Sector Applications and Case Analysis

Introduction

The Internet of Things (IoT) connects physical objects—machines, vehicles, buildings, medical devices, crops, and infrastructure—to digital systems that can sense, communicate, compute, and act. Artificial intelligence (AI) adds the ability to identify patterns, make predictions, optimize decisions, and support autonomous or semi-autonomous control. Their combination is often called IoT–AI or AIoT.

IoT–AI systems are not identical across sectors. A factory may prioritize millisecond-level control and worker safety; a smart city may prioritize interoperability and public legitimacy; healthcare may prioritize clinical validity and privacy; agriculture may prioritize low power and intermittent connectivity. Therefore, analyzing an application requires more than listing sensors and algorithms. It requires examination of architecture, stakeholders, operating conditions, performance measures, risks, governance, and the reasons implementations succeed or fail.

This chapter compares IoT–AI applications in smart manufacturing, cities, homes, healthcare, agriculture, energy, mobility, retail, and logistics. It also develops a reusable case-analysis framework and applies it to an illustrative predictive-maintenance scenario.

Learning Objectives

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

1. Explain how IoT and AI interact in sector-specific cyber-physical systems.

2. Compare edge, fog, cloud, and hybrid architectures.

3. Identify stakeholders, value propositions, risks, and governance requirements.

4. Select suitable performance measures for different applications.

5. Analyze a practical IoT–AI case using technical, organizational, and ethical criteria.

6. Distinguish successful deployment from a technically impressive but operationally ineffective pilot.

1. A General Model for Sector Applications

An IoT–AI application can be represented as a socio-technical loop:

physical environment → sensing → communication → data management → AI inference → decision → actuation or human action → physical environment

The loop is socio-technical because people, institutions, regulations, business processes, and technical components jointly determine outcomes.

1.1 Functional layers

A useful architecture contains the following layers:

• Physical and sensing layer: Sensors measure temperature, vibration, location, pressure, images, energy use, biometrics, or environmental conditions. Actuators change a physical state, such as opening a valve or controlling a motor.
• Connectivity layer: Wired Ethernet, industrial networks, Wi-Fi, Bluetooth, cellular, satellite, LPWAN, and other protocols transport data. Connectivity must match range, bandwidth, latency, reliability, and power requirements.
• Edge layer: Devices or gateways process data close to its source. Edge processing can reduce latency, bandwidth use, and exposure of raw data.
• Platform and data layer: Device identity, message brokers, storage, digital twins, data quality services, and APIs support integration.
• AI and analytics layer: Descriptive, diagnostic, predictive, and prescriptive analytics transform data into insight or action.
• Application and governance layer: Dashboards, workflow systems, human interfaces, access control, audit trails, policies, and regulatory controls determine how the system is used.

These layers are logical rather than necessarily separate products. A gateway may perform protocol conversion, storage, security functions, and AI inference at the same time.

1.2 Edge, cloud, and hybrid architectures

Cloud-centric architectures centralize storage and computation. They are useful for large-scale model training, cross-site comparison, long-term analysis, and centralized management. Their limitations include network dependence, latency, recurring data-transfer costs, and privacy concerns.

Edge architectures execute inference near the data source. They are appropriate when response time, resilience, privacy, or bandwidth is important. Edge devices may have limited memory, processing capacity, and energy.

Hybrid architectures train models centrally while deploying compact inference models at the edge. This is common in manufacturing, vehicles, healthcare devices, and energy systems. Model updates must be governed because an inaccurate or malicious update can affect many devices.

Architecture selection should follow requirements rather than fashion. A safety-critical control loop generally should not depend exclusively on a distant cloud service. Conversely, placing every computation at the edge may make fleet-wide learning and maintenance unnecessarily difficult.

1.3 AI functions in IoT systems

Common AI functions include:

• Classification: assigning an observation to a category, such as normal or defective.
• Regression: estimating a continuous value, such as remaining useful life or energy demand.
• Anomaly detection: identifying observations that differ from expected behavior, especially when failure examples are scarce.
• Forecasting: predicting demand, traffic, crop stress, or equipment load.
• Optimization: selecting actions subject to constraints such as cost, safety, comfort, and capacity.
• Computer vision: analyzing images or video for defects, occupancy, traffic, or disease symptoms.
• Natural-language interfaces: allowing operators or residents to query systems conversationally.

AI does not automatically create value. A model must be connected to a valid decision process. For example, a highly accurate failure prediction is of limited benefit if spare parts cannot be obtained before the predicted failure or if maintenance staff do not trust the alert.

2. A Framework for Comparing Sector Deployments

A systematic comparison should examine six dimensions.

2.1 Operational objective

What decision or physical outcome is being improved? Examples include reducing unplanned downtime, balancing electricity demand, reducing congestion, improving diagnosis, or minimizing delivery distance.

2.2 Architecture and criticality

Determine where sensing, inference, storage, and control occur. Classify the application as advisory, operational, or safety-critical. The more direct the connection to physical safety, the stronger the requirements for fail-safe behavior, validation, redundancy, and human oversight.

2.3 Stakeholders and incentives

Stakeholders may include owners, operators, workers, customers, patients, clinicians, residents, regulators, suppliers, insurers, and communities. Their incentives may conflict. A utility may value demand flexibility while a household values comfort and privacy.

2.4 Performance measures

Measures should include technical and outcome indicators:

• Latency: time between sensing and useful action.
• Availability: proportion of time the service is usable.
• Accuracy, precision, recall, and false-alarm rate: appropriate for classification tasks.
• Forecast error: for example, mean absolute error, defined as the average of absolute differences between predictions and observations. The units are the same as the measured quantity.
• Energy efficiency: energy consumed per unit of useful output.
• Economic value: avoided cost, increased revenue, or return on investment.
• Safety and reliability: incidents, near misses, mean time between failures, or recovery time.
• Equity and user outcomes: distribution of benefits and errors across groups.

No single metric is sufficient. A city traffic model may be accurate on average while systematically performing worse in a particular neighborhood. A healthcare model may have high accuracy because most cases are low risk while missing rare but serious conditions.

2.5 Risk and governance

Relevant risks include cyberattack, privacy loss, unsafe automation, biased decisions, model drift, vendor lock-in, poor data quality, and unclear accountability. Governance should define who can access data, who may change a model, how decisions are audited, and what happens during failure.

2.6 Scale and lifecycle

A pilot may work with ten devices but fail with ten thousand because of onboarding, network congestion, maintenance, data storage, or organizational complexity. Lifecycle analysis includes procurement, installation, calibration, software updates, training, decommissioning, and disposal.

3. Sector Applications

3.1 Smart manufacturing

Smart manufacturing uses connected machines, industrial controllers, robots, cameras, and enterprise systems. AI supports predictive maintenance, visual quality inspection, process optimization, demand planning, and worker assistance.

A typical architecture keeps deterministic control functions in industrial control systems, sends selected data to an edge gateway for rapid inference, and uses cloud or plant servers for model training and fleet-level analysis. A digital twin may represent equipment, production orders, operating conditions, and maintenance history.

Key stakeholders include production managers, maintenance engineers, operators, quality teams, equipment manufacturers, workers, and regulators. Important measures are unplanned downtime, overall equipment effectiveness, defect escape rate, changeover time, maintenance cost, inference latency, and false alarms.

Risks include connecting legacy operational technology to corporate networks, unsafe actuator commands, cyber-physical attacks, and worker surveillance. A successful system usually begins with a constrained use case, uses existing maintenance workflows, and gives workers interpretable alerts rather than replacing expertise without preparation.

3.2 Smart cities

Smart-city IoT includes traffic sensors, parking systems, air-quality monitors, lighting, water networks, waste containers, public transport, and emergency services. AI can forecast congestion, optimize traffic signals, detect leaks, predict flooding, and prioritize infrastructure maintenance.

City architectures are highly heterogeneous. Devices may be owned by different agencies and vendors, making open standards, common data models, and strong APIs important. Edge processing is valuable for traffic and video applications, while citywide planning benefits from centralized data integration.

Stakeholders include residents, elected officials, municipal departments, transport operators, businesses, emergency services, and civil society. Measures include travel time, emissions, response time, service reliability, water loss, energy consumption, cost, and distributional fairness.

Risks are substantial because urban data can reveal movement and social behavior. Facial recognition and predictive policing illustrate how technically feasible systems can create serious civil-liberties concerns. Public consultation, data minimization, impact assessments, retention limits, and independent oversight are essential.

3.3 Smart homes and buildings

Connected thermostats, meters, locks, appliances, occupancy sensors, cameras, and voice assistants can improve comfort, security, accessibility, and energy efficiency. AI predicts occupancy, learns preferences, detects unusual activity, and coordinates heating, cooling, lighting, and storage.

Home systems often use a hub or gateway, local automation rules, and cloud services. Local fallback is important when internet access fails. In multi-tenant or commercial buildings, building-management systems integrate HVAC, elevators, fire systems, and access control.

Performance measures include energy reduction, comfort complaints, response time, device availability, false alarms, and user retention. Risks include intimate behavioral profiling, insecure inexpensive devices, unauthorized access, and automation that conflicts with residents’ preferences. Users should receive understandable controls, meaningful consent choices, secure defaults, and the ability to override automation.

3.4 Healthcare

IoT–AI healthcare includes wearable monitors, connected inhalers, infusion pumps, imaging systems, hospital asset tracking, remote patient monitoring, and ambient sensors. AI may detect deterioration, classify images, predict readmission, support triage, or optimize bed and equipment allocation.

Healthcare architecture must preserve clinical safety and confidentiality. Devices may transmit to a hospital gateway or secure cloud, while alerts are integrated into electronic health-record and clinical workflow systems. A model should support—not silently replace—qualified clinical judgment unless it has been specifically validated and authorized for autonomous use.

Measures include sensitivity, specificity, positive predictive value, alert burden, time to intervention, calibration, patient outcomes, and subgroup performance. A model’s performance can change across hospitals because patient populations, equipment, coding practices, and workflows differ.

Risks include privacy breaches, biased diagnosis, automation bias, cybersecurity threats to medical devices, and liability uncertainty. Governance requires consent where appropriate, role-based access, auditability, clinical validation, post-deployment monitoring, and compliance with applicable medical-device and data-protection rules.

3.5 Agriculture

Precision agriculture uses soil-moisture sensors, weather stations, satellite imagery, drones, livestock tags, machine telemetry, and automated irrigation. AI estimates crop stress, detects disease, forecasts yield, identifies weeds, and recommends fertilizer or irrigation.

Agricultural systems often operate in remote areas with intermittent connectivity and limited power. Edge gateways, solar power, store-and-forward communication, and low-power wide-area networks are therefore useful. Models must account for local soil, climate, crop variety, and farm practices.

Measures include yield, water use per unit of production, fertilizer use, disease detection lead time, labor hours, and farmer income. Risks include inaccurate recommendations, unequal access for smallholders, dependence on proprietary platforms, and collection of land or production data without clear ownership. Human agronomic knowledge remains essential, especially under unusual weather conditions outside the training data.

3.6 Energy and utilities

IoT–AI is used in smart meters, grid sensors, renewable-energy forecasting, battery management, demand response, fault detection, and building energy management. AI forecasts electricity demand and solar or wind output, identifies abnormal consumption, and helps coordinate distributed energy resources.

Energy systems combine centralized control with increasingly distributed assets. Protection and stability functions require highly reliable and time-sensitive operation; predictive analytics may run in regional control centers or at substations. A system must distinguish advisory forecasts from automatic grid controls.

Measures include forecast error, outage duration, frequency and voltage quality, renewable curtailment, peak demand, energy losses, and customer cost. Cybersecurity is critical because compromised devices can affect grid stability. Fairness also matters: dynamic pricing or demand-response programs should not disproportionately burden households with limited flexibility.

3.7 Mobility and transportation

Connected vehicles, roadside units, fleet telematics, public transport systems, and traffic infrastructure support route optimization, driver assistance, predictive maintenance, parking management, and automated transport services.

Mobility systems are distributed and mobile. They require reliable localization, communication, low latency for safety functions, and careful handling of incomplete or conflicting sensor data. Fleet-level AI can run in the cloud, whereas immediate collision-avoidance functions require on-vehicle processing.

Measures include travel time, safety incidents, fuel or energy use, vehicle utilization, passenger waiting time, prediction error, and service equity. Risks include location surveillance, software compromise, unsafe model behavior, and unclear responsibility when automated recommendations influence human drivers.

3.8 Retail

Retail IoT includes shelf sensors, RFID, cameras, point-of-sale systems, beacons, smart carts, and cold-chain monitors. AI supports demand forecasting, inventory replenishment, personalized recommendations, queue management, price optimization, and loss prevention.

The central architecture integrates store-edge systems with enterprise cloud platforms. Edge processing can reduce video bandwidth and provide rapid alerts, while cloud systems learn from multiple stores. Data quality is challenging because promotions, stockouts, seasonal effects, and competitor actions distort observed demand.

Measures include stockout rate, inventory turnover, forecast error, conversion rate, basket value, shrinkage, queue time, and customer satisfaction. Risks include intrusive tracking, discriminatory pricing, excessive employee monitoring, and false accusations from loss-prevention systems. Transparency and proportionality are particularly important when biometric or behavioral data are considered.

3.9 Logistics and supply chains

Logistics applications use GPS, RFID, temperature sensors, telematics, warehouse robots, cameras, and shipment-management platforms. AI forecasts demand, selects routes, detects cold-chain violations, predicts arrival times, optimizes warehouse picking, and identifies bottlenecks.

The architecture crosses organizational boundaries: manufacturers, carriers, ports, warehouses, retailers, customs agencies, and customers may each own data. Shared identifiers, data contracts, access rules, and interoperability are therefore as important as algorithms.

Measures include on-time-in-full delivery, cost per shipment, utilization, damage rate, inventory days, energy use, estimated-time-of-arrival error, and cold-chain compliance. Risks include inaccurate tracking, cyberattack, worker displacement, and opaque algorithmic allocation of work. Resilience requires contingency routes and manual procedures when sensors or platforms fail.

4. Detailed Illustrative Case: Predictive Maintenance in a Manufacturing Plant

Background

Consider a fictional medium-sized plant producing electric motors. Its assembly line contains several automated test stations and conveyor drives. Unplanned failures cause production delays, overtime, and missed deliveries. Maintenance is mainly time-based: motors are inspected every six months even though their actual condition varies.

Problem

Management proposes vibration and temperature sensors connected to an AI model. The project team must determine whether the system will reduce downtime without creating excessive false alarms or unsafe dependence on automation.

Analysis

The team identifies three failure modes: bearing wear, overheating, and belt misalignment. Historical failure labels are incomplete, so anomaly detection is initially more realistic than a fully supervised failure classifier. Sensor readings are combined with operating speed, load, maintenance records, and production schedules.

The required response is measured in minutes rather than milliseconds, so edge inference at a plant gateway is suitable. Raw high-frequency vibration data are summarized locally, while selected features and events are sent to a central platform. This reduces bandwidth and keeps the line functioning during temporary cloud disconnection.

Stakeholders include maintenance technicians, operators, plant managers, the equipment supplier, the IT security team, and workers whose activities may be indirectly represented in production data. The system must not be used covertly to rank workers.

Approach and solution

Sensors are installed on a limited number of high-cost motors. The project establishes a baseline during normal operation and tests alerts against maintenance inspections. Each alert includes the affected asset, evidence such as an abnormal vibration trend, confidence information, and a recommended inspection—not an automatic shutdown.

The deployment uses staged validation:

1. Offline evaluation: historical and replayed data are used to check whether known events would have been detected.

2. Shadow mode: the model produces alerts that do not yet change operations.

3. Controlled operation: technicians act on alerts using documented procedures.

4. Scale-up: only after reviewing false alarms, missed events, usability, and security.

Model performance is assessed using precision and recall, but operational measures are primary: unplanned downtime, maintenance lead time, cost per avoided failure, and technician acceptance. Model drift is monitored because a new motor design, changed production speed, or sensor replacement can alter data patterns.

Outcome

In this illustrative case, the pilot identifies several genuine bearing problems earlier than routine inspection. However, it also generates frequent alerts during a temporary change in production speed. The team therefore adds operating-context features and requires a minimum persistence period before escalation. The revised system produces fewer nuisance alerts and is accepted by technicians.

The most important outcome is not the model’s numerical score. It is the integration of trustworthy alerts into a maintenance process with available staff, spare parts, escalation rules, and fallback procedures.

Learning points

• A narrowly defined business problem is more manageable than a general “smart factory” program.
• Sensor placement and maintenance labels can matter more than algorithm selection.
• Edge processing improves resilience but does not eliminate the need for centralized governance.
• Human expertise should be incorporated into alert interpretation and model improvement.
• Success must be measured by operational outcomes, not only model accuracy.
• A pilot is evidence of feasibility, not proof of scalable value.

5. Why IoT–AI Projects Succeed or Fail

Successful projects commonly have a measurable problem owner, reliable data-generation processes, interoperable integration, secure device management, clear human responsibilities, and a plan for maintenance after deployment. They begin with a modest scope and expand after evidence is collected.

Failure often results from treating the project as an algorithm demonstration. Other causes include poor sensor calibration, missing labels, unrepresentative training data, unanticipated network costs, lack of staff training, incompatible legacy systems, unclear data ownership, and procurement that ignores the total lifecycle cost.

A useful distinction is between technical validity and deployment validity. Technical validity asks whether the system works under tested conditions. Deployment validity asks whether it continues to work in the real environment, fits workflows, benefits intended users, and remains safe and governable over time.

6. Ethical, Legal, and Governance Considerations

IoT–AI expands data collection into physical and social spaces. Responsible deployment should address:

• Purpose limitation: collect data for a defined legitimate purpose rather than indefinite future use.
• Data minimization: avoid collecting more detail or retaining data longer than necessary.
• Security by design: use unique device identities, secure updates, encryption where appropriate, least-privilege access, network segmentation, logging, and vulnerability management.
• Explainability and contestability: affected people should receive understandable reasons and a route to challenge important decisions.
• Fairness: test performance and impact across relevant demographic, geographic, economic, and operational groups.
• Safety: apply risk assessment, fail-safe design, human oversight, validation, and incident reporting.
• Accountability: assign responsibility for data quality, model changes, alerts, decisions, and harm.
• Interoperability and portability: avoid unnecessary dependence on one vendor and support data and device migration.

Regulatory obligations vary by jurisdiction and sector. Healthcare, employment, transport, critical infrastructure, and personal data generally require stronger controls than low-risk consumer applications. Organizations should consult applicable privacy, cybersecurity, safety, medical-device, consumer-protection, and artificial-intelligence regulations rather than assuming that a general IoT policy is sufficient.

7. Contemporary Developments and Best Practices

Current development is moving toward smaller edge models, federated learning, digital twins, multimodal sensing, explainable decision support, and more automated device management. These developments create opportunities but also new risks. Federated learning can reduce centralization of raw data, but model updates may still leak information and require robust security. Generative AI can provide natural-language interfaces to operational data, but its outputs must not be treated as authoritative without validation.

Best practices include:

1. Start with a clearly defined decision and measurable outcome.

2. Map data flows, physical consequences, stakeholders, and failure modes before selecting technology.

3. Design for security, privacy, accessibility, and interoperability from the beginning.

4. Use representative data and document collection conditions, missingness, and labels.

5. Separate safety-critical control from experimental AI where necessary.

6. Validate in shadow mode and monitor after deployment.

7. Provide human override, fallback operation, and incident-response procedures.

8. Track total cost of ownership, including calibration, connectivity, updates, training, and disposal.

9. Review models for drift, bias, and changing context.

10. Engage affected workers, residents, patients, or customers in design and evaluation.

Chapter Summary

IoT–AI systems connect sensing and actuation with analytics, prediction, optimization, and human decision-making. Their architecture may be edge, cloud, or hybrid, and the appropriate choice depends on latency, connectivity, privacy, resilience, scale, and criticality. Manufacturing, cities, homes, healthcare, agriculture, energy, mobility, retail, and logistics apply similar technical building blocks to different operational problems and stakeholder environments.

A sound sector analysis compares objectives, architecture, stakeholders, performance measures, risks, governance, and lifecycle requirements. Model accuracy alone is inadequate: successful systems create measurable operational or social value, integrate with workflows, remain secure and maintainable, and provide accountability. The manufacturing case demonstrates that staged deployment, context-aware data, human expertise, and fallback procedures are central to trustworthy IoT–AI adoption.

Key Terms

IoT–AI: Integration of connected physical devices with artificial intelligence and analytics.

Edge computing: Processing data near the source where it is generated.

Cloud computing: Centralized, network-accessible computation and storage.

Hybrid architecture: Combination of edge, local, and cloud processing.

Actuator: Device that changes a physical condition in response to a command.

Digital twin: Digital representation of a physical asset, process, or system.

Predictive maintenance: Maintenance planned using predicted equipment condition or failure risk.

Anomaly detection: Identification of observations that differ from expected behavior.

Model drift: Change in data or relationships that reduces model performance over time.

Latency: Time between an event, its processing, and a useful response.

Availability: Proportion of time a service or system is operational.

Precision: Proportion of predicted positive cases that are actually positive.

Recall: Proportion of actual positive cases that a model detects.

Interoperability: Ability of systems from different sources to exchange and use information.

Digital privacy: Protection of information about identifiable people and their activities.

Data minimization: Limiting collection and retention to what is necessary.

Human-in-the-loop: Design in which a person reviews, guides, or can override AI outputs.

Total cost of ownership: All acquisition, operation, maintenance, training, and disposal costs.

Cyber-physical system: System combining computational components with physical processes.

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

Review Questions

Short Answer Questions

1. What are the main layers of a general IoT–AI architecture?

2. Why might a safety-critical application require edge processing?

3. Define model drift and explain why it matters.

4. Name two performance measures suitable for predictive maintenance.

5. What is the difference between technical validity and deployment validity?

Long Answer Questions

1. Compare cloud-centric, edge, and hybrid architectures for healthcare and smart manufacturing.

2. Discuss the major stakeholders, benefits, and risks in a smart-city IoT–AI deployment.

3. Explain why accuracy alone is insufficient for evaluating a healthcare AI system.

4. Analyze the ethical and governance requirements of smart-home monitoring.

5. Describe how IoT–AI can improve agricultural productivity while creating new forms of dependency or inequality.

Analytical / Application Questions

1. A logistics company wants to use sensors and AI to predict delivery delays. Propose an architecture, identify stakeholders, and select at least five performance measures. Justify each choice.

2. A hospital model has high overall accuracy but misses more deterioration cases among elderly patients than among younger patients. Analyze possible causes and recommend a validation and governance plan.

3. A city proposes automatic traffic enforcement using connected cameras. Evaluate the proposal using privacy, fairness, security, reliability, accountability, and public-value criteria.

Multiple Choice Questions with Answers

Question 1

Which architecture is generally most appropriate for a safety-critical control loop requiring very low latency?

A. Edge architecture

B. Cloud-only architecture

C. Manual spreadsheet architecture

D. Offline archival architecture

Correct Answer: A. Edge architecture

Question 2

What is the primary purpose of anomaly detection in many IoT applications?

A. To identify observations that differ from expected behavior

B. To guarantee that every device is physically secure

C. To replace all human decision-making

D. To eliminate the need for sensor calibration

Correct Answer: A. To identify observations that differ from expected behavior

Question 3

Which measure is especially relevant to evaluating a predictive-maintenance alert system?

A. Unplanned downtime

B. Number of social-media followers

C. Screen brightness

D. Office floor area

Correct Answer: A. Unplanned downtime

Question 4

What does model drift mean?

A. A change in data or relationships that reduces model performance

B. A permanent increase in network bandwidth

C. Physical movement of a sensor during installation only

D. Encryption of all historical records

Correct Answer: A. A change in data or relationships that reduces model performance

Question 5

Why is edge processing useful in remote agricultural IoT systems?

A. Connectivity may be intermittent and power may be limited

B. It guarantees perfect crop predictions

C. It removes the need for local agronomic knowledge

D. It makes all data legally public

Correct Answer: A. Connectivity may be intermittent and power may be limited

Question 6

Which healthcare measure indicates the proportion of actual positive cases detected by a model?

A. Recall

B. Latency

C. Availability

D. Inventory turnover

Correct Answer: A. Recall

Question 7

Which practice best supports responsible smart-home deployment?

A. Secure defaults and meaningful user control

B. Indefinite retention of all household recordings

C. Automatic sharing with unrelated advertisers

D. Removal of manual override controls

Correct Answer: A. Secure defaults and meaningful user control

Question 8

Why are open standards and common data models important in smart cities?

A. City systems are heterogeneous and often span multiple agencies and vendors

B. They guarantee that no cyberattack can occur

C. They eliminate all privacy obligations

D. They make sensors unnecessary

Correct Answer: A. City systems are heterogeneous and often span multiple agencies and vendors

Question 9

Which statement best describes deployment validity?

A. Whether a system works in real operational conditions and creates intended value

B. Whether a model has any mathematical equation

C. Whether a device has a screen

D. Whether training data are stored indefinitely

Correct Answer: A. Whether a system works in real operational conditions and creates intended value

Question 10

In the illustrative manufacturing case, why was shadow mode used?

A. To evaluate alerts before allowing them to change operations

B. To permanently disable all model outputs

C. To replace vibration sensors with cameras

D. To avoid involving maintenance technicians

Correct Answer: A. To evaluate alerts before allowing them to change operations