ZestyAI

The model is the entire product. What the company sells is a score, and the score is produced by models trained on loss data over imagery, parcel, permit and structural inputs. Strip the models and what remains is a pile of aerial photographs and public records with no product attached, which is the Oxane refinement answered at its clearest. Independent trade coverage describes it as built from inception around risk scoring rather than as a platform that added models, and nothing in its own material contradicts that.

The vendor supplies an input and the carrier decides, so the autonomy question sits mostly with the buyer, and what is published is better than most. Every score is delivered alongside its top risk drivers, aerial imagery of the property and a mitigation simulation, which puts the explanation in front of the underwriter before the decision rather than in a log afterwards. That is the right side of the standing test.

Held at B rather than A because the vendor states no threshold, no review structure and no position on how its output should be used, and independent commentary specifically cautions buyers against deploying the scores as a hard underwriting filter without validating against their own claims experience. The product is capable of being used with oversight and does not require it.

One of the strongest cases on this axis in the index, and most of it is externally compelled rather than voluntary, which is the rare and valuable form. Detailed technical documentation is published to support regulatory review and filings. The models are stated to be trained on actual loss experience rather than on hazard proximity alone, with the wildfire model built on the largest historical wildfire loss database the company knows of, covering two decades.

The scientific basis is named and external rather than proprietary and vague: structure ignition and fire behaviour research from the Insurance Institute for Business and Home Safety. The models are described as fully documented and explainable, and each score is delivered with its contributing drivers.

Crucially, a state regulator approving a model for use in rating has to be satisfied of its validity first, so the approvals are themselves evidence that documentation adequate to external technical review exists and has been examined. Held at A rather than qualified: what is still absent is published performance data, a validation report a buyer can read without filing, and any drift or versioning policy.

Graded A on the second route in the written bar rather than the first, and the reasoning is recorded so it is not relitigated. On the customer route this would be a B in the usual shape: Heritage Insurance and NEXT Insurance are named, and the numbers sit apart from them, namely roughly 40 percent of the California homeowners market and more than 511,000 previously uninsurable properties covered in 2024. What lifts it is the independent party route.

State insurance departments have reviewed these models technically and approved them for use in rate filings across sixteen or more states, which is validation by a party with public accountability at stake rather than by an analyst firm the vendor can pay. That is a stronger form of third party confirmation than anything else in this pocket, and it is repeatable and checkable state by state.

One phrase in the vendor's own description raises the whole question and nothing answers it: the models rest partly on validated loss signals derived from contributed and computed datasets. Contributed means carriers supply their own loss experience. Nothing states which carriers contribute, on what terms, whether contribution is a condition of purchase, or whether one carrier's loss history improves models sold to a competitor writing the same territory.

In property catastrophe risk that is commercially material, because loss experience in a peril and a geography is the scarce asset. Sixth instance this session of a claim about why the models are good standing in for a commitment about how data is handled.

Nothing published, and the shape of the exposure is unusual enough to state precisely. This company builds a detailed profile of an individual property, including roof condition, building materials, defensible space and structural vulnerability, from aerial imagery and public records, without the owner of that property being a party to anything. The subject of the analysis is not the customer and never interacts with the vendor. Nothing addresses retention, correction, or what a property owner may see or ask about the record held on their home.

No security page, trust portal or certification located. Written as a queued question rather than a silent refusal: this vendor sells to large admitted carriers and to insurers of last resort, and the procurement bar in that market makes some certification likely to exist. The specific gap worth noting is that a firm holding contributed loss data from multiple competing carriers is exactly the kind of custodian a buyer would want a security report from, and none is published.

A form of regulatory standing this index has not previously recorded, and it deserves its own entry on the ladder. The existing routes to an A are a supervised live test of the product by a financial regulator, or a licence granted to the firm by a competent authority. This is neither: it is regulatory approval of the model itself.

The wildfire model was the first AI based wildfire model approved as part of a carrier rate filing by the California Department of Insurance, and the severe convective storm suite holds approvals across sixteen or more states with further filings pending. The company maintains a dedicated regulatory team engaged with departments on emerging policy, and remains filing ready in California under the state's pre application determination process for catastrophe load factor models.

The distinction that makes this an A: the firm is unlicensed and unsupervised as an entity, but its models have been examined and admitted for use in a regulated pricing process by public authorities, repeatedly, in named jurisdictions a buyer can verify.

Rare on this axis and earned on the standard that separates a mechanism from a value. The vendor states that fairness, bias and discrimination testing is performed in alignment with state Department of Insurance guidance, which names both the practice and the external standard it is measured against rather than asserting a principle. That is the same reasoning that gave TCS its grade for shipping bias testing as a named component.

Held at B and not A because no results, methodology, protected class breakdown or independent review are published, and the underlying exposure remains real: property level scoring from structural and locational features can track historical patterns of housing quality and investment, and only the regulator currently sees whether it does.

Graded C, and it is the closest this index has come to an actionable recourse mechanism without being one. A high score can mean a declined application or a surcharge on a home, and the property owner is not a party to the analysis and has no published route to see the score, obtain the drivers behind it or dispute the underlying data.

What is genuinely notable is the mitigation simulation: the product can show what would change the outcome, which is the substance of recourse rather than the form of it. But it is delivered to the underwriter, not to the homeowner, and nothing commits a carrier to pass it on, states a correction process for wrong property data, or addresses who bears the consequence when a structure is scored on stale imagery or a permit record that is wrong.

No model architecture, family, framework or provider named, and the data supply chain is described only in categories rather than in sources: proprietary property data, aerial imagery, enriched permit and parcel data. Neither the imagery providers nor the parcel data vendors are named, which matters more here than the absence of a language model would, because for this product the imagery pipeline is the dependency that determines whether a score reflects the house as it stands today. The one named external input is a research institute supplying the science, which is credited under model risk and earns nothing here.

Two distinct routes into a carrier and both are described. A policy administration partnership with a named vendor puts scores into the underwriting workflow, and Z-VIEW is a browser application explicitly requiring no integration work, aimed at underwriters, actuaries and field teams from day one.

The more interesting artifact is commercial rather than technical: pre approved relativities and rating factors are offered so a carrier can drop the output into an existing rating plan without deriving its own. Held at B: no published interface documentation or connector inventory.

One useful statement and nothing else. The browser application is described as needing no integration, which tells a buyer something real about time to value, and beyond that there is no hosting model, no cloud named, no single tenant option, no customer responsibility split and no residency statement. Residency is admittedly less pressing for a United States only product scoring United States property, and the axis is still unaddressed.

No pricing, no commercial model, no indication of whether scores are sold per address, per policy in force or by subscription. Every route is a contact form or a demo request.

Real depth within a deliberately narrow scope. Buyers span admitted carriers, insurers of last resort, reinsurers and commercial small business insurers, and the company states that carriers writing roughly 40 percent of the California homeowners market use the wildfire model. Held at B rather than A because both the line of business and the geography are single: United States property, with no life, health, casualty, specialty or non United States coverage anywhere. That is a coherent strategy rather than a weakness, and the axis still measures reach.