Digital Insurance, Data Governance, Cybersecurity and AI

Introduction and Context

Insurance is increasingly delivered through digital ecosystems rather than through a single insurer–customer relationship. A policy may be purchased through a mobile application, priced using telematics, underwritten by a machine-learning model, administered in a cloud environment, and serviced through APIs connecting insurers, brokers, hospitals, repair networks, payment providers and public databases. This transformation creates efficiency and innovation, but it also changes the legal location of risk.

The central governance question is no longer merely whether an insurer has a valid policy form. It is whether the insurer can demonstrate that its digital systems produce legally valid contracts, lawful and fair decisions, secure processing, reliable records, explainable outcomes and effective accountability across its entire technology and outsourcing chain.

This chapter uses a jurisdiction-neutral approach. Specific obligations differ among legal systems, but common principles appear in data-protection legislation, electronic-transactions statutes, insurance supervision, outsourcing guidance, cybersecurity rules, consumer-protection law and emerging artificial-intelligence regulation.

Learning Objectives

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

1. Explain the legal foundations of electronic insurance contracting and digital records.
2. Evaluate InsurTech business models, APIs, cloud outsourcing and telematics.
3. Apply data-governance principles to insurance data throughout its lifecycle.
4. Assess cybersecurity, operational resilience and third-party technology risk.
5. Analyse automated underwriting and claims systems for bias, explainability and model risk.
6. Design governance arrangements preserving human oversight and regulatory accountability.
7. Critically evaluate a digital-insurance implementation through a practical case.

1. Digital Insurance as a Sociotechnical System

Digital insurance is not simply insurance conducted online. It is a sociotechnical system consisting of people, legal rules, data, software, infrastructure, business processes and institutional controls. A useful governance model has six layers:

The insurer remains responsible for outcomes even when it uses a vendor, cloud provider or external model. Delegation of performance is not necessarily delegation of legal accountability.

Important terms include InsurTech, meaning technology-enabled innovation in insurance; RegTech, technology used to support regulatory compliance; FinTech ecosystem, a network of financial-service providers and platforms; and embedded insurance, insurance offered within another digital customer journey, such as vehicle purchase or travel booking.

2. Electronic Contracting and Digital Evidence

2.1 Formation of an electronic insurance contract

An electronic contract normally requires the same substantive elements as a paper contract: legal capacity, offer or invitation to treat, acceptance, consideration or premium, lawful purpose, and sufficiently certain terms. Electronic form changes the method of communication and evidence; it does not automatically eliminate traditional contract principles.

A digital insurance journey may include:

1. product advertising and quotation;
2. disclosure of policy terms and exclusions;
3. proposal and risk information supplied by the applicant;
4. underwriting decision;
5. payment or authorization to collect premium;
6. acceptance by the insurer;
7. issuance of a policy and durable record; and
8. subsequent endorsements, renewals or claims communications.

A quotation is not always an offer. Whether it constitutes an offer depends on the wording, commercial context and applicable law. Interfaces should therefore clearly distinguish indicative pricing, application submission and binding acceptance.

2.2 Electronic signatures and authentication

An electronic signature may include a typed name, click acceptance, biometric confirmation, cryptographic signature or another method showing intention to authenticate a record. Legal effectiveness commonly depends on identification, intention, reliability, integrity and the ability to associate the signature with the signed document.

Risk increases where a customer clicks acceptance without reasonable access to terms, where the system permits account sharing, or where a signature cannot be linked to a reliable audit trail. Stronger transactions may require multi-factor authentication, time stamps, device information, certificate validation and tamper-evident storage.

2.3 Notice, incorporation and accessibility

A digital policy should make material terms reasonably available before or at contracting. Merely placing exclusions behind several links may be inadequate if the presentation prevents informed consent. Good practice includes:

• prominent presentation of significant exclusions, waiting periods and deductibles;
• downloadable terms in a durable format;
• version control showing which wording applied at formation;
• accessible language and formats for customers with disabilities;
• records of disclosures, consent and acknowledgements; and
• a process for correcting transmission or interface errors.

The insurer should preserve the exact policy version, proposal responses, premium calculation, underwriting rules, customer communications and relevant system logs. A database entry alone may not prove what the customer saw.

2.4 Electronic evidence

Digital records must be attributable, complete, retrievable and protected against unauthorized alteration. Evidential reliability is strengthened by synchronized clocks, immutable or append-only logs, cryptographic hashes, access controls, retention schedules and documented system controls. A hash can help demonstrate that a file has not changed, but it does not by itself prove that the original file was truthful or lawfully obtained.

3. InsurTech, APIs and Cloud Outsourcing

3.1 InsurTech operating models

InsurTech firms may provide digital distribution, comparison, automated advice, fraud analytics, claims platforms, parametric products, connected-device services, identity verification or infrastructure. The legal classification of a firm matters. A technology vendor may become an intermediary, claims administrator, adjuster, payment service or regulated risk carrier depending on its activities.

Managers should identify the regulated function, not rely on the label “technology provider.” Licensing, conduct, complaints, recordkeeping and capital requirements may attach to the function performed.

3.2 APIs and ecosystem risk

An application programming interface (API) permits systems to exchange structured requests and responses. APIs can improve speed and reduce duplicate data entry, but they create authorization and dependency risks. A secure insurance API programme should address:

• authenticated and authorized access;
• least-privilege permissions;
• encryption in transit and, where appropriate, at rest;
• input validation and rate limiting;
• consent and purpose controls;
• schema and version management;
• monitoring for unusual calls and data exfiltration;
• secure software development and testing; and
• rapid revocation of credentials.

An API may expose personal information even when the insurer’s core database is secure. Data minimization must therefore apply to fields, frequency and retention, not merely to the database as a whole.

3.3 Cloud outsourcing

Cloud services commonly involve infrastructure, platforms or software delivered by a provider. Benefits include elasticity, resilience, analytics capability and reduced capital expenditure. Risks include concentration, subcontracting, foreign access laws, insecure configuration, data commingling, service outages and difficult exit.

An insurer’s outsourcing framework should include due diligence, written allocation of responsibilities, information-security standards, audit and access rights, incident notification, business continuity, data-location analysis, subcontractor controls, regulatory cooperation and termination assistance. The contract should specify measurable service levels rather than broad promises of “industry-standard security.”

The shared-responsibility principle is essential: the cloud provider may secure physical infrastructure and core services, while the insurer remains responsible for identity management, configuration, data classification, application security and lawful processing. Responsibility must be documented for each control.

4. Data Governance in Insurance

4.1 Data as an asset and a liability

Insurance depends on information asymmetry reduction. Data can improve pricing, reserving, fraud detection, prevention and claims settlement. The same data can expose intimate details about health, location, behaviour, finances and family relationships. Data governance therefore combines commercial value with legal and ethical restraint.

A mature governance programme defines ownership, stewardship, quality standards, access rights, retention, lineage, classification, lawful purpose and disposal. It should cover structured data, documents, call recordings, images, sensor streams, model features and inferred attributes.

4.2 Core principles

Although terminology varies, effective data-protection regimes commonly require:

• Lawfulness, fairness and transparency: processing needs a valid legal basis and must not mislead or unfairly disadvantage individuals.
• Purpose limitation: information collected for one purpose should not be reused incompatibly without authorization or another lawful basis.
• Data minimization: collect information reasonably necessary for the defined purpose.
• Accuracy: correct material errors and provide mechanisms for challenge.
• Storage limitation: retain information no longer than justified by legal, contractual, actuarial, evidential or operational needs.
• Security and confidentiality: protect against unauthorized access, alteration, loss and disclosure.
• Accountability: maintain evidence that controls operate in practice.

Consent is not a universal solution. It may be invalid if bundled, uninformed, coerced or impossible to withdraw. An insurer must distinguish consent for data processing from acceptance of policy terms, marketing permission and authorization to access third-party records.

4.3 Sensitive and inferred data

Health, genetic, biometric, precise location and financial information often receive enhanced legal protection. Even apparently ordinary data can become sensitive through combination or inference. A driving score, for example, may reveal work patterns, medical conditions or religious observance without explicitly recording them.

Data protection impact assessments are appropriate where processing is large-scale, systematic, intrusive, innovative or likely to create significant effects. The assessment should identify purposes, data flows, affected persons, risks, safeguards and residual risk acceptance.

4.4 Data quality and lineage

Bad data produces bad underwriting and claims decisions. Data lineage records where information originated, how it was transformed, which model used it and where the result was stored. Controls should test completeness, validity, consistency, timeliness, duplication and representativeness.

For a variable x, a simple missingness rate may be expressed as:

Here, n_missing is the number of records missing the variable and n_total is the total number of relevant records. The metric is meaningful only when the population and time period are defined. A low missingness rate does not establish accuracy or fairness.

5. Cybersecurity and Operational Resilience

5.1 Threat landscape

Digital insurers face phishing, credential theft, ransomware, supply-chain compromise, denial-of-service attacks, malicious insiders, API abuse, fraudulent claims, model manipulation and data poisoning. Connected devices add risks involving insecure firmware, spoofed sensors and unauthorized tracking.

Cybersecurity objectives are often expressed through confidentiality, integrity and availability. Insurance adds authenticity, non-repudiation, privacy and evidential reliability because a corrupted record may affect coverage, liability or payment.

5.2 Governance controls

A risk-based security programme should include:

1. asset inventory and data classification;
2. identity governance and multi-factor authentication;
3. secure configuration and vulnerability management;
4. encryption and key management;
5. network segmentation and endpoint protection;
6. secure coding and independent testing;
7. continuous monitoring and anomaly detection;
8. tested incident response;
9. backups protected from destructive compromise; and
10. post-incident learning and regulatory reporting.

Security controls should be proportionate to impact. A public quotation page and a claims payment system do not have identical risk profiles, although both require protection.

5.3 Incident response and resilience

An incident plan should identify decision rights, legal notification triggers, customer communications, forensic preservation, containment, recovery and claims-handling continuity. “Availability” is not merely an IT issue: an outage may prevent policy purchase, stop emergency assistance or delay legally significant notification.

Business impact analysis should identify critical services, maximum tolerable disruption, recovery priorities and dependencies. Recovery objectives must be realistic and tested. A recovery point objective concerns how much data loss is tolerable; a recovery time objective concerns how quickly a service must be restored. These are management requirements, not universal technical constants.

6. Telematics and Connected Insurance

Telematics uses connected devices to capture information such as speed, braking, mileage, location or vehicle condition. In health or property insurance, connected devices may measure activity, environmental conditions or maintenance indicators.

Potential benefits include usage-based pricing, accident prevention, rapid assistance and more accurate risk segmentation. Risks include surveillance, device error, consent fatigue, discriminatory pricing, cybersecurity compromise and disputes about whether a sensor accurately represented behaviour.

A telematics programme should specify:

• what is measured and at what frequency;
• whether raw data or derived scores are retained;
• who can access the information;
• how device error and missing signals are treated;
• how customers challenge a score;
• whether pricing consequences are material;
• how tracking is disabled or withdrawn; and
• how data is deleted or anonymized.

The insurer should not treat a device-generated score as objective merely because it is numerical. Measurement design embeds choices about sampling, thresholds, missing data and interpretation.

7. Automated Underwriting and Claims

7.1 Automation levels

Automation ranges from decision support, where a human makes the decision using software, to fully automated decisions with no meaningful human intervention. Intermediate systems may automatically triage, request documents, calculate prices, flag fraud or recommend settlement values.

A useful control classification is:

7.2 Model risk

Model risk is the possibility that a model produces harmful consequences because of incorrect specification, poor data, implementation error, misuse or changed conditions. A model may be statistically accurate yet legally unacceptable if it uses unlawful variables or creates unjustified disparate effects.

Model governance should include inventory, ownership, purpose statement, data documentation, validation, performance monitoring, approval thresholds, change control, incident reporting and retirement. Independent validation should test conceptual soundness, data quality, coding, performance, stability and limitations.

Common performance measures include accuracy, precision, recall and calibration, but no single metric is sufficient. For a binary classifier:

Precision asks how many flagged cases were actually positive; recall asks how many actual positive cases were identified. The measures depend on the threshold and the definition of “positive.” In fraud detection, a false positive can cause reputational harm or unfair investigation, while a false negative may increase loss. Threshold selection is therefore a governance decision, not merely a technical optimization.

7.3 Bias and fairness

Bias can enter through historical decisions, unrepresentative samples, proxy variables, labels, measurement error, unequal access to data or deployment context. Removing a protected attribute does not necessarily remove discrimination because other features may act as proxies.

Fairness analysis should compare relevant outcomes across legally protected or otherwise vulnerable groups, while respecting privacy and statistical limitations. Possible indicators include selection-rate differences, error-rate differences and calibration. These measures can conflict; improving one may worsen another. The correct standard is determined by law, product purpose, actuarial justification, social impact and governance judgment—not by a universal fairness formula.

Insurers should document why each material feature is used, test alternative features, examine outcomes, provide correction routes and prohibit high-impact uses without appropriate controls. Actuarial relevance does not automatically defeat equality or consumer-protection obligations.

7.4 Explainability and adverse decisions

Explainability is context-dependent. A technical description of neural-network architecture is not a useful customer explanation. A meaningful explanation should identify the principal factors, relevant data, decision consequence, applicable rule, and route for correction or review.

Explanations must not reveal security-sensitive detection logic or confidential information about another person. Nevertheless, trade-secret concerns do not justify opaque, unreviewable decisions where law requires reasons or due process.

7.5 Human oversight

Human oversight must be substantive rather than ceremonial. A reviewer needs authority, time, information, competence and freedom to depart from the automated recommendation. Rubber-stamping is not meaningful oversight.

Escalation should be mandatory for vulnerable customers, unusual cases, material adverse decisions, low-confidence outputs, contradictory evidence, complaints and suspected data errors. The human decision and reasons should be recorded separately from the model output.

8. AI Regulatory Accountability

AI governance should allocate responsibility across the board, senior management, product owner, data-protection officer, chief information-security officer, actuarial function, model-risk function, compliance, internal audit and operational teams. Committees do not replace accountable individuals.

A practical accountability framework asks:

• What is the system intended to do?
• What legal and insurance function does it perform?
• Which persons may be affected?
• What data and vendors does it depend on?
• What could go wrong?
• What evidence demonstrates compliance?
• Who can stop, alter or override it?
• How are customers notified and heard?
• How will performance and fairness be monitored after deployment?

Documentation should be proportionate but sufficient to reconstruct the decision. Useful records include model cards, data sheets, impact assessments, validation reports, approval minutes, change logs, incident records, customer explanations and override analysis.

Generative AI creates additional risks: fabricated outputs, leakage of confidential claims information, inconsistent reasoning, prompt injection, copyright concerns and uncontrolled vendor retention. It should not autonomously issue legal or coverage conclusions without controlled workflows, verified sources and qualified review.

9. Illustrative Case: Horizon Motor’s Telematics Claims Platform

Background

Horizon Motor Insurance, a fictional regional insurer, introduced a mobile telematics product for young drivers. Customers received discounts based on acceleration, braking, night driving and mileage. Claims were submitted through an application that used image recognition to estimate vehicle damage. A cloud vendor hosted the platform, and an external fraud model ranked claims for investigation.

Problem

Within six months, complaints increased. Customers reported unexplained premium increases, inability to correct inaccurate journeys and delayed claims where the fraud score was high. An internal review found that phones shared by family members were sometimes attributed to the policyholder. The fraud model produced more investigation referrals for drivers from certain districts. The cloud vendor had also engaged a subcontractor without timely notice to Horizon.

Analysis

The insurer had treated the telematics score as an objective fact rather than an inference. It lacked a documented policy for missing or contradictory location data. The claims model’s output was a fraud indicator, but investigators treated it as proof. No group-impact testing had been performed. The outsourcing contract contained security commitments but weak audit and subcontractor provisions. Customer notices described “personalized pricing” without clearly explaining material data uses.

Approach and solution

Horizon took five steps. First, it paused automatic adverse premium changes where data quality was uncertain. Second, it created a correction and appeal process, allowing customers to submit evidence and receive a human review. Third, it separated fraud triage from claim denial and required corroborating evidence before adverse action. Fourth, it conducted a data-protection and fairness assessment, tested alternative features and introduced monitoring by customer segment. Fifth, it renegotiated vendor contracts to require subcontractor notification, audit access, incident reporting, deletion assistance and tested exit arrangements.

The board approved a model inventory and appointed accountable owners. A model-risk committee required pre-deployment validation, quarterly performance review and documented approval for material changes. Security teams implemented stronger API authentication, device binding and anomaly monitoring.

Outcome

The insurer reduced false-positive investigations, improved complaint resolution and identified several device-attribution errors. Pricing became slightly less granular but more defensible. The company also discovered that a lower-performing but more explainable model was more suitable for customer-facing pricing, while the complex model remained useful as a monitored advisory tool.

Learning points

1. Sensor data is not automatically accurate or fair.
2. A risk flag is not proof of fraud.
3. Human review requires real authority and evidence.
4. Vendor contracts must support regulatory supervision and exit.
5. Model performance, fairness and data quality require continuous monitoring.
6. A less complex model may produce better governance outcomes when decisions are consequential.

10. Implementation Framework for Insurers

A practical implementation sequence is:

Step 1: Map the service

Document customer journeys, decisions, data flows, systems, vendors, jurisdictions and critical dependencies.

Step 2: Classify impact

Identify low-, medium- and high-impact processing. Consider financial loss, denial of coverage, privacy intrusion, discrimination, physical safety and reputational harm.

Step 3: Establish lawful purpose and necessity

Define the purpose, legal basis, notice, retention and permitted reuse. Reject data fields that cannot be justified.

Step 4: Build controls before deployment

Complete security testing, privacy assessment, model validation, accessibility review, resilience testing and human-escalation design.

Step 5: Contract for accountability

Ensure vendors provide audit evidence, incident notification, subcontractor transparency, data return or deletion, continuity and cooperation with regulators.

Step 6: Monitor in production

Track accuracy, complaints, overrides, drift, disparate outcomes, security events, availability, data-quality failures and customer correction requests.

Step 7: Review and retire

Reassess legality, performance and necessity after material changes or incidents. Retire systems whose risks cannot be controlled.

Challenges and Limitations

Digital governance faces several tensions. More data may improve prediction but increase privacy intrusion. Greater automation may reduce cost but make errors harder to contest. Model transparency may conflict with security or intellectual property. Fairness metrics may conflict with one another. Cross-border cloud services may improve resilience while complicating jurisdiction and regulatory access.

Regulation also evolves unevenly. An insurer may face overlapping data-protection, insurance, outsourcing, consumer, cybersecurity and AI requirements. Compliance should therefore be designed around risk and accountability rather than around isolated legal checklists. Small insurers may lack specialist resources, while large firms may suffer from fragmented legacy systems and excessive complexity.

Best-Practice Principles

• Treat digital capability as a board-level risk and conduct issue.
• Assign one accountable owner to every material automated decision.
• Keep an inventory of models, data assets, APIs and critical vendors.
• Use privacy and security by design, not as post-deployment repair.
• Make explanations useful to affected customers.
• Test data quality, fairness and resilience in the real deployment context.
• Separate automated flags from final adverse decisions.
• Preserve durable records and evidence of human review.
• Require vendors to support audit, incident response and exit.
• Prefer controlled, explainable systems where consequences are high.

Chapter Summary

Digital insurance combines electronic contracting, data-intensive underwriting, cloud infrastructure, APIs, telematics and automated claims. Its benefits depend on lawful processing, reliable evidence, secure operations and fair treatment. Electronic contracts require clear formation, accessible terms, authentication and durable records. Cloud and API arrangements require risk-based due diligence, shared-responsibility controls, audit rights and exit planning.

Data governance covers purpose, necessity, quality, security, retention, lineage and accountability. Cybersecurity must protect confidentiality, integrity, availability and evidential reliability. Telematics and AI systems require particular attention to measurement error, sensitive inference, proxy bias, explainability and meaningful human review. Model risk includes specification, data, implementation, misuse and changing-environment risks.

The insurer remains accountable for regulated outcomes even when technology is supplied by a third party. Effective governance therefore connects board oversight, legal compliance, actuarial judgment, information security, privacy, model validation, customer redress and regulatory cooperation. The appropriate objective is not maximum automation, but trustworthy and demonstrably accountable digital insurance.

Key Terms

InsurTech: Technology-enabled innovation in insurance products, distribution, underwriting or claims.

Electronic contract: A contract formed or evidenced through electronic communications or records.

Electronic signature: Electronic data associated with a record to identify a person and indicate intention to authenticate or approve it.

API: An application programming interface that enables structured communication between software systems.

Cloud outsourcing: Use of an external provider’s remotely delivered computing infrastructure, platforms or software.

Telematics: Collection and transmission of behavioural, location or condition data from connected devices.

Data minimization: Limiting collection and use to information necessary for a defined purpose.

Data lineage: A record of data origin, transformation, movement and use.

Personal data: Information relating to an identified or identifiable individual.

Sensitive data: Personal information receiving enhanced protection because of its nature or potential impact.

Cybersecurity: Protection of information systems and data from unauthorized access, alteration, disruption or destruction.

Operational resilience: The ability to continue and recover important services during disruption.

Automated underwriting: Use of software or models to assess risk, eligibility or price.

Model risk: Risk of adverse consequences from model specification, data, implementation, use or change.

Explainability: The ability to provide understandable reasons for a system’s output.

Algorithmic bias: Systematic unfairness arising from data, design, implementation or use of an algorithm.

Human oversight: Competent and empowered human supervision of an automated system.

Data protection impact assessment: A structured assessment of privacy risks and safeguards for proposed processing.

Embedded insurance: Insurance offered within another product or digital customer journey.

Regulatory accountability: The demonstrable responsibility of an insurer and its officers for lawful, fair and controlled operations.

Review Questions

Short Answer Questions

1. What distinguishes an electronic signature from a mere electronic record?
2. Why does outsourcing technology not eliminate an insurer’s accountability?
3. Define data minimization and explain its relevance to telematics.
4. What is model risk?
5. Why should a fraud score not automatically result in claim denial?

Long Answer Questions

1. Explain the legal and evidential requirements for a robust electronic insurance-contracting process.
2. Evaluate the principal benefits and risks of cloud outsourcing and APIs in insurance.
3. Discuss how data governance supports privacy, underwriting accuracy and regulatory accountability.
4. Analyse the relationship between algorithmic bias, explainability and human oversight.
5. Design a governance framework for an insurer deploying automated claims settlement.

Analytical / Application Questions

1. An insurer uses location data to increase premiums for customers whose vehicles are frequently driven at night. Identify the privacy, fairness, data-quality and explainability issues that should be investigated.
2. A cloud-hosted claims model begins producing unusually high fraud referrals after a vendor changes its training data. What controls, evidence and decisions should the insurer require?
3. Compare two implementation choices for high-impact underwriting: a complex opaque model with slightly greater predictive accuracy and a simpler explainable model with slightly lower accuracy. Recommend a choice and justify it using legal, actuarial, ethical and governance considerations.

Multiple Choice Questions with Answers

Question 1

Which statement best describes electronic insurance contracting?

A. Electronic form removes the need for offer, acceptance and consideration.

B. Electronic form changes the method of communication but does not automatically remove substantive contract requirements.

C. A contract is valid whenever a customer opens a policy webpage.

D. Only cryptographic signatures can create valid insurance contracts.

Correct Answer: B. Electronic form changes the method of communication but does not automatically remove substantive contract requirements.

Question 2

Which control most directly supports proof of the policy wording accepted by a customer?

A. Deleting previous policy versions after renewal

B. Using a durable, version-controlled record of the terms and acceptance event

C. Allowing staff to edit the policy without an audit trail

D. Displaying terms only through a temporary pop-up window

Correct Answer: B. Using a durable, version-controlled record of the terms and acceptance event

Question 3

Under the shared-responsibility principle in cloud computing, which responsibility generally remains with the insurer?

A. Physical security of the provider’s data-centre building in every case

B. Manufacture of the provider’s servers

C. Customer identity, access configuration and lawful handling of insurance data

D. All responsibility for the provider’s underlying electricity supply

Correct Answer: C. Customer identity, access configuration and lawful handling of insurance data

Question 4

What is the principal purpose of data minimization?

A. To collect every potentially useful variable before defining a purpose

B. To limit collection and use to information reasonably necessary for a defined purpose

C. To retain all data permanently for future analytics

D. To prohibit insurers from using any personal data

Correct Answer: B. To limit collection and use to information reasonably necessary for a defined purpose

Question 5

A fraud model identifies 80 genuine fraud cases and 20 false alarms. What is its precision?

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 6

Why is a telematics score not automatically an objective fact?

A. Numerical data cannot be used in insurance.

B. The score depends on choices about measurement, thresholds, missing data and interpretation.

C. Telematics devices never produce useful information.

D. Only human observations can be relevant to underwriting.

Correct Answer: B. The score depends on choices about measurement, thresholds, missing data and interpretation.

Question 7

Which is the best example of meaningful human oversight?

A. A reviewer automatically approves every model recommendation.

B. A reviewer has authority, evidence, competence and time to depart from the model output.

C. A software label states that a human is responsible, although no review occurs.

D. A customer is denied access to the reasons for a decision.

Correct Answer: B. A reviewer has authority, evidence, competence and time to depart from the model output.

Question 8

Which statement about removing protected attributes from a model is correct?

A. It always eliminates discrimination.

B. It guarantees legal compliance.

C. Other variables may act as proxies, so fairness testing remains necessary.

D. It makes model validation unnecessary.

Correct Answer: C. Other variables may act as proxies, so fairness testing remains necessary.

Question 9

Which contractual provision is most important for regulatory accountability in critical outsourcing?

A. A prohibition on the insurer conducting audits

B. A clause allowing unlimited undisclosed subcontracting

C. Audit, incident notification, regulatory cooperation and exit-assistance rights

D. A promise that the service will always be inexpensive

Correct Answer: C. Audit, incident notification, regulatory cooperation and exit-assistance rights

Question 10

Which statement best characterizes a model-risk programme?

A. It tests only whether software runs without crashing.

B. It covers model purpose, data, validation, performance, changes, limitations and use.

C. It applies only to mathematical models used for reserving.

D. It ends once a model is approved for its first deployment.

Correct Answer: B. It covers model purpose, data, validation, performance, changes, limitations and use.