What problem does it solve?
A homeowners or commercial property insurer has to know the condition of the building it insures. The cheapest source is the application, which the applicant fills in from memory, and the usual check is a physical or drive by inspection ordered after the policy is bound or at renewal. Inspections cost money and time, they can leave customers with a poor experience, and they cover only part of the book. Florida Peninsula's Director of Underwriting describes the budget pressure: a third of its in force book is inspected again every year on a rotating basis, at around $35 a physical inspection.
The roof is the usual proxy for the rest of the structure, yet the standard rating input is roof age, which CAPE Analytics says is not an accurate measure of how a roof performs. Wildfire has a similar gap, because legacy wildfire models paint large regions as high risk, according to ZestyAI. AI that reads high resolution aerial imagery addresses both gaps: it rates each property from a recent image, without a visit, at the moment of the quote.
- Florida Peninsula's Director of Underwriting says a third of its in force book is inspected again every year on a rotating basis, and that the cost per physical inspection is around $35.AI-based Property Intelligence from CAPE Analytics Fuels A New Level of Underwriting Quality and Efficiency at Florida Peninsula (2021)
- ZestyAI says legacy wildfire models mark regions equal to 15% of all properties in California as high risk, while its Z-FIRE model estimates that 7% of properties in the state have a true high wildfire risk exposure.California Department of Insurance Approves Filings Using ZestyAI Z-FIRE Model for Both Underwriting and Rating (2021)
How does it work?
- Match the address to a building. At quote, renewal or on a portfolio refresh, the system sends the address to the imagery provider, which resolves it to a building footprint and the most recent aerial capture available for it.
- Extract property features. Computer vision models read the image for roof condition (missing shingles, tarps, discoloration), roof geometry and covering material, and in wildfire models for building materials and the vegetation around the structure.
- Score the risk. A model turns the features into a condition rating or a property level risk score. Wildfire scores also account for slope, climate history and mitigation work the owner has done, such as clearing vegetation or fitting fire resistant materials.
- Apply the insurer's rules. The underwriting system receives the score by API and compares it with the guideline. A good roof can skip an inspection, a poor one can send the file to an underwriter or an inspection order, and a severe one can make the property ineligible.
- Check for recent changes. Imagery lags reality. Florida Peninsula's Director of Underwriting says that if the score is poor but a permit for a new roof was pulled the day before, the home is inspected anyway.
- Explain and monitor. The score and the image are stored with the file so that the decision can be explained to the broker, the policyholder and the regulator. Some insurers rescan the portfolio to warn customers about changing conditions and to decide which renewals need a visit.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- API and system to system, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | at least 95% | 1 | 1 organization |
Value drivers: Lower cost to serve, Risk and loss reduction, Speed and cycle time, Customer experience.
Indicative value
A homeowners insurer with 300,000 policies in force
USD 93,000 to USD 439,500
Net annual inspection spend avoided per year
How this is calculated
Formula: policies * (inspectedShare * costPerInspection * avoidedShare - dataCost). The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Policies in force policies, policies | 300,000 | 300,000 | The reference insurer. |
| Share of the book that is physically inspected each year inspectedShare, fraction of policies | 0.33 | 0.33 | Florida Peninsula says a third of its book is inspected again every year (https://capeanalytics.com/florida-peninsula-qa/). Replace with your own inspection schedule. Source |
| Cost of one physical inspection costPerInspection, USD per inspection | 35 | 35 | The cost Florida Peninsula quotes for a physical inspection (around $35, https://capeanalytics.com/florida-peninsula-qa/). Replace with your vendor's price; a fuller inspection costs more, which raises the saving. Source |
| Share of those inspections the imagery score makes unnecessary avoidedShare, fraction of inspections | 0.2 | 0.3 | Editorial range, not a measured result. The only quantified pointer is a forecast. Florida Peninsula insures about 150,000 homes (2020 release) and inspects a third again of them each year, so about 50,000. Its Director of Underwriting says the roof ratings could easily lead to "another 10,000 properties we don't need to inspect each year", which implies roughly 20% of those inspections. The range starts at that forecast and allows for more; below about 17% the imagery fee outweighs the saving. Measure the share in a pilot on your own book. |
| Imagery and scoring fee per policy and year dataCost, USD per policy | 2 | 2 | Editorial assumption. Replace with the price in your contract. |
What it leaves out: Leaves out the loss prevention effect of finding bad roofs earlier, premium changes from better pricing, the cost of integration and of handling disputed scores, and the inspections the imagery adds when it flags a property that would not have been visited.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Zurich North America
United States · Insurance · 2025
In October 2025 Zurich North America and Nearmap announced that Nearmap's AI analyzed aerial imagery and roof condition scores are delivered through the Betterview platform directly into Zurich's U.S. Middle Market underwriting platform. Underwriters see regularly updated imagery, roof condition scores and signs of deferred maintenance while they assess a risk. In an initial limited rollout, underwriters found risks that conventional methods may have missed, and the results led Zurich to roll the integration out across its U.S. Middle Market underwriting offices. The release gives no figures.
No outcome disclosed.
California FAIR Plan Association
United States · Insurance · 2021
The California FAIR Plan, the state's insurer of last resort for property, selected ZestyAI's Z-FIRE wildfire risk model in 2021 after evaluating several models. Z-FIRE scores individual properties from features such as building materials, vegetation density and ease of fire suppression, with aerial imagery as one input. Under a five year agreement the FAIR Plan uses it to assess wildfire risk for every new quote and every property in its portfolio, to price more accurately and to reward homeowners who harden their homes. The plan's president says Z-FIRE outperformed the other models evaluated; no result figures were published.
No outcome disclosed.
Florida Peninsula Insurance Company
United States · Insurance · 2021
Florida Peninsula, a Florida homeowners insurer, chose CAPE Analytics in 2020 to rate roof condition, roof geometry and roof covering from geospatial imagery at quote, so it could validate rating data and cut unnecessary physical inspections. Its Director of Underwriting describes a vendor evaluation in which each provider analyzed the same properties and the team compared the results with its own physical inspections. About 12 months after integration he calls the result "phenomenal": the underwriting system receives the roof score instantly, some homes are screened out on condition before an inspection is ordered, and the insurer planned to extend the scores to its renewal book.
- Accuracy: at least 95%
"Every property with a good roof condition rating from CAPE was within our risk tolerance more than 95% of the time."
Claimed by: organization
CSAA Insurance Group
United States · Insurance · 2019
CSAA Insurance Group, a AAA insurer, partnered with CAPE Analytics in 2019 to assess the condition of member homes from geospatial imagery instead of ordering an inspection for each one. The insurer used the roof condition rating to help members repair damaged roofs, to price home policies more accurately at renewal and for new business, and to skip inspections of good roofs. No numbers were published; CSAA's managing principal of core innovation describes the aim as consistent and objective underwriting decisions.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Clean property addresses and geocodes for the in force book and for new quotes
- Written inspection and eligibility guidelines that can be encoded as rules, by roof condition and hazard class
- Historical inspection results for a sample of properties, to compare against the imagery score
- Records of permits, claims and prior repairs that can override a stale image
Systems to integrate
- Imagery and property intelligence provider, by API
- Quote, rating and underwriting workbench, where the score is shown and used
- Inspection ordering and vendor management system
- Policy administration system, for renewal and portfolio rescans
- Document storage for the image, the score and the rule version used
Complexity: Medium
The scoring comes from a specialist provider by API, so the build is the integration. It needs address matching, a write back into the quote and underwriting system, inspection and eligibility rules the underwriting committee signs off, and a way to keep the image and score with each file. The two reported rollouts describe an initial limited rollout (Zurich North America) and 12 months from integration to a review (Florida Peninsula).
- 1
Benchmark providers on your own properties
Send the same sample of properties to each provider and compare the ratings with your own physical inspections, as Florida Peninsula did. Decide the pass mark before you look at results, for example the share of good rated roofs that your inspectors also judge acceptable.
- 2
Start with inspection triage, not decisions
Use the score first to decide which properties need an inspection, in new business and at renewal. Keep declines and price changes with underwriters until the error rate on your book is known.
- 3
Encode the guideline as visible rules
Write what a good, fair, poor and severe rating does in each line and territory as rules that underwriting management can read and version. Add the override cases, such as a recent permit or a recent claim.
- 4
Fix the image age problem
Record the capture date next to every score and set a maximum age per use. A score from an image taken before a storm or a roof replacement should send the file to a human.
- 5
Run a controlled comparison before widening
Hold out a sample that is inspected regardless of the score, and compare claims and inspection findings for scored and unscored properties over a renewal cycle.
- 6
Add the portfolio view last
Once quote time scoring works, rescan the renewal book to rank properties for inspection and to warn policyholders about conditions they can fix.
Guardrails
- A policyholder or broker can ask for the image, the score and the rule that applied, and can dispute it
- No automatic non renewal or decline from an imagery score alone
- Every score stores its image capture date, provider and model version
- Wildfire and condition models tested for unfair discrimination, including proxy effects through geography, before use in rating
- Imagery limited to the insured address and its surroundings, with no use for other purposes
KPIs to instrument
- Share of quotes and renewals that skip an inspection, and the claims experience of that group
- Agreement between imagery rating and physical inspection on a held out sample
- Inspection spend per policy, before and after
- Quote to bind time for properties scored at quote
- Disputes and overrides per thousand scored properties
Human in the loop
Underwriters decide on referrals, ineligible ratings and disputed scores, and can order an inspection that overrides the score. Underwriting management owns the rating thresholds and reviews a monthly sample of properties that skipped inspection, compared with later inspection or claim findings.
Common failure modes
- Stale or wrong imagery
- A roof replaced last month still scores as poor, and a storm last week is not yet in the image. Show the capture date, check permits and claims, and send contradictions to a person.
- Rating drives non renewal without explanation
- An adverse decision that the customer cannot see or contest damages trust and invites regulatory attention. Keep the image and rule behind each decision and give a route to challenge it.
- Wildfire score used as a blunt exclusion
- A property level score is meant to separate good risks from bad ones inside a hazard zone. Applying it as a cut off for the whole zone brings back the problem of regional models.
- Provider lock in on the only copy of the data
- If the scores are not stored with the policy, changing provider restarts your inspection programme. Store the features and the version, and keep a held out sample for benchmarking.
What are the risks and rules?
EU AI Act
Minimal risk
Annex III point 5(c) lists risk assessment and pricing of natural persons in life and health insurance. Property condition and wildfire scoring for homeowners or commercial property is not listed. It would move up to high risk if the same scores were used to assess or price life or health cover for individuals.
Guidance
- Opinion on Artificial Intelligence governance and risk management (European Insurance and Occupational Pensions Authority, Europe). Sets out how existing insurance rules on governance, data quality, fairness and human oversight apply to AI systems that are not high risk under the AI Act, such as property risk scoring.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(c) names life and health insurance risk assessment and pricing of natural persons; property insurance is outside it.
- SB21-169, Protecting Consumers from Unfair Discrimination in Insurance Practices (Colorado Division of Insurance, North America). Requires insurers to test external data, algorithms and predictive models for unfair discrimination. Its work so far focuses on life, health and private passenger auto insurance, and shows the testing that regulators ask for from models that use external data.
Controls to put in place
- AI inventory entry per provider and model, with an accountable underwriting owner
- Documented rating thresholds and override rules with version history and sign off
- Stored image, capture date, score and rule version for every decision
- Periodic bias and proxy testing of condition and wildfire scores on the insurer's own book
- Complaint and dispute log reviewed with underwriting and compliance
- Third party risk assessment of the imagery provider under DORA
Frequently asked questions
- How accurate is roof condition scoring from aerial imagery?
Florida Peninsula compared several providers with its own physical inspections and reports that every property CAPE Analytics rated as having a good roof was within its risk tolerance more than 95% of the time. That is one insurer's evaluation reported on the vendor's site, so test it on your own book and your own definition of tolerance.
- Does imagery replace physical inspections?
Not in the cases on this page. CAPE Analytics and Florida Peninsula describe imagery as a way to reduce unnecessary inspections, with targeted in person checks where the score or other data calls for one. Zurich North America describes the imagery as information underwriters see alongside their other data while they assess a risk.
- Is wildfire scoring from imagery allowed for underwriting?
In California the department of insurance approved filings that use ZestyAI's Z-FIRE model for underwriting and rating in 2021, first for Metropolitan Direct and then for Farmers Insurance (underwriting only), according to ZestyAI. The California FAIR Plan later selected the same model for every new quote. Check the filing rules of each state or country before use.
- Is aerial imagery scoring high risk under the EU AI Act?
Property and commercial lines scoring is not on the Annex III list, which names risk assessment and pricing of natural persons in life and health insurance. Existing insurance rules on governance, data quality and fair treatment still apply, as EIOPA's 2025 opinion explains.
How to cite this page
Blits.ai AI Use Case Library, "AI aerial imagery risk scoring for property insurance underwriting", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/aerial-imagery-property-risk-scoring. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published