AI Roof Inspection Software: What Drone and Photo Tools Really Detect

AI roof inspection software can flag visible surface problems in clear photos, such as missing or lifted shingles, hail-like marks, exposed fasteners and granule loss. It cannot confirm the cause of damage, see under the surface, or sign a report, so a person must check every finding that reaches a customer or an insurer.

The vendor pages that rank for this keyword (I read the top four, plus a drone platform) describe what their tools find but print no detection accuracy figure, so a vendor's word is all you have until you test. Detection is also a different job from measurement: labelling a mark in a photo is not the same as getting an area or a pitch right, and each needs its own test.

Follow me on Instagram@sagnikteaches

What five vendor pages say, and what they leave out

I read five product pages in November 2026. The table records what each says about inputs, findings, checking and cost, and where it is silent. Prices are USD list prices.

Connect on LinkedInSagnik Bhattacharya

Read that for what it is. These are product pages, written to sell, and a missing accuracy figure does not mean a tool is poor. It means the claim cannot be checked from the page. You can read the claims yourself on the ScanShield AI page and Roof Hawk page. For each tool you are considering, find out whether it takes the kind of photos you actually take, whether it shows you a confidence level for each flag, and whether anyone has to approve a finding before it goes into a report.

Subscribe on YouTube@sagnikteaches

Detecting damage and measuring a roof are different jobs

Measurement is geometry: area, pitch and lengths worked out from imagery, with errors that come from scale, obstructions and imagery dates. Detection is classification: is that mark a hail hit, a shadow, a patch of moss or an old repair? Its errors are wrong labels, not wrong sizes. A tool can be good at one and poor at the other, and a page that talks confidently about measurements tells you nothing about damage. The measurement side has its own five-roof accuracy test, so this tutorial stays with detection.

The inputs differ as well. Measuring from the air usually needs overlapping images taken in a planned pattern. Spotting a damaged shingle needs a clear, close, well-lit view, and a wide shot from a height tends to lose small marks. That is why the capture method matters more than the brand. A drone gets to ridges and steep slopes without anyone climbing; a phone held from a ladder or the ground gets sharp close-ups of a low roof; ground-level photos catch gutters and fascias but not much else. DroneDeploy's roofing page also describes thermal capture, saying thermal maps help identify trapped moisture, wet insulation and leaks that colour photos do not show. A temperature map is a pointer, not a diagnosis, so I would treat a patch as a place to check with a moisture meter, not as a confirmed leak.

How a detection tool reaches a flag, and why it gets it wrong

The tool scans each photo, proposes regions that look like something it has learned, and attaches a label. Many tools also attach a confidence score, which is how sure the model is, and show only flags above a cut-off. Move the cut-off up and you get fewer false alarms but more misses. Move it down and you catch more but spend longer dismissing junk. You usually cannot change the cut-off, so you test the tool at the setting it ships with.

Illustrative output from a tool that does show scores looks like this. Not every product shows them.

Photo 14 | north slope, close-up | hail impact     | confidence 0.82
Photo 14 | north slope, close-up | granule loss     | confidence 0.41
Photo 22 | ridge                 | missing shingle  | confidence 0.67
Photo 31 | south slope, wide     | hail impact      | confidence 0.58
Photo 37 | east slope, close-up  | exposed fastener | confidence 0.74

A person looking at those five flags might keep photo 14's first flag (round marks in a cluster), drop its second (normal wear near the ridge on an old roof), keep 22 after checking a second angle, discard 31 as a tree shadow and keep 37 after zooming in on the nail head. Three of five survive. That filtering is the work the software does not do, and it is why the number of false flags matters as much as the number of hits.

The usual causes of false flags are worth knowing before you shoot a single photo: shadows from trees and chimneys, moss and algae streaks, dirt and stains, old repair patches, wet shingles, glare and low light. The usual causes of misses are small marks photographed from too far away, dark shingles, shadowed slopes and damage that sits on an edge or under a flashing. None of these tools can see decking, ventilation or moisture under the surface from a photo of the top.

A three-roof trial you can run before paying

Pick three roofs you already know. Roof A has storm damage that you or an adjuster confirmed on the roof. Roof B is clean, perhaps re-roofed in the last couple of years or inspected recently with no findings. Roof C is old and worn, with repair patches, moss and a few real defects, because that is the roof your crew meets most often. Shoot 40 to 60 photos of each, the way the tool will see them in real use: drone passes if you will fly, phone close-ups if you will not. Use a dry roof and even light, and start the vendor's trial clock only once the photos exist. ScanShield's Starter trial and RoofV's trial both run 7 days, so a week disappears quickly if you are still taking pictures.

Before the tool sees anything, have two people mark every photo independently, writing down each item they would put in a report. Then compare the lists. On Roof A, one inspector marked 26 items and the other 25. They agreed on 22, and the 7 that only one of them marked (4 and 3) went to a joint look, where 2 turned out to be real. The ground truth for Roof A is 24 items. That tells you something: each inspector found 23 of the 24 real items and added 3 and 2 false marks. People are not perfect, so score them the same way you will score the tool, and hold the tool to that standard rather than to flawless.

Now run the photos through the tool and record, for each roof, the number of flags, how many of them match a confirmed item, how many do not, and how many confirmed items it never flagged. The results, with every figure invented for the exercise:

The two right-hand columns have standard names, precision and recall, in case a vendor quotes them. Precision is the share of flags that were real; recall is the share of real items the tool found. Ask any vendor that quotes either figure what it was measured on, and how the real items were decided.

Reading the numbers: what the three roofs say

The clean roof is the most revealing row. A clean roof should produce nothing, and six flags means the tool will create work where there is none. The 26 false flags in the example had four causes: 10 shadows, 10 patches of moss, dirt or staining, 5 old repair patches and 1 sealed nail head. The 9 misses were not random either. On Roof A, two lifted shingles sat in shadow and two hail marks were on a dark slope; on Roof C, three small exposed fasteners, one cracked vent boot and one lifted tab went unflagged.

That pattern points to what you can fix and what you cannot. Shadows and dark slopes respond to how you shoot: reshoot one roof on an overcast day and see whether the false flags and misses fall. If they do, capture is part of the answer. If they do not, the limitation sits in the tool. Small fasteners and a cracked boot will always be easy to miss in a photo, whoever takes it.

The numbers also decide how the tool can be used. Nearly one real item in four (9 of 38) never reached the flag list, so a crew that skipped the unflagged photos would have left a quarter of the findings out of its reports. A tool like this cannot be your screen. It can be a second pair of eyes after your own inspection, or a way to sort photos so the likeliest ones are reviewed first, with every photo still seen by a person.

A capture checklist you can fill in for each trial roof

Poor capture ruins a trial and then gets blamed on the tool. Follow the vendor's own photo guidance first, and where it gives none, work to a routine you can repeat on every roof. Never step onto a roof for a trial; shoot from positions you would use anyway, and check the rules that apply to your flights before you fly.

1. One wide shot of each slope from the ground, then 8 to 10 overlapping close-ups along the slope in a grid, so no area is photographed only once.
2. Close-ups of every penetration (vents, flues, chimney), every flashing (valleys, walls, chimney) and the full ridge and eave line.
3. Name files by slope and number, such as N-01 to N-12 for the north slope, so tool flags and human marks can be matched photo by photo.
4. Record the conditions on a single line before you leave.

A filled-in line looks like this: "Roof C, 10:40, overcast, dry; drone pass at a consistent height plus 31 phone close-ups; last re-covered 2019, confirmed on the old invoice." The overcast note matters later, because shadows are the commonest cause of false flags and an overcast morning removes most of them. If you order flights from a service rather than flying yourself, ask who is responsible for the flight and what the images will be used for.

Where the tool stops and a person decides

Little of the right-hand column is something a photo can settle. The tool shortens the search for visible problems; it does not replace the person who stands in front of the customer and says what the roof needs.

Questions to put to a vendor before you upload a customer's photos

• Which photo types and resolutions does the tool accept, and which did your own accuracy test use?
• What is your detection accuracy, on how many roofs, and who decided what was real?
• Does each flag show a confidence level, and can I filter by it?
• Can a person edit or delete a flag before the report is made, and does the report show what was edited?
• Where are photos stored, for how long, and are they used to improve your models?
• What do your terms say about accuracy and liability?
• Can I export photos and findings if I leave?

The last three matter more than they sound. One page I read says its model "continuously learns from roofing imagery", and none of the pages says whether customers' photos are part of that. Ask in writing, and keep the reply with your contract notes.

Whose finding is it? Wording and liability

The report carries your name, not the vendor's. Treat anything the tool produces as a draft: a flagged mark is not a finding until a named person has confirmed it. That is especially true of cause. Words such as "storm damage" or "caused by hail on a given date" are conclusions, and on an insurance claim they belong to the insurer's adjuster or an engineer. A tool that drafts supplement letters raises the same issue. Roof Genius AI's page says "You approve", and the same applies to any text you did not write: read every line, and send nothing you could not defend yourself. For claim work, documenting damage without adjusting it covers where the line sits.

Keep three things for every inspection: the original photos untouched, the tool's output as exported, and your reviewer's marked-up version with a name and a date. Wording of the kind you might put in a report follows, to be checked with your insurer and a qualified adviser before you adopt anything like it:

Observations in this report were made by [name] from photographs taken on [date]. An AI tool was used to sort the photographs and suggest areas to examine. Each observation was reviewed by [name]. This report records what was seen. It does not state the cause of any damage or the date it occurred.

Here is how that can go wrong, as an illustration. A crew sends a homeowner a report with 14 machine labels, and two of them, "hail impact", are tree shadows. The homeowner passes it to an insurer, who asks for the date of loss and the storm record. Nobody has decided either, and the contractor spends a week explaining that the labels were suggestions. Compare the tool's labels with what a person writes for the same flags:

The person's version says what was seen, what it is consistent with and what is not known. The tool's version says only what it guessed. If you show homeowners a roof score, explain what it means or leave it out, because a single number reads as an expert verdict whatever its footnote says.

The time sum: does triage pay for the plan?

An inspection has 60 photos and a person takes 45 minutes to review them all (figures illustrative). If the tool's miss rate means you still look at every photo, as the trial says you should, the tool saves no review time at all. It adds about 9 minutes of dismissing false flags per roof (26 false flags over three roofs, at about a minute each). Its saving then has to come from elsewhere: sorting, labelling and laying out the report.

Put a plan at $99 a month against staff time at an illustrative $45 an hour. To break even, it must save 2.2 hours a month. Across 8 inspections that is about 17 minutes each, and because it also costs you 9 minutes of false-flag checking, it has to save roughly 26 minutes of sorting and writing per inspection to get there. That is a number you can test: time one inspection with the tool and one without, on the same photos, and compare. If the gap is nowhere near 26 minutes, the plan does not pay for itself, whatever the demo promised.

A rule for sorting the tool after the trial

I would put the tool in one of three boxes:

• Triage helper: it finds at least 9 in 10 confirmed items on the damaged roof and raises no more than one or two false flags on the clean roof. A person still looks at every photo before a report goes out. These thresholds are my starting points, not an industry standard.
• Photo organiser and drafting aid: it misses more than that or over-flags, but it measurably saves time on sorting, labelling and report layout. Nobody relies on its flags.
• Not for us: it does neither.

The example tool, with 76% of real items found and six flags on a clean roof, lands in the second box at best, and in the third if the timing test fails. That is a perfectly good result for a trial: you learned it in a week of testing, instead of finding out on a customer's roof. When a tool does earn a place, turning confirmed photos into a customer-ready report is the next job, and photo documentation tools for roofers compares the apps that hold the pictures. Repeat the trial after any major update, using the same three roofs, so you can see whether the numbers moved.

Roof inspection AI: questions to settle before a trial

Can AI tell hail damage from normal wear?

Sometimes it flags marks that look like hail, but it cannot know the cause or the date. Round marks can come from foot traffic, blistering or manufacturing flaws, and wear looks different on a young roof than an old one. Treat each flag as a place to look, and let a person decide what the mark is and whether it matters.

Do I need my own drone for AI roof inspection?

Not always. Some tools accept phone or camera photos, and one drone platform's page says you can order flights from a service network instead of flying yourself. Check each vendor's photo guidance, the rules that apply to flying near homes, and who is responsible for a flight, before you choose a route.

How many photos make a fair trial?

Aim for 40 to 60 per roof, covering every slope with close-ups and wide shots, taken on dry days in even light. Three roofs, one damaged, one clean and one old, give 120 to 180 photos. That is enough to see the pattern of false flags and misses without drowning in review.

Who is responsible if the tool misses damage?

Assume you are. The pages I read do not explain how accuracy or liability is handled, and the report carries your name. Keep a named person's sign-off on every report, keep the original photos, and ask your insurer and a qualified adviser how tool-assisted inspections sit with your cover and contracts.

Further reads

• Construction Photo Documentation App: What AI Adds, and Which Plan — What AI adds to site photo apps, and which plan covers it.
• Roofing RAMS Template: AI Drafts, a Competent Person Signs It Off — Who signs off risk assessments when AI drafts them.
• Which AI Tools Do Roofers Need? Measurements, Calls, Estimates — Which AI tools a roofer needs, from measurements to calls.
• Roofing Storm Leads: Triage and Follow Up After Hail With AI — Triage the inspection requests that follow a hail storm.
• AI Tools for Roofing Contractors in 2026: Estimating and Leads — AI tools for roofers across estimating and leads.
• Roofing Takeoff With AI: Squares, Waste and Bundles From a Report — Squares, waste and bundles from a measurement report.
• AI Home Inspection Software: What It Writes, Who Signs and What It Costs — What AI report tools actually write for a home inspector, why the inspector still owns every line, and the monthly cost per inspection at three volumes.
• ChatGPT for Roofing: 10 Prompts for Estimates, Inspections and Warranty — Ten ChatGPT prompts for a roofing company's inspections, estimates and warranty replies, each with a filled example and a check before it is sent.
• Free AI Tools for Land Surveyors: Drone Data, Research, Reports — Free tools for drone processing, deed research and report drafting, with each free tier's limits, what happens to client files and a closure-check example.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Roof Hawk, ScanShield AI, Roof Genius AI, RoofV and DroneDeploy product pages, all read in November 2026.