The setup was almost nothing
The video's claim up front: nobody has actually automated Xactimate estimating, and some companies that say they have are, per the video, sending your photos and notes offshore and piping the result back in as magic. Where the work happens between submit and delivery is the same question a vendor-managed estimating workflow raises. The experiment that follows is different in kind. It uses Claude Code, an agentic coding tool that runs in a terminal, plus two inputs: access to a CompanyCam project and one file of Xactimate codes.
The instruction was close to plain English. Navigate to the project, review the photos, use the code file, write the scope. The video shows the agent reasoning through the work on its own, going back to the CompanyCam photos after studying the codes, and asking surprisingly few qualifying questions along the way.
Graded against the human scope: 88 percent
After two more prompts and a couple of added details, the agent produced a full markdown list of line items, the target earlier photo-to-line-item work had been circling with trained models. The video then hands both versions to ChatGPT to grade: the agent's scope against the one written by the person who actually walked the loss.
The result came back 88 percent accurate on line items. The remaining 12 percent were not contradictions. They were job-specific expansions, items a person would have noted anyway if they had that information from the start. Per the video, that is a level of accuracy the builder never thought this approach would reach.
The missing 12 percent was information, not judgment.
The line the experiment refuses to cross
Could the agent push those line items into Xactimate itself? Yes, and the video is blunt about why it does not: that violates the end-user license agreement with Verisk. The line items come out for manual entry instead, which the video calls very close to automatic complete estimates without crossing that line.
That is the honest state of the experiment: scope generated from photos and a code file, judgment and entry still human. The production version of this idea runs through a legitimate file format instead. Axiom builds the room-organized draft your estimator reviews, then delivers a native ESX only when your team chooses to download it.
