What 20,537 Field Photos Taught an AI About Restoration Scope

There’s a difference between an AI that has read the internet and an AI that has seen 20,537 restoration field photos. The first can describe water damage in general terms. The second has seen what a Cat 2 loss in a 1980s tract home actually looks like at hour six, what gets missed on a pad-only carpet pull, and how a containment line reads in a real basement. Domain matters, and in restoration scoping it matters more than almost anywhere, because the cost of a generic answer is a thin scope and a fight with the desk adjuster.

Why generic models write thin scopes

A general-purpose model will give you a plausible restoration scope. It will also quietly leave items off, because it’s pattern-matching against everything it’s ever read rather than against what restoration files actually contain. It doesn’t know your region’s price list quirks, the line items carriers in your market fight, or the supplement that’s earned but routinely missed. Plausible isn’t the same as defensible, and a scope that isn’t defensible is a supplement you won’t collect. The broader picture of what AI is really doing in restoration is in AI in restoration: what’s real and what’s theater.

An AI trained on the open internet knows what water damage is. An AI trained on 200 real restoration projects and 20,537 field photos knows what your file is missing.

What real restoration data teaches a model

Training on 200 real projects and 20,537 field photos isn’t about volume for its own sake — it’s about teaching the model the patterns that only exist in actual restoration work. The relationship between a photo of standing water and the extraction, drying, and antimicrobial line items it implies. The way a sketch and a set of moisture readings should map to a scope. The line items that travel together, and the ones that get forgotten under deadline. That’s the difference between a draft you have to rewrite and a draft you QA and submit.

It’s also why the output is IICRC-referenced rather than just confident. A scope that cites S500 and the photos it read from is a scope your estimator can stand behind and your desk adjuster can’t easily wave off. That’s scope defense built into the draft — the same discipline covered in TPA management and scope defense.

The human stays in the loop

None of this replaces the estimator. The model drafts; the estimator reviews, corrects, and submits. Every correction your team makes is signal, and on a full engagement the engine gets tuned on your shop’s own approved scopes, so it learns your market and your standards over time. The point isn’t a black box that writes scopes you can’t see — it’s a fast, defensible first draft from a model that has actually seen the work.

Where to find the engine

The engine is R360 Scope. Feed it the CompanyCam photos and Xactimate sketch your crew already shoots; it hands back an IICRC-referenced scope narrative and a working ESX in minutes. It’s available standalone and separately priced at r360scope.com, and it ships inside a full engagement tuned on your own files. Same engine, two ways to run it.

What to do Monday

Take one recent file you scoped from scratch and ask what a domain-trained draft would have changed: the items you almost missed, the narrative you re-typed, the hour you spent assembling instead of judging. That gap — assembly you shouldn’t be doing by hand — is exactly what an engine trained on real restoration work is for.