14. AI, Automation and Data-Driven Personalisation

Introduction and Context

Insurance marketing has traditionally relied on broad segmentation, intermediary judgement, periodic campaigns and relatively standardised products. Digital channels have changed this arrangement. Insurers can now observe customer interactions across websites, mobile applications, contact centres, agency systems, claims platforms and connected devices. Artificial intelligence (AI), advanced analytics and marketing automation can transform these observations into predictions, recommendations and timely actions.

The significance of this transformation is not simply that an insurer can send more messages at lower cost. The deeper change concerns how marketing decisions are made. A conventional campaign may define a target group in advance and communicate a common offer. A data-driven system may estimate an individual’s likelihood of purchasing, recommend a suitable product, predict the risk of lapse, select the most appropriate channel and determine whether a human employee should intervene. In this sense, AI becomes part of the insurer’s decision architecture.

The opportunity is substantial. Predictive systems can improve lead allocation, recommendation engines can support needs-based selling, churn models can identify customers requiring attention, conversational AI can provide continuous service, and next-best-action systems can coordinate interactions across the customer life cycle. However, insurance is a high-trust and highly regulated industry. Personalisation can become intrusive; automation can make errors at scale; and a model that appears commercially successful may produce unfair outcomes, weak explanations or unacceptable privacy and cybersecurity risks.

This chapter evaluates the principal AI-enabled applications in insurance marketing and customer relationship management (CRM). It also considers the organisational capabilities, governance controls and human judgement required to use them responsibly.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the role of AI, automation and data-driven personalisation in insurance marketing and CRM.

2. Distinguish predictive lead scoring, recommendation engines, churn models, conversational AI and next-best-action systems.

3. Evaluate the data, modelling and performance considerations associated with these applications.

4. Design an appropriate marketing-automation workflow for an insurance use case.

5. Analyse model risk, explainability, privacy, cybersecurity, bias and human oversight.

6. Apply governance and managerial principles to the implementation of AI in insurance.

From Segmentation to Individualised Decisioning

Personalisation means adapting content, timing, channel, service or offer to the circumstances and inferred needs of a customer. It ranges from simple rules, such as sending a renewal reminder thirty days before expiry, to complex machine-learning systems that estimate a customer’s propensity to buy a particular protection product.

It is useful to distinguish four levels of decisioning. At the first level, the insurer uses static segmentation based on observable characteristics such as age band, location, policy type or distribution channel. At the second level, rule-based personalisation applies explicit business rules, for example offering a travel policy to a customer who purchases an international-flight add-on. At the third level, predictive analytics estimates an outcome, such as purchase, lapse or response probability. At the fourth level, prescriptive systems recommend an action by balancing predicted customer response, business objectives, compliance constraints and service capacity.

AI does not eliminate the need for marketing judgement. Rather, it changes the location of judgement. Humans define objectives, select permissible data, establish constraints and interpret exceptional cases, while algorithms process large volumes of information and identify patterns. The quality of the result therefore depends on both technical performance and the institutional choices surrounding the model.

Important Terminology

Artificial intelligence refers broadly to computational systems that perform tasks associated with human cognitive capabilities, including language interpretation, classification, prediction and decision support. Machine learning is a subset of AI in which systems learn patterns from data rather than relying solely on manually coded rules. Supervised learning uses labelled historical outcomes, such as whether a lead converted. Unsupervised learning identifies patterns without a predefined outcome, as in customer clustering.

A feature is an input variable used by a model, while a target or label is the outcome the model is trained to predict. A score is a numerical estimate, often representing probability or priority. A threshold is the point at which a score triggers an action. Model drift occurs when the relationship between inputs and outcomes changes over time. Data drift refers to changes in the distribution of input data. A champion model is the currently approved production model; a challenger is an alternative model evaluated against it.

Personalisation should be distinguished from automated decision-making. Personalisation may merely change the presentation of information, whereas automated decision-making can materially affect eligibility, pricing, service access or treatment. This distinction is important for governance and regulatory review.

Predictive Lead Scoring: Allocating Attention More Intelligently

A lead-scoring model estimates the likelihood that a prospect will undertake a desired action within a defined period. In insurance, the action may be requesting a quotation, completing an application, purchasing a policy or accepting an adviser appointment. The purpose is not to declare that a person is valuable in a general sense. It is to allocate scarce marketing or sales attention more effectively.

A typical workflow begins by defining the outcome and observation period. Historical leads are then assembled with information available before the outcome occurred. Potential features may include product interest, source channel, quotation activity, time spent in relevant digital journeys, previous relationship with the insurer, contact preferences and permitted demographic or geographic information. Data leakage must be avoided: information created after conversion, or information that would not be available when the decision is made, can make a model appear more accurate than it really is.

The model produces a score, which may be used to prioritise call-centre queues, trigger educational content or route a prospect to an adviser. A high score does not necessarily mean that the prospect should receive an aggressive sales approach. It may indicate that a timely, relevant explanation is appropriate. Business rules should also consider affordability, suitability, consent, capacity and the customer’s expressed preferences.

Evaluation should go beyond overall accuracy. Precision asks how many selected leads actually convert; recall asks how many eventual converters were identified. Calibration examines whether a predicted probability corresponds reasonably to observed outcomes. Lift compares performance with random selection or an existing process. Financial evaluation should include incremental conversion, cost per acquisition, adviser time, complaints and cancellation behaviour. An apparently accurate model may merely rank customers who would have purchased without additional intervention. Randomised holdout groups or controlled experiments are therefore valuable.

Recommendation Engines and Needs-Based Offers

A recommendation engine proposes products, content, coverage options or service actions. It can use content-based methods, which match attributes of an item with a customer’s known interests; collaborative methods, which learn from patterns across customers; knowledge-based methods, which apply product and suitability rules; or hybrid methods combining these approaches.

Insurance recommendation is more constrained than retail product recommendation. A policy is not merely an item to be promoted. Its suitability depends on risk exposure, existing coverage, exclusions, affordability, legal requirements and the customer’s understanding. A recommendation engine should therefore distinguish between an educational suggestion and a regulated advice or sales recommendation. Product governance, approved wording and suitability controls must be embedded in the design.

The system should account for the customer’s coverage inventory. Recommending an additional policy without recognising overlapping protection may create confusion or over-insurance. Similarly, recommending a low-premium product solely because it has a high conversion rate may conflict with fair customer outcomes. A useful architecture combines predictive relevance with constraints such as eligibility, product availability, underwriting rules, customer consent and compliance status.

Recommendation quality can be evaluated through click-through and quotation rates, but those measures are incomplete. The insurer should also examine completed purchase, persistency, claims understanding, cancellation, complaints, coverage gaps and outcomes across customer groups. A recommendation that generates immediate sales but increases early lapses may be commercially and ethically inferior to a less aggressive recommendation.

Churn and Lapse Models in the Customer Life Cycle

Churn modelling estimates the probability that a customer will discontinue a relationship, fail to renew or reduce engagement. In insurance, the relevant event may be policy cancellation, non-renewal, movement to another insurer or a decline in cross-product relationship. Lapse is not always a sign of dissatisfaction: a customer may experience a change in financial circumstances, sell an asset or no longer need the cover. Consequently, the model should be linked to an appropriate intervention rather than treating every predicted lapse as a sales opportunity.

Potential predictors include changes in payment behaviour, service interactions, unresolved complaints, premium changes, quotation activity, policy utilisation, renewal proximity and communication engagement. Sensitive proxies require careful review. A model may inadvertently learn that customers in a particular area or with a particular payment method are more likely to lapse because of structural disadvantage rather than genuine dissatisfaction.

The intervention should be matched to the likely cause. A customer showing payment difficulty may need flexible payment information or a human conversation. A customer who has not understood a renewal change may need a clear explanation. A customer who has experienced poor service may require remediation, not another promotional message. Retention models should therefore be integrated with service recovery and customer-value principles.

Model success should be measured against a control group. The relevant question is not simply whether contacted customers renewed, but whether the intervention increased renewal relative to what would otherwise have occurred. Insurers should also monitor whether retention efforts create unwanted outcomes, such as pressure selling, unsuitable discounts or unequal treatment.

Conversational AI in Insurance Service and Distribution

Conversational AI includes chatbots, voice assistants and agent-assist tools that interpret natural language and generate or retrieve responses. Its applications include answering policy questions, guiding quotation journeys, explaining documents, collecting preliminary information, scheduling appointments, supporting claims status enquiries and summarising conversations for employees.

There are two broad technological approaches. Retrieval-based systems select answers from an approved knowledge base. Generative systems produce language dynamically, often using a large language model. Retrieval systems may be more controllable for tightly defined questions, while generative systems can handle varied language and produce more natural interactions. In practice, a grounded hybrid architecture is often preferable: the model interprets the question, retrieves authorised information and generates a response within defined boundaries.

A safe conversational system requires identity and access controls, clear disclosure that the user is interacting with an automated system, escalation to a human, conversation logging, content filtering and protection against prompt injection or unauthorised data disclosure. It should not invent coverage terms, guarantee claim outcomes or provide an unapproved interpretation of policy wording. High-impact matters, complaints, vulnerability indicators, suspected fraud and complex claims should be routed appropriately.

Performance includes containment rate, response accuracy, first-contact resolution, transfer quality, customer effort, complaint rates and fairness across language or accessibility groups. A high containment rate is not inherently positive if customers cannot reach a human or if the system creates repeated interactions. Conversational AI should reduce friction, not merely reduce staffing costs.

Next-Best-Action Systems: Coordinating the Relationship

A next-best-action (NBA) system determines which action is most appropriate at a particular customer interaction. Possible actions include providing a renewal explanation, requesting missing information, offering a service remedy, recommending a product, scheduling a review or taking no action. The last option is important: responsible personalisation includes recognising when silence is better than another message.

An NBA system generally combines four elements. First, it uses customer and interaction data to estimate needs or likely responses. Second, it applies business and regulatory constraints, such as consent, contact frequency, product eligibility and vulnerability safeguards. Third, it ranks feasible actions according to an objective function. Fourth, it records the decision and its rationale for monitoring and audit.

The objective function should not be limited to short-term revenue. It may incorporate customer benefit, retention, service quality, risk, cost, fairness and long-term relationship value. In advanced settings, reinforcement learning can estimate the effect of sequences of actions, but such methods require caution because historical actions may reflect past bias and because experimentation on customers may have material consequences. Rule-constrained decisioning and controlled testing are generally more appropriate starting points.

An NBA system can operate across channels, but channel coordination is essential. A customer who has just spoken to an adviser should not receive an automated message repeating the same offer. A central decision log, event stream and contact policy can help prevent contradictory or excessive communications.

Marketing Automation as an Operating Process

Marketing automation connects data, decisions, content and execution. A basic workflow may proceed as follows: a customer provides consent; an event is captured; the customer is placed in an eligible audience; a model scores the customer; a decision engine selects an action; approved content is delivered through a channel; the response is recorded; and the outcome is used for evaluation or future model training.

A modern insurance architecture commonly includes a customer data platform or governed data layer, application programming interfaces connecting policy and CRM systems, an analytics environment, a model registry, a decision engine, content-management capability, campaign orchestration and monitoring dashboards. The architecture should support data lineage, role-based access and separation between development, testing and production.

Automation should be designed around customer journeys rather than isolated campaigns. For example, a home-insurance journey may begin with educational content, proceed to a needs assessment, provide a quotation, offer human assistance, issue policy documents and later deliver risk-prevention advice. Trigger rules, suppression rules, frequency limits and escalation conditions should be explicit.

Automation does not mean removing employees from the process. Employees design journeys, approve content, investigate exceptions, handle sensitive interactions and learn from customer feedback. Automation is most valuable when it makes human work more informed and timely rather than making the relationship impersonal.

Model Risk, Explainability and Validation

Model risk is the possibility that a model produces an unsuitable or unreliable result because of design error, poor data, misuse, implementation failure or changing conditions. In marketing, model risk may cause missed opportunities, discriminatory targeting, excessive contact, unsuitable recommendations or reputational damage. In more consequential decisions, it may affect access to insurance or the treatment of a claim.

A model-risk management process should document purpose, scope, data sources, assumptions, limitations, intended users, prohibited uses and performance thresholds. Independent validation should test conceptual soundness, data quality, outcome definition, stability, calibration, fairness and implementation. Models should be monitored after deployment, with defined triggers for review, retraining, suspension or retirement.

Explainability has several audiences. Customers may need a clear, meaningful explanation of why a communication or recommendation was made. Front-line employees need actionable reasons that help them serve the customer. Risk and compliance teams need technical documentation and evidence. Senior management needs to understand material risks and controls. A generic statement that an algorithm considered “many factors” is rarely sufficient.

Interpretability methods such as feature importance, partial-dependence analysis and local explanations can assist review, but they have limitations. An explanation is not necessarily a causal account, and a simple explanation can conceal complex interactions. The insurer should therefore combine technical explanations with plain-language communication, documented decision rules and human review.

Privacy, Cybersecurity and Responsible Data Use

Personalisation depends on data, but the ability to collect data does not establish a right to use it for every purpose. Privacy governance should address purpose limitation, data minimisation, lawful basis or consent where applicable, transparency, retention, access rights, correction, deletion and controls on sharing. Customers should be able to understand the significant categories of information used to shape their experience.

Particular care is required with inferred data. A model may infer financial stress, health-related circumstances, vulnerability or household composition even when a customer has not directly provided those facts. The insurer should assess whether such inferences are necessary, accurate, fair and appropriate for the intended use. Personalisation should not exploit a customer’s vulnerability or make it impossible to obtain a fair service without consenting to excessive tracking.

AI expands the cybersecurity attack surface. Threats include data poisoning, model theft, prompt injection, adversarial inputs, account takeover, insecure application interfaces, unauthorised employee access and leakage of confidential information through third-party AI services. Controls may include encryption, tokenisation, least-privilege access, secure software development, environment separation, vendor due diligence, red-team testing, monitoring and incident response.

Privacy impact assessments and algorithmic impact assessments can identify risks before deployment. Contracts with cloud and AI providers should address data use, retention, security, audit rights, breach notification, service continuity and responsibility for model changes.

Bias, Fairness and Human Oversight

Bias may enter through historical decisions, incomplete samples, proxy variables, measurement choices, labels, system design or the way an intervention is delivered. Removing a protected attribute does not remove bias if other variables act as proxies. Conversely, using a protected attribute in a controlled fairness analysis may be necessary to detect unequal outcomes.

Fairness is not a single statistical formula. Possible assessments include differences in selection rates, error rates, calibration, access to beneficial interventions and quality of service across groups. These measures can conflict. The appropriate assessment depends on the use case, legal context, customer impact and ethical rationale.

Human oversight should be substantive rather than ceremonial. Employees need authority, time, training and information to challenge an automated recommendation. They should know when a model is outside its validated scope and how to override or escalate it. A human merely clicking approval without meaningful review does not constitute effective oversight.

Governance should assign accountability. A product owner may be responsible for business purpose, a data scientist for development, an independent validator for challenge, an information-security team for technical controls, compliance and legal teams for regulatory interpretation, and senior management for risk appetite. Accountability remains with the insurer even when a vendor supplies the technology.

Illustrative Case: Meridian Mutual’s Renewal Intelligence Programme

Background

Meridian Mutual, a fictional medium-sized general insurer, sells motor and household insurance through agents, a website and a contact centre. Renewal rates had declined, while customers complained that communications were repetitive and that premium changes were not clearly explained. Management proposed an AI-enabled renewal programme combining a lapse model, conversational assistant and next-best-action engine.

Situation and Problem

The initial proposal ranked customers by predicted lapse probability and sent high-risk customers a discount offer. A pilot produced more renewals, but the compliance team identified concerns. Customers with payment difficulties were receiving promotional messages rather than support. The model performed less reliably for a minority language group. The chatbot occasionally described optional cover as if it were included. Moreover, the programme measured renewal rate but not whether the intervention caused the renewal.

Analysis

A cross-functional review first clarified the objective: improve informed renewal and service quality, not simply maximise retention. The data team reconstructed the training set using only information available before the renewal decision. It identified that complaint history and payment disruption were important predictors, but these signals required different interventions. The team also found that language and channel differences affected data completeness.

The model was recalibrated and evaluated separately across relevant customer groups. The chatbot knowledge base was restricted to approved policy content, and complex questions were transferred to trained employees. A contact policy prevented simultaneous messages from the website, agent and automated system.

Approach and Solution

Meridian replaced the single discount action with a constrained NBA framework. Customers predicted to be at risk because of confusion received a plain-language renewal explanation. Customers showing payment stress were offered information about available support and a human contact option. Customers with unresolved complaints were routed to service recovery. Only customers eligible under product and pricing rules could receive a commercial offer.

A randomised control group received the existing renewal process. The pilot dashboard included renewal, incremental renewal, complaint rate, early cancellation, customer effort, transfer rate, model calibration and outcome differences across groups. A model governance committee approved the pilot, required monthly monitoring and established suspension thresholds.

Outcome

The programme demonstrated that targeted service interventions could improve the renewal journey without relying exclusively on discounts. The control-group design also showed that some customers who would have renewed without intervention should not be counted as programme-generated successes. The chatbot was retained for routine information but excluded from complex coverage interpretation. Human advisers reported that the model’s reason codes helped them begin more relevant conversations, while customers received clearer explanations and fewer duplicate contacts.

Learning Points

The case illustrates five principles. First, a prediction is not an intervention; the action must be appropriate to the underlying need. Second, business success requires incremental evaluation rather than raw conversion measurement. Third, model performance and fairness must be examined across customer groups. Fourth, conversational AI requires grounded content and escalation. Finally, governance should be designed before scaling, not added after a problem occurs.

Managerial Choices, Limitations and Common Mistakes

Senior managers should begin with a customer or organisational problem, not with a fashionable technology. They should ask whether better data or automation will genuinely improve the outcome, whether the use is proportionate, and whether a simpler rule-based solution would be safer and easier to explain.

Common mistakes include training on biased historical outcomes, using post-event information, confusing correlation with customer need, optimising clicks instead of long-term value, ignoring customers with limited digital access, allowing vendors to determine the objective, and deploying a chatbot without a reliable human route. Another error is excessive personalisation. Customers may find highly specific messages unsettling, particularly where the insurer appears to know information that the customer did not knowingly disclose.

AI systems also have operational limitations. Data may be fragmented across legacy platforms; labels may be inconsistent; rare events may be difficult to predict; market conditions may change; and automated systems may amplify a small initial error. Generative AI can produce fluent but inaccurate answers. No model can remove uncertainty from insurance. It can only estimate patterns under stated assumptions.

A mature implementation therefore uses staged deployment: discovery, impact assessment, controlled pilot, independent validation, limited production, monitoring and periodic review. Clear stop conditions are as important as performance targets. The system should be capable of being corrected, rolled back or retired.

Emerging Developments

Insurance marketing is moving towards real-time decisioning, event-driven personalisation, multimodal conversational interfaces and more integrated customer data. Generative AI is likely to assist employees with research, summarisation, content adaptation and service quality monitoring. Synthetic data may support testing where direct customer data cannot be freely used, although synthetic data does not automatically remove privacy or bias risks.

Privacy-enhancing technologies, including tokenisation, federated approaches and secure computation, may enable some forms of analysis with reduced exposure of raw data. Causal and uplift modelling may improve the selection of interventions by estimating who is likely to benefit from an action, rather than merely who is likely to buy. Digital identity and consent-management systems may provide customers with greater visibility and control over data use.

These developments increase, rather than reduce, the need for governance. The strategic advantage will not belong simply to the insurer with the most sophisticated model. It will belong to the insurer that combines useful data, sound customer understanding, reliable technology, responsible experimentation and trustworthy human judgement.

Best-Practice Framework for Insurers

A practical governance framework can be organised around eight questions:

1. Purpose: What customer and business problem is being addressed?

2. Proportionality: Is AI necessary, and is the proposed use proportionate to its impact?

3. Data: Are the data relevant, accurate, permitted and sufficiently representative?

4. Decision: What exactly does the model predict or recommend, and what is outside its scope?

5. Controls: What constraints, approval rules, escalation paths and override mechanisms apply?

6. Evidence: How will incremental value, fairness, accuracy and customer outcomes be measured?

7. Security: How are access, vendors, interfaces, prompts, models and incidents controlled?

8. Accountability: Who owns the system throughout its life, and when will it be reviewed or retired?

Applied consistently, these questions turn AI from an isolated technology project into a governed capability within insurance marketing and CRM.

Chapter Summary

AI and automation enable insurers to move from broad campaign management towards data-driven, context-sensitive decisioning. Predictive lead scoring prioritises attention; recommendation engines match customers with relevant products or information; churn models identify possible lapse risks; conversational AI supports service and distribution; and next-best-action systems coordinate interventions across the relationship.

These applications require more than technical sophistication. Their value depends on reliable data, clear objectives, suitable interventions, controlled experimentation, model validation and continuous monitoring. Insurance recommendations must respect suitability, coverage needs and customer understanding. Conversational systems need grounded content and human escalation. Marketing automation must coordinate channels and avoid excessive contact.

Model risk, explainability, privacy, cybersecurity, bias and human oversight are central management issues. Responsible AI requires proportionality, transparency, security, fairness testing and clear accountability. The most effective systems augment employees and improve customer outcomes rather than treating automation as a substitute for judgement.

Key Terms

Artificial intelligence: Computational capability that performs tasks associated with human cognitive functions.

Machine learning: A method in which systems learn patterns from data to make predictions or decisions.

Data-driven personalisation: Adaptation of communication, service or offers using customer and contextual data.

Predictive lead scoring: Estimation of a prospect’s likelihood of taking a specified action.

Recommendation engine: A system that proposes products, content or actions based on data and constraints.

Churn model: A model estimating the probability that a customer will discontinue a relationship or policy.

Conversational AI: AI that interprets and generates natural-language interaction through text or voice.

Next-best action: The most appropriate permissible action selected for a customer at a particular point in time.

Marketing automation: Technology-enabled orchestration of data, decisions, content, channels and follow-up.

Feature: An input variable used by a predictive model.

Data leakage: Use of information in model development that would not be available at the time of prediction.

Calibration: The extent to which predicted probabilities correspond to observed frequencies.

Model drift: Decline in model relevance as relationships or operating conditions change.

Explainability: The ability to communicate how or why a model produced an output.

Model risk: The possibility of unsuitable results arising from model design, data, use or implementation.

Data minimisation: Limiting collection and use of personal data to what is necessary for a stated purpose.

Human oversight: Meaningful human ability to review, challenge, override or escalate automated outputs.

Uplift: The incremental outcome caused by an intervention compared with an appropriate control condition.

Prompt injection: An attempt to manipulate an AI system through crafted instructions or input.

Review Questions

Short Answer Questions

1. What is the purpose of predictive lead scoring in insurance marketing?

2. Distinguish a recommendation engine from a next-best-action system.

3. Why is calibration important when evaluating a predictive model?

4. Identify two risks associated with conversational AI in insurance.

5. What is meant by meaningful human oversight?

Long Answer Questions

1. Explain the uses, data requirements and evaluation measures for churn models in insurance CRM.

2. Discuss how recommendation engines should incorporate suitability, eligibility and customer understanding.

3. Examine the role of marketing automation in coordinating insurance customer journeys across channels.

4. Analyse model risk and explainability requirements for AI-enabled insurance marketing.

5. Discuss the relationship between privacy, cybersecurity and data-driven personalisation.

Analytical / Application Questions

1. An insurer’s lead-scoring model increases quotation requests but also increases complaints from customers who feel pressured. Diagnose the likely design and governance weaknesses and propose corrective actions.

2. Design a next-best-action workflow for customers approaching motor-policy renewal. Specify data inputs, permissible actions, control rules, human escalation and evaluation measures.

3. A chatbot gives accurate answers to routine questions but produces inconsistent responses about exclusions. Decide whether it should be expanded, restricted or withdrawn, and justify your decision using the chapter’s governance principles.

22. Multiple Choice Questions with Answers

Question 1

Which measure best determines whether a retention intervention caused additional renewals rather than merely identifying customers who would have renewed anyway?

A. Click-through rate

B. A randomised control-group comparison

C. Number of model features

D. Chatbot containment rate

Correct Answer: B. A randomised control-group comparison

Question 2

What is data leakage in predictive lead scoring?

A. Loss of data during a cybersecurity incident

B. Use of information that would not be available when the prediction is made

C. Sharing data with an authorised vendor

D. A decline in data volume over time

Correct Answer: B. Use of information that would not be available when the prediction is made

Question 3

Why should an insurance recommendation engine include product and suitability constraints?

A. To maximise the number of messages sent

B. Because insurance products are interchangeable retail items

C. Because relevance must be balanced with eligibility, coverage needs and customer understanding

D. To eliminate the need for customer consent

Correct Answer: C. Because relevance must be balanced with eligibility, coverage needs and customer understanding

Question 4

Which response is most appropriate when a churn model identifies a customer experiencing payment difficulty?

A. Send repeated promotional offers

B. Offer support information and, where appropriate, human contact

C. Automatically cancel the policy

D. Suppress all service communication

Correct Answer: B. Offer support information and, where appropriate, human contact

Question 5

Which control is particularly important for a generative conversational AI system used by an insurer?

A. Allowing it to answer all questions without restriction

B. Restricting it to approved or retrieved content with escalation for complex matters

C. Measuring only the number of conversations contained

D. Removing conversation logs to improve privacy

Correct Answer: B. Restricting it to approved or retrieved content with escalation for complex matters

Question 6

What does calibration assess in a predictive model?

A. Whether predicted probabilities correspond to observed outcomes

B. Whether the model has the largest possible number of variables

C. Whether customers prefer digital channels

D. Whether an employee agrees with every prediction

Correct Answer: A. Whether predicted probabilities correspond to observed outcomes

Question 7

Which statement best describes meaningful human oversight?

A. A human approves every output without reviewing it

B. A human can understand, challenge, override or escalate automated outputs

C. A vendor accepts responsibility for all decisions

D. A model is retrained every day

Correct Answer: B. A human can understand, challenge, override or escalate automated outputs

Question 8

An insurer’s renewal system sends an offer through the website immediately after an adviser has made the same offer. Which capability would most directly address this problem?

A. A central decision log and cross-channel contact policy

B. Increasing the model’s complexity

C. Removing all customer data

D. Maximising message frequency

Correct Answer: A. A central decision log and cross-channel contact policy

Question 9

Why can removing a protected attribute fail to eliminate bias?

A. Models cannot use any variables without it

B. Other variables may act as proxies for the protected attribute

C. Protected attributes are always irrelevant

D. Bias exists only in generative AI

Correct Answer: B. Other variables may act as proxies for the protected attribute

Question 10

Which is the most appropriate initial objective for a responsible AI renewal programme?

A. Maximise short-term retention at any cost

B. Improve informed renewal and service quality subject to fairness and compliance constraints

C. Replace all advisers with automation

D. Increase the number of customer contacts

Correct Answer: B. Improve informed renewal and service quality subject to fairness and compliance constraints