AI roof damage detection uses computer-vision models trained on millions of labeled hail, wind, and granule-loss photos to classify damage on a per-slope basis. In 2026 the best models report hail strikes per square, wind-lift signatures along ridge and rake lines, and granule-loss density heatmaps. Carriers accept AI evidence as supporting documentation alongside a licensed adjuster inspection — not as a replacement for it.
Every storm season produces the same argument on the roof: the adjuster sees three hail hits, the contractor sees thirty, and the homeowner just wants a straight answer. In 2026, AI roof damage detection is what finally closes that gap — but only if you understand what the models actually do and how carriers treat the output.
- →Modern damage-detection models classify hail, wind, and granule loss on a per-slope basis.
- →Best-in-class models hit 94%+ precision on hail strikes vs. licensed-adjuster ground truth.
- →Carriers treat AI output as supporting evidence, not as a standalone proof of loss.
- →Pair AI output with timestamped photos + storm-date NOAA report for fastest claim approval.
- →RoofGenius runs detection in 90 seconds from a 12-photo upload.
What AI damage detection actually does
Damage detection is a computer-vision classification problem. The model takes a roof photo, segments the image into shingles, ridges, valleys, flashings, and penetrations, then runs each segment through a classifier trained on labeled examples of hail strikes, wind-lift signatures, granule loss, and mechanical damage.
The 2026 generation of models — including ours — uses transformer-based backbones with hail-specific attention heads. That replaced the CNN-only architectures that struggled with granule-loss false-positives on aged 3-tab shingles.
The three damage categories the model classifies
- Hail impact — circular bruises with displaced granules and a soft mat behind. Distinguished from blisters by absence of UV halo and presence of fiberglass-mat compression.
- Wind damage — tab creasing, lifted edges, exposed nail heads, sealant strip failure. Often clustered along leading-edge ridges and rakes.
- Granule loss / mechanical — generalized weathering, foot-traffic scuffing, tree-limb abrasion. Important to classify so it's NOT mis-attributed to storm.
How accurate is it really?
The honest answer in 2026: best-in-class models report precision in the 92–95% range against licensed-adjuster ground truth on hail. Wind is harder (~88%) because tab lift can look identical to factory-curl on aged shingles. Granule loss is the easiest (~97%) because the visual signature is unambiguous.
No — and any vendor claiming it does is selling you a lawsuit. AI detection is supporting evidence that speeds up the inspection and prevents adjusters from missing damage. The licensed adjuster still writes the scope.
What confuses the models
- Heavy moss or algae growth — masks granule-loss signatures.
- Synthetic slate or polymer shingles — small training set, lower confidence.
- Wet roofs after rain — specular reflection mimics hail bruises.
- Sub-200 DPI imagery — anything below 4-megapixel close-ups is noise.
What carriers will and won't accept
Top-25 carriers have published guidance on AI damage reports in 2025 and 2026. The consensus: an AI report is welcomed as part of a supplement package, but it must be paired with timestamped field photos and a storm-date NOAA / SPC report to anchor causation.
| Carrier posture | What they accept | What they reject |
|---|---|---|
| Accepts as supporting evidence (most national carriers) | AI report + photos + NOAA storm date | AI report alone with no field inspection |
| Requires re-inspection regardless | AI report flags hits → carrier re-inspects | Any auto-approval of scope based on AI output |
| Pilot acceptance (USAA, Liberty Mutual) | Direct API integration for triage | Drone-only inspections in restricted airspace |
How to use AI detection on real claims
- Field tech captures 12–18 photos: 4 elevations, all ridges/valleys, every penetration, 3–4 close-ups of suspect damage.
- Photos uploaded to AI engine (60–120 seconds in 2026).
- Detection report comes back with per-slope hit counts, wind-lift map, and granule-loss heatmap.
- Contractor pairs the report with NOAA storm date and county hail-size data.
- Supplement letter includes the AI report as Exhibit A, photos as Exhibit B, NOAA report as Exhibit C.
Manual documentation: 35–55 min on the roof + 25 min in the office. AI workflow: 12 min on the roof + 90 seconds for the report. Net savings: 45–60 min per inspection — and the photo coverage is more complete because the tech is following a shot list, not their memory.
What to look for when choosing a vendor
- Per-slope output — anything that returns 'damage: yes/no' for the whole roof is useless on a supplement.
- Confidence scores — the model should tell you when it's guessing.
- Hit-count detail — 'moderate hail' is not data; '17 hits on south slope, 3 on north' is.
- Exportable evidence file — PDF with annotated photos for the carrier file.
- NOAA storm-date overlay — automatic match to the closest recorded hail event.
Cost in 2026
Standalone damage-detection services run $25–$95 per inspection in mid-2026. Bundled platforms (RoofGenius included) deliver detection plus measurements plus supplement drafting in a single monthly plan — $149–$497/mo for unlimited inspections, which is the model most active storm crews choose.
If you're running 8+ inspections a week, the bundled model breaks even in week one. Standalone services only make sense if you're inspecting fewer than 4 roofs a month.
The honest limits
AI damage detection is not magic. It's pattern recognition trained on labeled photos. It will miss the rare exotic damage signature, it will occasionally false-positive on heavy granule-loss roofs, and it cannot tell you whether a hit is from this storm or three storms ago. That last point is why the NOAA storm-date overlay matters more than the model itself for claims work.
Used correctly — as the second set of eyes on every roof — it eliminates the two failure modes that cost contractors the most: the missed damage that loses a job, and the over-claimed damage that gets a supplement denied.