AI-assisted roof and climate data for property insurance underwriting

An underwriter looking at a building needs three answers: what the roof looks like, what the site is exposed to, and what the official records say. HOMAI puts all three in one report per address. It is built remotely from aerial, satellite, street-level and LiDAR imagery and Norway's public registers, without a site visit.

The report fits three points in the policy cycle:

• New business. Check the roof and the hazard bands before you quote, and see which values are official and which are estimated.
• Renewal. Run the address again before renewal and compare the result with the property's analysis history.
• Book review and prevention. Rank properties by risk band and follow up the right buildings before a claim, not after it.

An analysis runs as a background job and can take minutes, so HOMAI fits referral, renewal and prevention work better than an instant online quote.

Every report combines four layers of property data for insurance underwriting. Each derived value carries a provenance label (official, calculated, AI-estimated, AI-inferred or estimated), the provider, a confidence between 0 and 1 and a short basis.

The roof layer is a remote roof condition assessment per address: a multimodal AI model describes roof type, material, condition, defects, pitch and penetrations, and the formula turns those observations into factor scores. The climate layer gives the flood, landslide, precipitation and storm exposure of the exact location from NVE's hazard maps and MET Norway's station data.

The overall property score is banded Low (under 40), Medium (40–69) or High (70 and above). By default the climate score counts for 60% of it and the roof risk index for 40%. The report can be read on screen or exported as an A4 PDF.

"AI underwriting" often means a model that makes the call. HOMAI works the other way round. A multimodal AI model describes what is visible on the roof, in a fixed structured format with randomness turned off (temperature 0), and it is not allowed to estimate age, footprint or area. A text model then writes the summary from the computed facts. The score itself is plain arithmetic over weighted factors. The AI never sets it.

• Reproducible. The same property with the same imagery and the same weights gets the same score.
• Traceable. Each report shows the formula with the actual numbers substituted, and every factor shows its source and confidence.
• Tunable, on the record. The weights are defaults that an administrator can change per deployment, so the settings behind a score are a deliberate choice.
• Honest about gaps. Missing data shows as "Not assessed" or "Not derivable", never as a guess, and low confidence raises the risk index.

When a price, a referral or a declined risk is questioned by a customer, internal audit or a model review, you can show which inputs moved the score and where each one came from. See how the deterministic score is calculated and the provenance labels on every value.

Use HOMAI as remote screening first and send people where it matters. Signals in the report worth acting on include:

• an overall score in the High band, or a roof risk index driven by visible defects, drainage or flat sections;
• low-confidence values, or counts marked ~N because the imagery was hard to read;
• a "Roof not yet assessed" banner: no usable imagery was found, so the report covers climate and registry data only;
• flood or landslide bands at the top of the scale.

The imagery is only as recent as its capture date. What you get is a consistent, documented first look at every address, so inspections go where they change the decision.

If your terms or underwriting rules refer to roof age, it helps to know exactly what the data shows.

• Age uses the official byggeår from Matrikkelen: the year the building was built, not the year the roof was last replaced. HOMAI uses it as a proxy for roof age, and the AI never overrides it.
• Historical change is a separate factor, inferred from historical orthophotos. It shows visible change over time; it does not date the roof.
• Images are never used to estimate age, footprint or area.

Condition comes from what is visible: material, defects, flat sections, drainage, chimneys, roof windows, vents and pitch, cross-checked against LiDAR. By default, age weighs 0.25 in the roof risk index and material and visible defects 0.15 each. See the 10 roof risk factors and their default weights.

Import properties from a CSV file, or add them by address search, map click or a drawn area; they appear as pins on the map. The portfolio view summarises the latest analysis of each property: the count per risk band, the highest-scoring properties and the analysis history per address. That lets you review your whole book by risk band and plan prevention work or inspections where the bands are highest.

The view aggregates the analyses you have run. It does not monitor properties continuously or send alerts; you re-run a property to refresh it.

• It does not replace a site inspection or a tilstandsrapport. It is remote screening.
• It does not set premiums or estimate loss costs, repair costs or property values. There is no monetary model.
• It does not handle claims, including claims under the natural perils scheme. It supplies risk data about the building.
• It does not score snow load. The four climate exposures are flood, landslide, precipitation and storm.
• It does not publish accuracy percentages. Each value carries a confidence and a basis instead.
• It covers Norwegian addresses only, and an analysis takes minutes, not seconds.

Claims and maintenance concern the same roofs. If you work with owners on prevention, see roof data for maintenance planning, or see all industries.