EagleView launched its Horizon engine on April 21, 2026, CAPE Analytics has folded EagleView's imagery archive into its property signals, and ZestyAI has crossed 200 state rate-filing approvals. Three views of the same shift.

Aerial imagery analysis has moved from pilot technology to rate filing infrastructure. A carrier choosing among them is not picking an analytics vendor. It is picking the data architecture its homeowners filings will compound on.

Key Takeaways

  • EagleView's archive holds more than 3.5 billion high-resolution property images from over 100 aircraft across more than 25 years, covering 96 percent of the U.S. population.
  • CAPE derives more than 80 property-level signals per property from aerial imagery, satellite data, weather records, and public records, delivered through an API built for high-volume underwriting.
  • ZestyAI's Roof Age cross-validates more than 20 years of imagery against building permit records with confidence scores across 97 percent of U.S. properties.
  • Z-STORM is accepted in 32 states and total ZestyAI approvals exceed 200, which is the asset a new entrant cannot buy.
  • Imagery reaches a filing through three channels, and only the third, catastrophe model input, puts the reliance burden on the actuary signing the indication.

Three Assets, None of Them the Software

Each platform guards something a competitor cannot rebuild on a funding round, and the assets are different in kind.

EagleView's is time. The archive holds more than 3.5 billion high-resolution images captured by a fleet of over 100 aircraft across more than 25 years of continuous flight, covering 96 percent of the U.S. population, with more than 300 patents across the geospatial portfolio.

Horizon does not add imagery. It replaces a structured API with a natural language interface over more than 20 integrated tools, and adds agent-to-agent access through Model Context Protocol so an external underwriting system can query the geospatial analysis directly. Change detection, flagging properties that have physically changed between captures, is the feature that matters for renewal underwriting, because it moves the assessment from point-in-time to continuous. It opened invitation-only on June 1, 2026, aimed at a $1.05 trillion P&C market, with homeowners losses up more than 30 percent in the first half of 2025.

CAPE's is calibration. Its models run against aerial imagery, satellite data, weather records, and public records to produce more than 80 property-level signals: roof condition and material, vegetation proximity and yard debris, outbuildings, pool and trampoline indicators, environmental exposures. Because they arrive through an API built for volume, a carrier running a straight-through workflow applies all 80-plus variables to every submission rather than to the ones that clear a manual inspection threshold. The July 2024 EagleView collaboration extended change detection further back and brought top-50 metro annual refresh into the pipeline.

ZestyAI's is the permit layer and the approval record. Roof Age identifies replacement events from observable changes in rooftop reflectance and material, cross-validated against county building permit records across more than 20 years of imagery, with confidence scores on 97 percent of U.S. properties. Permit records are county-level administrative data, and the ingestion and normalization work behind them is not something more aircraft can substitute for. Z-FIRE was the first AI-based wildfire model approved in a California rate filing and continues to clear the CDI's PRID process; Z-STORM is accepted in 32 states.

Where the Data Enters, and Who Owns the Reliance

The regulatory burden is not uniform. It depends entirely on which channel the imagery takes into the filing.

As a rating variable, the carrier must show the reviewing actuary that the variable is credible as a loss predictor, derived from a defined and reproducible methodology, and free of proxy discrimination. The vendor's filing documentation handles credibility and methodology. The proxy-discrimination test is carrier-level work in the specific territory, and no vendor discharges it on the carrier's behalf.

As an underwriting tier variable, the score has to be filed as a classification variable with a disclosed derivation. A condition signal that triggers extra inspection below a threshold is a representation to the regulator that the trigger is objective and consistently applied, described precisely enough that an examiner could reconstruct the logic from the filing alone.

The third channel carries the weight. When a condition score adjusts modeled average annual loss or probable maximum loss, the actuary signing the indication has to understand how the adjustment is derived, the conditions under which it behaves as expected, and its failure modes.

The clearest failure mode is timing. Captures taken shortly after a major storm record post-loss conditions. Feeding them in unadjusted to assign pre-loss condition scores injects a systematic bias into any retrospective validation of the score against loss experience, and it biases in the direction that makes the model look better than it is. Passing a vendor condition adjustment straight through to a rate indication as a black box is exactly the reliance that fails under challenge.

State posture is converging on that objectivity and reproducibility standard rather than case-by-case review. The NAIC's 13-state AI bulletin guidance and the CDI's PRID framework set the same baseline. Imagery clears the objectivity bar more cleanly than most alternatives, because the underlying observation is a photograph of a building rather than a policyholder-reported field.

The Moat Is Switching Cost, Which Cuts Both Ways

What compounds is not the feature set. It is the calibration and the filing history, and both are hostage to the vendor that holds them.

Years of production use generate labeled datasets linking a vendor's signal values to a specific carrier's actual loss outcomes in a specific territory. That mapping does not transfer. A rival can produce similar raw signals and cannot produce signals calibrated against a decade of one carrier's loss experience, so the carrier that switches runs a less accurate model for the years it takes to rebuild the calibration.

The filing history compounds the same way. A carrier two or three homeowners revisions into an imagery variable tied to one methodology is re-certifying an established approach each cycle. Switching mid-lifecycle means re-documenting the derivation, running a fresh credibility study of new data against its own loss experience, and explaining to state actuaries why the methodology changed between filings, to reviewers already stretched by AI-based submissions.

The dependency runs the other direction too, and it belongs in a vendor risk assessment rather than a procurement scorecard. A successful regulatory challenge to a ZestyAI methodology in one state would propagate quickly across the 32-state Z-STORM footprint, because state actuaries follow each other's precedents. A carrier whose imagery-dependent filings across multiple states and perils all rest on one provider's approvals inherits that correlation whole.

Some carriers split it: the approved wildfire and storm models for filing purposes, a broader platform for condition assessment and inspection triage. That builds filing history with the approved model and operational history with the platform, and it is the only structure that preserves the option to move without surrendering the vendor-specific credibility the filings were built on.

Further Reading

Sources