InsurTech, AI and the Future of Underwriting and Claims

Introduction

Insurance is being reshaped by connected devices, cloud computing, artificial intelligence (AI), application programming interfaces (APIs), digital distribution and distributed ledgers. These technologies do not eliminate the fundamental functions of insurance—risk selection, pricing, pooling, contract administration and claims payment—but they change how information is collected, decisions are made and customers are served.

For insurers, the central challenge is not simply to automate more activities. It is to automate appropriate activities while preserving fairness, explainability, security, professional judgment and legal accountability. A fast decision based on poor data can produce worse underwriting than a slower decision based on reliable evidence. Similarly, a fully automated claim may be efficient for a simple, low-value loss but inappropriate for a disputed bodily injury claim or a catastrophe involving vulnerable policyholders.

This chapter evaluates digital distribution, automated underwriting, telematics, the Internet of Things (IoT), remote assessment, computer vision, generative AI, straight-through claims processing and blockchain. It also examines model risk, explainability, cybersecurity, human oversight, climate risk and the implications for the future insurance workforce.

Learning objectives

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

1. Explain how digital technologies affect insurance distribution, underwriting and claims.
2. Distinguish rules-based automation, predictive models and generative AI.
3. Evaluate telematics, IoT, remote assessment and computer vision in practical insurance settings.
4. Apply expected-loss reasoning to automated underwriting and claims decisions.
5. Identify model, data, cyber, ethical and regulatory risks.
6. Design an appropriate human-in-the-loop control framework.
7. Assess climate-risk and workforce implications of InsurTech.

1. Digital distribution and the changing insurance value chain

Digital distribution includes insurer websites, mobile applications, comparison platforms, embedded insurance, digital brokers, bancassurance APIs and partner ecosystems. A customer may purchase travel cover during an airline booking, device insurance at the point of sale or usage-based motor insurance through a mobile application.

Digital channels can reduce transaction costs, improve accessibility and produce structured data. They may also create new risks: customers can misunderstand short digital disclosures, comparison tools can encourage price-only decisions, and platform algorithms can steer customers toward products that are commercially attractive rather than suitable.

A digital insurance journey commonly contains these stages:

• customer identification and consent;
• quotation and product comparison;
• data collection and risk assessment;
• payment and policy issuance;
• servicing, renewal and cancellation;
• first notification of loss (FNOL), assessment and settlement.

The quality of the journey depends on interoperability. APIs can connect insurers with payment providers, vehicle databases, weather services, repair networks and identity-verification systems. However, data sharing must have a lawful purpose, appropriate consent or another valid legal basis, retention limits and controls over onward use.

2. Automated underwriting

Automated underwriting uses software to determine eligibility, rating, limits, exclusions or referral requirements. It ranges from simple rules—such as declining an application outside an approved territory—to statistical and machine-learning models that estimate claim frequency, severity or lapse probability.

Rules-based and predictive underwriting

A rules engine is relatively transparent: the insurer can identify which rule produced a result. Predictive models may detect nonlinear relationships and interactions that are difficult to observe manually. Their usefulness depends on representative training data, stable relationships and effective monitoring.

A basic expected-loss model illustrates the economic logic. Suppose possible outcomes are collectively exhaustive: no claim, a partial loss, or a total loss. If the probabilities sum to one, expected loss is:

Here, L is loss, pᵢ is the probability of outcome i, and Lᵢ is the loss associated with that outcome. The no-claim outcome must be included when probabilities are intended to cover all possible outcomes; its loss is zero.

For example, an insurer estimates a 0.92 probability of no claim with loss $0, a 0.06 probability of a $2,000 partial loss and a 0.02 probability of a $20,000 total loss:

Expected loss is not automatically the premium. A simplified technical premium may be represented as:

where P is the indicated premium, C is expected claims-handling and acquisition expense, R is a risk or capital margin, and M is other permitted margin or adjustment. Actual pricing also considers credibility, reinsurance, expenses, regulation, competition and portfolio effects.

Automated decision design

An underwriting system should define at least four outcomes: accept, accept with modified terms, refer to an underwriter, and decline. Referral is important because uncertainty is not equivalent to high risk. A model may be unfamiliar with a new construction type, have missing data or operate outside its validated population.

Automation thresholds should be explicit. A confidence score is not necessarily a probability of correctness, and a probability estimate is not necessarily calibrated. If an insurer wants fewer automatic decisions, it should state whether it will raise the minimum confidence required for automation, reduce the maximum eligible-value limit, narrow the risk classes permitted for automation, or use a combination of these measures.

3. Telematics and the Internet of Things

Telematics combines telecommunications, sensors and analytics. In motor insurance, devices or smartphone applications may record distance, time of travel, acceleration, braking, cornering and location. IoT applications also include water-leak sensors, fire alarms, smart meters, industrial equipment monitors and wearable health devices.

The potential benefits are improved risk differentiation, prevention, faster evidence collection and behavioural feedback. A telematics programme may reward safer driving or alert a homeowner before a pipe failure causes extensive damage.

The risks include privacy intrusion, inaccurate sensors, device tampering, unequal access to technology and the possibility that behavioural data will be reused beyond the original purpose. Location data may reveal religious attendance, medical visits or employment patterns. Data minimisation and purpose limitation are therefore central controls.

Telematics should not be treated as automatically objective. A braking score can be affected by traffic density, road design, weather, vehicle type and emergency avoidance. Fair underwriting requires testing for disparate impacts and assessing whether apparently predictive variables act as proxies for protected characteristics.

4. Remote assessment and computer vision

Remote assessment uses video calls, mobile photographs, drones, satellite imagery, geospatial data and digital measurement tools. It can reduce travel time, accelerate catastrophe response and provide evidence when physical access is unsafe.

Computer vision can classify damage in photographs, estimate the location and extent of damage, compare images with repair-cost databases and identify inconsistencies. Typical pipeline stages are image capture, quality checking, object or damage detection, severity estimation, human or automated validation, and claims-record integration.

Performance must be evaluated by class, not only by overall accuracy. Relevant measures include false-positive and false-negative rates, calibration, processing time and settlement variance. A model that performs well on clear daylight photographs may fail on night images, low-bandwidth uploads, culturally different building styles or unfamiliar forms of damage.

Computer vision should support, not silently replace, evidential judgment. Customers need an opportunity to submit additional evidence or challenge an assessment. Images also create security and privacy risks because metadata may reveal location and time.

5. Generative AI in insurance

Generative AI produces text, summaries, code, images or dialogue from prompts. In insurance it can assist with policy explanations, call-centre support, claims summarisation, document classification, knowledge retrieval and drafting correspondence.

Its main hazards are hallucinated facts, inaccurate citations, leakage of confidential information, inconsistent outputs, automation bias and inappropriate tone. A language model should not independently invent coverage terms or make an unreviewed adverse decision. Retrieval-augmented systems can constrain responses to approved policy and procedural sources, but retrieval does not guarantee correct interpretation.

A controlled implementation should include approved use cases, protected data environments, prompt and output logging, source grounding, testing, escalation rules, access controls and post-deployment monitoring. Staff must understand that fluent language is not evidence of factual reliability.

6. Straight-through claims processing

Straight-through processing (STP) means that a claim moves through defined digital stages without manual intervention when eligibility and risk conditions are satisfied. It is most suitable for simple, low-value, well-documented claims, such as a delayed-flight benefit or a minor repair with verified coverage.

A sound claims workflow is sequential:

Figure: Sequential claims workflow from FNOL through assessment to settlement or human review.

STP requires more than an automated payment. It requires reliable identity verification, policy and coverage matching, evidence validation, fraud screening, reserving logic, payment controls, customer notification and audit trails. A claim should be referred when evidence conflicts, the value exceeds a defined limit, vulnerability is identified, coverage is ambiguous, fraud indicators are material or the model is outside its validated range.

7. Blockchain and distributed ledgers

A blockchain is a distributed ledger in which transactions are recorded through a consensus mechanism and linked cryptographically. Potential insurance applications include parametric products, shared records among authorised participants, reinsurance data exchange and automated payment through smart contracts.

A smart contract is software that executes specified logic; it is not necessarily a complete insurance contract and cannot resolve every question of interpretation. A parametric policy may pay when an objectively measured trigger—such as rainfall below a specified level—occurs. This can accelerate payment, but it creates basis risk: the measured index may not reflect the policyholder’s actual loss.

Blockchain does not make input data truthful. An inaccurate weather feed remains inaccurate when immutably recorded. Privacy, scalability, governance, key management, legal recognition and correction of erroneous data must be addressed. Permissioned networks may be more suitable than open networks where participants and data access require control.

8. Model risk, explainability and human oversight

Model risk is the possibility that a model produces harmful outcomes because of incorrect specifications, poor data, implementation errors, instability or inappropriate use. It includes data drift, concept drift, proxy discrimination, feedback loops and misuse outside the development population.

A model-risk framework should cover:

• documented purpose, scope and ownership;
• data lineage, quality and representativeness;
• independent validation and challenge;
• performance, calibration and fairness testing;
• version control and change approval;
• monitoring thresholds and incident response;
• records sufficient for audit, complaints and legal review.

Explainability has several meanings. Global explanation describes general drivers of model behaviour; local explanation describes why a particular case received a result; procedural explanation describes how a customer can challenge or correct the decision. Feature importance alone may not be a legally or ethically adequate explanation, especially when correlated variables obscure the true basis of a decision.

Human oversight must be substantive. A reviewer needs authority, time, training, relevant evidence and protection from performance targets that encourage rubber-stamping. The system should record who reviewed a decision, what information was considered and why the final outcome was reached. Human involvement is not a safeguard if the human cannot override the system or lacks the competence to do so.

9. Cybersecurity, privacy and operational resilience

Insurers hold identity, financial, health, vehicle, property and behavioural data, making them attractive targets. Threats include ransomware, credential theft, API abuse, supply-chain compromise, model poisoning, adversarial inputs and denial-of-service attacks.

Core controls include least-privilege access, multifactor authentication, encryption in transit and at rest, secure software development, network segmentation, vulnerability management, immutable backups, supplier assurance, anomaly monitoring and tested recovery plans. AI systems additionally require protection against prompt injection, sensitive-data disclosure and unauthorised model extraction.

Security must be linked to resilience. An insurer should identify critical services, define recovery objectives, maintain alternate channels and test catastrophe scenarios. During a major storm, for example, customers may need telephone or assisted digital reporting if applications fail.

10. Climate risk and catastrophe technology

Climate change affects underwriting through physical risk—flood, wildfire, storm, heat and drought—and transition risk arising from changes in technology, regulation and asset values. Historical claims data may understate future hazard frequency or severity. Models therefore need forward-looking hazard information, scenario analysis and careful treatment of uncertainty.

Satellite imagery, weather feeds, digital elevation models and IoT sensors can improve prevention and rapid response. Yet automated settlement after a catastrophe requires conservative controls. If image quality deteriorates or loss patterns move outside validation data, the insurer should reduce the eligible-value limit or raise the confidence threshold for automatic settlement; saying merely that it “lowers the automation threshold” is ambiguous because it could imply more automation.

Climate models should be stress-tested across plausible scenarios rather than presented as precise forecasts. Underwriting decisions also have social consequences: withdrawing cover from high-risk areas may protect solvency but reduce affordability and economic resilience. Mitigation incentives, risk-prevention grants, public-private pools and parametric products may complement, but not replace, responsible availability strategies.

11. An illustrative case: automated storm claims

Background

A fictional regional insurer introduces a mobile claims process for residential storm damage. Customers submit identity information, policy details, photographs and a short description. Weather data confirms whether a storm occurred in the reported location.

Problem

The insurer wants rapid settlement, but early testing shows that photographs taken in darkness and claims involving older roofs are frequently misclassified. A purely automated system would risk both overpayment and unfair rejection.

Analysis

The insurer segments claims by value, image quality, roof type and uncertainty score. It discovers that overall accuracy conceals a substantially higher false-negative rate for older roofs. It also identifies a cyber risk: attackers could reuse photographs from earlier claims.

Approach

The insurer adopts the following controls:

1. automated identity, policy and weather checks;
2. image-quality testing and metadata review;
3. computer-vision triage rather than unconditional settlement;
4. automatic settlement only below a defined value limit and above a validated confidence threshold;
5. mandatory human review for low-quality images, older roofs, conflicting evidence or vulnerability indicators;
6. duplicate-image detection, payment controls and audit logging;
7. customer notification explaining the evidence used and the appeal route.

After a severe storm, the insurer raises the confidence threshold and lowers the maximum eligible-value limit for auto-settlement because conditions differ from the training data. It does not simply claim to have “lowered the automation threshold.”

Outcome

Simple, well-evidenced claims are paid quickly. Ambiguous cases are reviewed by trained examiners, and the insurer monitors error rates by property type and neighbourhood. The process is slower for some claims but more reliable and defensible.

Learning points

The case demonstrates that automation is a controlled operating range, not an all-or-nothing choice. Catastrophe conditions require adaptive thresholds, human escalation, data-security controls and fairness monitoring.

12. Future workforce and professional practice

Automation is likely to reduce repetitive data entry, document matching and routine status enquiries while increasing demand for data literacy, model validation, cyber risk, customer advocacy, claims investigation, climate analytics and technology governance. Underwriters and claims professionals will increasingly supervise portfolios of automated decisions rather than process every case individually.

Future competencies include interpreting model outputs, questioning data quality, explaining decisions to customers, recognising vulnerable circumstances, documenting overrides and collaborating with data scientists. Professional judgment remains valuable where evidence is incomplete, interests conflict or the consequences of error are severe.

Insurers should invest in reskilling, clear accountability and multidisciplinary teams. A responsible transformation programme measures not only cost and processing speed, but also complaint rates, correction rates, fairness outcomes, cyber incidents, customer comprehension and claims quality.

Best-practice framework

A practical governance framework can be organised around six questions:

Chapter Summary

InsurTech changes how insurers distribute products, gather evidence, assess risk and pay claims. Automated underwriting can improve consistency and speed, but expected loss is only one component of pricing and must be based on collectively exhaustive outcomes when probabilities sum to one. Telematics and IoT enable prevention and behavioural pricing but raise privacy, accuracy and fairness concerns. Remote assessment and computer vision accelerate evidence handling but require subgroup testing and human challenge routes. Generative AI is useful for assistance and summarisation but must be constrained by data, security and review controls.

Straight-through claims processing is appropriate for defined, low-complexity cases and must follow a complete sequence from FNOL through coverage, loss and fraud assessment to settlement or human review. Blockchain can support shared records and parametric products but cannot guarantee truthful inputs. Model risk, explainability, cybersecurity, climate uncertainty and human oversight are central governance issues. The future workforce will combine insurance judgment with data, technology, ethics and resilience skills.

Key Terms

InsurTech: Technology-enabled innovation in insurance.

Digital distribution: Sale and servicing of insurance through digital channels and platforms.

Automated underwriting: Software-supported determination of risk, terms or referral.

Telematics: Use of telecommunications and sensors to collect and analyse usage or behaviour data.

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

Computer vision: AI methods that interpret images or video.

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

FNOL: First notification of loss made to an insurer.

Straight-through processing: End-to-end processing without manual intervention within defined controls.

Model risk: Risk of adverse outcomes from model error, misuse or instability.

Calibration: Agreement between predicted probabilities and observed frequencies.

Data drift: Change in the distribution of input data over time.

Explainability: Ability to communicate how and why a model or decision operates.

Human-in-the-loop: Design in which human expertise can review or control automated decisions.

Blockchain: Distributed ledger using cryptographically linked records and a consensus process.

Smart contract: Software that executes predefined conditions and actions.

Parametric insurance: Cover triggered by an objectively measured index rather than direct loss adjustment.

Basis risk: Risk that the trigger or index does not match the policyholder’s actual loss.

Climate risk: Physical and transition risks arising from climate change and response measures.

Operational resilience: Capacity to continue and recover critical services during disruption.

Review Questions

Short Answer Questions

1. What is digital distribution, and what are two risks associated with it?
2. Why must the no-claim outcome be included in a complete expected-loss model?
3. Distinguish telematics from the broader Internet of Things.
4. What is meant by straight-through claims processing?
5. Why is human oversight important in automated insurance decisions?

Long Answer Questions

1. Explain the benefits and limitations of automated underwriting.
2. Discuss privacy, fairness and data-quality issues in telematics insurance.
3. Evaluate the use of computer vision in remote claims assessment.
4. Explain model-risk governance and meaningful explainability.
5. Assess the role of blockchain and smart contracts in insurance.

Analytical / Application Questions

1. Design referral rules for an automated household-claims system that receives photographs of varying quality.
2. An insurer observes increasing catastrophe losses despite stable historical model predictions. Analyse possible causes and recommend controls.
3. Develop a human-in-the-loop framework for a generative-AI claims assistant, including data, security, performance and customer-appeal controls.

Multiple Choice Questions with Answers

Question 1

In a collectively exhaustive expected-loss model, which outcome must be included when it is possible?

A. A no-claim outcome with zero loss

B. Only the largest possible claim

C. Only outcomes that generate premiums

D. An outcome with a negative probability

Correct Answer: A. A no-claim outcome with zero loss

Question 2

Which sequence correctly represents the core claims workflow described in the chapter?

A. Settlement, FNOL, fraud screening, coverage check

B. FNOL, coverage and eligibility, loss assessment, fraud screening, settlement or human review

C. Underwriting decision, settlement, FNOL, loss assessment

D. Fraud screening, settlement, FNOL, policy issuance

Correct Answer: B. FNOL, coverage and eligibility, loss assessment, fraud screening, settlement or human review

Question 3

If an insurer wants fewer claims to qualify for automatic settlement after a severe storm, which action is unambiguous?

A. Lower the confidence requirement and raise the value limit

B. Raise the confidence requirement or lower the eligible-value limit

C. Remove image-quality checks

D. Automatically settle all weather-confirmed claims

Correct Answer: B. Raise the confidence requirement or lower the eligible-value limit

Question 4

What is a principal limitation of blockchain in insurance?

A. It cannot record transactions sequentially

B. It guarantees that external data entered into it is truthful

C. It does not guarantee the accuracy of input data

D. It cannot be used by authorised participants

Correct Answer: C. It does not guarantee the accuracy of input data

Question 5

Which activity is most appropriate for generative AI under controlled governance?

A. Inventing coverage terms without source material

B. Issuing every adverse claim decision without review

C. Drafting a claim summary grounded in approved records

D. Disclosing confidential prompts to external users

Correct Answer: C. Drafting a claim summary grounded in approved records

Question 6

What does calibration assess in a predictive model?

A. Whether predicted probabilities correspond to observed frequencies

B. Whether the model has the newest user interface

C. Whether all variables have identical units

D. Whether a database is physically encrypted

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

Question 7

Which is an example of basis risk in parametric insurance?

A. The index trigger occurs but the policyholder’s actual loss is small

B. A claim is paid only after manual data entry

C. A policy has an identified beneficiary

D. A ledger contains a timestamp

Correct Answer: A. The index trigger occurs but the policyholder’s actual loss is small

Question 8

Which control best addresses the risk of unauthorised access to insurer systems?

A. Multifactor authentication and least-privilege access

B. Removing audit logs

C. Sharing administrator credentials

D. Disabling software updates

Correct Answer: A. Multifactor authentication and least-privilege access

Question 9

Why can overall computer-vision accuracy be misleading?

A. It may conceal poor performance for particular groups or conditions

B. Accuracy is unrelated to evidence quality

C. Images never contain useful claims information

D. All image models have identical error rates

Correct Answer: A. It may conceal poor performance for particular groups or conditions

Question 10

Which workforce capability is increasingly important as routine insurance processing becomes automated?

A. Ignoring model outputs and data limitations

B. Model interpretation, customer explanation and professional judgment

C. Eliminating all human escalation

D. Avoiding multidisciplinary collaboration

Correct Answer: B. Model interpretation, customer explanation and professional judgment