Stratenity POV, Insurance Industry Scan

Pricing a future that no longer resembles the past

Insurance is the business of pricing risk you cannot yet see, and every foundational assumption in that business is destabilizing at once. Property carriers priced decades of policies on catastrophe models calibrated to a stationary climate, and those models are now systematically under-forecasting. US insured catastrophe losses have exceeded 100 billion dollars in multiple recent years, and secondary perils, severe convective storms, wildfire, and flood, now drive a growing share that traditional hurricane and earthquake models never captured well.

On the casualty side the problem is social inflation: jury awards and litigation financing are pushing loss severity up faster than economic inflation, with some commercial liability lines seeing severity trends of 7 to 10 percent annually. The result is a structural squeeze. A carrier that misprices by a few points does not lose margin, it loses its combined ratio, and in a soft market it can write itself into insolvency while growing premium.

The combined ratio is the whole game

Underwriting profitability lives and dies on the combined ratio, the sum of the loss ratio and the expense ratio. Below 100 percent the carrier makes an underwriting profit, above 100 percent it is paying out more than it takes in and relying on investment income to survive. For the past decade a low-yield environment meant investment income could not rescue a bad underwriting year, which is why discipline on the numerator matters more than ever.

The worked example: on a book with a 72 percent loss ratio and 30 percent expense ratio, the combined ratio is 102 percent, an underwriting loss. Cutting claims leakage and straight-through processing to shave 4 points off the expense ratio, plus 2 points of loss ratio from better risk selection, turns a losing book into a 96 percent, profitable one. That 6-point swing is the difference between a book you grow and a book you exit.

Actuaries, adjusters, and the data-science bridge

The workforce challenge is the collision of a graying, deeply specialized actuarial and claims workforce with the new demand for data science and machine-learning talent, and the two groups often do not speak the same language.

• Actuaries own the regulatory-facing pricing and reserving models and must ultimately sign off on any AI-augmented rate, so bridging actuarial rigor with ML flexibility is the critical integration.
• Experienced claims adjusters carry judgment that is hard to codify, and a wave of retirements risks a knowledge cliff in complex-claims handling just as severity is rising.
• Underwriters are shifting from manual risk assessment to model supervision, curating and overriding algorithmic recommendations rather than pricing each risk by hand.
• Data scientists who understand insurance regulation, particularly what a model can and cannot legally use as a rating variable, are scarce and decisive.

Rich data, brittle systems, and the legacy core

Insurers sit on some of the richest longitudinal data in any industry, decades of policy, claims, and loss history, yet much of it is trapped in mainframe policy administration systems written in COBOL that resist integration. The paradox is that a carrier can have a century of claims data and still be unable to feed it cleanly into a modern pricing model.

AI's highest-value applications here are concrete: computer vision for property and auto damage triage that cuts claims cycle time, natural language processing to extract exposure detail from unstructured submissions, and gradient-boosted models that materially sharpen risk segmentation over traditional generalized linear models. But every one of these creates a governance obligation. A model that improves the loss ratio by using a variable the regulator later deems a proxy for a protected class is not an asset, it is a liability with a compliance action attached. Provenance on every model input is the price of admission.

NAIC, state DOIs, and the model-governance frontier

Insurance is regulated primarily at the state level in the US through the individual departments of insurance, coordinated by the NAIC, and by Solvency II for European operations. The NAIC has issued a model bulletin on the use of AI by insurers that puts explicit governance, testing, and documentation obligations on algorithmic decision-making, and states are adopting it. Colorado's SB21-169 already requires insurers to test their models and external data for unfairly discriminatory outcomes.

• NAIC AI model bulletin: mandates a governance framework, ongoing testing for bias, and documentation of AI and third-party model use.
• State rate filing: rates must be filed and, in many states, pre-approved, and must not be excessive, inadequate, or unfairly discriminatory.
• Solvency II: risk-based capital requirements, the ORSA, and the three-pillar framework for European operations.
• Fair-lending and anti-proxy-discrimination rules: increasing scrutiny of any variable that correlates with protected characteristics.

The through-line is that a rate you cannot explain is a rate you cannot legally use. Every pricing decision must be reconstructable: which model version, which variables, which assumptions, and a demonstrable test that it does not produce a disparate outcome.

The claim is the moment of truth

Every other interaction in insurance is a promise, the claim is the delivery, and it is where the customer relationship is won or lost. Yet the industry's incentives can pull against the customer at exactly that moment, which is why claims handling is both a service imperative and a regulatory tripwire around bad-faith and unfair-claims-practices statutes.

The opportunity is real: AI-assisted triage can settle a simple auto claim in hours rather than weeks and route complex claims to senior adjusters faster, improving both satisfaction and loss-adjustment expense. The risk is equally real: an automated denial that a customer cannot get explained, or that turns out to be systematically biased, converts a claims decision into a market-conduct examination. Speed without transparency is a liability, not a service improvement.

From standalone carrier to embedded risk platform

The distribution and data landscape is fragmenting, and carriers increasingly compete less on the policy and more on how they plug into a broader risk ecosystem.

• Embedded insurance places coverage at the point of sale inside other platforms, shifting distribution away from agents toward API partnerships.
• InsurTech partnerships provide underwriting data, telematics, and IoT sensor feeds that carriers rarely build in-house.
• Reinsurers are strategic capital and modeling partners, and their appetite increasingly dictates what primary carriers can write in catastrophe-exposed markets.
• Third-party data and catastrophe-model vendors are core dependencies, which makes vendor model governance a first-order compliance concern, not a procurement footnote.

The pricing model as a governed, auditable artifact

In most sectors explainability is a feature. In insurance it is the product, because the regulator, the reinsurer, and the courts all demand to see the reasoning behind a rate. Stratenity's position is that the underwriting and pricing model should be treated as a typed, versioned decision artifact, where every rate carries its full provenance: model version, input variables, assumptions, bias-test results, and the human actuary who approved it.

This reframes AI adoption entirely. The goal is not the most predictive model, it is the most defensible one, a model whose every output can be reconstructed and justified in a rate filing or a market-conduct exam. AI agents can propose pricing and reserving updates continuously, but the consequential output passes an actuarial approval checkpoint and never reaches a customer without its reasoning attached.

Where to focus the next twelve months

• Stand up a formal model-governance function aligned to the NAIC AI bulletin, with mandatory bias testing and documentation before any model touches a rate.
• Re-underwrite catastrophe-exposed books against updated, secondary-peril-aware models and exit or reprice the tail exposures the old models mispriced.
• Attack the expense ratio with straight-through processing on simple policies and claims, targeting a measurable point reduction rather than incremental automation.
• Load casualty reserves explicitly for social inflation and instrument reserving with early-warning signals on adverse development.
• Build a vendor-model governance program so third-party catastrophe and data models are auditable to the same standard as internal models.

Sector-specific levers with a metric on each

• Sharpen risk selection with governed gradient-boosted models, measured as loss-ratio improvement on the newly segmented book versus the GLM baseline.
• Automate straight-through processing on low-complexity policies and claims, measured as expense-ratio points removed.
• Deploy computer-vision claims triage, measured as reduction in claims cycle time and loss-adjustment expense per claim.
• Reprice catastrophe-exposed portfolios on secondary-peril-aware models, measured as reduction in modeled tail-loss exposure per premium dollar.
• Instrument reserving for social inflation, measured as reduction in adverse reserve development on long-tail casualty lines.

Related reading

Put this sector view to work with the cross-cutting Stratenity frameworks.