What AI Can and Cannot Reliably Detect From Property Photos

AI photo analysis is genuinely useful for estimating window cleaning jobs — but it is not magic, and a tool that pretends otherwise sets both the business and the customer up for a bad surprise. This is a plain-language rundown of where it tends to be reliable, where it isn’t, and why the honest answer matters more than a confident-sounding wrong one.

What it’s generally good at

• Counting clearly visible windows across a well-framed, well-lit photo of a full elevation — this is the core task, and it’s the one modern vision models handle well when the photo itself is good.
• Recognizing common window types (double-hung, casement, picture windows, sliders) from a reasonably clear, unobstructed view.
• Flagging its own uncertainty — a well-built system can distinguish “I counted this confidently” from “this photo doesn’t show enough of the property to be sure,” which matters more for a usable estimate than raw accuracy alone.

Where it struggles — and why

• Obstructed views. A tree branch, a parked car, a fence, or a porch roof can hide a window entirely, or make a partial view look like a different count than reality. AI can only analyze what’s actually visible in the frame.
• Glare and reflections. Bright sun on glass can wash out a whole section of a photo, and reflections can occasionally make a window look like two, or make a shadow look like a window.
• Sides of the property not photographed. This sounds obvious, but it’s the most common real-world gap: no photo was taken of a side yard, a back elevation, or a detached structure, so there’s nothing for the AI (or a human) to go on.
• Subtle condition details at a distance. Hard water etching, hairline cracks, or how dirty a track actually is can be hard to judge confidently from a photo taken a few feet back, even when the window itself is clearly visible.
• Scale and depth. A single photo doesn’t always make it obvious whether a window is standard-sized or oversized, or exactly how many stories up a window sits, without other cues in the frame (a doorway, a roofline, a person).

Why the gap matters more than the average

A system that’s often right but confidently wrong the rest of the time is worse for a business than one that’s a bit less accurate overall but clearly says “I’m not sure about this part” when it doesn’t know. The second system lets a human step in exactly where it matters; the first one hides the problem until a quote is already in the customer’s inbox. This is the reasoning behind Tallyvis’s estimator: the AI proposes characteristics, surfaces what it’s uncertain about, and a human (the customer, or the business) confirms or corrects before a price is ever calculated — the AI never determines the price on its own.

Getting better results in practice

Most of the accuracy gap above is closed not by a smarter model, but by better photos: full elevations, shot straight-on, in even light, with nothing important cropped out of frame. See how to quote window cleaning jobs from photos for the specifics.