By Axel Taylor (ERNI Spain)
The insurance industry has always been built on data, rules and human judgement. For decades, investments have been going into automation to improve efficiency across underwriting, claims management, policy administration and customer service. Yet how many projects have moved beyond isolated pilots into something that genuinely reshapes how the business operates?
Why the profound transformation has not even started
According to the 2024 digitalisation report from the European Insurance and Occupational Pension Authority (EIOPA), half of non-life insurers and roughly a quarter of life insurers across Europe were already deploying AI across their value chain. Yet the gap between “using AI” and “turning AI agents into something tangible” is where the industry’s real story begins, and closing that gap is considerably harder than building a proof of concept.
Where agents are creating value
The most fertile ground for AI agents in insurance is not customer-facing, but operational. According to scnsoft.com, one in three insurers reported at least one AI agent operating in production environments by late 2025. Three areas stand out:
Claims processing
The complexity of a mid-to-large claim involves pulling information from multiple systems, matching policy terms against submitted documentation, cross-referencing fraud indicators, and coordinating across adjusters, assessors and legal teams. An AI agent can orchestrate that workflow end-to-end, surfacing what a human reviewer needs to make a decision rather than requiring the reviewer to chase it down manually. The efficiency gains are measurable: industry data shows routine claims resolution has compressed from seven to ten days down to 24 to 48 hours in deployments that have reached production.
Underwriting
Traditional underwriting models use structured data and predefined rules. Agentic systems can work across unstructured sources, analysing survey reports, financial filings, news feeds and third-party databases simultaneously to form a more complete risk picture before a human underwriter ever opens a file. This does not replace underwriting judgement. It changes the quality of the inputs that judgement is applied to.
Fraud detection
Insurance fraud costs the European industry an estimated €13 billion annually, according to Insurance Europe. AI agents that can correlate claims behaviour across time, geography and policy history, flagging anomalies that no human team could identify at volume, represent a meaningful shift in the detection toolkit.
The issue nobody wants to lead with
Deploying an AI agent in insurance is not primarily a technology challenge. It is a governance challenge, and in Europe, it is increasingly a regulatory one. The EU AI Act, which entered into force in the summer of 2024, establishes a risk-based framework for all AI systems across the economy. Insurance is directly in scope, and not at the low end of the risk spectrum. Life and health insurance AI is explicitly classified as high-risk under Annex III of the Act, triggering comprehensive requirements around data governance, transparency, human oversight and auditability. Claims management AI used to make consequential decisions about policyholders sits in the same bracket.
EIOPA’s response has been measured but clear. In August 2025, EIOPA published an Opinion on AI Governance and Risk Management, which, while formally addressed to national supervisors, sets out detailed supervisory expectations that insurers cannot reasonably ignore. The Opinion does not add new rules. It clarifies how existing frameworks – Solvency II, the Insurance Distribution Directive and Digital Operational Resilience Act (DORA) – apply to AI systems that were not in widespread use when those frameworks were written. Its six governance pillars, covering fairness, data governance, documentation, transparency, human oversight and cybersecurity, amount to a compliance baseline that any serious AI deployment needs to meet.
The practical implication is this: an AI agent that makes autonomous claims decisions, or contributes materially to an underwriting outcome, must be explainable to regulators and to policyholders. It must have traceable decision logic. It must have defined human escalation paths. And ultimate accountability must remain with the insurance undertaking, not with the AI vendor or the model itself. That is not a constraint imposed on top of AI deployment; it is a design requirement that has to be built into it from the start.
Why most pilots stall
Legacy core systems in insurance were not built to expose the data that AI agents need at the latency they require. A claims agent that must wait 48 hours for a batch data feed to sync is not an agent; it is a workflow addition. Connecting agentic systems to live operational data requires integration work that most organisations underestimate, particularly in environments where the core platform is decades old and lightly documented.
The second stall point is accountability design. How does the human factor stay in the loop of AI-supported claims assessment? Who owns the output of an AI agent’s recommendation? How does a claims manager override it? How is that override logged, and what happens to the model when override rates exceed a threshold? These questions need answers before deployment, not after the first regulatory enquiry. EIOPA is explicit that human oversight must be ensured throughout the system lifecycle, with clear roles, escalation procedures and training in place.
The third stall is cultural. Underwriters and claims professionals who have built careers on expert judgement do not readily cede decision authority to a system they do not understand. The most effective deployments treat AI agents as decision support infrastructure that makes the expert faster and better informed, rather than positioning them as replacements. That framing is not just softer; it produces better outcomes, because human judgement remains the crucial check on edge cases that no model handles well.
What good looks like
Insurers that have successfully scaled AI agents in operations share a few practices that are worth making explicit.
They built governance infrastructure before they built the agent. Model risk management frameworks, explainability tooling, audit logging and bias monitoring were in place and tested before any agent touched a live operational decision. This investment looked expensive at the pilot stage but looked prudent at the scale stage.
They treated data as a first-class dependency, not a problem to resolve later. The data pipelines feeding the agent were scoped, cleaned and documented with the same rigour as the model itself. In several cases, this triggered broader data quality programmes that proved valuable well beyond the AI initiative.
They staffed for the intersection. The teams that succeeded were neither pure AI teams nor pure insurance operations teams. They were hybrid groups with genuine expertise in both domains, capable of translating between the technical requirements of the model and the operational realities of the process it was embedded in. This kind of cross-functional capability is rare and tends to be the binding constraint.
Conclusion
The future of insurance operations will not be fully human nor fully autonomous. It will be collaborative, intelligent and increasingly driven by AI agents operating alongside experienced professionals.
