12. AI, Automation and Emerging Experience Technologies

Introduction and Context

Customer experience (CX) is increasingly shaped by digital systems that interpret customer needs, make decisions, generate responses and execute service processes. Chatbots answer questions, recommendation engines personalise offers, robotic process automation (RPA) completes repetitive administrative work, and predictive analytics anticipates customer behaviour. More recently, generative artificial intelligence (AI) has made it possible to produce fluent text, images, summaries and conversational responses at scale.

These technologies can improve speed, availability, consistency and personalisation. They can also produce serious quality failures. A chatbot may give a technically correct but insensitive answer; a recommendation system may reinforce discrimination; an automated process may repeat an error across thousands of transactions; and a generative model may confidently invent information. Consequently, AI adoption in customer experience is not simply a technology decision. It is a quality-management, governance and service-design decision.

For a Total Quality Management (TQM) professional, the central question is not whether an organisation uses advanced technology. It is whether the technology reliably creates value for customers and other stakeholders while controlling variation, risk and unintended consequences. This chapter examines major AI and automation technologies, their effects on service quality, and the human-in-the-loop controls required for responsible deployment.

Learning Objectives

After studying this chapter, students should be able to:

1. Distinguish among chatbots, virtual assistants, RPA, recommendation systems, predictive analytics, sentiment analysis and generative AI.

2. Explain how these technologies influence customer experience and quality management.

3. Evaluate AI systems using accuracy, reliability, fairness, transparency, safety and service-quality criteria.

4. Design human-in-the-loop controls for high-risk customer interactions.

5. Identify sources of bias, hallucination, privacy risk, automation failure and process deterioration.

6. Apply suitable metrics, monitoring methods and governance practices to AI-enabled services.

7. Analyse an organisational case involving automation and recommend an improvement strategy.

From Digital Service to Intelligent Experience

Traditional digital service systems followed explicit rules. A customer selected an option on a website, entered information into a form and received a predetermined response. Intelligent experience technologies add three capabilities: perception, prediction and generation. Perception involves interpreting text, speech, images or behavioural signals. Prediction estimates what may happen or what a customer may need. Generation produces a new response or recommendation rather than selecting only from a fixed menu.

The customer experience produced by an AI system depends on more than the model itself. It is the result of a socio-technical system comprising data, algorithms, interfaces, employees, policies, processes and customers. A highly accurate model can still create poor experience if it is embedded in a confusing interface or if customers cannot reach a human representative. Similarly, a modest model may perform well when its scope is narrow, its knowledge base is controlled and escalation is immediate.

Quality management therefore requires attention to the complete service process. The relevant unit of analysis is often the customer journey rather than the individual AI interaction. A bot may reduce average handling time while increasing repeat contacts, complaints or customer effort. An RPA tool may process transactions rapidly while transferring exception work to employees without adequate training. Performance must be judged across the entire flow.

Important Technologies and Their Service Roles

Chatbots and Virtual Assistants

A chatbot is a software application that conducts text-based interaction with users. Rule-based chatbots match inputs to predefined intents and responses. AI-enabled chatbots use natural language processing and machine-learning models to interpret varied language. A virtual assistant generally has broader capabilities, including speech recognition, task execution, personalisation and integration with enterprise systems.

Their typical applications include answering frequently asked questions, checking order status, scheduling appointments, guiding product selection and collecting preliminary information. Their value lies in availability and rapid response. However, they must recognise uncertainty and transfer customers when the issue is complex, emotional, financially significant or outside the system’s authority.

Robotic Process Automation

RPA uses software robots to perform structured, repetitive, rule-based tasks through existing applications. Examples include copying data between systems, validating fields, generating standard reports and initiating account updates. RPA is not necessarily intelligent; it often follows deterministic instructions. Intelligent automation combines RPA with optical character recognition, machine learning or natural-language processing.

RPA can reduce manual error and processing time, but it may automate a defective process. If an underlying workflow contains unnecessary approvals or poor data definitions, automation can make waste occur faster. RPA bots are also vulnerable to interface changes, missing data and exceptions that were not included in the original design.

Recommendation Systems

Recommendation systems estimate which product, service, content or action is most relevant to a user. Collaborative filtering uses patterns among users or items; content-based systems compare item characteristics with a user’s preferences; hybrid systems combine several methods. Recommendations may improve discovery and personalisation, but they can narrow choice, amplify popularity bias or exploit behavioural data without adequate transparency.

Quality evaluation should include not only prediction accuracy but also diversity, novelty, relevance, fairness and long-term customer value. A recommendation that increases immediate conversion but causes dissatisfaction or unwanted purchases may be locally successful and strategically harmful.

Predictive Analytics

Predictive analytics uses historical and current data to estimate future outcomes. In CX, organisations may predict churn, service demand, payment default, delivery delay, complaint likelihood or next-best action. Predictive models support proactive service, such as contacting a customer before a failure occurs.

Prediction is not certainty. A churn score identifies statistical association, not an inevitable customer decision. Managers must consider false positives, false negatives, changing conditions and the consequences of intervention. A retention offer sent to a customer incorrectly classified as at risk may waste resources; failure to identify a genuinely dissatisfied customer may result in preventable loss.

Sentiment Analysis

Sentiment analysis classifies expressed language as positive, negative, neutral or more detailed emotional categories. It can process surveys, social-media posts, complaint narratives and call transcripts. It is useful for identifying themes and prioritising investigation, but sentiment is context-dependent. Sarcasm, cultural expression, mixed emotions and domain-specific language can reduce accuracy.

Sentiment scores should therefore be treated as screening signals rather than final judgements. Quality teams should sample the underlying comments and compare automated classifications with human coding.

Generative AI

Generative AI produces new content from learned patterns. Large language models can summarise conversations, draft replies, search approved knowledge sources, classify cases and support agents. Image and voice models can create additional forms of interaction. Retrieval-augmented generation (RAG) connects a generative model to a controlled knowledge base so that responses are grounded in organisational information.

The principal quality risk is hallucination: the production of plausible but unsupported content. Generative systems may also reproduce bias, disclose confidential information, follow malicious instructions embedded in documents or provide inconsistent answers to similar questions. Fluency must never be confused with accuracy.

A Quality Framework for AI-Enabled Experience

A useful evaluation framework examines six connected dimensions.

Accuracy concerns whether the system’s classification, prediction or response is correct. Accuracy alone is insufficient when classes are unequal; precision, recall, false-positive rates and false-negative rates may be more informative.

Reliability concerns consistent performance under normal and unusual conditions. It includes uptime, latency, resilience to changes in data and predictable escalation behaviour.

Service quality concerns customer effort, responsiveness, empathy, resolution, accessibility and continuity across channels. A fast answer that fails to solve the problem is not high-quality service.

Fairness concerns whether outcomes or error rates differ unjustifiably across relevant customer groups. Fairness requires examination of data, features, outcomes and business rules.

Transparency and explainability concern whether customers, employees and managers can understand what the system does, what information it uses and when a human decision is involved. Explanations should be meaningful rather than merely technical.

Safety, privacy and security concern protection from harmful advice, unauthorised disclosure, manipulation, cyberattack and inappropriate use of personal data.

These dimensions may conflict. A highly personalised model may require more data, creating privacy concerns. A strict fraud model may reduce financial loss but wrongly reject legitimate customers. A fully automated service may reduce cost while weakening empathy. Quality decisions must make these trade-offs explicit.

Designing Human-in-the-Loop Controls

Human-in-the-loop (HITL) design assigns people a defined role in monitoring, reviewing, correcting or overriding an AI system. It is not satisfied by placing a nominal approval button at the end of an opaque process. Human involvement must be timely, informed and capable of changing the outcome.

Three patterns are common. In human-in-the-loop, a person approves or rejects an AI recommendation before action. This suits high-impact decisions such as complaint resolution, credit-related communication or safety incidents. In human-on-the-loop, the system acts automatically while employees monitor performance and intervene when thresholds are breached. This suits lower-risk, high-volume transactions. In human-in-command, leaders define the system’s scope, authority, escalation rules and shutdown conditions.

A practical HITL design includes:

• clearly defined tasks the system may and may not perform;
• confidence thresholds and rules for escalation;
• access to evidence supporting the recommendation or answer;
• an easy transfer to a trained employee;
• logging of prompts, inputs, outputs, decisions and overrides;
• independent quality sampling;
• a process for correcting the knowledge base or model;
• periodic review of customer outcomes and group differences; and
• a tested fail-safe or manual operating mode.

Escalation should be triggered not only by low confidence but also by topic. Legal threats, self-harm, abuse, vulnerable customers, identity disputes, large financial consequences and repeated failed interactions generally require human attention even if the model appears confident.

Measuring Performance and Managing Variation

A balanced measurement system combines operational, customer, model and governance indicators. Operational measures include response time, containment rate, transfer rate, abandonment and cost per contact. Customer measures include customer effort, first-contact resolution, satisfaction, complaint recurrence and trust. Model measures include precision, recall, calibration, drift and hallucination rate. Governance measures include override frequency, unresolved incidents, audit completion and access violations.

Measurement must distinguish between efficiency and effectiveness. For example, chatbot containment—the percentage of conversations ending without transfer—may rise because the bot prevents customers from reaching employees. It should therefore be interpreted with resolution, repeat contact and complaint data.

A control-chart mindset is useful. Baseline performance should be established, expected variation defined and special causes investigated. Monitoring should be continuous for critical services and periodic for lower-risk applications. Models can deteriorate because customer language changes, products change, regulations change or the relationship between variables and outcomes shifts. This is known as data or concept drift.

Testing should occur before and after deployment. Pre-deployment tests include representative scenarios, adversarial prompts, accessibility checks, subgroup analysis and failure-mode analysis. Post-deployment tests include random conversation audits, shadow evaluation, customer feedback, incident review and champion–challenger comparisons between the current and proposed systems.

Governance, Ethics and Professional Responsibility

AI governance allocates accountability across business owners, data scientists, technology teams, frontline employees, legal specialists and senior leadership. Responsibility must not be transferred to “the algorithm.” Organisations should maintain an inventory of AI use cases, classify them by risk, document intended purposes and specify accountable owners.

Privacy requires purpose limitation, data minimisation, retention controls, access management and appropriate consent or lawful basis. Sensitive information should not be entered into public or unapproved generative tools. Security controls must address prompt injection, unauthorised retrieval, model extraction and compromised integrations.

Customers should receive appropriate notice when interacting with automation, particularly where a synthetic voice or generated response is used. They should have a reasonable route to human assistance and a method for contesting consequential decisions. Employees also require transparency: monitoring systems should not create unreasonable surveillance or punish staff for overriding unsafe recommendations.

Ethical quality includes respect, autonomy and non-manipulation. Personalisation becomes problematic when it exploits vulnerability, conceals important alternatives or makes it difficult to refuse. The organisation should ask not only “Can we automate this?” but also “Should we automate it, for whom, and under what safeguards?”

An Applied Case: Northstar Utilities’ Customer Service Transformation

Background

Northstar Utilities, a fictional regional energy provider, handled approximately 60 percent of customer contacts through telephone channels. Customers commonly called about bills, meter readings, payment arrangements and service interruptions. Management introduced a chatbot, RPA-based account updating and a predictive model intended to identify customers at risk of payment difficulty.

Situation and Problem

After six months, average digital response time had fallen substantially. However, complaints increased. Customers reported that the chatbot repeated scripted answers, failed to recognise vulnerability and transferred conversations only after several unsuccessful attempts. The payment-risk model generated many alerts for customers in particular geographic areas, leading to aggressive reminders. An RPA bot also entered incorrect meter readings when customers used an unfamiliar format.

Analysis

The organisation had optimised speed and automation volume rather than end-to-end service quality. The chatbot’s containment rate was treated as its principal success measure, while repeat contact and escalation delay were not monitored. The predictive model used historical payment patterns that reflected unequal access to payment facilities. This created a risk of reproducing structural disadvantage. The RPA process lacked exception handling and validation against plausible meter ranges.

Approach and Solution

Northstar classified use cases by risk. General information remained eligible for automated handling, while vulnerability, payment distress, disputes and service interruption required rapid human escalation. The chatbot was redesigned with explicit uncertainty messages and a visible “speak to an adviser” option. A controlled knowledge base and retrieval system were introduced for generated answers, with blocked topics for which the system could not respond.

The predictive model was reviewed using subgroup error analysis and replaced, for high-impact interventions, with a broader assessment that included customer self-identification and human review. Employees received guidance on treating model scores as signals rather than decisions. The RPA bot gained input validation, exception queues and daily reconciliation. A quality dashboard combined first-contact resolution, repeat contact, complaints, transfer time, subgroup outcomes and override rates.

Outcome

The redesigned service accepted a lower level of automatic containment but improved resolution and reduced repeated unsuccessful interactions. Human advisers received clearer context when cases were transferred. The organisation also created a monthly AI quality forum involving operations, compliance, customer representatives and technical staff.

Learning Points

The case illustrates four principles. First, automation must be evaluated across the customer journey. Second, confidence scores do not remove the need for judgement. Third, historical data may encode social and organisational bias. Fourth, a controlled reduction in automation can produce better quality when it protects customers and employees from failure.

Common Failure Modes and Limitations

Organisations frequently automate before stabilising the underlying process. This creates “automated chaos,” in which errors are reproduced at greater speed. Another failure is inadequate problem definition: a company deploys a chatbot to reduce calls when the real issue is unclear billing information. Poor training data, unrepresentative testing, weak integration and ambiguous ownership are additional causes of failure.

Generative AI introduces distinctive limitations. It can fabricate citations, misunderstand negation, lose conversational context and produce different answers to equivalent prompts. Prompt engineering may improve performance but is not a substitute for governance. Models may also be difficult to audit because their internal representations are not easily interpretable.

Automation can deskill employees if human judgement is rarely exercised. Conversely, employees may experience automation bias, accepting system recommendations because they appear objective. Training must therefore include challenge behaviour, escalation and error recognition.

Emerging Directions

Experience technologies are moving toward multimodal interaction, in which systems process text, speech, images and video together. Agentic systems may plan and execute multi-step tasks rather than merely respond to a prompt. Digital twins and process-mining tools can simulate service operations and identify bottlenecks. Smaller, specialised models may offer better control, lower cost and reduced data exposure than general-purpose models.

These developments increase, rather than reduce, the importance of quality management. Organisations will need stronger evaluation datasets, provenance records, model cards, incident-reporting processes and continuous assurance. The likely competitive advantage will not come from using the most fashionable tool, but from integrating technology with trustworthy processes, skilled employees and disciplined learning systems.

Best-Practice Guidance

A responsible implementation should begin with a customer and process problem, not a technology demonstration. Establish a baseline, define the quality objective and identify affected stakeholders. Select the least complex technology that can meet the need. Pilot in a controlled environment, test ordinary and adversarial cases, and involve frontline employees and customers in evaluation.

Before scaling, define authority limits, escalation rules, performance thresholds and rollback procedures. Use approved data sources, protect personal information and retain sufficient logs for investigation. Monitor outcomes by customer segment and review both benefits and harms. Finally, treat deployment as an ongoing quality-improvement cycle: plan the service, test it, measure performance, learn from variation and improve the system.

Chapter Summary

AI and automation can reshape customer experience through faster service, personalisation, prediction and task execution. Chatbots and virtual assistants support interaction; RPA performs structured work; recommendation systems personalise choices; predictive analytics anticipates outcomes; sentiment analysis identifies expressed attitudes; and generative AI creates responses and content. Each technology can improve quality only when embedded in a well-designed service process.

Evaluation must extend beyond accuracy and cost to include reliability, customer effort, resolution, fairness, transparency, privacy, safety and resilience. Human-in-the-loop controls are essential for high-impact or uncertain interactions. Effective governance defines ownership, authority, escalation, monitoring, documentation and customer redress. The Northstar case demonstrates that lower automation can sometimes produce higher overall quality when it prevents repeated failure and protects vulnerable customers. The strategic task is therefore to combine intelligent technology with human judgement and continuous improvement.

Key Terms

Artificial intelligence: Computational techniques that perform tasks associated with human intelligence, such as learning, prediction or language interpretation.

Automation: The use of technology to perform tasks with limited direct human intervention.

Chatbot: A software system that conducts conversational interaction, usually through text or voice.

Virtual assistant: An AI-enabled system that interprets requests and may perform tasks across connected services.

Robotic process automation: Software-based automation of structured, repetitive and rule-governed tasks.

Recommendation system: A system that predicts and presents items or actions likely to be relevant to a user.

Predictive analytics: Analysis that estimates future events, behaviours or outcomes from data.

Sentiment analysis: Automated classification or interpretation of attitudes and emotions expressed in language.

Generative AI: AI that produces new text, images, audio, code or other content.

Hallucination: A plausible but unsupported or factually incorrect output generated by an AI system.

Human-in-the-loop: A design in which a person reviews, approves or can directly alter an AI-supported decision.

Human-on-the-loop: A design in which humans monitor an operating system and intervene when necessary.

Retrieval-augmented generation: Generation grounded in information retrieved from a specified knowledge source.

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

Precision: The proportion of positive model classifications that are correct.

Recall: The proportion of relevant positive cases that the model successfully identifies.

Automation bias: The tendency to accept an automated recommendation without adequate independent judgement.

Explainability: The extent to which a system’s operation or output can be meaningfully understood.

Algorithmic fairness: The principle that automated systems should not produce unjustified unequal treatment or error across groups.

Review Questions

Short Answer Questions

1. Distinguish between a rule-based chatbot and an AI-enabled chatbot.

2. What is the principal purpose of RPA in a service process?

3. Why is chatbot containment an insufficient measure of service quality?

4. Define hallucination in generative AI.

5. What is the difference between human-in-the-loop and human-on-the-loop control?

Long Answer Questions

1. Explain how recommendation systems and predictive analytics can improve and weaken customer experience.

2. Discuss the six dimensions of quality used to evaluate AI-enabled services.

3. Describe a governance framework for deploying generative AI in customer service.

4. Explain how data and concept drift affect the reliability of predictive systems.

5. Analyse the role of employees in responsible automation and continuous improvement.

Analytical / Application Questions

1. A bank’s chatbot has high containment but rising repeat contacts and complaints. Design a revised measurement and escalation framework.

2. A retailer’s recommendation engine produces lower conversion rates for a minority customer group. Identify possible causes and propose an investigation.

3. Design a human-in-the-loop process for an AI system that identifies customers at risk of financial hardship. Specify authority limits, evidence, escalation and monitoring requirements.

22. Multiple Choice Questions with Answers

Question 1

Which measure best complements chatbot containment when evaluating service quality?

A. Number of bot messages generated

B. Repeat contact and first-contact resolution

C. Size of the chatbot’s vocabulary

D. Number of automated greetings

Correct Answer: B. Repeat contact and first-contact resolution

Question 2

What is the defining risk of a generative AI hallucination?

A. The system always refuses to answer

B. The system produces a fluent but unsupported answer

C. The system reduces processing speed

D. The system uses a deterministic workflow

Correct Answer: B. The system produces a fluent but unsupported answer

Question 3

Which task is most suitable for conventional RPA?

A. Making an empathetic judgement about vulnerability

B. Copying validated invoice data between systems

C. Interpreting sarcasm in a complaint

D. Negotiating a complex legal dispute

Correct Answer: B. Copying validated invoice data between systems

Question 4

A predictive model identifies customers likely to churn. What is the most appropriate managerial interpretation?

A. The prediction proves that each identified customer will leave

B. The score is a statistical signal requiring suitable action and review

C. The model eliminates the need for customer research

D. The model is automatically fair because it uses historical data

Correct Answer: B. The score is a statistical signal requiring suitable action and review

Question 5

Which situation most clearly requires immediate human escalation?

A. A request for store opening hours

B. A request to reset a routine password

C. A customer reporting financial hardship and vulnerability

D. A request for a product specification

Correct Answer: C. A customer reporting financial hardship and vulnerability

Question 6

What does retrieval-augmented generation primarily seek to achieve?

A. Removing all human review from AI systems

B. Grounding generated responses in an approved knowledge source

C. Increasing the number of unsupported responses

D. Replacing organisational data with random internet content

Correct Answer: B. Grounding generated responses in an approved knowledge source

Question 7

Which example illustrates automation bias?

A. An employee independently checks a model recommendation

B. A manager rejects a recommendation after reviewing evidence

C. An adviser accepts an AI decision solely because it appears objective

D. A system records an employee override

Correct Answer: C. An adviser accepts an AI decision solely because it appears objective

Question 8

Why can historical data create unfair predictive outcomes?

A. Historical data is always too small to analyse

B. It may reflect previous unequal access, treatment or organisational decisions

C. It cannot contain customer information

D. It prevents models from producing any prediction

Correct Answer: B. It may reflect previous unequal access, treatment or organisational decisions

Question 9

Which metric combination gives the strongest indication that automation is harming customers?

A. Higher containment with higher repeat contacts and complaints

B. Lower response time with higher first-contact resolution

C. Higher accuracy with lower false-negative rates

D. Lower processing time with stable satisfaction

Correct Answer: A. Higher containment with higher repeat contacts and complaints

Question 10

What is the main purpose of a fail-safe or rollback procedure?

A. To ensure the system cannot ever be improved

B. To permit controlled suspension when significant failure or risk occurs

C. To increase the system’s autonomy without monitoring

D. To conceal errors from customers and employees

Correct Answer: B. To permit controlled suspension when significant failure or risk occurs