AI in Insurance: Why the Technology Is Ready but the Industry Is Not

AI in insurance: where the industry stands today

Insurance has the kind of work AI does well: document-heavy intake, probabilistic decisions, repetitive servicing tasks, and large claims and call-center operations. McKinsey estimates that generative AI could unlock $50 billion to $70 billion in insurance industry revenue, with the biggest gains in underwriting, claims, distribution, and customer service.[1] ↗ The technology fits the work. The economic case is clear. And yet, when you look at how insurers are actually using AI today, the picture is less impressive than the headlines suggest.

An April 2026 AM Best survey of more than 150 carriers and managing general agents captured the gap clearly. Nearly 60% of insurers expect AI to significantly transform their business model within one to three years, but only 41% said AI is already active across their core business areas, and just one in five said their organization has reached an advanced stage of implementation.[2] ↗ That is the central contradiction in insurance AI today. The models are better, the use cases are clearer, and the economics are increasingly visible. Yet most carriers are still moving from proofs of concept to small, tightly bounded deployments rather than redesigning their decision systems end to end. The question is no longer whether AI can create value in insurance. It is whether insurers can modernize fast enough, govern well enough, and change deeply enough to capture it.[3] ↗

Where AI is already being used in insurance

AI is not new to insurance. It is already running across underwriting, pricing, claims, fraud detection, and customer service. The National Association of Insurance Commissioners (NAIC) notes that insurers use AI for routine chatbots, image-based damage estimation, fraud detection, and pricing,[4] ↗ while McKinsey describes AI as active in sales productivity, underwriting accuracy, claims management, voice-based service, and back-office work.[5] ↗ In other words, the surface-level adoption story looks healthy.

Look closer, though, the adoption is uneven. A 2024 Deloitte survey of 200 U.S. insurance executives, published in its 2025 outlook, found that 76% had implemented generative AI in at least one business function, with life and annuity carriers ahead of P&C at 82% versus 70%. Larger organizations were further along than smaller ones, but most projects were still in the scoping stage rather than full deployment.[6] ↗ This is where it pays to separate experimentation from real industrialization. Boston Consulting Group found that insurance has outpaced many other sectors in early AI adoption, yet only 7% of insurers have brought AI to scale — about two-thirds remain stuck in piloting.[7] ↗ The IAIS reaches the same conclusion from the supervisor's seat: gen AI is growing, but in most countries it is still limited to a small number of insurers and a defined set of activities.[8] ↗

When AI does scale, the upside is real. McKinsey reports that domain-level AI rewiring has produced 10–20% improvement in new-agent success, 10–15% premium growth, 20–40% lower customer onboarding costs, and 3–5% better claims accuracy. Deloitte adds that AI-powered fraud analytics could save P&C insurers $80 billion to $160 billion by 2032. These are meaningful results, but they share an important quality: they come from redesigning workflows and operating models, not from bolting a chatbot onto a legacy process. That distinction explains why headline adoption rates and headline business impact are not yet matching up.[9] ↗

Why AI adoption in insurance is slow: four structural barriers

If AI works and the rewards are clear, why is scaling so hard? The answer is not the model. It is the environment around the model. Four barriers come up again and again across the major industry surveys, and they tend to compound rather than offset each other.

1. Legacy systems and fragmented data

AM Best names data readiness, security and privacy, and legacy integration as the three biggest barriers to AI deployment, noting that legacy systems store data in inconsistent, poorly standardized formats — which makes AI integration much harder than executives expected.[10] ↗ KPMG's 2025 U.S. Insurance CEO Outlook makes the same point: older operating models still hold data in silos, and that fragmentation is slowing AI down.[11] ↗ Deloitte's research likewise lists weak data foundations and aging IT among the top reasons AI implementations stall.[12] ↗ The problem is bigger than "clean up your data." Modern AI gets much of its power from unstructured data — documents, images, adjuster notes, submissions — and in insurance, those live everywhere: old policy systems, claims platforms, email threads, third-party tools. Bringing them together is real engineering work, not a one-time cleanup. Many insurers discover that the hard part is not building the AI model itself. The harder challenge is connecting policy systems, claims platforms, emails, PDFs, and third-party tools that were never designed to work together.

2. ROI is hard to measure with old yardsticks

Many carriers still measure AI like a typical automation project, which tends to understate long-term value and overstate short-term disappointment. PwC notes that most insurance gen-AI efforts remain experimental, and ROI has been slow to show up because pilots are too small to deliver enterprise-wide impact.[13] ↗ AM Best is blunter: respondents said the cost benefits will likely take years to materialize.[14] ↗ In other words, insurers are trying to justify foundational transformation using tools built for incremental process improvement. The math rarely works under that framing, and projects that could pay off over three to five years end up cancelled at the twelve-month review.

3. A culture built to defend decisions, not to experiment

Insurance is not a cautious industry by accident. Carriers are built to control uncertainty, document reasoning, and defend decisions after the fact. That culture has value, but it clashes with AI systems, which are probabilistic, adaptive, and sometimes opaque. BCG says many insurance AI programs stall because of organizational resistance, weak business engagement, and unclear roles, combined with the probabilistic nature of AI itself.[15] ↗ The tension is sharpest in underwriting and claims, where carriers must combine speed with defensible judgment.

The talent gap makes the cultural problem worse. KPMG's 2025 CEO Outlook found that 72% of insurance CEOs see AI workforce readiness and upskilling as a top constraint on growth, even as 73% list AI as a top investment priority. Confidence in near-term AI returns jumped from 21% in 2024 to 67% in 2025, which shows ambition is rising faster than capability.[16] ↗ Accenture's 2025 underwriting survey adds field-level color: 81% of underwriting executives believe AI will create new roles, 60% see it augmenting the workforce, and 65% see urgent need for upskilling.[17] ↗ Insurers want AI-led productivity, but most do not yet have the workforce model to deliver it. There is also a quieter failure mode: the gap between business teams and tech teams. Deloitte found that lack of business-line support was the single biggest reason AI projects fail, while close cross-functional collaboration was the single biggest reason they succeed.[18] ↗ Production AI in insurance is not a standalone tech asset — it is a decision system embedded in underwriting, claim routing, service scripts, pricing, audits, and customer communications. If the business owners are not co-designing it, scale rarely follows.

4. Regulation, compliance, and privacy

Insurance is a regulated industry before it is a digital one, and AI inherits all of that complexity. The NAIC's 2023 Model Bulletin requires that any AI-supported decision affecting consumers comply with existing insurance law — including unfair trade practice and unfair discrimination rules — and expects carriers to maintain a written AI program with governance, risk controls, internal audit, and accountability up to senior management and the board.[19] ↗ By April 2026, the NAIC implementation map listed 25 states that had adopted the bulletin, with separate insurance-specific guidance in California, Colorado, New York, and Texas. Explainability is one of the hardest constraints. The bulletin explicitly calls out transparency, explainability, human involvement in final decisions, and oversight of third-party dependencies,[20] ↗ and the IAIS expects model risk and explainability to see the largest jump among AI-related risks over the next two years.[21] ↗ "Black box" performance will not pass review, especially in decisions that affect access, price, or claims.

Europe is going further. The EU AI Act takes a risk-based approach, and EIOPA's 2024 factsheet says it classifies AI used for risk assessment and pricing in life and health insurance as high-risk.[22] ↗ EIOPA's 2025 opinion clarified that insurance-sector laws still apply on top of the AI Act, with data governance, fairness, cybersecurity, explainability, and human oversight as core supervisory concerns. Privacy rules also vary by line of business: HHS notes that HIPAA covers health insurers but not life insurers.[23] ↗ That means a single insurance group cannot run AI under one uniform privacy playbook — health AI sits under stricter health-information rules, while life and P&C lines face different state-level privacy and unfair-practice obligations. For carriers operating across multiple lines, AI controls have to be line-aware from day one, not retrofitted after a compliance review.

Why each major insurance AI use case is hard

The four structural barriers above show up differently in each function. Some functions look like easy wins on paper but expose the deepest problems in practice. Here is how the picture looks across the four most active areas of insurance AI.

Underwriting

Underwriting looks like an ideal AI target because it depends on extracting signals from submissions, loss histories, inspections, and external data. In practice, it exposes nearly every adoption problem at once. Accenture's underwriting survey found ineffective systems were the top challenge for underwriters at 65%, followed by lack of information and analytics at 42% and poor information access at 40%. Insurers expect underwriting AI adoption to climb from 14% today to 70% within three years, which shows ambition rather than readiness.[24] ↗ In life and health insurance, regulatory scrutiny is even tighter because underwriting and pricing decisions directly affect consumer access, pricing fairness, and discrimination risk.

Claims processing

Claims is one of the most active AI domains, and also one of the hardest to fully automate. Claims files mix unstructured documents, photos, adjuster notes, policy language, repair estimates, and edge cases that do not fit neat categories. McKinsey says modern AI can improve claims by analyzing all of those inputs together, while the IAIS lists claims handling as one of the leading current gen-AI use cases.[25] ↗ Yet the IAIS also notes that the industry is still cautious about autonomous AI in critical consumer touchpoints. That caution is well placed: a wrong claims decision is not just an operational miss, but legal exposure, customer harm, and reputational damage in a single event.

Customer experience and AI agents

External chatbots are among the leading gen-AI use cases identified by supervisors, and customers increasingly expect human-like, on-demand interactions. But customer-facing insurance AI has a trust problem. The NAIC warns that large language models can produce output that sounds correct but is wrong and should be reviewed carefully, especially in important decisions.[26] ↗ Deloitte adds that hallucinations can create multi-state compliance errors and that black-box outputs can deepen the trust deficit with customers, regulators, and employees.[27] ↗ Virtual agents work for service, triage, and information retrieval, but carriers still need strong containment, review, and escalation rules whenever the conversation drifts toward advice or adjudication.

Pricing and risk modeling

Pricing is where AI's analytical power runs straight into regulatory sensitivity. The NAIC model bulletin states clearly that rates developed using AI or predictive models still must not be excessive, inadequate, or unfairly discriminatory.[28] ↗ EIOPA classifies life and health risk assessment and pricing as high-risk under the AI Act, and the IAIS reports that supervisors are watching algorithmic bias, unlawful discrimination, data privacy, and model risk closely. Insurers can use AI to find finer-grained risk signals, but they cannot escape the burden of showing that the decision logic is fair, lawful, and explainable enough for both regulators and consumers.

The gap between AI capability and industry readiness

AM Best's core insight is the right frame: the technology may be ready, but most insurers are not. Two-thirds of respondents in its 2026 survey want to increase AI investment in the next 12 to 24 months, and most companies already have a formal AI policy. Yet data readiness, security, privacy, and legacy integration remain the biggest barriers, and productivity gains so far are mostly modest rather than transformative.[29] ↗ That is the gap in one sentence: high strategic conviction, modest operational readiness.

Closing it requires a shift in mindset. Insurers have to stop treating AI as a portfolio of separate experiments and start treating it as an operating-model redesign. McKinsey's prescription is domain-level transformation built on reusable components and a modern data layer; Deloitte's is stronger discipline around resources, responsibility, and returns, anchored in data foundations, talent, and risk management. KPMG points to enterprise-grade data governance, scalable infrastructure, and board-level ownership. PwC adds a useful angle: responsible AI is not just a compliance overlay but a scaling enabler, with 58% of executives saying it improves ROI and efficiency and 55% saying it improves customer experience and innovation.[30] ↗ The pattern across all of these sources is the same — modernization, human-AI collaboration, and governance are not side topics, they are the adoption strategy.

Put plainly, the next step for insurance is not more pilots. It is the move from AI as a point solution to AI as a decision system — embedded in workflows, monitored continuously, paired with humans where judgment matters, and supported by auditable governance and modern data architecture. That is the difference between a claims assistant that drafts text and a claims operation that routes, prioritizes, documents, and escalates consistently at scale. It is also the difference between a pricing model that scores risk and a pricing system that can survive a regulatory examination.

Conclusion: AI in insurance is inevitable, but not instant

AI will reshape insurance, but it will not happen overnight, and not because the models are weak. Insurance is a complicated mix of legacy technology, fragmented data, regulated decisions, distributed accountability, and risk-averse culture. The carriers most likely to lead in the near term will be the ones that modernize their data and core systems, redesign work around human-AI collaboration, and build explainable governance into production workflows — instead of layering AI on top of broken processes.[31] ↗ The winners will not just automate tasks. They will move carefully from traditional process chains to AI-assisted decision systems that are fast enough to matter and controlled enough to trust.

One caveat on the data: cross-market AI adoption is hard to measure consistently. The IAIS notes that many supervisors still lack detailed data on insurer AI use cases, and its 2025 market report finds gen AI adoption still limited in most jurisdictions even as broader use is clearly rising.[32] ↗ That uncertainty does not weaken the conclusion — it sharpens it. The obstacle is not a shortage of AI capability. It is a mismatch between capability, enterprise readiness, and the regulatory conditions under which insurers actually operate.