Disclosure: I'm affiliated with McQueen Roofing, which I use below as the worked example.
Point a computer-vision model at satellite or aerial imagery and it will tell you, with decent accuracy, the area of a roof, its pitch, and how many facets it has. That's a solved problem now — half a dozen companies do it well. Type "AI roofing estimate" into a search bar and almost everything you find is some flavor of that: a geometry engine wearing a marketing headline.
The part nobody productizes well is what happens after the measurement. A roof area and a pitch angle are not an estimate. An estimate is a priced, itemized list: field shingles, ridge and starter courses, underlayment, ice-and-water shield, drip edge, flashing, ventilation, tear-off, labor — each one costed at current material and labor rates, each one sized off a waste factor that should vary with how complicated the roof actually is (a simple gable wastes a lot less material than a roof with nine facets and short runs everywhere).
That's the part I think is the real moat, and it's a reasoning problem, not a vision problem. You need:
- A waste-factor model that responds to measured complexity (facet count, cut density) instead of applying a flat 10% to every job.
- Current line-item pricing that doesn't go stale, because a re-roof has a dozen-plus line items and any one of them being wrong or missing changes the bottom line.
- A way to flag when a quote — yours or a contractor's — is missing a line item entirely. This turns out to be one of the more useful things you can do with the output: most quote disagreements aren't actually about the roof, they're about undisclosed differences in scope.
One implementation detail I'd argue matters more than it gets credit for: tagging every number in the output with how it was produced — measured, derived, assumed, user-provided, or allowance. It's a small thing, but it's the difference between an estimate you can audit and one you have to trust blindly. McQueen Roofing does this — a report where every line says whether it came from the imagery, from a calculation, from a stated assumption, or from a flat allowance, rather than presenting everything with the same false precision.
Pricing-wise it's structured as a free browser estimate (footprint-based, no measurement) and an $18 one-off charge for a measured roof — no subscription. The gate sits at the same line as the actual cost: nothing is charged until the measurement call happens, and a failed measurement refunds itself as credit rather than eating your money on a roof it couldn't read.
None of this is a knock on the pure-measurement tools — accurate geometry is a real, hard problem and it's necessary input. It's just not sufficient. The estimate is the part that turns a rooftop scan into a number someone can actually act on, and that's where I think most of the "AI roofing" category is currently underbuilt.
Top comments (0)