13. Digital Operational Excellence, Automation and AI

Introduction: From Process Control to Intelligent Operations

Operational excellence has traditionally been associated with stable processes, waste reduction, statistical control, standardisation and continuous improvement. Digital technologies extend these capabilities by making processes more visible, connected, adaptive and increasingly autonomous. Process mining can reveal how work is actually performed; workflow platforms can coordinate people and systems; the Internet of Things (IoT) can provide real-time operational data; digital twins can simulate alternative decisions; robotic process automation (RPA) can execute repetitive rules-based tasks; and advanced analytics and artificial intelligence (AI) can support prediction, optimisation and decision-making.

Digital operational excellence, however, is not achieved by purchasing technology or automating every activity. Technology can accelerate a good process, but it can also accelerate waste, defects, non-compliance and poor customer experiences. The central managerial question is therefore not “What can be automated?” but “What outcome should the process deliver, how should the process be redesigned, and where can digital capabilities create superior value?”

This chapter examines digital tools as components of an integrated improvement system. It emphasises process redesign before automation, data quality, cybersecurity, governance and human oversight. These issues are particularly important in total quality management because quality is not merely conformance to a specification; it also includes reliability, responsiveness, safety, fairness and the ability to learn from evidence.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the meaning and dimensions of digital operational excellence.

2. Distinguish process mining, workflow management, RPA, IoT, digital twins, advanced analytics and AI.

3. Evaluate technologies for prediction, optimisation, automation and decision support.

4. Apply a structured method for redesigning and automating processes.

5. Diagnose data-quality, cybersecurity, ethical and governance risks.

6. Design appropriate human oversight and performance measures.

7. Recommend a digitally enabled improvement approach for a real or illustrative organisation.

The Architecture of Digital Operational Excellence

Digital operational excellence can be understood as the coordinated capability to design, execute, measure, improve and govern processes through digital information and intelligent technologies. It has five mutually dependent dimensions.

Process discipline provides clear objectives, ownership, standard work, controls and improvement routines. Data visibility connects operational events, resources, customers, equipment and outcomes. Decision intelligence converts data into descriptive, diagnostic, predictive and prescriptive insight. Automation executes suitable tasks consistently and rapidly. Organisational capability ensures that people can interpret results, challenge systems, manage exceptions and improve the operating model.

These dimensions form a socio-technical system. A technically sophisticated solution may fail if employees do not trust it, if incentives encourage workarounds, or if process ownership is unclear. Conversely, a modest digital intervention can create considerable value when it addresses a well-defined bottleneck and is supported by disciplined management.

Important concepts and terminology

A digital process is a process whose activities, decisions or records are substantially supported by digital systems. Event data are time-stamped records of activities, such as an order being approved or a machine generating an alarm. A process variant is a distinct route through a process. Straight-through processing completes a transaction without manual intervention. An exception is a case that falls outside defined rules or normal patterns.

Automation means using technology to perform tasks with limited human intervention. RPA uses software agents, often called bots, to interact with applications through user interfaces or system connections. Workflow orchestration coordinates activities, information, responsibilities and approvals across systems and people. Analytics includes methods for describing what happened, diagnosing why it happened, predicting what may happen and prescribing what should be done.

Machine learning is a class of computational methods that identify patterns from data and use them for classification, prediction or other tasks. Artificial intelligence is the broader field concerned with systems that perform activities commonly associated with human intelligence, such as perception, reasoning, language processing and learning. Explainability concerns whether the basis of a model’s output can be understood by relevant users. Model drift occurs when changes in the environment reduce a model’s performance over time.

A digital twin is a digitally represented model of a physical asset, process or system that is connected to relevant operational data and can be used for monitoring, simulation or control. Data lineage records the origin, transformation and use of data. Data governance establishes accountability, standards, controls and decision rights for data. Human-in-the-loop means that a person reviews, approves or intervenes in an automated decision or action.

Seeing the Process as It Is: Process Mining

Traditional process mapping usually represents how a process is intended to work. Process mining reconstructs how it actually works by analysing event logs from enterprise resource planning, customer relationship management, service, production or financial systems. A minimum event log generally contains a case identifier, an activity and a timestamp; richer logs include resources, locations, costs, materials and outcomes.

Process mining has three principal uses. Process discovery generates a model from event data. Conformance checking compares actual behaviour with a prescribed model, identifying deviations, rework, skipped controls or unauthorised routes. Enhancement adds performance or contextual data to explain delays, defects, variation or cost.

Its value lies in revealing hidden complexity. A process described as “order to cash” may contain dozens of variants, repeated approvals and long waiting periods. A process-mining study can distinguish processing time from waiting time, identify the activities associated with late delivery and show whether a policy is causing unnecessary handoffs.

Process mining does not automatically explain causation. An observed correlation between a particular route and poor performance may reflect product type, customer segment or workload. Consequently, process mining should be combined with interviews, operational knowledge, quality data and controlled improvement experiments. Data completeness is also critical: undocumented work, offline activity and inconsistent case identifiers can produce a misleading model.

Workflow Platforms and Intelligent Automation

Workflow platforms coordinate tasks across employees, applications, documents and rules. They can route work, enforce segregation of duties, send alerts, collect approvals, maintain audit trails and expose service-level performance. Low-code and no-code tools make process configuration accessible to business users, but governance is required to prevent uncontrolled “shadow systems.”

RPA is particularly appropriate when a task is repetitive, rule-based, high-volume, digitally accessible and relatively stable. Examples include extracting data from standard forms, reconciling records, transferring information between legacy systems and generating routine notifications. RPA is less suitable when the process changes frequently, requires complex judgement, depends on ambiguous documents or involves poorly structured exceptions.

Automation should be assessed at the level of the complete process rather than the individual task. Automating a defective data-entry step may increase transaction volume without improving the customer outcome. A better sequence is:

1. Define the customer, regulatory and operational outcome.

2. Map the current process using evidence, including exceptions and informal work.

3. Remove unnecessary approvals, duplication, handoffs and data capture.

4. Standardise policies, definitions and decision rules.

5. Select the appropriate technology and redesign roles.

6. Pilot the solution with controls and measurable acceptance criteria.

7. Monitor results, exceptions, drift and unintended consequences.

This sequence reflects the principle “simplify before standardising, standardise before automating, and automate before optimising with AI.” It is not an absolute rule—some AI tools can help discover opportunities—but it is a valuable safeguard against digitising waste.

IoT, Advanced Analytics and Digital Twins

IoT connects sensors, devices, machines and other physical objects so that data can be collected and exchanged. In operations, sensor data may concern temperature, vibration, pressure, location, energy consumption, speed or cycle time. IoT supports condition monitoring, traceability, predictive maintenance, energy management and real-time quality control.

The usefulness of IoT depends on more than sensor installation. Measurement validity, calibration, sampling frequency, connectivity, time synchronisation and contextual metadata determine whether readings can support a decision. Excessive data can also overwhelm operators. A useful design begins with the decision to be improved and identifies the minimum reliable data needed.

Advanced analytics develops progressively. Descriptive analytics reports events and performance. Diagnostic analytics investigates relationships and causes. Predictive analytics estimates future outcomes, such as equipment failure or late delivery. Prescriptive analytics evaluates possible actions, often subject to constraints such as capacity, cost, safety and service requirements.

Digital twins extend operational visibility by linking a model with current or historical data. A production twin may represent equipment, material flows, inventories, cycle times and quality conditions. It can support “what-if” analysis before changing a layout or production schedule. A twin is not simply a three-dimensional visualisation; it must represent relevant relationships and be sufficiently connected to reality for its intended purpose. A model used for strategic simulation may not be suitable for real-time control.

AI for Prediction, Optimisation and Decision Support

AI can contribute to operational excellence in several ways. Classification models can identify defective products or prioritise service cases. Forecasting models can estimate demand, workload or failure probability. Natural-language systems can assist with document analysis, knowledge retrieval and service interaction. Optimisation algorithms can allocate resources, sequence jobs or balance competing objectives.

A model should be judged by its operational decision, not only by statistical accuracy. A highly accurate prediction may have little value if no action is available. Conversely, a model with moderate accuracy may be valuable when it enables early, low-cost intervention. Evaluation should therefore include precision, recall or forecast error where appropriate, as well as financial impact, quality, safety, response time, fairness and user acceptance.

The analytics life cycle normally includes problem definition, data preparation, feature or variable selection, model development, validation, deployment, monitoring and review. Data should be separated appropriately for training and evaluation to avoid leakage. Validation should reflect the conditions under which the model will operate, including seasonal changes and rare but important events.

AI decision support must distinguish recommendation from authority. In a safety-critical or high-impact context, the system may identify risk while a qualified person decides the response. In a low-risk repetitive context, automated execution may be appropriate within predefined limits. These limits should include escalation thresholds, permitted actions, rollback procedures and accountability for errors.

Redesigning Before Automation: A Quality-Centred Method

A practical redesign begins by defining value from the customer and stakeholder perspective. The team should identify the output, critical-to-quality requirements, applicable laws, failure modes and service expectations. It should then establish a baseline using measures such as lead time, first-pass yield, rework, cost of poor quality, exception rate, availability and customer complaints.

Next, the team analyses demand and variation. Some variation is caused by customer requirements and should be accommodated; other variation is created by inconsistent work methods, defective information or unnecessary policy. Root-cause tools such as Pareto analysis, cause-and-effect diagrams, failure mode and effects analysis, statistical process control and structured experimentation remain relevant in digital environments.

The redesigned process should specify decision rights, data definitions, controls, exception paths and performance feedback. Automation candidates can then be prioritised using criteria such as value, feasibility, risk, data readiness, process stability and reversibility. A portfolio view is preferable to isolated projects because automation in one process can shift workload or risk to another.

A balanced performance system should include outcome, process and risk indicators. For example, a claims operation may track customer resolution time, percentage of claims correctly settled at first decision, manual review rate, model error by customer segment, complaint rate and unauthorised access attempts. Optimising one measure, such as processing speed, can damage quality if it increases erroneous decisions.

Data Quality, Cybersecurity and Governance

Data quality is multidimensional. Accuracy concerns correctness; completeness concerns whether required values exist; consistency concerns agreement across systems; timeliness concerns availability when needed; validity concerns conformity to permitted definitions; and uniqueness concerns the absence of inappropriate duplication. Poor data quality can create false process variants, unreliable forecasts and unfair automated decisions.

Data governance should assign ownership for critical data, establish common definitions, specify retention and access rules, manage metadata and monitor quality. Master data—such as customer, product, supplier and asset records—requires particular attention because errors propagate across processes. Data-quality controls should operate at source, during transformation and at the point of use.

Digital operations expand the attack surface through connected equipment, application programming interfaces, cloud services, mobile devices and third-party platforms. Cybersecurity controls should include identity and access management, least privilege, encryption, network segmentation, secure configuration, vulnerability management, logging, backup, incident response and recovery testing. Operational technology may require special care because patching or shutting down equipment can affect safety and continuity.

Governance also covers AI-specific risks. Organisations should document the purpose and scope of a model, training data, performance limits, approval status, version history, monitoring arrangements and retirement criteria. Sensitive personal data require lawful processing, purpose limitation, minimisation and appropriate protection. Automated decisions should be contestable where individuals may experience significant consequences. Human oversight must be meaningful rather than a nominal approval step that merely rubber-stamps system outputs.

People, Roles and Organisational Change

Automation changes work rather than simply eliminating it. Employees may move from data entry to exception management, from inspection to process improvement, or from routine scheduling to judgement-intensive coordination. New risks can arise when operators lose practical knowledge or become overdependent on systems.

Successful implementation therefore involves process owners, subject-matter experts, data specialists, information-security professionals, frontline employees and affected customers where appropriate. Training should cover not only system operation but also model limitations, escalation responsibilities and methods for reporting anomalies. Workers should be able to challenge an output without suffering inappropriate penalties for identifying a system problem.

Leadership must align incentives with total performance. If a team is rewarded solely for throughput, it may bypass controls. If employees are evaluated only on adherence to a model, they may suppress valuable professional judgement. A mature operating model combines standardisation with structured discretion and learns from exceptions rather than treating them only as failures.

Illustrative Case: Intelligent Maintenance in a Food-Processing Plant

Background. An illustrative food-processing company operates several packaging lines. Unplanned stoppages cause missed delivery commitments, overtime and product waste. Maintenance is mainly preventive: components are replaced according to calendar intervals, while operators record some abnormalities on paper.

Situation and analysis. The company first created an event log by integrating machine alarms, production orders, maintenance work orders and quality records. Process mining showed that stoppages were often followed by repeated short interventions rather than a single effective repair. Analysis also revealed inconsistent asset identifiers and missing timestamps. A data-quality team standardised equipment codes, synchronised clocks and introduced structured failure categories.

Engineers then installed vibration and temperature sensors on selected high-risk motors. A predictive model estimated the probability of failure within a defined planning window. The model was not connected directly to machine shutdown. Instead, it generated a maintenance recommendation when confidence exceeded a threshold and required a technician to verify the condition.

Approach. Before automation, the plant redesigned the maintenance workflow. It removed duplicate approvals, defined emergency and planned-maintenance routes, established escalation rules and linked spare-parts availability to the work order. A workflow platform assigned tasks and recorded completion. A digital twin of the selected line was used to test alternative maintenance windows and assess the effect on production schedules.

Outcome. The pilot improved visibility of recurring faults and reduced repeated call-outs in the selected area. More importantly, it exposed the conditions under which the prediction was unreliable, including recently modified equipment and sensors affected by cleaning activities. The plant therefore added model-monitoring rules and required review after engineering changes. The outcome was not treated as proof that predictive maintenance should be deployed everywhere; expansion depended on data readiness, safety assessment and demonstrated operational value.

Learning points. The case illustrates that process mining can diagnose workflow failure, IoT can provide condition data, AI can support prediction, and workflow technology can coordinate action. None of these technologies substitutes for reliable identifiers, process redesign, engineering judgement or human accountability. The most important intervention was the integration of technical and managerial controls.

Challenges, Limitations and Common Mistakes

Digital transformation projects commonly fail through unclear problem definition, excessive scope, weak ownership and insufficient attention to adoption. Other frequent mistakes include automating unstable processes, measuring activity rather than outcomes, treating dashboards as improvement, ignoring exceptions, using biased or incomplete training data, and assuming that a vendor’s model will remain valid indefinitely.

Technology can also create new forms of operational fragility. Dependence on a single platform, connectivity failure, vendor lock-in, cyberattack or model outage may interrupt critical services. Automation bias can cause users to accept an incorrect recommendation because it appears objective. Alert fatigue can lead operators to ignore genuine warnings. Digital systems may intensify surveillance or deskill work if governance is weak.

These limitations call for resilience. Critical operations should have fallback procedures, manual alternatives where feasible, tested recovery arrangements and clear communication during incidents. Pilot designs should include failure scenarios, not only successful demonstrations. Benefits should be verified after implementation and compared with the baseline.

Current Developments and Best Practices

Generative AI is increasingly being explored for natural-language search across procedures, summarisation of service records, code assistance, conversational workflow interfaces and generation of draft responses. Its outputs require controls because fluent language does not guarantee factual accuracy, confidentiality or appropriate reasoning. Retrieval from approved knowledge sources, access controls, human review and output monitoring are important safeguards.

Edge computing enables analysis close to machines and devices, reducing latency and limiting unnecessary transfer of sensitive data. Industrial platforms are increasingly integrating IoT, analytics, workflow and digital-twin capabilities. Responsible AI frameworks are placing greater emphasis on transparency, accountability, robustness, privacy and fairness. Regulatory expectations are also moving toward risk-based controls, documentation and oversight for high-impact automated systems.

Sound practice can be summarised as follows:

• begin with a measurable quality or operational problem;
• redesign the end-to-end process before automating tasks;
• establish data ownership, definitions and quality controls;
• choose technology according to process characteristics and risk;
• test technical performance and operational outcomes separately;
• involve users and affected stakeholders in design;
• maintain human oversight for consequential decisions;
• secure systems throughout their life cycle;
• monitor exceptions, drift, fairness, cybersecurity and unintended effects; and
• scale only after evidence from a controlled pilot.

Chapter Summary

Digital operational excellence combines process discipline, data visibility, analytics, automation and organisational capability. Process mining reveals actual process behaviour through event logs; workflow platforms coordinate work; RPA executes suitable repetitive tasks; IoT supplies real-time operational evidence; digital twins support simulation; and AI contributes to prediction, optimisation and decision support. Their value depends on redesigning processes before automation, defining meaningful outcomes and maintaining reliable data.

Digital systems introduce risks involving data quality, cybersecurity, privacy, bias, model drift, automation dependence and loss of human judgement. Governance must therefore cover ownership, access, model documentation, monitoring, escalation and accountability. A quality-centred organisation uses technology to strengthen learning and control, not to replace critical thinking. The best implementation is incremental, evidence-based, secure and designed around both operational performance and human responsibility.

Key Terms

Digital operational excellence: Coordinated use of digital capabilities to improve, control and govern organisational processes.

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

Event log: Time-stamped record of activities associated with a process case.

Process variant: A distinct route or sequence through a process.

Workflow orchestration: Coordination of tasks, information, people and systems across a process.

Robotic process automation (RPA): Software-based execution of repetitive, rule-based digital tasks.

Internet of Things (IoT): Network of connected physical devices that collect and exchange data.

Digital twin: Data-connected digital representation of a physical asset, process or system.

Advanced analytics: Use of statistical and computational methods for description, diagnosis, prediction and prescription.

Machine learning: Methods that learn patterns from data to perform predictive or classificatory tasks.

Artificial intelligence: Technologies that perform activities associated with human cognitive capabilities.

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

Explainability: Ability to understand the basis of a model’s output.

Data lineage: Record of data origin, transformation and use.

Data governance: Rules, responsibilities and controls for managing data.

Human-in-the-loop: Arrangement in which a person reviews, approves or can override an automated output.

Straight-through processing: Completion of a transaction without manual intervention.

Exception: A case requiring treatment outside normal rules or process paths.

Automation bias: Tendency to accept an automated recommendation without adequate critical review.

Prescriptive analytics: Analysis that recommends actions while considering objectives and constraints.

Review Questions

Short Answer Questions

1. What is the difference between process discovery and conformance checking?

2. Identify two conditions under which RPA is appropriate.

3. Why is process redesign necessary before automation?

4. Define model drift and explain why it matters.

5. What are two dimensions of data quality?

Long Answer Questions

1. Explain how process mining, workflow platforms and RPA can be combined in an operational improvement programme.

2. Discuss the role of IoT, advanced analytics and digital twins in improving maintenance or production performance.

3. Describe a structured method for evaluating an AI application for operational decision support.

4. Examine the cybersecurity and data-governance requirements of connected operations.

5. Discuss why digital operational excellence is a socio-technical rather than purely technical capability.

Analytical / Application Questions

1. A hospital has automated appointment scheduling, but complaints and rework are increasing. Explain how you would investigate the process and decide whether further automation is appropriate.

2. A logistics company uses a predictive model to prioritise delayed shipments. Design a performance and governance framework that addresses accuracy, fairness, human oversight and model drift.

3. A manufacturer wants to install sensors on every machine immediately. Evaluate this proposal and recommend a phased approach using process value, data readiness, cybersecurity and operational risk.

22. Multiple Choice Questions with Answers

Question 1

Which activity is the primary purpose of process mining?

A. Replacing all process workers with software

B. Reconstructing and analysing actual process behaviour from event logs

C. Designing a physical layout without operational data

D. Encrypting data held in enterprise systems

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

Question 2

A company wants to automate a highly variable process involving ambiguous documents and frequent judgement. Which conclusion is most appropriate?

A. RPA is automatically the best solution

B. The process should be assessed and redesigned before selecting automation

C. The company should remove all exception handling

D. Automation should proceed without baseline measures

Correct Answer: B. The process should be assessed and redesigned before selecting automation

Question 3

Which example best represents prescriptive analytics?

A. Reporting last month’s machine downtime

B. Identifying that downtime is associated with one production shift

C. Estimating the probability that a motor will fail

D. Recommending a maintenance schedule subject to capacity constraints

Correct Answer: D. Recommending a maintenance schedule subject to capacity constraints

Question 4

Why might an IoT project fail to improve operational decisions despite collecting large volumes of data?

A. Sensors always produce too little information

B. Data may lack validity, context, synchronisation or a defined decision use

C. IoT data cannot be used for maintenance

D. Connected devices eliminate the need for process knowledge

Correct Answer: B. Data may lack validity, context, synchronisation or a defined decision use

Question 5

In the illustrative maintenance case, why was the predictive model not allowed to shut down equipment automatically?

A. The model had no access to any data

B. The plant had no workflow platform

C. Human verification was retained because the decision had operational and safety consequences

D. Predictive models cannot be used in manufacturing

Correct Answer: C. Human verification was retained because the decision had operational and safety consequences

Question 6

Which control most directly addresses model drift?

A. Deleting all historical data

B. Monitoring performance over time and reviewing the model after operating conditions change

C. Increasing the number of user interface screens

D. Removing human review from high-impact decisions

Correct Answer: B. Monitoring performance over time and reviewing the model after operating conditions change

Question 7

What is the principal risk of automating a defective process without redesigning it?

A. The process will necessarily become slower

B. Technology may increase the speed or scale of waste and defects

C. Data quality will automatically improve

D. Cybersecurity risks will disappear

Correct Answer: B. Technology may increase the speed or scale of waste and defects

Question 8

Which measure is an example of data completeness?

A. Whether a customer address follows the approved format

B. Whether the recorded value is correct

C. Whether all required fields are populated

D. Whether two systems use the same value for a customer

Correct Answer: C. Whether all required fields are populated

Question 9

What does meaningful human-in-the-loop governance require?

A. A person approving every output without seeing the evidence

B. A person having authority, information and competence to review or override relevant outputs

C. Removing all automated recommendations

D. Allowing only system developers to make decisions

Correct Answer: B. A person having authority, information and competence to review or override relevant outputs

Question 10

Which sequence best reflects a quality-centred automation approach?

A. Purchase technology, automate tasks, then define the outcome

B. Define the outcome, analyse and redesign the process, standardise it, then automate and monitor

C. Collect unlimited data, remove controls, then optimise

D. Deploy an AI model before identifying process ownership

Correct Answer: B. Define the outcome, analyse and redesign the process, standardise it, then automate and monitor