AI can help an insurer price risk, review claim images or flag possible fraud. It cannot, by itself, make the underlying process reliable: the records still need to be accurate, systems and teams must hand work off correctly, exceptions need an owner, and decisions must meet legal and customer obligations. The gap is between a model performing a task and an insurer executing the full process around it.
Where insurers use AI—and what adoption does not prove
Insurance AI is already in use for bounded tasks. The National Association of Insurance Commissioners (NAIC) describes machine learning for pricing, including risk scoring and rate-factor relativities, and for claims tasks such as analyzing accident images, estimating ultimate settlement values and detecting fraud.
As an Amazon Associate I earn from qualifying purchases.
The European Insurance and Occupational Pensions Authority (EIOPA), reporting results from its 2023 Digitalisation Market Monitoring Survey in 2024, found that 50% of non-life insurer respondents and 24% of life insurer respondents used AI. These are survey findings from European respondents, not global adoption rates. EIOPA noted that many surveyed systems were developed in-house for simpler tasks, with more explainable algorithms and human oversight.
Those figures show that some insurers are applying AI; they do not show that end-to-end workflows are integrated, that input records are sound, or that customer cases are handled correctly. A model can support one decision point while the surrounding process remains fragmented.
Why data and workflow problems survive a model deployment
Incomplete data maps and weak lineage
A model can only work with the information it receives. If records are incomplete, inconsistent or difficult to trace to their source, the model does not automatically repair them. Data lineage—the ability to follow data from its origin through transformations and onward use—matters when an insurer needs to explain a result, find an error or establish which information informed a decision.
France’s Autorité de contrôle prudentiel et de résolution (ACPR) summarized a 2024 survey of 224 insurance organizations on data used for prudential calculations; the summary was published in 2026. It describes inventories that may be limited to actuarial data, fail to involve all relevant departments or lack clear definitions of criticality. Data-flow maps had expanded, but the summary still identified limits in their completeness, manual processes and weak traceability. It also raised concerns about managing data from external providers.
This is evidence about prudential-calculation data at surveyed French insurers, not a finding that every insurer has the same weaknesses. It does, however, illustrate why adding a model cannot establish the completeness or provenance of the records feeding it.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Manual handoffs and unclear ownership
Insurance work often moves across teams, legacy systems, vendors and customer channels. If a case leaves one system but does not arrive in another, or an exception sits in a queue without an accountable owner, automating a single step may simply move the bottleneck. The operational question is not only whether a model produces an output, but whether that output reaches the right person or system and is acted on in time.
ACPR also reported persistent first-line control weaknesses even where second-line control and internal-audit coverage had progressed. Its summary noted that senior leadership did not routinely receive even brief information about periodic data-quality indicators. A model’s technical performance cannot substitute for controls performed by the teams closest to the process, oversight, or a clear path to escalate problems.
Skills and legacy technology
EIOPA’s 2024 survey release said, “Insurers see acquiring adequate talent and skills as central to a successful digital transformation.” Its account also described differences in digital readiness among European insurers. In a 2025 discussion of scaling AI, EIOPA identified legacy IT, fragmented data sources and limited AI-specific skills as structural constraints. These are supervisory observations, not a controlled estimate of how much each factor changes performance.
Why AI does not remove governance or legal duties
AI-assisted decisions remain insurance decisions. The NAIC says its Model Bulletin on the Use of AI by Insurance Companies was adopted in December 2023 and reminds insurers that decisions or actions made or supported by AI must comply with applicable insurance laws and regulations. That is a U.S. regulatory position; legal requirements vary by jurisdiction, product and use case.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →EIOPA’s 2025 supervisory discussion identifies explainability, potential discrimination and bias, cybersecurity vulnerabilities, and concentration in the AI supply chain among the risks that can arise as insurers scale AI—particularly when decisions affect access to insurance or claim outcomes. It advocates governance proportionate to each application’s characteristics and risks, including human oversight, data-quality controls and documentation.
Human oversight is meaningful only if the reviewer can understand the case, challenge an output and act on an exception. Documentation and testing likewise need owners and a place in the operating process. AI may improve a particular step; model capability alone does not establish compliant decisions, accurate input data, clear control ownership or reliable integration.
Third-party dependency also deserves attention. EIOPA reported in 2024 that nearly 80% of its respondents used BigTech companies for cloud storage. That is context about cloud reliance, not a measure of dependence on AI vendors; the two should not be conflated.
As of March 2026, the NAIC described its AI Systems Evaluation Tool as being piloted by 12 states, with adoption anticipated at the 2026 Fall National Meeting. That is a dated status and anticipated milestone, not confirmation that adoption later occurred.
Free tools Windows power users keep installed
One-click scans. No signup required.
What customer-facing failures reveal about operations
Customer outcomes offer a practical test of whether an insurer’s processes work beyond a model demonstration. In December 2024, the Australian Securities and Investments Commission (ASIC) summarized a review of 11 Australian general insurers, representing about 86% of the general insurance market by premium. In a sampled exercise, ASIC found that one in six customer complaints had not been identified. This was not a census of all complaints across all insurers.
ASIC also reported that every insurer reviewed failed to meet one or more mandatory internal dispute resolution obligations. One in five delay notifications did not meet mandatory content requirements, with non-compliance ranging from 7% to 75% between insurers; all reviewed insurers missed required timeframes for delay notifications. These findings concern the specified Australian insurers and review, not the entire global market. They show why complaint capture, communication, escalation and timing must be assessed as operating controls, not inferred from a model’s sophistication.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to read insurance outcome metrics without overclaiming
The UK Financial Conduct Authority (FCA) publishes general insurance value measures for retail products, including claims frequency, acceptance, average payouts and complaints as a share of claims. Its 2026 publication of 2025 data reported home claims acceptance rates of 62–71%, travel at 83–86% and motor at 99%. The FCA cautioned that home claims reporting is inconsistent, so those figures need particular care; they are not a straightforward ranking of insurers.
The same FCA release reported that average home insurance claim payouts rose 17% and that combined home buildings and contents claims costs were 48% of premiums in 2025. These are market measures, not direct measures of operational efficiency or evidence that AI caused a result. The FCA says its measures should be considered alongside factors such as business mix, product age, target market, claim volumes, climate and inflation. The historical data is not intended to help consumers choose policies directly.
In short, outcome metrics can reveal questions worth investigating, but they do not explain the cause on their own. A high or low figure may reflect product and customer mix, reporting practices or external conditions as well as process performance.
How to assess the process around an insurance model
Before treating an AI deployment as an operational improvement, examine the full path from input to customer outcome. Use the same checks for a manual process, an AI-assisted one or a mixed workflow:
- Data at the point of use: Are inputs complete and current? Can their source, transformations and permitted uses be traced?
- Handoffs: Does work move reliably across legacy systems, teams, external providers and customer channels, or do people have to bridge gaps manually?
- Exceptions: Are unusual cases visible, routed to an accountable person and resolved within the required time?
- Controls and accountability: Who performs first-line checks, who provides independent oversight, what is recorded for audit, and who owns escalation?
- Customer outcomes: Are complaints captured and answered, decisions communicated on time, and claims and payouts monitored in context?
- Model and supplier risks: Are explainability, bias, cybersecurity and third-party concentration considered for this specific use?
- Long-term ownership: Does the organization have the skills and assigned responsibility to maintain the workflow after deployment?
If those conditions are weak, AI may still improve a defined task, but its output is not proof that the insurer’s broader operation is dependable. The useful comparison is not simply manual versus automated; it is whether the whole process—with its data, handoffs, controls and customer consequences—works better and remains accountable.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
