Digital Lean, Industry 4.0 and Artificial Intelligence

Introduction

Lean manufacturing and service blueprinting were developed to make value-creating work visible, remove waste, stabilize flow, and improve customer outcomes. Digital technologies extend these purposes by creating faster access to operational information, connecting processes and assets, and supporting prediction and decision-making. However, digitization is not automatically lean. A poorly designed digital system can make waste faster, more expensive, and less visible. It can also create new forms of waste, including excessive data collection, duplicate applications, alert fatigue, cybersecurity exposure, and dependence on opaque algorithms.

Digital Lean therefore means the disciplined use of digital technologies to improve customer value, process flow, quality, flexibility, and employee capability. The central principle remains the same as in conventional Lean: first understand the process and its purpose; then simplify and stabilize it; only afterward automate or digitize where technology creates verifiable value.

Industry 4.0 refers broadly to the integration of cyber-physical systems, industrial networks, cloud computing, advanced analytics, automation, and artificial intelligence. In services, equivalent developments include digitally enabled service platforms, customer journey analytics, robotic process automation, and intelligent decision support. Total Quality Management provides the governance perspective: quality must be designed into the system, measured through meaningful indicators, and improved through learning rather than inspection alone.

Learning Objectives

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

1. Explain the relationship between Lean, Industry 4.0, digital Lean, and service blueprinting.
2. Evaluate process mining, IoT, digital twins, robotics, cloud platforms, and electronic Kanban.
3. Interpret real-time analytics and AI-supported forecasting, inspection, and root-cause analysis.
4. Identify technology-led waste and assess data quality, cybersecurity, privacy, explainability, and human oversight.
5. Design a controlled roadmap for implementing digital Lean in manufacturing or service operations.
6. Critically evaluate the organizational, ethical, and managerial consequences of intelligent automation.

1. From Lean Thinking to Digital Lean

Traditional Lean is based on several connected ideas: value is defined by the customer; the value stream should be identified; flow should be improved; production should be pulled by demand; and improvement should be continuous. Common wastes include defects, overproduction, waiting, unnecessary transport, excess inventory, unnecessary motion, overprocessing, and underused human capability.

Digital Lean adds a further question: what information and decision latency prevent the process from flowing? A production line may stop because a component is unavailable, but it may also stop because the replenishment signal is delayed or inaccurate. A hospital may have sufficient beds, but patient flow may be impaired because status information is fragmented across systems. A digital intervention is valuable only when it reduces a significant cause of process instability or improves a decision without introducing greater risk.

A useful distinction is between three layers:

Digital Lean is strongest when these layers are designed together. A dashboard cannot compensate for an unstable process, and an AI model cannot compensate for undefined decision rights.

Technology-enabled and technology-led waste

Technology-enabled waste occurs when a digital tool removes an existing waste. Technology-led waste is created or amplified by the technology itself. Examples include:

• Automating a redundant approval instead of eliminating it.
• Requiring operators to enter the same information into multiple systems.
• Generating alerts that are not linked to an action.
• Building dashboards with hundreds of measures but no management routine.
• Storing excessive personal or operational data without a legitimate purpose.
• Increasing system complexity so that small changes require specialist intervention.
• Replacing a visible queue with an invisible digital queue.

The appropriate test is not whether a system is technologically sophisticated, but whether it improves a validated value-stream outcome such as lead time, first-pass yield, service reliability, safety, or employee capability.

2. Process Mining and Digital Value-Stream Analysis

Process mining reconstructs actual process behavior from event logs. An event log commonly contains a case identifier, activity name, timestamp, and sometimes information about the resource, location, product, or outcome. In manufacturing, a case may be a production order; in a service organization, it may be a customer request, claim, admission, or invoice.

Process mining has three principal uses:

1. Process discovery: deriving a process model from recorded events.
2. Conformance checking: comparing actual behavior with a prescribed process.
3. Performance analysis: examining throughput time, waiting, rework, bottlenecks, and variation.

The basic elapsed time for a case can be expressed as:

Here, T_case is the total case duration, while t_completion and t_start are timestamps measured in the same time zone and time unit. The measure includes both processing and waiting unless the data are explicitly separated. It is therefore not equivalent to touch time.

For a set of n cases, mean cycle time is:

The mean is useful but can be distorted by exceptional cases. Median, percentile, and dispersion measures should also be examined, particularly where customers experience the tail of the distribution.

Process mining is valuable because it reveals variants that workshops may miss. For example, a service blueprint may show one intended path, whereas event data may reveal that 28% of cases bypass a validation step, 16% are reopened, and a small group of cases accounts for most delays. Yet process mining has limitations. Event logs may be incomplete, timestamps may represent administrative rather than actual events, and activity names may be inconsistent. A visually impressive process model based on poor logs can mislead management.

Data preparation for process mining

Before analysis, practitioners should establish:

• A unique and stable case identifier.
• A consistent activity vocabulary.
• Reliable timestamps and time-zone rules.
• Rules for duplicate, cancelled, and incomplete cases.
• A clear distinction between system events and real work.
• Data ownership and access controls.

The analysis should be connected to direct observation. A discovered delay is a hypothesis about the process, not automatically an explanation of why the delay occurs.

3. IoT, Edge Computing, and Real-Time Analytics

The Internet of Things connects physical objects, sensors, machines, products, and environments so that data can be collected and exchanged. Typical industrial measurements include temperature, vibration, pressure, current, position, cycle count, and energy consumption. Service applications include location, queue status, equipment availability, and interaction events.

An IoT architecture commonly includes sensing, connectivity, edge processing, platform storage, analytics, and action. Edge computing processes data near the source. It is useful when response time, bandwidth, resilience, or data confidentiality makes continuous cloud transmission unsuitable. Cloud platforms are useful for scalable storage, cross-site comparison, model training, and collaboration.

Real-time analytics should be interpreted carefully. “Real time” is not a universal technical standard; it means that information arrives rapidly enough for the decision at hand. A safety interlock may require milliseconds, while a weekly capacity decision may need only daily data.

Useful operational measures include availability, performance, and quality. A common overall equipment effectiveness expression is:

where:

• Availability = operating time ÷ planned production time;
• Performance = (ideal cycle time × total units) ÷ operating time;
• Quality = good units ÷ total units.

All three factors are dimensionless ratios, normally expressed as percentages. The formulation assumes that the time definitions, unit counts, and ideal cycle time are consistently specified. OEE is a diagnostic measure, not a complete measure of business performance. Maximizing it may be harmful if it encourages overproduction, excessive inventory, or production of items that customers do not need.

A real-time dashboard should therefore show not only current status but also decision thresholds, responsible roles, and prescribed responses. An indicator without an action rule is often digital decoration.

4. Digital Twins

A digital twin is a dynamic digital representation of a physical asset, process, product, or system, linked through data over its life cycle. It differs from a static simulation or a three-dimensional model because it is intended to remain connected to the evolving real-world object.

Digital twins can support:

• Condition monitoring and predictive maintenance.
• Production scheduling and scenario testing.
• Energy and resource optimization.
• Product configuration and traceability.
• Service capacity and customer-flow analysis.

A useful maturity progression is:

1. Descriptive: reports the current or historical state.
2. Diagnostic: identifies relationships and likely causes.
3. Predictive: estimates future states or failures.
4. Prescriptive: compares possible actions and recommends one.
5. Autonomous: executes bounded actions under defined controls.

The higher levels require stronger data quality, model validation, safeguards, and governance. A twin that represents an asset inaccurately may create false confidence. Model drift can occur when equipment, products, operating conditions, or customer behavior changes. Consequently, a twin requires configuration management, validation against physical measurements, and clear ownership.

5. Robotics and Automation in Lean Systems

Robotics includes physical robots, collaborative robots, automated guided vehicles, autonomous mobile robots, and software robots that perform rule-based administrative tasks. Automation can reduce ergonomic risk, improve repeatability, shorten handling time, and support continuous flow. It can also create capital cost, maintenance dependency, integration complexity, and skill displacement.

The Lean question is whether the task should be eliminated, simplified, combined, or redesigned before it is automated. Automating a poor method stabilizes the poor method. A practical evaluation should consider:

• Customer and process value.
• Safety and ergonomic exposure.
• Variation and task frequency.
• Changeover requirements.
• Failure modes and recovery procedures.
• Total cost of ownership.
• Workforce capability and redeployment.
• Cybersecurity and physical safety.

Collaborative robots do not remove the need for risk assessment. The human-robot workspace, speed, force, tooling, foreseeable misuse, and emergency behavior must be evaluated according to applicable safety requirements. Software robots similarly require access controls, exception handling, audit trails, and separation of duties.

6. Cloud Platforms and Electronic Kanban

Cloud platforms provide shared computing resources, storage, integration services, and applications delivered through networks. They can support multi-site visibility, rapid deployment, elastic capacity, and common master data. Their risks include vendor dependence, service outages, data location concerns, misconfiguration, and unclear responsibility for security.

Electronic Kanban translates pull signals into digital transactions. A signal may be generated when a consumption event occurs, when inventory reaches a defined trigger, or when a downstream process requests replenishment. Compared with paper cards, electronic Kanban can provide traceability, automated notification, status visibility, and integration with enterprise systems.

However, electronic Kanban does not eliminate the need to set appropriate replenishment parameters. A simple safety-stock relationship is often written as:

where ROP is the reorder point in units, d is average demand per time unit, L is replenishment lead time in the same time unit, and SS is safety stock in units. This expression assumes relatively stable definitions and does not itself calculate safety stock. If demand or lead time is variable, SS must be determined using an appropriate service-level and variability model.

For example, if average demand is 40 units per day, lead time is 3 days, and safety stock is 30 units, then:

The result is valid only if demand and lead time are measured consistently and the replenishment policy actually responds at the trigger point. A digital signal cannot correct inaccurate inventory records, unrecorded scrap, or unreliable suppliers.

7. AI-Supported Forecasting, Inspection, and Root-Cause Analysis

Artificial intelligence includes computational methods that perform tasks commonly associated with human intelligence, such as classification, prediction, language processing, and pattern recognition. Machine learning systems infer patterns from data; they do not automatically understand operational purpose or ethical consequences.

Forecasting

AI-supported forecasting may use historical demand, promotions, weather, prices, calendar effects, lead times, and external signals. Forecast accuracy must be evaluated against a suitable baseline, such as the previous-period forecast or a moving average. Mean absolute error is:

where yᵢ is observed demand, ŷᵢ is forecast demand, and n is the number of observations. MAE has the same units as demand and is easy to interpret. The evaluation period should be separated from the data used to train the model. Randomly mixing future information into training data creates leakage and produces overoptimistic results.

A more accurate forecast can still produce worse outcomes if it encourages excessive safety stock, unstable schedules, or frequent expedites. Forecast value must therefore be assessed through operational outcomes, not accuracy alone.

AI-supported inspection

Computer vision can classify surface defects, verify assembly, read labels, and measure dimensions. The confusion matrix distinguishes:

• True positive: a defect is correctly detected.
• False positive: a non-defect is incorrectly rejected.
• True negative: a non-defect is correctly accepted.
• False negative: a defect is missed.

Precision and recall are:

Here, TP, FP, and FN represent the respective counts. Precision matters when unnecessary rejection is costly; recall matters when missed defects create safety or customer risk. The appropriate threshold depends on the consequences of each error. A model should be tested across product variants, lighting conditions, equipment states, and relevant populations.

Root-cause analysis

AI can identify associations among process variables, detect anomalies, cluster failure modes, and rank candidate factors. Association is not proof of causation. Root-cause analysis should combine statistical evidence, process knowledge, designed experiments where appropriate, and confirmation of corrective action.

A responsible workflow is:

1. Define the problem operationally.
2. Verify the measurement system and data quality.
3. Establish the baseline and segment the process.
4. Generate candidate causes using process knowledge.
5. Use analytics to prioritize investigation.
6. Test causes through observation or controlled intervention.
7. Implement and monitor corrective action.

AI should accelerate disciplined inquiry, not replace it.

8. Service Blueprinting in the Digital Context

A service blueprint maps customer actions, frontstage employee or system interactions, backstage processes, support processes, and physical or digital evidence. Digital services require additional attention to interfaces, APIs, identity management, automated decisions, notification timing, and handoffs between human and machine actors.

For example, an online insurance claim may involve customer submission, automated document extraction, fraud screening, human review, payment authorization, and customer notification. A blueprint can reveal that the customer experiences a “silent” period while several internal systems exchange information. Process mining can quantify this delay; IoT or platform data can provide status information; AI can prioritize cases; and human review can handle exceptions.

The blueprint should identify failure points and control points, including:

• Where data are captured.
• Who owns the data.
• Which decision is automated.
• What happens when confidence is low.
• How the customer can challenge an outcome.
• How errors are corrected across connected systems.

Digital service quality includes not only speed but also accessibility, reliability, transparency, privacy, and recovery from failure.

9. Data Quality, Cybersecurity, Privacy, and Explainability

Digital Lean depends on trustworthy data. Important dimensions of data quality include accuracy, completeness, consistency, timeliness, validity, uniqueness, and traceability. Data quality should be managed at the point of capture rather than repaired indefinitely downstream.

Cybersecurity protects confidentiality, integrity, and availability. A connected factory or service platform should apply least-privilege access, strong identity management, network segmentation, secure configuration, vulnerability management, backups, incident response, and supplier-risk controls. Operational technology requires special care because a security action that is suitable for an office system may interrupt a safety-critical process.

Privacy requires purpose limitation, data minimization, appropriate retention, access rights, and protection against unauthorized disclosure. Employee monitoring is particularly sensitive. Collecting detailed movement or performance data without a legitimate, transparent purpose can damage trust and create discriminatory effects.

Explainability means that affected users can understand the relevant basis, limitations, and consequences of a model-supported decision. The required explanation depends on context. A maintenance engineer may need contributing sensor patterns; a customer may need the factors behind a service decision and a route for review. Technical interpretability does not by itself establish fairness or correctness.

Human oversight should be designed rather than assumed. Governance should specify:

• Which decisions may be automated.
• Confidence or risk thresholds for escalation.
• Who can override a recommendation.
• How overrides are recorded.
• How models are monitored for drift and unequal performance.
• How incidents are investigated.
• When a model must be suspended or retired.

10. Implementation Roadmap and Performance Governance

A practical digital-Lean program can follow six stages:

1. Select a value-stream problem. Define customer impact, operational pain, and boundaries.
2. Observe and stabilize. Map the current state, verify standard work, and remove obvious manual waste.
3. Assess data readiness. Identify sources, ownership, quality gaps, latency, and security constraints.
4. Design the smallest useful intervention. Pilot one workflow, signal, model, or automation with explicit success criteria.
5. Validate operational and human effects. Compare baseline and pilot results, including unintended consequences.
6. Scale with governance. Standardize interfaces, training, controls, model monitoring, and continuous improvement.

A balanced scorecard should include several categories:

Technology benefits should be evaluated over the full life cycle. Total cost includes acquisition, integration, training, support, cybersecurity, upgrades, downtime, and retirement. A low purchase price may conceal high dependence on consultants or proprietary data formats.

Illustrative Case: Digital Pull and Predictive Quality in a Components Plant

Background

The following is a fictional illustrative case. Meridian Components produces precision housings for industrial equipment. It operates three machining cells and supplies several customers with short lead times. Management planned to install sensors, electronic Kanban, and AI-based visual inspection after frequent late deliveries and customer complaints.

Problem

An initial review found that the visible problems were not purely technological. Inventory records were updated in batches, operators used different defect codes, and planners frequently expedited orders. The proposed system risked automating inaccurate information.

Analysis

A value-stream study and process-mining analysis were conducted using production-order logs. The analysis showed that average order duration was 5.2 days, but median touch time was only 3.1 hours. The large difference represented queues, movement, waiting for inspection, and rescheduling. Direct observation showed that 35% of recorded “machine downtime” was actually waiting for material or approval.

A data-quality audit found that 12% of serial numbers were duplicated and that two inspection stations used incompatible defect categories. The plant also identified a cybersecurity risk: proposed sensors were to share a network with office devices.

Approach and solution

The team first standardized defect codes, introduced barcode scans at consumption points, and separated the operational network from the corporate network. An electronic Kanban pilot was then installed for one high-volume family. Average demand was 40 units per day, lead time was 3 days, and safety stock was set at 30 units, giving a reorder point of 150 units under the stated policy.

A small vision model was trained using images from normal operating variation. It was not permitted to release product automatically. Low-confidence cases and all safety-critical defects were routed to trained inspectors. The model’s performance was reviewed by product family and lighting condition rather than only through an overall average.

A dashboard displayed material status, queue age, inspection backlog, and model exceptions. Each measure had an owner and a response routine. Operators were trained to challenge incorrect signals, and overrides were recorded for improvement analysis.

Outcome

After the pilot, late orders declined, but the most important improvement came from eliminating unrecorded waiting and clarifying replenishment responsibility. The vision system reduced routine inspection effort, while human inspectors retained authority for ambiguous cases. The plant postponed full-scale rollout until sensor reliability and supplier connectivity had been improved.

Learning points

The case demonstrates five principles. First, process understanding precedes technology selection. Second, data standards are part of the process, not merely an IT concern. Third, digital pull requires accurate transactions and disciplined replenishment. Fourth, AI inspection must be evaluated using the consequences of false acceptance and false rejection. Fifth, scaling should follow evidence from a controlled pilot rather than enthusiasm for a technology portfolio.

Critical Challenges and Contemporary Developments

Digital Lean programs face several tensions. Standardization improves integration but may reduce local flexibility. Centralized cloud platforms improve visibility but may increase concentration risk. Automation can improve safety while reducing opportunities for skill development if work is poorly redesigned. Predictive models may improve planning but reproduce historical inequities or encode past workarounds as if they were ideal practice.

Emerging developments include edge AI for low-latency inspection, synthetic data for rare-defect training, industrial foundation models, digital product passports, energy-aware scheduling, and federated learning that reduces the need to centralize sensitive data. These developments remain uneven in maturity. Managers should distinguish a tested capability from a demonstration and should demand evidence of robustness under realistic operating conditions.

Sustainability must also be included. Digital equipment consumes energy and requires materials, maintenance, and eventual disposal. Data centers and connected devices have environmental impacts. A technology decision should consider energy per transaction, equipment life, repairability, and whether the intervention reduces total resource consumption rather than merely shifting it.

Best-Practice Principles

1. Start with customer value and a specific process problem.
2. Go to the workplace and validate digital records against actual work.
3. Simplify and stabilize before automating.
4. Define data ownership, quality rules, and retention requirements.
5. Design alerts with explicit actions and escalation paths.
6. Use baseline measures and evaluate variation, not only averages.
7. Treat cybersecurity, privacy, and safety as design requirements.
8. Keep humans accountable for consequential decisions.
9. Test models for drift, bias, false positives, and false negatives.
10. Scale only after a pilot demonstrates operational, financial, and human value.

Chapter Summary

Digital Lean applies Industry 4.0 technologies to the Lean objectives of customer value, flow, pull, quality, and continuous improvement. Process mining reveals actual process variants and delays from event data, while IoT and edge computing provide timely information from physical operations. Digital twins support monitoring, prediction, and scenario analysis, but their usefulness depends on validated models and reliable data. Robotics and cloud platforms can improve repeatability and coordination, yet they also create integration, safety, dependency, and cybersecurity risks. Electronic Kanban extends pull control but cannot correct inaccurate inventory or unstable replenishment.

AI can support demand forecasting, visual inspection, anomaly detection, and root-cause investigation. Forecast and classification metrics must be interpreted in relation to operational consequences. Association is not causation, and model accuracy does not replace process knowledge or human judgment. In digital service blueprinting, interfaces, data ownership, automated decisions, exception paths, and customer recourse must be made visible.

Successful digital Lean requires disciplined sequencing: understand and stabilize the process, assess data readiness, pilot a focused intervention, measure intended and unintended effects, and scale under governance. Data quality, cybersecurity, privacy, explainability, sustainability, and human oversight are not secondary compliance concerns; they are conditions for trustworthy quality improvement.

Key Terms

Digital Lean: Use of digital technologies to improve Lean value, flow, quality, and learning.

Industry 4.0: Integration of connected cyber-physical systems, automation, analytics, and intelligent technologies.

Process mining: Analysis of event logs to discover, compare, and evaluate actual process behavior.

Event log: Structured record of activities, cases, timestamps, and related attributes.

Internet of Things: Network of physical objects that sense, communicate, and exchange data.

Edge computing: Processing data near its source rather than sending all data to a central system.

Digital twin: Dynamic digital representation linked to a physical asset, process, product, or system.

Electronic Kanban: Digital pull signal used to trigger replenishment or production.

Cloud platform: Network-delivered computing, storage, integration, and application services.

Robotic process automation: Software automation of structured, rule-based digital tasks.

Artificial intelligence: Computational methods that perform tasks involving prediction, classification, or other forms of intelligent behavior.

Machine learning: Methods that infer patterns or decision rules from data.

Model drift: Decline in model performance as data, conditions, or relationships change.

Data quality: Fitness of data for use, including accuracy, completeness, consistency, and timeliness.

Explainability: Ability to communicate how and why a model-supported result was produced.

Human oversight: Defined human authority to review, challenge, override, or stop automated decisions.

False positive: Incorrect identification of a non-event or acceptable item as an event or defect.

False negative: Failure to identify a real event or defect.

Technology-led waste: Waste created or amplified by a digital or automated intervention.

Service blueprint: Visual representation of customer actions, frontstage work, backstage work, and support processes.

Review Questions

Short Answer Questions

1. What is meant by digital Lean, and how does it differ from simply digitizing existing work?
2. Distinguish process discovery from conformance checking in process mining.
3. State the three components of OEE and explain what each represents.
4. Why is human oversight necessary in AI-supported inspection?
5. Identify four important dimensions of data quality.

Long Answer Questions

1. Explain how IoT, edge computing, cloud platforms, and real-time analytics can be integrated into a Lean operating system.
2. Critically evaluate the benefits and limitations of digital twins for manufacturing or service operations.
3. Discuss how electronic Kanban can improve pull control and why it cannot compensate for poor master data.
4. Explain the role of AI in forecasting and root-cause analysis, including the difference between association and causation.
5. Develop a governance framework for privacy, cybersecurity, explainability, and human oversight in an AI-enabled service process.

Analytical / Application Questions

1. A process has an average demand of 60 units per day, a replenishment lead time of 4 days, and safety stock of 50 units. Calculate the reorder point and state the assumptions required for the calculation.
2. A visual-inspection model detects 90 defective units correctly, incorrectly rejects 30 good units, and misses 10 defective units. Calculate precision and recall, and explain which operational risks each measure highlights.
3. Design a digital-Lean pilot for a hospital outpatient service or a manufacturing cell. Specify the baseline problem, data sources, technology intervention, human controls, success measures, and possible technology-led wastes.

Multiple Choice Questions with Answers

Question 1

Which activity is a primary use of process mining?

A. Replacing all process observation with simulations

B. Discovering actual process behavior from event logs

C. Increasing inventory without measuring demand

D. Encrypting every physical machine automatically

Correct Answer: B. Discovering actual process behavior from event logs

Question 2

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

A. Supplier speed

B. Standard service

C. Safety stock

D. System stability

Correct Answer: C. Safety stock

Question 3

If availability is 0.90, performance is 0.80, and quality is 0.95, what is OEE?

A. 0.684

B. 0.855

C. 0.950

D. 2.650

Correct Answer: A. 0.684

Question 4

Which statement best describes a digital twin?

A. A static paper map of a process

B. A dynamic digital representation linked to a physical system

C. A replacement for all physical measurements

D. A cybersecurity policy only

Correct Answer: B. A dynamic digital representation linked to a physical system

Question 5

A false negative in AI-supported inspection occurs when the system:

A. Rejects a good item

B. Accepts a defective item

C. Correctly accepts a good item

D. Correctly rejects a defective item

Correct Answer: B. Accepts a defective item

Question 6

Why should a Lean team stabilize a process before automating it?

A. Automation always reduces quality

B. Unstable methods can be automated and amplified

C. Digital systems cannot record stable processes

D. Stabilization eliminates the need for measurement

Correct Answer: B. Unstable methods can be automated and amplified

Question 7

Which is an example of technology-led waste?

A. Removing duplicate data entry

B. Reducing customer waiting time

C. Generating alerts with no defined response

D. Improving traceability of a safety-critical product

Correct Answer: C. Generating alerts with no defined response

Question 8

What is the principal purpose of edge computing?

A. To process data close to its source

B. To eliminate all cloud services

C. To increase manual paperwork

D. To prevent sensor data from being measured

Correct Answer: A. To process data close to its source

Question 9

Why is overall model accuracy insufficient for evaluating an AI inspection system?

A. Accuracy has no mathematical meaning

B. It does not show the consequences of false positives and false negatives

C. It can only be calculated for financial data

D. It automatically proves causation

Correct Answer: B. It does not show the consequences of false positives and false negatives

Question 10

Which governance practice most directly supports human oversight?

A. Removing all override rights

B. Allowing undocumented automatic decisions

C. Defining escalation thresholds and recording overrides

D. Using only historical data without review

Correct Answer: C. Defining escalation thresholds and recording overrides