Artificial Intelligence, Excellence Models and Future Quality

Introduction

Quality management is entering a period in which inspection, process control and organisational learning are increasingly supported by artificial intelligence (AI). Cameras can identify surface defects at production speed; machine-learning models can predict equipment failure; natural-language systems can classify complaints and retrieve corrective-action knowledge; and decision-support tools can integrate process, supplier and customer data. These capabilities do not replace quality management. They change its scope, speed and risk profile.

The central managerial question is not whether an organisation can deploy AI, but whether AI produces trustworthy decisions that improve customer value, process capability, compliance, resilience and learning. A technically accurate model may still damage quality if its data are biased, its recommendations are not understood, or its use weakens human accountability. Conversely, an apparently simple rule-based system may create substantial value when it is well governed and embedded in a mature quality system.

This chapter examines AI-enabled inspection, predictive quality and decision support; relates them to excellence and maturity models; and evaluates human oversight, algorithmic bias, resilient quality systems and emerging industrial trends.

Learning objectives

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

1. Explain the role and limitations of AI in inspection and quality assurance.
2. Distinguish descriptive, diagnostic, predictive and prescriptive quality analytics.
3. Evaluate classification models using appropriate quality metrics.
4. Relate AI capability to quality, excellence and digital-maturity models.
5. Design proportionate governance, human oversight and bias controls.
6. Apply the concepts to a realistic quality-management case.
7. Assess future trends and their implications for quality professionals.

1. From inspection to intelligent quality

Traditional inspection is commonly positioned at the end of a process. Its principal question is whether a product conforms to specification. Modern quality management seeks earlier intervention: understanding variation, preventing defects, learning from incidents and optimising the whole value stream. AI extends this movement by finding patterns in large, fast or complex datasets.

Four analytical purposes are useful:

AI includes machine learning, computer vision, natural-language processing, optimisation and knowledge-based systems. Machine learning learns relationships from examples or feedback; it does not automatically establish causation. A correlation between humidity and defects, for example, may be useful for prediction but does not prove that humidity is the only or direct cause.

A quality application should therefore begin with a defined decision and risk, not with a fashionable algorithm. The organisation should specify the critical-to-quality characteristic, acceptable error, response time, data requirements, escalation path and evidence needed for release or regulatory compliance.

2. AI-enabled inspection

2.1 Computer vision

A vision system typically contains illumination, optics, an image sensor, processing software and an interface to the production or quality system. It may detect scratches, missing components, incorrect assembly, dimensional deviations or label errors. Models can be trained using labelled images of conforming and nonconforming products, or can identify departures from a learned normal condition.

The physical design is as important as the model. Variable lighting, lens contamination, vibration, product orientation and camera position can create false alarms. A robust validation plan must test representative materials, colours, shifts, operators, environmental conditions and defect severities.

2.2 Classification performance

For a binary inspection, a confusion matrix distinguishes actual condition from model decision:

Important measures are:

Precision is the proportion of flagged items that are genuinely defective. Recall, also called sensitivity, is the proportion of actual defects detected. Specificity is the proportion of conforming items correctly accepted. These measures are dimensionless and apply when the relevant cells of the confusion matrix are defined consistently.

A further summary is the F1 score:

It is the harmonic mean of precision and recall and is useful when both are important. It can conceal business priorities, however. In a safety-critical application, a false negative may be much more serious than a false positive, so recall and risk-weighted consequences should receive priority rather than relying on F1 alone.

Worked example: inspection performance

An inspection model examines 1,000 units. It correctly identifies 72 defective units, misses 8 defects, and incorrectly rejects 30 conforming units. Thus TP = 72, FN = 8, FP = 30 and TN = 890.

Recall = 72 ÷ (72 + 8) = 0.90, or 90%.

Precision = 72 ÷ (72 + 30) ≈ 0.706, or 70.6%.

Specificity = 890 ÷ (890 + 30) ≈ 0.967, or 96.7%.

The model detects most defects, but approximately three of every ten flagged units are actually conforming. The operational cost of reinspection and unnecessary scrap must be compared with the cost of missed defects before setting a decision threshold.

2.3 Validation and deployment

A responsible inspection lifecycle includes:

1. Problem definition: identify the defect, unit of analysis and release consequence.
2. Data preparation: establish representative images, labels, traceability and access controls.
3. Model development: separate training, validation and final test data. Avoid placing near-identical images in different sets, because this produces over-optimistic performance.
4. Process validation: test under realistic production conditions and compare with competent human inspection.
5. Controlled release: define approval authority, fallback inspection and change control.
6. Performance monitoring: monitor drift, false alarms, missed defects and subgroup performance.
7. Retirement or retraining: revise the model when products, equipment, suppliers or environments change.

A model should not silently make a disposition decision when uncertainty is high. An effective design uses a confidence band: automatic acceptance or rejection for clear cases, and human review for ambiguous cases. The threshold must be established from risk and process economics, not chosen solely to maximise an abstract score.

3. Predictive quality and decision support

Predictive quality combines historical quality records with process, equipment, environmental, supplier and usage data. Applications include predicting dimensional nonconformance, remaining useful life, warranty claims, complaint escalation and supplier delivery risk.

The quality professional must distinguish prediction from control. A prediction that a defect is likely does not identify the best intervention. Decision support should connect the prediction to a controllable cause, an authorised action and a way to verify the result.

A typical architecture has five layers:

• Sensing: machines, laboratory systems, inspection devices and people generate observations.
• Integration: identifiers, timestamps, units and genealogy connect data across systems.
• Analytics: statistical models, machine learning or rules estimate risk.
• Decision: a person or authorised workflow selects an action.
• Learning: the result of the action is recorded and used to improve the process and model.

Data quality is a quality issue. Missing values, inconsistent units, delayed entries, changing definitions and poor master data can undermine an advanced model. Data lineage should show where a value originated, how it was transformed and which model version used it.

Process monitoring and drift

A model trained under one operating regime may become unreliable after a new supplier, recipe, tool, software version or workforce pattern is introduced. This is known as data or concept drift. Monitoring should include input distributions, prediction distributions, confirmed outcomes, calibration and performance by relevant product or operating group.

Calibration asks whether predicted probabilities correspond to observed frequencies. If a model assigns a 0.8 defect probability to a group, approximately 80% of comparable cases should be defective over a sufficiently large and stable sample. A high ranking ability without reliable probabilities may be inadequate for resource allocation.

Decision-support interfaces should show the evidence used, uncertainty, model version, recommended action and escalation route. Explanations are not merely a user-experience feature: they support challenge, investigation, auditability and learning. However, a simple explanation must not falsely imply that a complex model has established causation.

4. Excellence models and digital maturity

AI should be evaluated within an organisational system rather than as an isolated technology. Excellence models provide a broader benchmark. The EFQM Model, the Baldrige Excellence Framework and national quality awards differ in terminology, but commonly examine leadership, strategy, customers or stakeholders, people, processes, results, learning and improvement. ISO 9001 is a requirements standard for a quality management system, whereas excellence models generally provide a holistic framework for assessing organisational performance and maturity.

Maturity models usually describe progression from ad hoc practice to managed, measured, integrated and optimised capability. The labels vary, so organisations should define observable evidence for each level. A useful AI-quality maturity scale is shown below.

This scale is not a substitute for a recognised external model. It is a diagnostic aid that can be mapped to leadership, strategy, people, process and results criteria. A high technical maturity with weak leadership accountability is not organisational excellence.

Benchmarking should compare more than model accuracy. Relevant dimensions include:

• customer and safety outcomes;
• defect prevention and cost of poor quality;
• process capability and response time;
• workforce competence and adoption;
• data governance and cybersecurity;
• supplier and ecosystem integration;
• fairness, accessibility and explainability;
• resilience under disruption; and
• evidence of sustained improvement.

Benchmarking must also control for context. A hospital, semiconductor plant and public service cannot be ranked fairly using a single defect metric. Comparisons should state scope, definitions, time period, product complexity and risk environment.

5. Human oversight, ethics and governance

Automation bias occurs when people accept a system recommendation too readily. The opposite error, algorithm aversion, occurs when users reject useful evidence because they distrust automation. Both risks require designed roles rather than vague instructions to “use judgement”.

Human oversight should specify:

• who may approve, override or suspend the system;
• which decisions require two-person or specialist review;
• what evidence must be recorded;
• how disagreements are resolved;
• when the model is automatically placed in a safe mode; and
• how affected customers, employees or suppliers can seek correction.

The principle of meaningful human control is especially important when decisions affect safety, employment, access to services or legal rights. Human involvement is not meaningful if the reviewer lacks time, competence, authority or access to the relevant evidence.

Algorithmic bias

Bias may enter through unrepresentative sampling, historical discrimination, inconsistent labels, proxy variables, unequal measurement quality or a business process that treats groups differently. Removing a sensitive attribute does not necessarily remove bias because correlated variables may remain.

Controls include representative sampling, subgroup performance analysis, label review, independent validation, impact assessment, documented limitations and post-deployment monitoring. Fairness measures must be selected for the application; no single metric resolves every ethical trade-off. For example, equal false-negative rates may conflict with equal positive predictive values when base rates differ.

Governance controls

A practical governance framework includes:

1. an inventory of AI systems and their risk classification;
2. a named accountable owner and technical custodian;
3. documented purpose, data sources, limitations and intended users;
4. validation evidence and acceptance criteria;
5. cybersecurity, privacy and access controls;
6. versioning, change control and audit logs;
7. incident reporting, rollback and business continuity;
8. periodic independent review; and
9. training for users and affected personnel.

Applicable law varies by jurisdiction and sector. Organisations must consider privacy and data-protection law, product-safety obligations, employment rules, intellectual-property rights, contractual duties and emerging risk-based AI regulation. Compliance is a minimum condition, not proof that a system is ethical or effective.

6. Resilient quality systems

Resilience is the ability of a quality system to anticipate, absorb, adapt to and recover from disruption while maintaining acceptable outcomes. Disruptions may include cyberattacks, extreme weather, pandemics, supplier failure, data loss, equipment breakdown or abrupt demand changes.

AI can improve resilience by detecting weak signals, simulating scenarios and prioritising scarce inspection or maintenance resources. It can also reduce resilience if the organisation depends on one opaque vendor, one data pipeline or continuous connectivity. A resilient design therefore includes redundancy, manual fallback, tested recovery procedures, offline records where necessary and transparent supplier arrangements.

Quality risk management should connect failure modes to controls. A conventional failure mode and effects analysis (FMEA) considers severity, occurrence and detectability, but organisations should avoid treating a numerical priority score as an objective truth. The assessment is a structured judgement requiring evidence, cross-functional review and periodic revision. AI may assist with evidence retrieval and pattern detection, but accountable experts remain responsible for risk acceptance.

7. Illustrative case: AI inspection in a medical-device assembly plant

Background

A fictional medical-device plant assembles sterile connectors. Manual visual inspection is slow and fatigue-sensitive. Defects include incomplete bonding, contamination and incorrect component orientation. Because a missed defect could affect patient safety, the plant cannot treat the AI system as an unrestricted replacement for qualified release personnel.

Problem

The plant experiences inconsistent inspection results across shifts, rising reinspection cost and delayed release. Management proposes a camera-based classifier connected directly to product disposition.

Analysis

A cross-functional team maps the process, identifies critical defects and reviews historical images. It discovers that most training images come from day shift and one product colour. Defect labels also vary between inspectors. A model trained on this dataset would risk poor performance on night-shift lighting and less frequent but severe contamination defects.

The team defines acceptance criteria separately for each defect type. For safety-critical defects, the cost of a false negative is treated as substantially higher than the cost of a false positive. Images are collected across shifts, equipment states, lots and defect severities. A separate test set is held back by production lot rather than by random image alone.

Approach and solution

The system is deployed in stages. Initially it provides a second opinion while human inspectors retain authority. High-confidence conforming images are accepted only after validation; uncertain cases and all suspected critical defects go to qualified review. The system records image, model version, confidence, decision, reviewer action and final disposition.

A dashboard tracks recall, precision, false-negative investigations, subgroup performance, lighting conditions and drift. A change in camera position automatically triggers a validation check. If image quality falls below a defined threshold, the system enters a safe state and manual inspection continues.

Outcome

In the illustrative scenario, reinspection effort declines and consistency improves, but the plant does not claim that AI alone guarantees quality. The most important gains arise from standardised defect definitions, improved lighting, traceability and disciplined escalation. The system is accepted as a controlled aid within the quality management system, not as an independent quality authority.

Learning points

1. The business and safety decision must be defined before model selection.
2. Representative data and reliable labels are essential.
3. Human oversight must include authority, competence and time.
4. Performance must be monitored after deployment.
5. Process redesign may create more value than algorithmic sophistication.

8. Emerging trends and professional implications

Several developments are likely to influence future quality practice:

• Edge AI: processing data near the machine can reduce latency and dependence on cloud connectivity, but increases device-management and cybersecurity requirements.
• Generative AI: language systems can draft reports, search procedures and summarise complaints. They may hallucinate, expose confidential data or reproduce incorrect procedures, so outputs require verification and controlled retrieval sources.
• Digital twins: connected models of products and processes support scenario analysis and predictive maintenance, provided that model assumptions and data validity are understood.
• Autonomous laboratories and adaptive process control: experimentation and adjustment may become faster, but release boundaries and safe operating limits must remain explicit.
• Federated and privacy-preserving analytics: organisations may learn across sites without centralising sensitive data, although differences in local data quality can still bias results.
• Traceable and sustainable quality: carbon, resource use, repairability and circularity will increasingly become quality characteristics alongside conformance and reliability.
• Human–AI collaboration: future competence will combine metrology, statistics, process knowledge, data literacy, ethics and communication.

Quality professionals will therefore act as system designers, risk interpreters and governance leaders. They must be able to challenge a model, explain uncertainty to decision-makers and ensure that improvement remains aligned with customer and societal value.

Best-practice checklist

Before approving an AI-enabled quality application, ask:

• Is the quality problem and decision clearly defined?
• Are data representative, traceable, secure and legally usable?
• Are labels, units and sampling procedures controlled?
• Has performance been tested on genuinely independent and realistic data?
• Are false-positive and false-negative consequences understood?
• Is there a human escalation and manual fallback process?
• Are model drift, cybersecurity and supplier dependency monitored?
• Can an auditor reconstruct the decision and model version?
• Are affected people treated fairly and given a route for correction?
• Does the application improve the wider quality system rather than merely automate an isolated task?

Chapter Summary

AI can strengthen quality by enabling rapid inspection, predictive risk detection, pattern recognition and decision support. Its value depends on the quality of the process, data, governance and human response surrounding it. Confusion-matrix measures such as precision, recall and specificity describe different aspects of inspection performance and must be interpreted in relation to risk and cost.

Excellence models provide a system-wide perspective covering leadership, strategy, people, processes, stakeholders and results. Maturity assessment helps organisations progress from reactive pilots to integrated and adaptive capability, but technical sophistication alone is not excellence. Human oversight, bias controls, explainability, cybersecurity, change control and fallback arrangements are essential. Resilient quality systems use AI where it improves anticipation and adaptation while preserving accountability and continuity. Future quality professionals will combine domain expertise with data literacy, critical evaluation and ethical governance.

Key Terms

Artificial intelligence: Computational systems that perform tasks involving learning, reasoning, perception or decision support.

Algorithmic bias: Systematic and unfair performance differences arising from data, design or use.

Calibration: Agreement between predicted probabilities and observed frequencies.

Computer vision: AI-based analysis of images or video.

Concept drift: Change in the relationship between inputs and outcomes over time.

Confusion matrix: Table comparing predicted classifications with actual conditions.

Decision support: Information or recommendations that assist, but do not necessarily replace, human decisions.

Digital twin: A connected digital representation of a physical product, process or system.

Excellence model: A framework for assessing holistic organisational performance and improvement.

False negative: An actual defect incorrectly classified as conforming.

False positive: A conforming item incorrectly classified as defective.

Human oversight: Organised human authority to review, challenge, approve or stop an automated system.

Maturity model: A staged framework describing progression in capability.

Predictive quality: Use of data and models to estimate future quality outcomes.

Precision: The proportion of positive model decisions that are truly positive.

Recall: The proportion of actual positive cases detected by a model.

Resilience: The ability to maintain and recover acceptable performance during disruption.

Specificity: The proportion of actual conforming cases correctly identified.

Traceability: The ability to follow data, decisions, products or changes through their lifecycle.

Review Questions

Short Answer Questions

1. Distinguish predictive quality from prescriptive quality analytics.
2. What is the difference between precision and recall?
3. Why is representative training data important in AI inspection?
4. State two functions of human oversight.
5. How can a quality system become less resilient through inappropriate AI deployment?

Long Answer Questions

1. Explain how a computer-vision inspection system should be validated and monitored after deployment.
2. Compare the purposes of ISO 9001, excellence models and maturity models in evaluating AI-enabled quality.
3. Discuss the main sources of algorithmic bias and propose controls for reducing them.
4. Explain how AI can support predictive maintenance and process improvement without establishing causation automatically.
5. Discuss the governance requirements for using generative AI in a regulated quality environment.

Analytical / Application Questions

1. An inspection model produces TP = 90, FN = 10, FP = 20 and TN = 880. Calculate recall, precision and specificity, and recommend which metric should dominate if missed defects create a serious safety risk.
2. Design a staged implementation plan for AI inspection in a food-processing plant where lighting, product appearance and seasonal raw materials vary.
3. An organisation has excellent model accuracy but frequent outages, no manual fallback and unclear accountability. Evaluate its maturity and recommend three priorities for creating a resilient quality system.

Multiple Choice Questions with Answers

Question 1

Which measure is the proportion of actual defects detected by an inspection model?

A. Precision

B. Recall

C. Specificity

D. Calibration

Correct Answer: B. Recall

Question 2

A model records TP = 80 and FN = 20. What is its recall?

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 3

Which statement best describes concept drift?

A. The model has too many variables at training time

B. The relationship between inputs and outcomes changes over time

C. A camera produces higher-resolution images

D. A user forgets a password

Correct Answer: B. The relationship between inputs and outcomes changes over time

Question 4

Which activity is most appropriate before connecting an AI inspection model to automatic product release?

A. Remove all human review

B. Validate it under representative production conditions

C. Maximise the number of model parameters

D. Use only data from the easiest production shift

Correct Answer: B. Validate it under representative production conditions

Question 5

What does precision measure?

A. The proportion of flagged items that are truly defective

B. The proportion of conforming items correctly accepted

C. The proportion of defects that are missed

D. The time required to train a model

Correct Answer: A. The proportion of flagged items that are truly defective

Question 6

Which characteristic most clearly distinguishes an excellence model from ISO 9001 requirements?

A. An excellence model provides a holistic performance and improvement framework

B. An excellence model eliminates the need for documented processes

C. ISO 9001 is exclusively an AI standard

D. An excellence model applies only to laboratory measurements

Correct Answer: A. An excellence model provides a holistic performance and improvement framework

Question 7

Which is an example of meaningful human oversight?

A. A reviewer sees only the final recommendation and cannot override it

B. A trained reviewer can examine evidence, challenge the output and stop the process

C. A manager signs every output without reviewing any evidence

D. A system hides uncertainty to prevent disagreement

Correct Answer: B. A trained reviewer can examine evidence, challenge the output and stop the process

Question 8

Which condition is most likely to create algorithmic bias?

A. Representative sampling across operating conditions

B. Independent testing by production lot

C. Training data that underrepresent an important product group

D. Documented model version control

Correct Answer: C. Training data that underrepresent an important product group

Question 9

Which design feature most directly supports resilience when an AI service fails?

A. A manual fallback process that is tested periodically

B. Dependence on a single unverified data feed

C. Removing audit logs

D. Allowing unreviewed automatic changes

Correct Answer: A. A manual fallback process that is tested periodically

Question 10

A plant has TP = 72, FN = 8, FP = 30 and TN = 890. Approximately what is its specificity?

A. 70.6%

B. 90.0%

C. 96.7%

D. 98.9%

Correct Answer: C. 96.7%