncertain cases should be referred to a person or a secondary control.

Cameras can also create a surveillance environment. Organisations should define the purpose of data collection, restrict access, establish retention periods and communicate policies clearly. Video collected for stock verification should not automatically be repurposed for continuous worker-performance scoring.

5. Predictive Maintenance

Corrective maintenance repairs equipment after failure. Preventive maintenance services equipment at scheduled intervals. Predictive maintenance uses condition data to estimate deterioration and support intervention before failure.

Sources of data include vibration, temperature, motor current, error codes, operating hours, battery condition and maintenance history. A system may identify an abnormal pattern in a conveyor motor and schedule inspection during a planned shutdown.

For a binary failure-prediction task, a confusion matrix distinguishes true positives, false positives, true negatives and false negatives. A useful maintenance measure is precision:

and recall:

where TP is correctly predicted failure, FP is an unnecessary failure alert, and FN is a missed failure. High recall may produce many inspections; high precision may miss developing faults. The appropriate balance depends on equipment criticality, inspection cost, safety consequences and spare-part availability.

Predictive maintenance is valuable only when the organisation can act on alerts. A model that predicts failure but cannot obtain parts, allocate technicians or stop equipment safely does not create reliable availability. Maintenance data should be linked to asset identity, operating context and confirmed failure outcomes.

6. Digital Twins and Operational Simulation

A digital twin is a digital representation of a physical asset, process or facility that is linked to relevant data and updated over time. It is more than a static three-dimensional model. A warehouse digital twin may represent storage locations, inventory, equipment states, orders, routes, labour capacity and environmental conditions.

Uses include:

• testing layout changes before construction;
• evaluating robot fleet size;
• simulating peak-season order volumes;
• identifying bottlenecks;
• testing emergency or equipment-failure scenarios;
• comparing alternative picking strategies;
• training supervisors and operators.

A digital twin is only as credible as its assumptions and data links. If travel times, process rules or equipment capacities are inaccurate, simulation results can mislead decision-makers. Models should be validated against observed performance, documented with assumptions and periodically recalibrated.

Digital twins support what-if analysis but do not eliminate uncertainty. They should be treated as decision-support tools rather than exact forecasts of the future.

7. Autonomous Mobile Robots and Collaborative Automation

AMRs use sensors, maps, software and onboard control systems to navigate without fixed tracks. They may transport totes, bring shelves to workers, perform inventory scans or move completed orders. Unlike many automated guided vehicles, which often follow fixed markers or paths, AMRs can calculate routes around obstacles within a controlled operating environment.

A typical AMR system includes:

1. a fleet manager that assigns missions;
2. a navigation and localisation system;
3. obstacle-detection and safety sensors;
4. a battery and charging system;
5. interfaces to the WMS or warehouse-control system;
6. human–robot operating rules.

AMRs can reduce walking and improve throughput, but they do not remove the need for process discipline. Incorrect item master data, unavailable inventory, blocked locations or poor task sequencing can cause robot congestion. Battery charging, network availability, traffic rules and manual recovery procedures must be designed from the beginning.

Collaborative robots, or cobots, are designed to work in proximity to people, subject to appropriate risk assessment and safeguards. “Collaborative” does not mean inherently safe in every application. Payload, speed, tooling, impact energy, reach and foreseeable misuse must be assessed. Safety systems must comply with applicable national requirements and recognised machinery-safety practices.

8. Data Quality, Integration and Cybersecurity

Data is the foundation of intelligent warehousing. Critical data domains include item identification, dimensions, weight, units of measure, location master data, supplier lead time, inventory status, order history, equipment identity and employee or contractor access rights.

Data quality has several dimensions:

Integration commonly involves the enterprise resource-planning system, WMS, warehouse-control system, transportation systems, IoT platforms and robotics software. Interfaces should define ownership, event timing, error handling, reconciliation and fallback procedures.

Cybersecurity risks include compromised credentials, ransomware, manipulation of inventory records, unauthorised robot commands, sensor spoofing and interruption of operational technology. Controls should include least-privilege access, network segmentation, multifactor authentication where practical, patch management, backups, monitoring, incident response and tested manual operating procedures.

A key principle is graceful degradation. If the AI service or network fails, the warehouse should enter a safe and controlled mode rather than stop unpredictably or issue unsafe commands.

9. Ethics, Privacy and Responsible Governance

Responsible warehouse technology requires more than legal compliance. It requires consideration of fairness, dignity, accountability and safety.

Privacy and surveillance

Worker location, productivity, biometric information and video may constitute personal data under applicable law. Organisations should establish a legitimate purpose, collect only necessary data, limit retention, secure access and provide appropriate notice. Biometric identification should receive especially careful scrutiny because biometric characteristics cannot easily be changed after compromise.

Fairness and bias

Algorithms may allocate undesirable shifts, performance ratings or disciplinary attention unevenly if training data reflects historical bias. Models should be tested across relevant worker groups and job conditions. Decisions affecting employment should normally include human review, an explanation and an appeal route.

Accountability and explainability

A manager must be able to identify who approved a model, what data it used, which version produced a recommendation and who is responsible for the final action. High-impact decisions should not be treated as unquestionable merely because they are generated by software.

Safety and human control

Safety-related systems require conservative design, validation, emergency stops, clear responsibilities and regular drills. Human override should be meaningful, not merely a button that is technically present but operationally inaccessible.

Environmental responsibility

Automation can reduce travel, damage and energy consumption, but robots, servers, batteries and sensors have embodied environmental impacts. Evaluation should consider lifecycle energy, repairability, battery disposal, equipment reuse and electronic waste.

A practical governance process includes an approved use case, risk classification, data-impact assessment, safety assessment, cybersecurity review, pilot criteria, performance monitoring, incident reporting and retirement procedures.

10. Workforce and Management Implications

Technology changes tasks before it changes jobs. Repetitive walking, scanning, lifting and data entry may decline, while exception handling, equipment supervision, data stewardship, troubleshooting and coordination become more important.

Managers should conduct a task-based impact assessment:

1. identify tasks affected;
2. classify them as eliminated, assisted, redesigned or newly created;
3. assess safety, skill and workload consequences;
4. consult employees and representatives;
5. provide training and transition support;
6. monitor actual outcomes rather than relying on implementation assumptions.

Relevant skills include WMS operation, data interpretation, robot recovery, basic cybersecurity, process mapping, equipment safety and ethical decision-making. A warehouse diploma graduate does not need to become a machine-learning engineer, but should understand model limitations, data dependencies and operational controls.

Performance measurement should avoid creating harmful incentives. If workers are measured only by picks per hour, they may skip quality checks or safety practices. A balanced scorecard may include service level, inventory accuracy, order quality, safety incidents, equipment availability, worker wellbeing and energy use.

11. A Practical Adoption Framework

A disciplined adoption sequence is preferable to purchasing technology first and searching for a problem later.

Step 1: Define the operational problem

Specify the service failure, cost, risk or capacity constraint. “We need AI” is not a problem statement; “peak-season picking capacity is insufficient while error rates exceed target” is more useful.

Step 2: Establish the baseline

Measure current travel, throughput, inventory accuracy, stockouts, labour hours, equipment downtime, safety incidents and exception rates. Without a baseline, benefits cannot be demonstrated.

Step 3: Check process and data readiness

Standardise item, location and transaction data. Remove unnecessary process variation. Determine whether the proposed technology can access reliable, timely data.

Step 4: Select the least complex effective intervention

A better location master, barcode discipline or revised replenishment rule may solve a problem more cheaply than an AI platform. Technology complexity should be justified by operational value.

Step 5: Pilot safely

Use a representative area, defined success measures, worker involvement and a rollback plan. Do not pilot only under ideal conditions.

Step 6: Validate outcomes

Compare performance with the baseline and, where possible, a control area. Examine unintended effects such as increased replenishment work, worker stress or upstream errors.

Step 7: Scale with governance

Document ownership, training, cybersecurity, maintenance, model monitoring, change control and audit requirements before expansion.

Step 8: Review continuously

Models drift as products, customers, layouts and processes change. Revalidation and periodic review are therefore essential.

12. Illustrative Case: Northbridge Distribution Centre

Background

Northbridge Distribution Centre is an illustrative case involving a regional wholesaler with 18,000 active stock-keeping units. It experienced rising e-commerce orders, long picker travel distances, frequent stock discrepancies and unplanned conveyor stoppages.

Problem

Management proposed purchasing a large autonomous system immediately. Initial investigation found that 11% of item records had missing dimensions, several locations had duplicate identifiers, and maintenance work orders did not consistently record confirmed failure causes. These data weaknesses could have caused an automated system to make poor decisions.

Analysis

The project team established a baseline for four weeks. It found that fast-moving products were distributed across several zones, replenishment tasks conflicted with picking during peak periods, and conveyor motors with repeated overheating had no common asset identifiers. Worker consultation also revealed that existing productivity reports were viewed as punitive and did not distinguish system delays from individual performance.

Approach and solution

Northbridge first corrected item and location master data, introduced barcode confirmation at receiving, and created a common asset register. It then implemented a limited forecasting and slotting pilot for 1,200 high-volume items. The pilot considered demand, product dimensions, replenishment frequency and ergonomic restrictions.

Computer vision was used at dispatch to detect missing labels and visibly damaged cartons, with human review for uncertain cases. Sensors on critical conveyor motors generated maintenance alerts, but technicians retained authority to confirm the fault. Two AMRs were introduced for tote transport in a defined zone rather than across the whole facility.

The company established a governance committee including operations, IT, safety, HR and worker representatives. It prohibited the use of camera data for individual performance scoring without separate consultation and approval. Model performance, false alerts, safety events and employee feedback were reviewed monthly.

Outcome

The pilot reduced unnecessary picker travel and improved label detection, while maintenance staff reported earlier awareness of overheating motors. However, the project also exposed a problem: frequent dynamic slot changes increased replenishment confusion. The company therefore reduced the frequency of changes and added clearer location communication. The AMRs improved transport capacity in the pilot zone but required a manual recovery procedure for obstructed routes.

Learning points

1. Data correction preceded advanced automation.
2. A targeted pilot produced more useful evidence than a full-site launch.
3. Human review remained necessary for uncertain or safety-critical decisions.
4. Worker consultation identified operational and trust risks not visible in system dashboards.
5. Technology benefits depended on process redesign, governance and maintenance capability.

Challenges and Limitations

Intelligent warehousing has several structural limitations:

• Data drift: product ranges, demand patterns and layouts change.
• Automation bias: users may accept an algorithmic recommendation without adequate review.
• Integration cost: legacy systems may lack suitable interfaces.
• Vendor dependence: proprietary platforms can create switching and continuity risks.
• Uncertain return on investment: benefits may be offset by integration, training and maintenance costs.
• Exceptional conditions: fires, outages, labour shortages and unusual demand can invalidate normal operating assumptions.
• Cyber-physical risk: a digital attack may produce physical consequences.
• Human factors: poorly designed automation can increase cognitive load and reduce trust.
• Accessibility and inclusion: interfaces and work redesign must accommodate diverse abilities.

A sound business case should include total cost of ownership: equipment, software, integration, training, cybersecurity, support, energy, spare parts, downtime and eventual replacement. It should also quantify service, safety and quality benefits rather than focusing only on labour reduction.

Contemporary Developments

Current development is moving toward edge analytics, where some data processing occurs near cameras or sensors to reduce latency and network dependence. Generative AI is being tested for natural-language interaction with WMS data, maintenance assistance and procedure search. Its outputs must be grounded in approved operational data because fluent language does not guarantee factual accuracy.

Other developments include multi-robot coordination, battery-aware fleet scheduling, reusable packaging recognition, automated dimensioning and increasingly integrated supply-chain control towers. These developments strengthen the need for common data standards, interoperable systems and clear accountability.

Best-Practice Checklist

Before approving an AI or robotics project, management should ask:

• Is the business problem clearly defined?
• Is the baseline measured?
• Are data ownership and quality responsibilities assigned?
• Have safety, privacy and cybersecurity risks been assessed?
• Are affected workers involved and trained?
• Are model limitations and escalation rules documented?
• Is there a human review and appeal process for high-impact decisions?
• Can the system fail safely and operate in a degraded mode?
• Are benefits measured across cost, service, safety, quality and sustainability?
• Is there a plan for maintenance, model drift, vendor exit and retirement?

Chapter Summary

AI and robotics are transforming warehouses into connected decision-and-execution systems. AI supports demand forecasting, replenishment, slotting, computer vision and predictive maintenance, while digital twins enable simulation and AMRs support flexible material movement. These technologies are complementary: forecasting informs inventory decisions, accurate master data supports slotting, sensors enable maintenance models, and integrated systems coordinate robots and people.

The value of intelligent warehousing depends on data quality, process discipline, integration and human judgement. Forecast measures and optimisation models must be interpreted within their assumptions. Computer vision and predictive maintenance require risk-sensitive performance measures. Digital twins require validation, and AMRs require safe operating rules and recovery procedures.

Responsible adoption requires privacy protection, cybersecurity, fairness, explainability, human oversight, worker participation and safe failure modes. The best implementation begins with a defined operational problem, establishes a baseline, corrects data, pilots the least complex effective solution and scales only after evidence supports it. The future warehouse is therefore not merely automated; it is governed, adaptable, measurable and human-centred.

Key Terms

Artificial intelligence: Computational methods that perform tasks involving prediction, classification, reasoning or pattern recognition.

Autonomous mobile robot: A mobile robot that navigates and performs missions with limited direct human control.

Computer vision: Use of cameras and algorithms to interpret visual information.

Data drift: Change in data patterns that reduces model reliability over time.

Digital twin: A data-linked digital representation of a physical asset, process or facility.

Edge analytics: Processing data near the point where it is generated.

Forecast accuracy: The degree to which predicted demand corresponds with actual demand.

Human oversight: Deliberate monitoring, review or intervention by authorised people.

Intelligent slotting: Data-supported assignment of products to warehouse locations.

Machine learning: Methods that learn patterns from data to make predictions or classifications.

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

Predictive maintenance: Maintenance based on evidence of developing equipment condition or failure risk.

Reorder point: Inventory position at which replenishment is initiated.

Safety stock: Additional inventory held to protect against demand or supply uncertainty.

Smart warehouse: An integrated warehouse using connected data, analytics, automation and governed decision-making.

Total cost of ownership: All significant costs incurred during a technology’s useful life.

Warehouse-control system: Software that coordinates automated material-handling equipment.

Warehouse-management system: Software that manages inventory, locations, tasks and warehouse transactions.

Review Questions

Short Answer Questions

1. Distinguish automation from artificial intelligence in warehouse operations.
2. What is the purpose of a reorder point?
3. Identify three applications of computer vision in a warehouse.
4. Why is master-data quality important for AMR deployment?
5. State two privacy risks associated with warehouse cameras.

Long Answer Questions

1. Explain how AI-supported forecasting can improve replenishment while still requiring human judgement.
2. Discuss the main design and governance requirements for predictive maintenance.
3. Explain the role of digital twins in warehouse layout and capacity decisions.
4. Compare fixed automation, AMRs and collaborative robots in terms of flexibility, risk and implementation requirements.
5. Evaluate the workforce implications of intelligent warehousing and propose a responsible transition plan.

Analytical / Application Questions

1. A warehouse has frequent stockouts, but its sales history contains many periods in which items were unavailable. Explain why a model trained only on sales may underforecast demand and recommend additional data or controls.
2. A slotting algorithm places heavy, fast-moving cartons in high-level locations because this minimises travel distance. Analyse the error and propose constraints that should be added.
3. Design a pilot plan for introducing computer vision at dispatch. Specify the baseline, performance measures, privacy safeguards, human-review process and criteria for scaling.

Multiple Choice Questions with Answers

Question 1

Which condition most directly supports reliable AI forecasting in a warehouse?

A. Accurate and sufficiently complete historical and contextual data

B. Maximum camera coverage in every aisle

C. A larger warehouse building

D. Elimination of all human review

Correct Answer: A. Accurate and sufficiently complete historical and contextual data

Question 2

In the expression ROP = μ_L + SS, what does SS represent?

A. Supplier standard

B. Safety stock

C. Seasonal sales

D. Storage surface

Correct Answer: B. Safety stock

Question 3

Why can MAPE be unsuitable for intermittent-demand items?

A. It cannot use percentages

B. It requires equipment sensor data

C. It becomes unstable or undefined when actual demand is zero or near zero

D. It always produces values in kilograms

Correct Answer: C. It becomes unstable or undefined when actual demand is zero or near zero

Question 4

Which is a typical computer-vision application in warehousing?

A. Negotiating supplier contracts

B. Detecting damaged cartons or missing labels

C. Setting national tax rates

D. Replacing all inventory records

Correct Answer: B. Detecting damaged cartons or missing labels

Question 5

What is the main purpose of predictive maintenance?

A. To repair equipment only after complete failure

B. To use condition information to plan intervention before likely failure

C. To eliminate maintenance technicians

D. To increase equipment failure frequency

Correct Answer: B. To use condition information to plan intervention before likely failure

Question 6

Which statement best describes a digital twin?

A. A static warehouse photograph

B. A data-linked digital representation of a physical asset, process or facility

C. A duplicate employee identity

D. A paper-based stock ledger

Correct Answer: B. A data-linked digital representation of a physical asset, process or facility

Question 7

What distinguishes an AMR from many fixed-path automated vehicles?

A. An AMR can navigate dynamically within defined operating constraints

B. An AMR never requires sensors

C. An AMR cannot communicate with a WMS

D. An AMR is always safe without risk assessment

Correct Answer: A. An AMR can navigate dynamically within defined operating constraints

Question 8

Which data-quality dimension asks whether information is current enough for a decision?

A. Uniqueness

B. Timeliness

C. Validity

D. Completeness

Correct Answer: B. Timeliness

Question 9

Which governance practice is most appropriate when camera data could affect workers?

A. Use it for every possible purpose without notice

B. Delete all operational records immediately

C. Define purpose, restrict access, limit retention and communicate the policy

D. Allow unrestricted access to all supervisors

Correct Answer: C. Define purpose, restrict access, limit retention and communicate the policy

Question 10

What should normally occur before scaling a warehouse AI pilot?

A. The baseline should be measured and outcomes validated against defined criteria

B. All manual procedures should be permanently removed

C. The model should be accepted without testing

D. Worker consultation should be cancelled

Correct Answer: A. The baseline should be measured and outcomes validated against defined criteria