Agentic AI in Insurance: Achievements, Challenges, and the Road to Scale

Insurers and technology providers are already generating tangible benefits from AI.

QBE has accelerated claims handling through the introduction of an AI-powered intelligent document processing solution for emails and attachments submitted during claims notification. Managing General Agent (MGA) Dale has implemented data capture and processing in underwriting, using a third-party AI solution to analyse and summarize attachments and emails.

Celent's research[3] highlights how insurers themselves are already putting agentic capabilities into production:

• Allstate uses a specialized GenAI model to analyse and authenticate claims documents – such as invoices, police reports, and receipts – for claims fraud detection.
• QBE’s Cyber Underwriting AI Assistant reads and summarizes broker submissions, checks completeness, identifies key risk factors, and determines whether the risk falls within QBE’s appetite.
• State Farm is scaling GenAI across its operations, including digital knowledge assistants for employees, AI-powered customer service chatbots, and expedited contract analysis using natural language processing and computer vision.
• Allianz[4] offers one of the clearest examples of this shift toward multi-agent orchestration. In July 2025, the insurer launched Project Nemo in Australia – an agentic AI solution built to automate food spoilage claims following power outages caused by natural catastrophes. Rather than a single model handling the task, Nemo runs seven specialized agents in sequence: a planner agent that initiates the workflow, a cyber agent overseeing data security, a coverage agent verifying the policy, a weather agent confirming the triggering event, a fraud agent screening for irregularities, a payout agent calculating the amount, and an audit agent summarizing the process for human review. A claim that previously took up to four days to process – because claims teams were focused on more complex cases during a catastrophe event – is now resolved in under five minutes of AI processing time, with Allianz reporting an 80% reduction in overall claim processing and settlement time. Crucially, the payout decision itself is never automated: a human claims professional always makes the final call, a human-in-the-loop principle Allianz describes as central to trust and governance across all its AI applications. Built and deployed in under 100 days, Nemo is now being explored for other high-frequency, low-complexity use cases such as travel delays and simple auto claims.

These examples highlight the industry's evolution from AI assistants to AI agents capable of executing business processes.

Despite strong momentum, many insurers struggle to move beyond isolated use cases. The most significant challenges fall into six categories: data, governance, skills, regulation, cost, and organizational readiness.

1. Data and connectivity

Agentic AI depends on high-quality data, yet insurers continue to work with fragmented and unstructured information found in claims notes, broker emails, call transcripts, and numerous other documents. Data silos within legacy systems remain widespread. While AI can extract information from emails, submissions are often incomplete, requiring underwriters and claims handlers to repeatedly request additional details from brokers. The real opportunity lies in improving ecosystem connectivity through stronger integration between insurer and broker systems. Industry middleware providers such as ACORD and WebConnectivity aim to improve interoperability across the market.

Solution

• Establish robust data quality processes
• Modernize data management capabilities
• Invest in APIs and integration platforms
• Reduce system fragmentation
• Enable real-time data exchange across the insurance ecosystem

Without reliable data, AI agents cannot deliver reliable outcomes.

2. Governance and accountability

As AI agents become more autonomous, governance becomes increasingly important. A fundamental question remains: who is accountable when an AI agent makes an incorrect decision? The governance gap is considerable. Although more than 80% of UK MGAs reportedly use AI, fewer than half have formal governance frameworks in place. Gartner predicts[5] that by 2027, 40% of enterprises will scale back autonomous AI agents due to governance failures discovered after production incidents.

Solution

Implement a managed autonomy model that includes:

• Clear accountability structures
• Human oversight for high-risk decisions
• Defined escalation procedures
• Controls aligned with each agent's level of autonomy
• Continuous performance monitoring

Governance must be established before scaling, not after something goes wrong.

3. Skills shortages

Very few professionals combine deep insurance expertise with practical experience managing AI systems. Demand for AI-related skills continues to rise, particularly in the UK and the London Market, where organizations compete for talent capable of operating AI in production environments.

Solution

• Improve AI literacy across the workforce
• Upskill underwriters and claims specialists
• Develop AI governance expertise
• Create cross-functional teams that combine business and technical knowledge

Organizations that invest in people will be better positioned to scale AI successfully.

4. Regulation and compliance

Insurance operates in a highly regulated environment. Requirements such as FCA Consumer Duty, UK GDPR, and PRA model risk expectations make explainability and auditability essential. Compliance has become one of the fastest-growing priorities in insurance technology recruitment, reflecting the need to manage AI responsibly.

Solution

Build compliance into AI initiatives from day one through:

• Explainable AI practices
• Audit trails and documentation
• Transparent decision-making
• Regular model risk reviews
• Regulatory controls embedded within workflows

5. Cost management

Every AI interaction incurs a cost. As adoption grows, managing these costs becomes a business priority. The true cost of agentic AI extends far beyond model usage and token consumption. Monitoring, governance, infrastructure, security, and human oversight all contribute to the total cost of ownership.

Solution

Measure:

• Cost per claim handled
• Cost per underwriting submission
• Productivity gains
• Overall business impact

Effective monitoring, observability, and guardrails are just as important as controlling model costs.

6. Organizational readiness

The biggest misconception about agentic AI is that technology alone drives transformation. In reality, implementation challenges are often linked to integration, change management, operational processes, and workforce adoption.

Solution

Successful insurers:

• Adapt operating models to AI capabilities
• Promote AI literacy
• Encourage controlled experimentation
• Focus on measurable business outcomes
• Regularly evaluate costs versus benefits
• Recognize that not every problem requires AI

In many cases, simpler automation solutions deliver greater value than autonomous agents.

Agentic AI is already delivering measurable value in underwriting, claims handling, employee services, and software development. Examples from Aviva, Intact, Allstate, QBE, State Farm, and Allianz demonstrate that AI is becoming embedded in the core business of insurance. Yet the largest barriers to success are not technological. Data quality, connectivity, governance, skills, compliance, cost control, and organizational readiness will determine whether insurers can unlock the full potential of agentic AI. The winners will not necessarily be the organizations with the most AI agents, but those that can effectively integrate, govern, and operationalize them at scale.

[1] K. Raymond, C. Beattie, A. Grealish, AI Governance: Trends, Drivers, and Best Practices. Insurance Edition, Celent, April 2026.

[2] K. Raymond, Getting to “Yes” with AI: Overcoming the Barriers to Adoption in Insurance Organizations, Celent, January 2026.

[3] Ibidem.

[4] Allianz launched its first agentic AI to automate claims, Allianz.

[5] Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner.