U.S. severe convective storm losses reached at least $43 billion through September 2026, the sixth-costliest January-to-September run on record.
Drones and computer vision now take over much of the climbing, photo sorting and measuring on hail-damaged roofs.
The new bottleneck sits further down the line, in estimates and claim files that insurers increasingly read with their own AI.
By June, parts of the insurance industry were ready to call 2026 a soft year. Texas barely saw its usual spring hail, the Southeast stayed oddly calm, and a few catastrophe modelers said as much in public.
August changed that. Four days of storms, from August 9 to 12, tore through the Plains, the Midwest and a large stretch of the Mid-Atlantic. When Gallagher Re released its Q3 figures, U.S. insurers had absorbed at least $43 billion in severe convective storm losses through September. That makes four straight years above $40 billion.
Roofing companies didn’t need the report to know it. Their voicemail boxes said the same thing.
The more interesting story is how the work got done. For many roofing shops, 2026 was the first season with some form of AI sitting between the inspection and the signed contract. Some of it saved real hours. Some of it mostly created more files to manage.
Tornado footage leads the evening news, but hail fills a roofer’s calendar. Allianz Commercial puts hail at roughly 50% to 80% of all severe convective storm losses, and the U.S. is where most of that money is paid out.
2026 had one quirk. Guy Carpenter’s July update showed total storm reports 36% above the 2016-2025 average through June, yet hail reports came in slightly under normal. Roofs didn’t get off easy, though. Where hail fell, it fell hard, and residents near Chicago were picking up record-size stones as early as March.
This is exactly why AI fits the problem. Hail damage is visual, and it repeats. A bruised shingle on one house looks a lot like a bruised shingle three doors down, and software trained on enough examples gets fairly good at spotting it.
A storm job breaks down into four steps: capture, classify, cost and close. AI helps a great deal with some of them and very little with others.
Aerial roof measurement isn’t new. EagleView and Hover have sold imagery-based roof reports for years, and DroneDeploy lists roofing alongside construction and mining. What has sped up is the work after the flight. A drone can shoot 400 photos of a single house, and someone, or something, has to review all of them.
Analytics Insight has covered how AI-powered drones shifted from surveillance work toward inspection and maintenance, and roofing turned out to be a natural fit.
Roofers usually bring up safety first. Walking a wet 12/12 pitch at seven in the morning is how crews get hurt.
Contractors are borrowing the computer vision methods already used for defect and quality inspection on construction sites and pointing them at shingles. A capable model counts probable strikes in a test square, judges whether a mark came from hail, blistering or a boot, and flags which slope deserves a closer look.
Older 3-tab roofs still trip these systems up, even ones that have already shed half their granules. Many adjusters also still trust a chalk circle they drew themselves over any heat map. Honestly, it’s hard to fault them.
Estimating changed more than anything else this year. In January, XBuild closed a $19 million Series A and launched a residential roofing proposal tool that, according to the company, produces a finished estimate in under 15 minutes.
On the commercial side, Coruzant has examined how AI is pulling takeoff data out of static PDF drawings so estimators stop measuring by hand.
Dalton Mills, for instance, offers its latest AI for roofing as a set of no-code tools covering proposals, pricing, crew dispatch and job reports. The idea at Dalton Mills, and at a handful of similar platforms, is that software should fit how a roofing company already runs rather than the reverse.
The specific vendor matters less than one habit: measurements should never be typed in twice. Each retype is a chance to fat-finger the pitch multiplier or drop the waste factor, and on a 34-square job that single slip can erase the profit.
Carriers spent the year automating too. Analytics Insight has reported on how property and casualty insurers use machine learning to score claim severity and route files. In practice, a contractor’s photos and line items may be read by a model before any adjuster opens the folder.
That shifts how an estimate should be built. Photos need consistent labels. Documentation should go slope by slope. Line items should match what the carrier’s estimating platform expects. Shops that treated AI as a record-keeping tool, rather than a way to impress homeowners, appear to have been paid faster this season.
The next hail season is roughly five months away. Fixing the process now is far easier than in the middle of 60 open inspection requests.
Elevations, every slope, test squares, soft metals, gutters, collateral damage. Same order at every house. Both the software and the adjuster benefit.
Whether the link runs through Dalton Mills, JobNimbus, AccuLynx or a spreadsheet maintained since 2019, numbers from the measurement report should land in the estimate automatically.
Reviewing every 2026 supplement shows what kept getting left off: drip edge, starter strip, ice and water shield, code items. Those belong in the template.
Imagery-based triage helps sort roofs into likely replacement, likely repair and probably nothing before anyone is dispatched. Some calls will be wrong. It still beats working jobs in the order people happened to phone in.
Vendors rarely volunteer this list:
Borderline damage calls. If the model counts six hits per square and the carrier wants eight, someone should go up and look.
Anything under the shingles. No camera at 120 feet can tell that the decking is soft.
Homeowner conversations. People are stressed after a storm and often confused about deductibles. That talk belongs with a person, not a chatbot.
Code questions. Requirements differ by county, and a general-purpose model has no idea what a local inspector flagged last month.
After a $43 billion year, AI’s place in roofing looks clearer. It has taken over much of the climbing, photo sorting and measuring, plus a good share of first-draft estimating. It hasn’t taken over judgment, and probably won’t soon. The slow part of a storm job used to be on the roof. Now it sits mostly in the claim file and the pricing behind it.
Whether a crew runs EagleView, XBuild, Dalton Mills or something launched next March, the contractors who come out ahead will be the ones whose tools pass data cleanly from one step to the next. Time spent on that over the winter will make the spring surge a lot less painful.