Digital Operations, Automation and Artificial Intelligence

Introduction and context

Digital operations use connected information systems, sensors, software, automation and analytical methods to design, execute, monitor and improve processes. In total quality management (TQM), digitalisation is not an end in itself. Its value lies in improving conformity, reducing variation, strengthening traceability, shortening response time and supporting evidence-based decisions.

Contemporary operations commonly combine enterprise resource planning (ERP), cloud platforms, industrial Internet of Things (IoT), robotics, digital twins, process mining, intelligent automation, machine learning (ML) and generative artificial intelligence (AI). These technologies can create a highly visible and responsive operation, but they also introduce risks: inaccurate data, cyberattack, opaque decisions, excessive dependence on vendors, unsafe automation and weak human accountability.

A useful quality principle is that digital systems should be managed as part of the process system. A poorly designed process can be automated into a faster, more expensive failure. Therefore, technology selection should follow process understanding, customer requirements, risk analysis and governance.

Learning objectives

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

1. Explain the role and integration of ERP, cloud, IoT, robotics, digital twins and process mining.
2. Distinguish rule-based automation, intelligent automation, ML and generative AI.
3. Evaluate digital technologies using quality, cost, responsiveness, risk and people criteria.
4. Assess data quality, cybersecurity, privacy, human oversight and implementation risk.
5. Recommend a governed implementation approach for a digitally enabled process.

1. The digital operations architecture

A digital operation can be viewed as five connected layers:

The layers should not be treated as independent. For example, a temperature sensor may send data through an industrial gateway to a cloud platform; an analytics model may identify a likely equipment fault; the ERP system may create a maintenance order; and a technician may approve and complete the intervention. Quality depends on the entire chain, including sensor calibration, data interpretation and human response.

Systems integration and the digital thread

Integration means that relevant data and actions can move between systems without unnecessary manual re-entry. Common mechanisms include application programming interfaces (APIs), message queues, event streams and standard data formats. The digital thread is the traceable connection between requirements, design, procurement, production, inspection, delivery, service and improvement records.

Integration should preserve:

• a unique identity for products, assets, customers and suppliers;
• consistent definitions of units, status codes and timestamps;
• event history and audit trails;
• authorisation for data changes;
• reconciliation between source and receiving systems.

Point-to-point interfaces may be quick to create but can become difficult to maintain as the number of systems increases. An integration platform, governed master data and documented interfaces usually provide greater long-term control.

2. ERP integration and cloud platforms

Enterprise resource planning

An ERP system integrates core transactions such as sales orders, purchasing, inventory, production planning, finance, quality and maintenance. Its principal contribution to TQM is process visibility and consistency. A nonconformance can, for example, be linked to a batch, supplier, inspection result, customer order and corrective action.

ERP implementation is not merely software installation. It requires process harmonisation, role design, data migration, configuration, testing, training and change management. A critical distinction is between configuration and customisation. Configuration uses supported system options; customisation changes the software and may increase upgrade, security and support costs.

ERP integration questions include:

• Which system is the authoritative source for each data object?
• How are duplicate records prevented?
• What happens when an interface fails?
• Can transactions be traced from origin to outcome?
• Are segregation of duties and approval controls preserved?

Cloud platforms

Cloud computing provides remotely hosted computing, storage and software services. Software as a service (SaaS) supplies an application; platform as a service (PaaS) provides development and runtime capabilities; infrastructure as a service (IaaS) supplies configurable computing infrastructure. Cloud services may improve scalability, availability and collaboration, but they do not eliminate responsibility for access control, data classification, continuity or compliance.

A sound cloud assessment considers service-level agreements, data location, encryption, identity management, backup and recovery, portability, subcontractors, incident notification and exit arrangements. The shared-responsibility principle is important: the provider secures parts of the service, while the customer remains responsible for correct configuration, users, data and business processes.

3. Industrial IoT and connected quality

The IoT links physical objects to networks so that they can identify, sense, communicate and sometimes act. Industrial IoT commonly includes sensors, edge devices, gateways, networks, platforms and analytical applications.

A typical quality application is condition monitoring. Vibration, temperature, pressure, current or acoustic measurements can reveal an emerging abnormal condition. Edge processing can filter or analyse data close to the machine, reducing latency and bandwidth use; cloud processing can support wider historical analysis and coordination.

IoT value depends on measurement validity. A sensor that is incorrectly installed, uncalibrated or exposed to unsuitable environmental conditions can create false confidence. Data should therefore have known provenance, timestamp accuracy, calibration status, sampling conditions and units.

IoT can support:

• real-time process control;
• traceability of batches and assets;
• predictive or condition-based maintenance;
• energy and environmental monitoring;
• remote service and visual inspection;
• automatic alerts for control-limit breaches.

A key distinction is between a specification limit and a statistical control limit. A specification limit expresses what the customer or design permits. A control limit is estimated from process behaviour and helps identify unusual variation. They should not be treated as interchangeable.

A useful operational measure: OEE

Overall equipment effectiveness (OEE) combines availability, performance and quality for a defined equipment and time period:

where:

All three factors are ratios expressed as decimals or percentages. The calculation assumes that the period, equipment boundary and definitions are consistent. OEE is diagnostic rather than a complete measure of productivity; it does not automatically capture labour cost, maintenance cost, energy, schedule value or customer service.

Worked example. A machine is planned for 480 minutes. It operates for 432 minutes, produces 800 units, and has an ideal cycle time of 0.5 minutes per unit. Of the 800 units, 760 are good.

Availability = 432 ÷ 480 = 0.90.

Performance = (0.5 × 800) ÷ 432 = 400 ÷ 432 ≈ 0.926.

Quality = 760 ÷ 800 = 0.95.

Therefore, OEE ≈ 0.90 × 0.926 × 0.95 = 0.792, or approximately 79.2%. The largest improvement opportunity should be investigated using loss data rather than assumed from the OEE value alone.

4. Robotics and automation

Robotics uses programmable machines to perform physical tasks such as movement, welding, assembly, picking, inspection or packaging. Traditional industrial robots often operate in safeguarded areas, whereas collaborative robots are designed to work near people under specified safety conditions. “Collaborative” does not mean inherently safe; risk assessment, speed limits, guarding, tooling assessment and validated operating modes remain necessary.

Automation may be:

• fixed, with dedicated equipment for stable high-volume work;
• programmable, with software-controlled changes for product variation;
• flexible, with rapid changeover and integrated sensing;
• software-based, such as workflow or robotic process automation (RPA).

Automation is attractive where work is repetitive, hazardous, ergonomically difficult or highly consistent. It is less suitable where requirements are unstable, exceptions dominate or the business case depends on unrealistic utilisation.

Before automation, the process should be stabilised. Waste, unnecessary approvals and unclear specifications should be removed first. Otherwise automation may increase the speed at which defects, rework and customer dissatisfaction are produced.

5. Digital twins

A digital twin is a digital representation of a physical object, process or system that is linked to the physical entity through data and, where appropriate, supports analysis or control. A static CAD model is not necessarily a digital twin because it may lack operational connection and updating.

Digital twins may represent:

• a component or machine;
• a production line;
• a facility or supply network;
• a service process or customer journey.

They can be used for simulation, virtual commissioning, maintenance planning, capacity analysis, energy optimisation and operator training. Their credibility depends on model purpose, assumptions, data quality, update frequency and validation against observed behaviour. A highly detailed model is not automatically more useful; unnecessary complexity can reduce transparency and increase maintenance effort.

6. Process mining and process intelligence

Process mining analyses event logs to discover, compare and improve the actual paths followed by cases such as orders, claims, repairs or purchase requests. Each event should normally include a case identifier, activity, timestamp and, often, resource or organisational information.

The main forms are:

1. Process discovery: derives a process model from event data.
2. Conformance checking: compares actual behaviour with a reference model or policy.
3. Enhancement: uses additional data to explain performance or predict outcomes.

Process mining can reveal rework loops, bottlenecks, unauthorised routes, excessive waiting and variation between sites. It must be interpreted carefully. Missing events, inconsistent case identifiers, merged cases and timestamp errors can produce misleading models. Privacy and employee trust are also important where logs are used to assess individual behaviour.

A practical process-mining study should define the improvement question, assess event-log fitness, protect personal information, validate findings with process owners and test whether proposed changes improve outcomes.

7. Intelligent automation, machine learning and generative AI

Rule-based and intelligent automation

Rule-based automation follows explicit logic: “if condition X occurs, perform action Y.” It is transparent and appropriate when rules are stable and exceptions are limited. Intelligent automation combines workflow, document understanding, decision support, analytics or ML to handle greater variation.

RPA can copy information between applications, validate fields, generate notifications and initiate workflows. It is useful where existing systems lack integration, but it may create fragile “digital spaghetti” if interfaces change or underlying process weaknesses are ignored.

Machine learning

ML learns patterns from data for tasks such as classification, prediction, anomaly detection, recommendation and image inspection. Supervised learning uses labelled examples; unsupervised learning seeks structure without target labels; reinforcement learning learns through feedback and rewards.

A model should be evaluated on data that was not used to train it. Important concerns include overfitting, class imbalance, drift, false positives, false negatives, explainability and fairness. A model with high overall accuracy may perform poorly for a rare but safety-critical defect. Confusion-matrix measures should be selected according to the cost of errors.

For binary classification:

Precision answers, “When the model flags a case, how often is it correct?” Recall answers, “Of the actual positive cases, how many did it find?” These measures assume clearly defined positive and negative classes and reliable labels.

Generative AI

Generative AI produces text, code, images, summaries or other content from prompts and context. In operations it may draft work instructions, summarise incidents, assist service agents, search controlled knowledge bases, translate procedures or generate process documentation.

Its outputs are probabilistic and may contain fabricated statements, outdated information, confidential-data leakage or unintended bias. Retrieval-augmented generation can connect a model to approved documents, but retrieved information must still be current, authorised and correctly interpreted. Generative AI should not independently release safety-critical instructions, approve regulated decisions or alter master data without appropriate controls.

8. Data quality and information governance

Data quality is fitness for a defined use. Common dimensions include accuracy, completeness, consistency, timeliness, validity, uniqueness and traceability. Quality is contextual: a location rounded to the nearest kilometre may be adequate for strategic planning but unsuitable for robot navigation.

A data-quality programme should establish:

• data owners and stewards;
• definitions and metadata;
• validation rules at capture;
• reference and master-data governance;
• error correction and escalation;
• lineage and retention requirements;
• periodic quality measurement.

A simple completeness rate may be calculated as:

This ratio is meaningful only when required fields are correctly identified and missingness is measured consistently. A high completeness rate does not prove accuracy.

Data governance also includes privacy, lawful processing, retention, access rights and responsible secondary use. Data minimisation is particularly important: collecting more data than necessary increases exposure without guaranteeing better decisions.

9. Cybersecurity, resilience and safety

Digitally connected operations expand the attack surface. Threats include stolen credentials, ransomware, malicious software updates, denial of service, unauthorised remote access, manipulation of sensor data and attacks on operational technology. A cyber incident can become a quality, safety and continuity incident.

Core controls include:

• asset inventory and network segmentation;
• least-privilege access and multi-factor authentication;
• secure configuration and patch management;
• encryption in transit and at rest where appropriate;
• backup, recovery and tested business continuity;
• vulnerability management and monitoring;
• supplier-security requirements;
• incident response exercises;
• secure development and change control.

Operational technology requires special care because immediate patching or rebooting may be unsafe. Security controls must be coordinated with availability, production and functional safety requirements. Safety-critical automation should have independent protective mechanisms rather than relying solely on an AI prediction or network connection.

10. Human oversight, ethics and accountability

Automation changes work, authority and skill requirements. Effective design treats employees as participants in the control system, not merely as sources of data or obstacles to efficiency. Workers often possess tacit knowledge about exceptions and weak signals that are absent from databases.

Human oversight should specify:

• which decisions may be automated;
• which decisions require review;
• who can override a system;
• what evidence the reviewer receives;
• how overrides are recorded;
• how disputes and appeals are handled;
• when the system must fail safely or stop.

The human-in-the-loop model requires active approval before action. Human-on-the-loop permits automated action with monitoring and intervention. Human-in-command places strategic responsibility with people while allowing broad automation. The appropriate model depends on risk, reversibility, time pressure and legal requirements.

Ethical governance includes fairness, explainability, privacy, accessibility, non-discrimination, accountability and respect for worker dignity. A technically accurate system can still be unacceptable if it systematically disadvantages a group or makes consequential decisions without recourse.

11. Implementation framework and risk management

A disciplined implementation proceeds through stages:

1. Define value and risk. Identify customer requirements, quality problems, constraints and measurable outcomes.
2. Map the current process. Establish the baseline, variation, failure modes, handoffs and information flows.
3. Assess readiness. Review data, skills, infrastructure, cybersecurity, governance and supplier dependence.
4. Select the smallest viable use case. Prefer a contained problem with observable benefits and manageable risk.
5. Design controls before deployment. Include validation, access, auditability, fallback, human review and change control.
6. Pilot and validate. Compare performance with the baseline using representative conditions, including exceptions.
7. Train and involve users. Explain purpose, limitations, responsibilities and escalation routes.
8. Scale by standard. Reuse architecture, controls and lessons rather than copying untested configurations.
9. Monitor continuously. Review quality, drift, incidents, security, benefits and unintended consequences.

A business case should include total cost of ownership, not only licence or equipment cost. Relevant costs include integration, migration, cybersecurity, validation, training, downtime, support, model monitoring and eventual replacement.

Implementation risks can be organised as follows:

Benefits should be monitored using a balanced set of measures: first-pass yield, defect escape, cycle time, schedule adherence, downtime, energy, safety events, customer outcomes, employee experience and total cost. Measuring only automation volume can reward harmful behaviour.

12. Illustrative case: connected quality in a packaging plant

Background

The following fictional case is illustrative. A medium-sized packaging plant produces several product formats. Customer complaints concern incorrect labels and occasional seal failures. The plant has an ERP system, separate inspection software and ageing production equipment.

Problem and analysis

The quality team initially proposed installing AI cameras across all lines. Process mapping showed that the larger problem was not only inspection. Product and label data were manually re-entered from ERP into line systems, changeovers were poorly documented, and inspection images lacked reliable batch identifiers. An exploratory review of event logs also found that many holds were released without a recorded reason.

A risk assessment identified label mismatch as a high-severity failure. The team therefore defined a controlled use case: verify label identity at changeover and record the result against the batch. It did not initially automate final release.

Approach and solution

The plant created a master-data ownership policy for product, label and batch identifiers. An ERP interface supplied approved production information to the line system. A barcode reader checked the physical label, while a vision model identified orientation and print defects. The model’s confidence threshold was established using a validation set containing normal products, known defects and realistic lighting variation.

Low-confidence cases were routed to a trained inspector. The inspection decision, image reference, operator identity, equipment status and batch identifier were recorded in an audit trail. Network segmentation separated production equipment from general office traffic. A fallback checklist allowed safe operation during system unavailability.

Before scale-up, the team compared four weeks of baseline data with a pilot. It reviewed false accepts, false rejects, changeover time, unplanned stops, operator overrides and customer complaints. The team also conducted failure-mode analysis for sensor failure, wrong master data, network loss and unauthorised changes.

Outcome

The pilot reduced label-related escapes and improved traceability. Some false rejects increased at first because lighting changed between shifts. Rather than lowering the threshold blindly, the team corrected lighting, retrained the model with representative images and retained human review for uncertain cases. The project was expanded only after process owners accepted the evidence and controls.

Learning points

The principal lesson is that technology did not create the result alone. Value came from master-data governance, process clarification, controlled integration, representative validation, cybersecurity and human accountability. The plant also avoided a common error: treating a model’s confidence score as proof of product conformity.

13. Contemporary developments and professional implications

Operations managers increasingly work with edge AI, low-code automation, digital engineering, connected supply chains and sustainability analytics. Foundation models may make knowledge work more accessible, but their governance requirements are becoming more important. Regulatory approaches are also moving toward risk-based controls, transparency, documentation, data protection and accountability, although exact legal duties vary by jurisdiction and application.

Professionals in TQM need a hybrid capability. They must understand process capability, root-cause analysis and audit evidence while also being able to question data lineage, model validation, access controls and system boundaries. Quality roles are expanding from inspection of outputs to assurance of the digital process that produces those outputs.

Best practice can be summarised in seven principles:

1. Start with customer and process value.
2. Make data ownership and definitions explicit.
3. Validate technology under real operating conditions.
4. Design cybersecurity and safety from the beginning.
5. Keep consequential accountability with appropriately authorised people.
6. Measure outcomes, unintended effects and total cost.
7. Treat digital systems as continually changing processes requiring audit and improvement.

Chapter Summary

Digital operations integrate physical work, data, applications and governance. ERP systems coordinate enterprise transactions; cloud platforms provide scalable services; IoT connects assets and processes; robotics automates physical tasks; digital twins support connected representation and simulation; and process mining reveals how work actually occurs. Intelligent automation combines workflows and analytics, while ML supports prediction and classification and generative AI assists with controlled content and knowledge tasks.

Technology creates quality value only when process requirements, data quality, cybersecurity, human capability and governance are addressed together. OEE can diagnose equipment losses, but it is not a complete business-performance measure. ML must be evaluated with suitable error measures, and generative AI requires strong controls because plausible output may be incorrect or confidential. Implementation should be staged, risk-based and measurable, with human oversight for consequential decisions. The central TQM lesson is that digital transformation is a socio-technical improvement programme, not simply an equipment or software project.

Key Terms

ERP: An integrated system for managing enterprise transactions and resources.

Cloud computing: Delivery of computing services over remotely managed infrastructure or platforms.

Industrial IoT: Networked sensing, communication and control for industrial assets and processes.

Edge computing: Processing data near its source rather than exclusively in a central cloud.

Digital thread: Traceable data connection across the life cycle of a product or process.

Robotic process automation: Software automation that performs structured tasks across digital applications.

Collaborative robot: A robot designed for specified operation in proximity to people under safety controls.

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

Process mining: Analysis of event logs to discover, check or improve process behaviour.

Machine learning: Methods that infer patterns from data to support predictions or decisions.

Generative AI: AI that creates new content such as text, code or images from learned patterns and prompts.

Data lineage: The documented origin, movement and transformation of data.

Master data: Shared reference data about core entities such as products, customers and suppliers.

Data drift: Change over time in the data or relationships on which a model depends.

Precision: The proportion of model-positive predictions that are true positives.

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

Human-in-the-loop: A control arrangement requiring human participation before a consequential action.

Operational technology: Hardware and software that monitors or controls physical processes.

Total cost of ownership: All significant acquisition, operation, support, risk and replacement costs.

Review Questions

5 Short Answer Questions

1. What is the main quality contribution of ERP integration?
2. Distinguish a digital twin from a static digital model.
3. Name the three components of OEE.
4. What is process mining used to discover?
5. Why is data completeness alone insufficient to establish data quality?

5 Long Answer Questions

1. Explain how ERP, cloud platforms and IoT can be integrated into a quality-controlled digital operation.
2. Compare fixed automation, robotics, RPA, ML and generative AI in terms of suitable applications and risks.
3. Discuss the importance of human oversight in AI-enabled operational decisions.
4. Explain the cybersecurity and resilience controls required for connected operational technology.
5. Develop a risk-based implementation framework for a digital transformation project.

3 Analytical / Application Questions

1. A plant reports high OEE but increasing customer complaints. Analyse how this situation could occur and recommend additional measures.
2. Design a process-mining investigation for delayed service claims. Specify the event data required, possible findings and privacy safeguards.
3. A manager proposes using generative AI to approve safety-critical maintenance instructions without human review. Evaluate the proposal and design a safer alternative.

Multiple Choice Questions with Answers

Question 1

Which system is primarily designed to integrate transactions such as purchasing, inventory, production and finance?

A. ERP

B. Digital twin

C. Vision sensor

D. Process-mining model

Correct Answer: A. ERP

Question 2

In the OEE calculation, which factor is calculated as Good Units ÷ Total Units?

A. Availability

B. Performance

C. Quality

D. Utilisation cost

Correct Answer: C. Quality

Question 3

A digital twin differs from a static CAD model mainly because it is normally:

A. Printed in three dimensions

B. Connected to operational data and the physical entity

C. Limited to financial information

D. Used only for product advertising

Correct Answer: B. Connected to operational data and the physical entity

Question 4

Which process-mining technique compares actual event-log behaviour with a reference process model?

A. Process discovery

B. Conformance checking

C. Data encryption

D. Image segmentation

Correct Answer: B. Conformance checking

Question 5

A model has 90 true positives and 10 false positives. Its precision is:

A. 80%

B. 90%

C. 91%

D. 100%

Correct Answer: C. 91%

Question 6

Which statement best describes edge computing?

A. It processes data close to where it is generated

B. It eliminates the need for sensors

C. It stores every record only on paper

D. It prevents all cyberattacks

Correct Answer: A. It processes data close to where it is generated

Question 7

Which is a major risk of generative AI in operational documentation?

A. It cannot produce any text

B. It may generate plausible but incorrect information

C. It always requires a physical robot

D. It automatically guarantees regulatory compliance

Correct Answer: B. It may generate plausible but incorrect information

Question 8

Which control most directly supports least-privilege cybersecurity?

A. Giving every employee administrator access

B. Providing users only the access required for their roles

C. Removing audit trails

D. Disabling all backups

Correct Answer: B. Providing users only the access required for their roles

Question 9

Why should a process generally be stabilised before automation?

A. Automation always reduces quality

B. Unstable processes can automate waste and defects

C. Stable processes cannot use technology

D. Automation is unrelated to process design

Correct Answer: B. Unstable processes can automate waste and defects

Question 10

Which measure answers the question, “Of the actual positive cases, how many did the model find?”

A. Recall

B. Precision

C. Availability

D. Completeness

Correct Answer: A. Recall