12. Digital Quality, Automation and Artificial Intelligence

Introduction and Context

Quality management has traditionally depended on documented procedures, sampling, inspection, audits, statistical analysis, and the experience of employees. These practices remain important, but connected technologies are changing how quality is designed, measured, controlled, and improved. Digital quality management extends the quality management system (QMS) by using integrated data, automated workflows, sensors, advanced analytics, artificial intelligence (AI), and real-time decision support.

The central promise of digital quality is not simply faster inspection or fewer paper records. It is the movement from detecting nonconformity after an event to understanding process behaviour continuously and preventing failure before it occurs. This requires alignment among process design, information architecture, people, technology, governance, and improvement methods. A poorly designed digital system can automate waste, produce unreliable alerts, or create new risks such as cyberattack, opaque decision-making, and excessive dependence on algorithms.

For a Total Quality Management (TQM) practitioner, digitalisation must therefore be evaluated as an organisational and socio-technical change. Technology supports customer focus, process management, fact-based decisions, employee involvement, and continual improvement, but it does not replace leadership, professional judgement, or a quality culture.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the meaning and scope of digital quality management.

2. Distinguish among process mining, IoT sensing, digital twins, robotic process automation, predictive quality, and AI-assisted inspection.

3. Analyse how digital technologies support quality planning, assurance, control, and improvement.

4. Design a responsible implementation approach involving data governance, validation, cybersecurity, explainability, and human oversight.

5. Evaluate the managerial, ethical, operational, and technical risks of automated quality decisions.

6. Apply these ideas to a practical quality-improvement project.

From Conventional Quality Control to Digital Quality Management

A conventional QMS usually contains policies, procedures, work instructions, records, corrective-action systems, audits, and performance indicators. Its effectiveness depends on whether information is accurate, available, timely, and connected to decision-making. In many organisations, however, data remain divided among enterprise systems, laboratory applications, spreadsheets, maintenance records, inspection devices, and paper documents. Such fragmentation makes it difficult to see the complete process.

Digital quality management integrates quality information with operational and business systems. It may connect customer complaints with production batches, supplier records, machine conditions, inspection results, training histories, and corrective actions. The objective is a reliable flow of information across the quality lifecycle. A digital QMS can support electronic approval, document control, nonconformance management, audit planning, change control, risk registers, and management review.

Digital maturity should not be confused with the quantity of software purchased. A mature organisation has clearly defined processes, controlled data, capable users, appropriate measures, and governance over automated decisions. Digitising an unstable process without first understanding it may merely make failure occur more quickly.

A useful quality architecture contains four connected layers:

• Physical layer: products, equipment, materials, people, and environmental conditions.
• Data layer: sensor readings, transaction records, images, laboratory results, and customer information.
• Analytical layer: statistical methods, process mining, simulation, machine learning, and rules engines.
• Decision and action layer: alerts, workflow, operator response, maintenance, containment, and improvement projects.

Value is created only when the layers are connected through defined responsibilities and feedback loops.

Essential Concepts and Terminology

A digital thread is the connected flow of information about a product or process across its lifecycle. A digital QMS is a technology-enabled system for managing quality processes, records, risks, and improvement activities. Data lineage describes the origin, transformations, movement, and use of data. Master data are consistently defined reference data, such as product codes, equipment identifiers, or supplier identities. Data quality concerns attributes such as accuracy, completeness, consistency, timeliness, validity, and uniqueness.

The Internet of Things (IoT) refers to connected physical objects that sense, communicate, and sometimes act on conditions. Edge computing processes data near the source rather than sending everything to a central platform. A digital twin is a dynamic digital representation of a physical asset, process, or system that is updated with relevant data and used for monitoring, prediction, simulation, or optimisation.

Process mining uses event logs to discover, compare, and analyse how processes actually operate. Robotic process automation (RPA) uses software robots to perform structured, rule-based tasks through digital applications. Artificial intelligence refers broadly to computational systems that perform tasks associated with human cognitive abilities, including classification, prediction, language processing, and pattern recognition. Machine learning is a subset of AI in which models learn relationships from data rather than relying entirely on manually specified rules.

Predictive quality uses historical and real-time information to estimate the probability of a future defect, failure, or nonconformance. Computer vision applies image analysis to identify features, defects, or assembly conditions. Explainability concerns the extent to which the reasoning or factors behind an AI output can be understood. Human-in-the-loop means that a person participates in reviewing, approving, correcting, or acting on an automated result.

Process Mining: Seeing the Process as It Is

Traditional process mapping describes how a process is intended to operate. Process mining analyses event logs to reveal how it actually operates. An event log normally includes a case identifier, activity name, timestamp, and often information about the resource, product, location, or outcome. In a complaint process, for example, a case may be a complaint number and activities may include registration, technical assessment, customer response, corrective action, and closure.

Process mining commonly supports three activities. Process discovery constructs a model from event data. Conformance checking compares actual behaviour with a prescribed process or standard. Process enhancement uses performance and quality data to improve the model or identify causes of delay and variation. A process-mining study may reveal rework loops, unauthorised skips, excessive handoffs, or long waiting periods between approval stages.

The method is powerful only when event logs are complete and correctly structured. Missing timestamps, inconsistent case identifiers, or activities recorded outside the system can produce misleading conclusions. Process mining should therefore be combined with interviews, observation, value-stream analysis, and statistical investigation. The event log shows behaviour; it does not automatically explain why people behaved in that way.

IoT Sensing and Real-Time Quality Control

IoT sensors can measure temperature, pressure, vibration, humidity, torque, speed, energy use, location, or environmental contamination. In manufacturing, sensor data may be used to determine whether a process remained within validated conditions. In healthcare, connected devices can support monitoring of equipment or patient-care processes. In logistics, sensors can verify cold-chain conditions.

A quality application should begin with the decision that the data must support. Installing sensors without identifying a decision often creates high-volume, low-value data. The organisation must define measurement specifications, sampling frequency, calibration requirements, communication reliability, storage rules, alarm limits, and response responsibilities.

A threshold alarm is not the same as quality intelligence. A threshold may detect that a temperature has already exceeded a limit, while a predictive model may identify a developing pattern before the limit is breached. Nevertheless, predictive tools must not eliminate basic control charts, standard work, or direct investigation. A sensor can be technically accurate yet operationally useless if no one receives or acts on its signal.

Digital Twins and Simulation-Based Improvement

A digital twin links a physical object or process to a digital model that is refreshed by operational data. Unlike a static computer-aided design model, a useful twin represents current or changing conditions. It may be used to monitor equipment health, simulate production schedules, test process changes, estimate remaining useful life, or examine the effects of variation.

For quality management, digital twins support virtual experimentation. A team can examine how altered machine settings, maintenance intervals, material properties, or demand patterns might affect capability and risk before making a physical change. This can reduce experimentation time and prevent disruption. However, a twin is only as reliable as its assumptions, data, model structure, and validation. A model that represents nominal conditions but excludes operator practices, abnormal states, or supply variation may provide false confidence.

A disciplined digital-twin project defines its purpose, system boundary, variables, data sources, update frequency, model assumptions, validation evidence, and users. The twin should be periodically reviewed because equipment, software, products, and operating conditions change.

Automation, RPA, and Intelligent Workflows

Automation can improve quality by reducing transcription errors, enforcing sequence controls, creating traceability, and ensuring that required approvals occur. RPA is particularly suitable for repetitive, rule-based work such as transferring approved information between systems, generating routine reports, checking document completeness, or issuing reminders. It is less suitable where decisions require ambiguous evidence, negotiation, or professional interpretation.

Automation should be assessed using the same quality principles applied to any process change. The organisation should identify failure modes, define exception handling, test interfaces, control access rights, and maintain audit trails. A software robot that copies an incorrect value consistently can create systematic rather than random error. Automation also changes work. Employees may need to monitor exceptions, interpret trends, manage data quality, and investigate unusual cases rather than perform routine transactions.

Predictive Quality and AI-Assisted Inspection

Predictive quality models may estimate defect probability, process capability, equipment failure, customer dissatisfaction, or the likelihood that a corrective action will be ineffective. Inputs can include process parameters, material characteristics, environmental conditions, operator or shift information, historical defects, and maintenance records. The model output is not a conclusion; it is evidence to support a decision.

AI-assisted inspection is commonly used with images, acoustic signals, text, or sensor patterns. A computer-vision model may classify a surface as acceptable or suspect. A language model may help classify complaint descriptions or identify recurring themes. A diagnostic model may rank likely causes of a machine fault.

Model performance must be evaluated with appropriate measures. In classification, precision is the proportion of predicted positives that are truly positive, while recall is the proportion of actual positives that the model detects. A model with high precision but low recall may miss many defects; a model with high recall but low precision may overwhelm staff with false alarms. The acceptable balance depends on the consequences of each error. In safety-critical settings, a missed defect may be far more serious than an additional manual review.

Validation should use data that represent actual operating conditions, including variation across products, equipment, shifts, suppliers, and time. A random split of highly similar observations can produce an overly optimistic result. The model should be tested on a separate data set and monitored after deployment for drift, changing defect patterns, and changes in measurement systems.

A Governance Framework for Trustworthy Digital Quality

A practical governance framework can be organised around six questions.

Purpose: What quality decision or customer outcome is the technology intended to improve? The use case should be linked to a recognised risk, objective, or improvement opportunity.

Data: Are the data relevant, traceable, representative, secure, and legally usable? Data definitions, ownership, retention, access, and correction procedures should be documented.

Model or automation: Is the method appropriate, validated, and proportionate to the decision? Simpler statistical rules may be preferable to complex AI when they provide sufficient performance and better transparency.

Decision rights: Who may accept, reject, override, or suspend an automated recommendation? Responsibilities must remain clear even when a system acts automatically.

Controls: How are cybersecurity, change control, auditability, bias, privacy, and business continuity managed? Controls should cover both normal and exceptional conditions.

Learning: How will performance, user feedback, incidents, and unintended effects be reviewed? Digital systems should enter the continual-improvement cycle rather than be treated as finished projects.

Explainability is especially important where an automated result affects product release, employee evaluation, customer treatment, or safety. Explanations may include influential variables, comparable examples, confidence levels, decision rules, or a traceable chain of evidence. An explanation should be understandable to the person responsible for acting, not merely technically impressive.

Human oversight must be meaningful. A person who can only click “accept” without adequate time, information, authority, or training is not exercising effective oversight. Human review should be focused on high-risk cases, uncertainty, exceptions, and decisions with significant consequences.

Cybersecurity, Validation, and Regulatory Discipline

Connected quality systems enlarge the attack surface of an organisation. Threats include unauthorised access, ransomware, manipulation of sensor data, theft of intellectual property, disruption of production, and compromise of supplier or customer information. Cybersecurity is therefore a quality concern because compromised information can lead directly to nonconforming product or unsafe decisions.

Important controls include identity and access management, least-privilege permissions, network segmentation, encryption, secure backups, patch management, vulnerability assessment, incident response, logging, and supplier-security requirements. Operational technology should not be connected casually to corporate or public networks. Contingency procedures must specify how quality decisions will be made if digital systems are unavailable.

Validation demonstrates that a system is fit for its intended use. It should address requirements, installation or configuration, functional operation, performance under realistic conditions, data integrity, user access, changes, and periodic review. For AI systems, validation also includes training-data suitability, performance by relevant subgroups or operating conditions, threshold selection, false-positive and false-negative consequences, and post-deployment monitoring.

Digital records should be attributable, legible, contemporaneous, original or appropriately preserved, accurate, complete, consistent, enduring, and available. Industry-specific legal and regulatory requirements may impose additional controls, particularly in pharmaceuticals, medical devices, aviation, food, finance, and critical infrastructure.

Applied Case: Predictive Quality in a Medical-Device Assembly Line

Background

The following is an illustrative case. MedSure Devices produces a disposable fluid-control component. Final inspection identified intermittent leakage after assembly. The defect rate was low but costly because units had to be disassembled, tested, and sometimes scrapped. Quality records, machine settings, supplier-lot information, and test results were stored in separate systems.

Situation and Analysis

A cross-functional team first mapped the process and performed a measurement-system study. It found that the leak-test equipment was capable of distinguishing acceptable from defective units, although maintenance intervals needed tighter control. Process mining then showed that rework was concentrated on two assembly cells and occurred more frequently after changeovers. Sensor data indicated that insertion force and alignment varied during the first fifteen minutes after some changeovers. The initial proposal was to deploy an AI model to stop the line automatically whenever a defect was predicted.

The team rejected automatic stoppage as the first intervention. The historical data were incomplete, and the cost of a false alarm could interrupt supply. Instead, engineers standardised changeover verification, improved fixture maintenance, and created a controlled data set linking insertion force, alignment, component lot, operator confirmation, and leak-test outcome.

Approach and Solution

A predictive model was developed to classify assemblies requiring additional verification. It was tested on later production data rather than only on randomly selected historical records. The model displayed higher recall than the existing rule-based screen, but it also generated false positives during one material-lot transition. The model was therefore used as a decision-support tool: suspect units were diverted to enhanced testing, while final release remained under authorised quality personnel.

The interface displayed the principal factors associated with each alert, the model confidence, and the required response. Operators could record an override reason. A weekly review examined false alarms, missed defects, sensor availability, and changes in material or equipment. Access to the model and its data was restricted, and a manual procedure was prepared for system downtime.

Outcome

The combined intervention reduced reliance on end-of-line discovery and improved visibility of changeover risk. More importantly, the organisation learned that the predictive model was not the improvement by itself. The improvement came from measurement discipline, process understanding, standardised changeover work, controlled data, and appropriately bounded automation.

Learning Points

The case demonstrates five principles: validate the measurement system before modelling; connect event and sensor data to a defined quality decision; address process causes rather than merely predicting symptoms; use human oversight where consequences are significant; and monitor the deployed system for drift and unintended effects.

Implementation Through a Continuous-Improvement Project

A digital quality project can follow a modified DMAIC or PDCA logic. During Define, the team states the customer or process problem, risk, scope, decision owner, and expected benefit. During Measure, it assesses the current process, data-generating systems, measurement capability, and baseline performance. During Analyse, it combines process knowledge with statistical analysis, process mining, or exploratory modelling. During Improve, it may redesign work, add sensing, automate a transaction, or deploy a validated model. During Control, it establishes monitoring, response plans, change control, access management, and periodic review.

The project should include operations, quality, information technology, cybersecurity, data specialists, process owners, and affected employees. A benefits case should include not only labour savings but also defect prevention, reduced risk, improved traceability, customer outcomes, resilience, and the cost of false alarms or system failure.

Common mistakes include selecting technology before defining the problem, treating all data as reliable, ignoring process variation, measuring model accuracy without considering consequences, automating an unstable process, overlooking cybersecurity, and failing to train users. Another mistake is to create dashboards with many indicators but no decision rules. Visualisation is useful only when it leads to timely and responsible action.

Current Developments and Emerging Directions

Several developments are shaping digital quality. Edge AI allows some analysis to occur close to equipment, reducing latency and dependence on continuous connectivity. Synthetic and augmented data may support model development when rare defects are difficult to obtain, although such data must not be assumed to represent reality without validation. Generative AI is being explored for searching quality records, drafting investigation summaries, assisting root-cause analysis, and interacting with procedures. Its outputs require verification because fluent language does not guarantee factual accuracy.

Interoperability is becoming increasingly important. Quality information must move across machines, enterprise applications, laboratories, suppliers, and customers through controlled interfaces and common definitions. Organisations are also giving greater attention to algorithmic accountability, climate and energy data, product traceability, and the quality implications of autonomous systems.

The likely direction is not complete replacement of quality professionals. It is a redistribution of effort: fewer routine searches and manual transfers, and more work involving judgement, system design, risk interpretation, investigation, ethics, and improvement leadership.

Chapter Summary

Digital quality management uses connected data, automation, analytics, and AI to strengthen quality planning, assurance, control, and improvement. Process mining reveals actual process behaviour from event logs; IoT provides real-time information about physical conditions; digital twins support monitoring and simulation; RPA automates structured transactions; and predictive quality and AI-assisted inspection support earlier detection and diagnosis.

Successful application depends on more than technical performance. Data quality, measurement capability, process understanding, validation, cybersecurity, explainability, human oversight, and change management are essential. AI outputs should support accountable decisions rather than obscure them. A continuous-improvement project should begin with a meaningful quality problem, assess the current system, validate the technology, define decision rights, and monitor outcomes after deployment. Digital transformation is therefore best understood as a disciplined extension of TQM rather than a substitute for its principles.

Key Terms

Digital quality management: Use of digital systems and data to manage and improve quality processes.

Digital QMS: Technology-enabled system for controlling quality records, workflows, risks, and improvement activities.

Digital thread: Connected flow of information across a product or process lifecycle.

Data lineage: Record of where data originate, how they are transformed, and where they are used.

IoT: Network of connected physical objects that sense, communicate, or act.

Edge computing: Processing data near the point at which they are generated.

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

Process mining: Analysis of event logs to discover and evaluate actual process behaviour.

RPA: Software automation of structured, repetitive, rule-based digital tasks.

Artificial intelligence: Computational capability associated with tasks such as prediction, classification, and pattern recognition.

Machine learning: Method in which models learn relationships from data.

Predictive quality: Use of data to estimate future defects, failures, or nonconformities.

Computer vision: Automated analysis of images or video for recognition or inspection.

Explainability: Ability to understand the factors or reasoning behind an automated output.

Human-in-the-loop: Arrangement in which people review, approve, correct, or act on automated results.

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

Data integrity: Assurance that data remain accurate, complete, consistent, traceable, and protected from improper alteration.

Conformance checking: Comparison of observed process behaviour with a prescribed process model.

Validation: Evidence that a system is suitable and performs as intended for its specified use.

Cybersecurity: Protection of systems, data, and operations from unauthorised access, disruption, or manipulation.

Review Questions

Short Answer Questions

1. What distinguishes digital quality management from merely converting paper records into electronic files?

2. Define process mining and identify its three principal uses.

3. Why is human oversight important in AI-assisted quality decisions?

4. What is the difference between precision and recall in classification?

5. Identify four cybersecurity controls relevant to connected quality systems.

Long Answer Questions

1. Explain how IoT sensing, process mining, and digital twins can work together in a continuous-improvement project.

2. Discuss the conditions under which RPA is appropriate for quality management and the risks of automating an unstable process.

3. Explain a governance framework for trustworthy predictive-quality applications.

4. Analyse the role of validation in deploying an AI-assisted inspection system.

5. Discuss how digitalisation changes the responsibilities and capabilities required of quality professionals.

Analytical / Application Questions

1. A factory receives thousands of sensor alerts but experiences no improvement in defect rates. Diagnose possible causes and propose a corrective project.

2. Design a human-oversight and validation plan for an AI model that recommends whether a batch should receive additional inspection.

3. An organisation wants to connect supplier data, production records, and customer complaints into a digital thread. Identify the data, governance, cybersecurity, and change-management issues it should address.

22. Multiple Choice Questions with Answers

Question 1

Which activity is the primary purpose of process mining?

A. Replacing all process documentation with artificial intelligence

B. Discovering and analysing actual process behaviour from event logs

C. Calibrating physical sensors automatically

D. Encrypting quality records during transmission

Correct Answer: B. Discovering and analysing actual process behaviour from event logs

Question 2

A predictive-quality model has high recall but low precision. What is the most likely operational consequence?

A. It will miss most actual defects

B. It will generate many false alarms requiring review

C. It will be unable to process any new data

D. It will eliminate the need for human oversight

Correct Answer: B. It will generate many false alarms requiring review

Question 3

Which statement best describes a digital twin?

A. A static procedure stored in a document-control system

B. A duplicate employee account used for testing

C. A dynamic digital representation updated with relevant physical-system data

D. A database containing only historical inspection results

Correct Answer: C. A dynamic digital representation updated with relevant physical-system data

Question 4

Why should an organisation assess its measurement system before developing a predictive model?

A. Poor measurement can make model relationships unreliable

B. Measurement systems are unrelated to model performance

C. Models can correct every measurement error automatically

D. Validation is unnecessary when sensor data are abundant

Correct Answer: A. Poor measurement can make model relationships unreliable

Question 5

Which use case is most suitable for robotic process automation?

A. Negotiating an ambiguous customer complaint

B. Making an ethical decision about employee discipline

C. Transferring approved quality data between structured applications

D. Determining the root cause of an unprecedented failure without evidence

Correct Answer: C. Transferring approved quality data between structured applications

Question 6

Meaningful human oversight requires that the reviewer should have

A. Only the ability to click an approval button

B. Authority, relevant information, time, and competence to intervene

C. No knowledge of the process being assessed

D. A requirement to accept every high-confidence prediction

Correct Answer: B. Authority, relevant information, time, and competence to intervene

Question 7

Which control most directly reduces the risk that compromised sensor data will affect quality decisions?

A. Removing all audit trails

B. Using least-privilege access, logging, and data-integrity controls

C. Increasing the number of dashboard colours

D. Allowing unrestricted access to the sensor network

Correct Answer: B. Using least-privilege access, logging, and data-integrity controls

Question 8

In the illustrative medical-device case, why was automatic line stoppage rejected initially?

A. The organisation had no quality problem

B. The historical data were incomplete and false alarms could disrupt supply

C. Predictive models cannot be used in manufacturing

D. Operators refused to use any digital technology

Correct Answer: B. The historical data were incomplete and false alarms could disrupt supply

Question 9

What is model drift?

A. A deliberate change to a documented work instruction

B. A decline in model performance as data or operating conditions change

C. A failure of an operator to follow a calibration procedure

D. The movement of data from edge devices to cloud storage

Correct Answer: B. A decline in model performance as data or operating conditions change

Question 10

Which sequence best reflects a disciplined digital quality project?

A. Purchase technology, automate immediately, then define the problem

B. Define the quality problem, measure and analyse, validate improvement, then control performance

C. Collect unlimited data, remove human review, then publish a dashboard

D. Deploy an AI model, ignore exceptions, and revise procedures later

Correct Answer: B. Define the quality problem, measure and analyse, validate improvement, then control performance