How insurers and reinsurers are applying AI across claims, underwriting and fraud – and why consequential decisions still remain under human control.
Artificial intelligence is no longer new to the insurance sector.
For more than a decade, insurers have used predictive models to assess risk, detect suspicious claims, estimate loss severity and identify customers likely to renew. Optical character recognition has been used to extract information from documents, robotic process automation to transfer data between systems, and chatbots to answer routine customer questions.
These technologies have created meaningful efficiencies, but most have remained narrow in scope. A model may produce a fraud score. A document-processing system may extract a claimant’s name and loss amount. A chatbot may retrieve a policy FAQ. Each performs a defined task, but none necessarily understands or coordinates the wider insurance process.
The next phase is more consequential.
Insurance organisations are beginning to introduce systems that can work towards a defined objective, decide which intermediate tasks are required, retrieve information from multiple sources, use approved business tools, coordinate workflow stages and escalate exceptions. In carefully controlled circumstances, some can also complete predefined actions without requiring a human to initiate every individual step.
This is the emerging transition from AI assistance to agentic execution.
However, the transition should not be overstated. Most public disclosures frequently group predictive AI, workflow automation, generative AI and agentic systems under the same broad label. Many solutions described as “agents” are still copilots, analytics tools or conventional automation with a conversational interface.
Fully autonomous insurance – where AI independently accepts complex risks, changes material pricing, denies claims or determines fraud—remains uncommon. The more credible pattern is bounded agency where AI performs and coordinates defined work, while consequential decisions remain with authorised professionals.
The future of insurance is therefore unlikely to be a choice between humans and machines. It is more likely to involve redesigning work so that specialised AI systems manage information and routine execution while underwriters, claims professionals, brokers, investigators and compliance officers retain accountable judgment.
Separating Genuine Agentic Adoption from AI Marketing
This analysis focuses on publicly disclosed initiatives from insurers, reinsurers and insurance intermediaries.
A deployment is described as agentic only where there is evidence that the system:
- coordinates several connected tasks;
- decides which approved action to perform next;
- maintains context across a workflow;
- uses tools, systems or external information;
- or possesses bounded authority to execute an action.
Systems that retrieve information, summarise documents or generate recommendations are classified as copilots or decision-support tools unless stronger evidence of workflow autonomy is available.
The distinction is necessary because the industry is adopting AI rapidly, but the level of operational authority delegated to it varies widely.
AI adoption is widespread, but agentic maturity is uneven
The underlying adoption of AI across insurance is already substantial.
The US National Association of Insurance Commissioners reported that 88% of responding auto insurers, 70% of home insurers, 58% of life insurers and 92% of health insurers either used, planned to use or planned to explore AI or machine-learning models.
The reported applications included underwriting, pricing, marketing, claims adjudication, fraud detection, prior authorisation, policy issuance and risk management. These figures cover AI broadly rather than agentic AI specifically, but they show that the data and modelling foundation for more advanced automation is already being established.
Evidence of autonomous execution is much more limited.
Aon reported in its Q3 2025 Global Insurance Market Insights that insurers were accelerating the use of AI and automation for repetitive claims activities. However, most major insurers speaking with Aon still considered agentic AI operating without human interaction too risky for significant lines of business.
The market is therefore progressing at different speeds:
- Digital-native insurers are achieving broader automation because their customer, policy and claims systems were designed to work together.
- Highly integrated Asian financial groups are deploying AI at enormous scale across service, underwriting and fraud.
- Established global insurers are pursuing narrower, controlled workflows with explicit human approval.
- Reinsurers are applying AI to document-intensive specialist work.
- Brokers and risk advisers are using AI around pricing, claims intelligence, risk information and client advice.
There is no single established model of agentic insurance adoption.
Not Every Insurance Copilot Is an AI Agent
A practical maturity model often helps distinguish genuine agency from ordinary automation.
Predictive AI estimates outcomes such as fraud probability, claim severity, policy lapse or underwriting risk, while rule-based automation executes predefined actions when known conditions are met.
Generative-AI assistants go further by retrieving information, summarising documents and supporting professional judgment, but they generally remain dependent on human prompts.
Workflow agents can coordinate several connected activities – such as identifying claim documents, retrieving the relevant policy, checking completeness and preparing the case for review – without requiring a person to initiate each step. Multi-agent systems divide this work among specialised agents for coverage, fraud, external verification, calculation, communication and audit. At the highest practical level, bounded autonomous systems may complete clearly defined low-risk actions within explicit financial, regulatory and operational limits, while escalating anything outside their authority.
The real measure of agentic maturity is therefore not the sophistication of the underlying model, but the extent to which the system can coordinate work, use approved tools, respond to exceptions and exercise delegated authority while preserving human accountability.
Comparing global insurance firms
The following comparison reflects publicly available evidence for each insurance firm.
| Firm | Position in the insurance value chain | Disclosed AI applications | Maturity indicated by public evidence | Human role |
|---|---|---|---|---|
| Allianz | Global insurer | Low-value claims orchestration | Production multi-agent workflow | Human claims professional makes the final payout decision |
| Lemonade | Digital-native insurer | Distribution, policy servicing, claims and fraud | Integrated AI-enabled automation with bounded claims and servicing authority | Claims outside the system’s authority or raising concerns are escalated to human specialists |
| Swiss Re | Reinsurer and corporate insurer | Complex claims intelligence and life-underwriting support | Production claims intelligence and underwriting decision support | Claims professionals and underwriters retain decision authority |
| Ping An | Integrated insurer and financial group | Customer service, smart underwriting, claims assessment and fraud detection | AI deployed at very large operational scale; the degree of agentic autonomy varies and is not always publicly detailed | Human involvement varies by workflow |
| AIA | Pan-Asian life and health insurer | Advisor training, contact-centre copilots and document processing | Production copilots and assistive generative AI | Employees retain responsibility for customer and insurance actions |
| Zurich | Global insurer | Underwriting-guideline retrieval and comparable historical-case analysis | Publicly disclosed proof of concept | Designed to support rather than replace underwriter decision-making |
| Aviva | Composite insurer | Application and claims-fraud detection | AI-enabled fraud analytics and investigation support | Specialist human investigation remains central |
| Generali Switzerland | Insurance carrier | Multilingual policy-selection, customer-service and claims assistance | Production conversational assistant | Public disclosure describes customer assistance, not autonomous insurance decisions |
| Aon | Broker and risk adviser | Pricing technology, claims and litigation analytics, and insurance-market research | Decision-support and analytics capabilities; end-to-end agentic execution is not publicly detailed | Professional advisers, claims specialists and insurer teams retain decision responsibility |
Allianz represents the bounded incumbent model: a genuine multi-agent workflow applied to a narrow, carefully selected claims category.
Lemonade represents the digital-native model: automation spanning distribution, servicing and claims, enabled by a more integrated operating foundation.
Ping An represents the industrial-scale model: AI deployed across enormous volumes of customer service, underwriting, claims assessment and fraud detection.
Swiss Re, AIA, Zurich, Aviva, Generali Switzerland and Aon illustrate the broader near-term pattern across insurers, reinsurers and intermediaries: AI structures information, accelerates workflows and supports expert judgment, while consequential authority generally remains with human professionals.
Claims: Where Agentic AI Is Most Mature
Claims is currently the most developed area for agentic AI in insurance. The function combines high transaction volumes, repeated document handling, external-data checks and relatively clear escalation rules. It also provides measurable outcomes through settlement time, cost per claim, fraud savings, recovery rates and customer experience.
A practical agentic AI-enabled claims workflow can be represented as follows:
(Image made from AI Tool)
AI can support almost every stage, but the appropriate degree of autonomy depends on the claim’s value, complexity, evidence quality and potential impact on the customer.
Allianz Project Nemo: Multi-Agent Claims Orchestration
Allianz’s Project Nemo is one of the clearest publicly documented examples of a genuine multi-agent claims workflow.
Launched in Australia in July 2025, it handles food-spoilage claims below AUD500 that often increase following severe weather and power outages. Seven specialised agents coordinate the process: planning, cybersecurity, policy coverage, weather verification, fraud screening, payout calculation and audit.
The system prepares the recommendation, but a human claims professional retains the final payment decision.
Allianz reports that Nemo reduced processing and settlement time by approximately 80%. The agent workflow can prepare a case for human review in under five minutes, while overall customer processing has fallen from several days to one day or even a few hours.
The significance of Nemo lies not only in its speed. It shows how an insurer can divide a defined claims process among specialised agents while keeping consequential authority with an accountable professional.
Its boundaries are equally important. The claims are low in value, relatively standardised and supported by independently verifiable weather events. This makes them particularly suitable for controlled agentic execution.
Lemonade: Automation Across an Integrated Insurer
Lemonade demonstrates a broader digital-native approach.
Its customer acquisition, policy servicing, claims and fraud systems were developed as connected parts of the same operating environment. This allows AI to work across a larger part of the customer journey rather than remaining confined to one claims task.
As of December 31, 2025, AI Jim received first notification of loss without human intervention in 96% of claims. Approximately 55% of claims were automated, while cases outside the system’s authority or displaying possible concerns were routed to human claims specialists. Eligible claims can be paid almost instantly, although more complex property-damage and liability cases continue to require human involvement.
The main lesson is not that traditional insurers can simply reproduce Lemonade’s model. Its level of automation is partly enabled by an integrated digital foundation.
Where policy, billing, claims, customer and fraud information is fragmented across different legacy systems, an agent may understand the case but still be unable to complete the workflow safely.
Swiss Re ClaimsGenAI: Finding Value Inside Complex Claims
Swiss Re addresses a different type of claims problem.
Large corporate claims can generate extensive reports, emails, invoices and technical documents. In these cases, the greatest value may not come from automatic settlement but from identifying information that claims professionals could otherwise overlook.
Swiss Re Corporate Solutions handles more than 40,000 claims annually. ClaimsGenAI was introduced to triage documents, extract relevant information and identify possible irregularities or third-party recovery opportunities, while human claims professionals retain decision authority.
During its first year after deployment in mid-2024, the system generated more than 1,000 alerts concerning possible irregularities and contributed to a fraud-savings pipeline potentially worth millions of dollars. It also identified hundreds of recovery opportunities beyond those already found by claims handlers.
This shows that claims AI can create value through more than administrative efficiency. It can help identify leakage, improve recovery, prioritise complex cases and preserve the evidence needed for professional action.
What the Claims Evidence Suggests
The global examples point to three different adoption paths.
High-volume personal-lines claims are suitable for bounded automation where the claim is low in value, evidence is complete and escalation rules are clear.
Complex commercial claims are more likely to benefit from document intelligence, coverage support, litigation analysis, fraud detection and recovery identification than from fully automated settlement.
High-consequence claims involving bodily injury, disputed liability, vulnerable customers, significant business interruption or suspected fraud will continue to require extensive professional judgment.
Aon’s benchmarking of more than 100 claims operations similarly identified contact, investigation and settlement as major areas where AI could improve quality and efficiency. Its research highlights applications across first notification, document handling, coverage assessment, fraud detection, valuation and payment.
The future claims organisation is therefore unlikely to apply one automation model to every case. Claims will increasingly be segmented according to value, complexity, evidence quality, emotional sensitivity, legal uncertainty and decision risk.
The most successful systems will not necessarily be those that automate the final decision. They will be those that remove the greatest amount of routine work while ensuring that complex and consequential cases reach the right professional with better evidence and greater speed.
Underwriting: From Document Review to Submission Orchestration
Underwriting is the next major area where agentic AI can create value. The core judgment may belong to the underwriter, but a significant amount of time is spent before that judgment begins – collecting documents, identifying missing information, extracting exposures and checking the submission against appetite and referral rules.
A practical agentic AI-enabled underwriting workflow can be represented as follows:
Commercial and life-insurance submissions may include proposal forms, loss runs, financial statements, medical records, asset schedules, inspection reports and broker correspondence. AI agents can organise this information and prepare the case without necessarily taking underwriting authority.
Swiss Re: Preparing Complex Life-Insurance Referrals
Swiss Re’s MagnumXP Underwriting Assistant supports life-insurance applications that cannot be completed through straight-through automated underwriting and must be referred to a human underwriter.
The assistant extracts and organises structured and unstructured information, highlights missing evidence and potential impairments, and brings relevant medical and financial details into a common view. It can also work alongside underwriting manuals, referral rules and automated assessment engines.
Swiss Re has deployed the capability in North America and conducted proof-of-concept work with insurers in Europe and Asia. The company reports significant reductions in review effort, although it has not published one standard performance figure across all deployments.
The value lies in reducing the detective work required to reconstruct an applicant’s history. The underwriter still makes the risk decision, but receives a more complete and structured case.
Zurich: Combining Guidelines with Historical Cases
Zurich has explored a different model through a reinsurance-underwriting proof of concept.
The system combines retrieval of relevant underwriting guidance with machine-learning methods that identify comparable historical cases. The objective is to reduce manual document review, improve consistency and help underwriters focus on more complex analysis.
Zurich has been careful to describe the initiative as a proof of concept requiring further validation before production deployment in a regulated environment.
Historical cases can be useful, but they must be treated cautiously. A previous underwriting decision may reflect a different risk appetite, policy form, regulation or market cycle. Similarity should inform professional judgment rather than determine it automatically.
Ping An: Automation at Scale
Ping An reported that 93% of its life-insurance policies were underwritten within seconds during 2024.
This demonstrates the scale possible when product rules, data, workflows and customer channels are highly integrated. However, the figure should not be interpreted as evidence that a generative agent independently made every underwriting decision.
Ping An’s wider environment combines predictive models, automated decision engines, workflow systems and large AI models. The exact degree of agentic autonomy is not publicly detailed for each process.
What the Underwriting Evidence Suggests
The strongest near-term opportunity is not a universal autonomous underwriter.
It is a submission-orchestration agent that can:
- classify incoming documents;
- extract material exposures;
- identify missing information;
- check appetite and referral rules;
- retrieve relevant guidance;
- prepare clarification questions;
- and generate an evidence-linked underwriting memo.
Pricing, material exclusions, capacity allocation, risk acceptance and binding should remain subject to formal underwriting authority.
The practical benefit is not merely faster document processing. Better-prepared submissions can reduce clarification cycles, improve consistency and allow underwriters to spend more time on portfolio judgment, negotiation and complex risks.
The underwriting function is therefore likely to evolve through greater automation around the decision, rather than immediate automation of the decision itself.
Fraud: From Risk Scores to Investigative Intelligence
Fraud detection is one of insurance’s oldest applications of AI, but the role of the technology is expanding. Traditional systems primarily assign anomaly or fraud-risk scores. More advanced systems can now connect identities, compare claim narratives, examine documents and images, map relationships across cases and prepare structured evidence for investigators.
A practical agentic AI-supported fraud-investigation workflow can be represented as follows:
The most credible role for agentic AI is not to declare that fraud has occurred, but to assemble, prioritise and organise the evidence required for a professional investigation.
Aviva: Responding to AI-Generated Evidence
Aviva reported that it identified more than 18,400 suspect claims worth £233 million during 2025 and stopped more than 105,000 fraudulent insurance applications.
The insurer has also warned that fraudsters are increasingly using generative AI to create or alter supporting evidence, including accident images and documents. Aviva says it is responding through advanced analytics, AI-enabled tools, behavioural insight and specialist investigation teams.
These figures should not be attributed to AI alone, but the case illustrates how generative AI is changing both sides of insurance fraud. Insurers increasingly need to assess not only what a document or image appears to show, but also whether it is authentic, internally consistent and connected to the wider body of evidence.
Ping An, Swiss Re and Lemonade: Different Forms of Fraud Intelligence
Ping An reported RMB11.94 billion in claims savings from smart fraud detection during 2024.
Swiss Re’s ClaimsGenAI identifies possible irregularities inside complex corporate-claim files and routes them to professionals for further investigation.
Lemonade’s Forensic Graph Network uses behavioural and connected-data signals to identify suspicious patterns across customer and insurance activity.
Together, these examples show that modern fraud intelligence can combine:
- transaction and behavioural scoring;
- network and relationship analysis;
- document and image checks;
- narrative comparison; and
- contextual review of previous claims and applications.
What the Fraud Evidence Suggests
The strongest fraud agent is an investigation-support system, not an autonomous accuser.
It can:
- prioritise suspicious cases;
- connect related people, addresses, devices and claims;
- compare documents, images and narratives;
- identify inconsistencies and unusual patterns; and
- prepare an investigator-ready evidence file.
Human investigators should remain responsible for interpreting intent, testing the evidence and deciding what action is appropriate.
That boundary is essential because false positives can delay legitimate claims, create hardship for customers and expose insurers to legal, regulatory and reputational risk.
The practical objective is therefore not to maximise the number of fraud alerts. It is to improve the quality, speed and precision of investigations while reducing unnecessary friction for genuine policyholders.
Taken together, claims, underwriting and fraud reveal both the promise and the present limits of agentic AI in insurance. The strongest systems are already coordinating information, managing workflow stages, identifying exceptions and preparing decisions. However, where the outcome may materially affect a customer, pricing decision, claim settlement or fraud allegation, consequential authority generally remains with accountable human professionals.
The emerging model is therefore not one of unrestricted autonomy. It is one of bounded execution: specialised agents operating within defined workflows, evidence requirements, authority limits and escalation rules. This is where agentic AI is currently most credible—and where insurers are beginning to generate measurable operational value without removing professional accountability.
In Part 2, we examine how this transition is unfolding across policy servicing, customer support, distribution, broking and back-office operations – and the data, governance and economic foundations required to move agentic systems from experimentation into production.
About Poniak Labs
Poniak Labs works with enterprises to identify, design and implement governed AI workflows across document-intensive and operational processes. In insurance, this may include claims intelligence, underwriting submission orchestration, fraud-investigation support, knowledge retrieval, workflow integration and human-in-the-loop decision systems.
Our focus is not on automating consequential decisions without oversight, but on building secure, evidence-linked systems that reduce routine work, improve operational visibility and help professionals act with greater speed and consistency.