McKinsey puts a number on what’s at stake: generative AI could add up to $1.1 trillion in annual value across the global insurance industry, according to research cited by analytics firm Tredence in June 2026. Gartner goes further, naming generative AI the top investment priority for insurers heading into 2026. The appetite is not in question. Every carrier board has approved a budget line for it.
What’s in question is whether that budget produces a working system or another pilot that quietly stalls.
The evidence so far points to the same failure mode across the industry: insurers are layering AI on top of fractured legacy infrastructure instead of solving the underlying enterprise design problem — and getting stuck in what more than one analyst has started calling pilot purgatory.
As Tredence’s analysts put it in their June 2026 review of the sector, insurers are not competing against a future threat. They are already behind, and the gap widens every quarter they delay.
That is a strong claim. On the evidence, it is also a fair one — but not for the reason most carriers assume.
The Model Was Never the Hard Part
Ask most insurance executives where their AI initiative stalled, and the honest answer is rarely that the model underperformed.
It is usually some version of the same problem: the model needed data that lived in six different systems, none of which were built to communicate with one another. By the time the data was clean enough to trust, the pilot had already missed its budget cycle.
This is the same lesson defense programs have been relearning since at least 2018. AI does not fail because the algorithm is wrong. It fails because it gets deployed onto a foundation that was never built to support it.
We have examined the same pattern in mission-critical defense operations, where the data problem — not the model — is the real barrier.
Insurance is the same story expressed through different terminology: policy administration systems instead of program-of-record silos, adjuster notes instead of maintenance logs, claims platforms instead of command systems.
Where Insurance Data Fragmentation Actually Lives
Fragmentation inside an insurance carrier is not abstract. It has specific, well-known addresses, usually inherited rather than intentionally designed.
Line-of-Business Silos
Auto, property, life and commercial insurance frequently operate on entirely separate policy administration systems.
These platforms may have been acquired at different points in the company’s history and often have no shared data model connecting them. A customer represented consistently in one system may appear completely differently in another.
M&A Residue
Carriers that have reached scale through mergers and acquisitions routinely operate parallel claims and underwriting platforms years after a transaction has closed.
Migrating a live book of business is expensive, operationally disruptive and risky enough that the work is repeatedly postponed. The result is an enterprise that may function as one company commercially while still operating as several companies technically.
Rating Engines Frozen in Time
For years, underwriting logic lived inside static risk tables built for a slower and less data-rich environment.
These systems are the opposite of the dynamic pricing infrastructure now enabled by telematics, connected devices, behavioural data and continuously updated risk signals.
Unstructured Data Nobody Pipelines
Adjuster notes, claim photographs, call transcripts and inspection reports contain some of the richest fraud and risk signals available to an insurer.
In many carriers, none of this information reaches a model. It remains stored inside document-management systems designed for archiving rather than operational use.
None of these constraints is fundamentally a model problem. They are engineering problems embedded inside the insurance industry’s legacy stack.
The Regulatory Layer Raises the Cost of Getting This Wrong
Insurance carries a constraint that makes poorly designed AI adoption more dangerous than it is in many other industries: regulators are already watching, and the rules are no longer hypothetical.
The NAIC’s model bulletin on the use of AI systems by insurers requires carriers to document how automated systems reach their decisions. The EU AI Act classifies certain AI systems used in life and health insurance risk assessment as high-risk, triggering additional conformity, governance and documentation obligations before deployment.
That regulatory reality changes what “working” means for an insurance AI system.
A claims model that cannot explain why it flagged a policyholder for fraud is not merely a technical inconvenience. It is a system that a regulator may require the carrier to suspend.
An underwriting model trained on biased historical data is not only a fairness concern. It represents a compliance exposure with formal reporting and governance requirements attached to it.
Explainability is not an optional feature in this industry. It is an operating condition.
What “Ready” Actually Requires for an Insurance AI System
A production AI system operating inside an insurance carrier must clear a significantly higher bar than a demonstration or isolated pilot.
Cross-System Data Unification
Policy, claims, underwriting and customer data must be brought into a single trustworthy operational view.
This does not necessarily require replacing every core system. It requires an integration and data layer capable of resolving inconsistent formats, identities and business definitions across the systems already running.
Explainability by Design
Every underwriting or claims decision must be traceable to the specific data, rules and model outputs that produced it.
That traceability must be designed to satisfy an NAIC or EU AI Act inquiry from the beginning. It cannot be retrofitted only after a regulatory request arrives.
Human Review for Adverse Decisions
Denials, non-renewals and fraud flags should be routed to a qualified person before they become the policyholder’s problem.
The cost of an incorrect automated decision in insurance is not merely user frustration. It can result in a regulatory complaint, financial harm and material reputational exposure.
Integration With Existing Core Systems
The system must work with the policy administration, underwriting and claims platforms that already operate the business.
A production deployment has to meet those systems where they are. Most carriers cannot pause operations or accept the disruption associated with a complete core-platform replacement.
This is the difference between adding an AI feature to an insurance stack and re-engineering how that stack converts raw policy and claims data into a decision someone can defend.
Questions Every Carrier Should Ask Before Scaling a Pilot
Before scaling an AI initiative across claims, underwriting or fraud operations, carriers should be able to answer four questions clearly:
Can the system explain, in terms a regulator would accept, exactly why it reached a specific underwriting or claims decision?
Does it retrieve data from every line-of-business system that touches the policy, or only from the platform used during the original pilot?
What happens to a denial or fraud flag before it reaches the policyholder? Does a qualified person review it first?
Are we resolving data fragmentation across the organization, or building a model on top of that fragmentation and hoping the foundation holds?
A pilot that cannot answer the first two questions is not ready to scale.
It is ready to become another example of what pilot purgatory looks like from inside an insurance carrier.
The Opportunity Is Real. So Is the Reason Most Carriers Will Not Reach It.
McKinsey’s $1.1 trillion estimate and Gartner’s ranking of generative AI as a leading investment priority both describe the scale of the opportunity in front of the insurance industry.
Neither describes what happens automatically once a budget is approved.
The carriers that capture that value will be the ones that treat data unification, integration, governance and explainability as the actual engineering project.
They will not be the carriers that purchase the most capable model and assume the rest of the enterprise will reorganize itself around it.
That principle is central to our broader approach to enterprise AI transformation: we build the infrastructure underneath the decision, rather than adding another disconnected layer on top of it.
Explore how we approach regulated, high-stakes industries, or talk to our team about an AI initiative across claims, underwriting or fraud operations.