Type "Xactimate AI" into a search bar and you get two different things wearing one name. There is XactAI, the AI add-on Verisk built into the Xactware products you already license. Then there is a growing pile of third-party tools that use AI to read claim documents. Both are real, and they are not the same job.
I work the claim side of this at ESXPress, where we turn carrier scope PDFs into import-ready Xactimate ESX files. Here is where AI actually helps on a carrier scope, where it makes the estimate worse, and the one rule we never break: the price printed on your PDF stays the price on your line.
Does Xactimate have AI? What XactAI actually is
Yes. Verisk ships XactAI, an AI layer that runs inside the Xactware products you already license. It bundles an AI Estimator in XactRestore and XactRemodel, a Line Item Advisor and Sketch Scan in Xactimate, automated photo labeling and descriptions, and XactRebuild. It is help that lives inside an estimate you are already writing, and Xactimate Pro bundles parts of it.
That framing matters more than the feature list. XactAI is not a product you point at a folder of carrier documents. It is assistance inside the estimating environment, which is a different job from reading a file an adjuster sent you.
Where the XactAI pieces run
- The AI Estimator runs in XactRestore and XactRemodel, not the carrier-facing Xactimate app.
- Line Item Advisor and Sketch Scan run inside Xactimate.
- Photo labeling, descriptions, and XactRebuild touch the wider Xactware set.
What XactAI does not do
XactAI is an assistant inside an estimate, not a converter. Verisk's own Line Item Advisor documentation notes it cannot supply every item needed for a described job, and it does not read a carrier PDF and hand you an ESX. It also does not read a scope the carrier already wrote and tell you what is missing from it. That last part is the job that decides whether a claim funds the work, and it is the reason we built our own reading pass. The completeness check is a good look at how that works.
Claude plus Xactimate: what the connector really is
In May 2026 Verisk connected XactAI to Anthropic's Claude through a Model Context Protocol connector. The detail a lot of coverage skipped is that the connector runs against XactRestore, not the Xactimate desktop application. The headline "AI writes Xactimate estimates" overstates what shipped.
None of that makes the connector useless. It makes it specific. A tool that helps build an estimate is not the same as a tool that reads the scope an adjuster already sent, pulls its line items into structure, and flags the code-required items that never made it onto the page.
The three reading jobs AI does well on a carrier scope
AI earns its keep by reading. Deciding stays with you. On a carrier scope the reading breaks into three jobs: turning a scan or image-only page into text, pulling line items, quantities, and unit prices into a structured list, and finding rooms on a drawing. Each one produces a list you can hold against the source page. AI in roofing claims covers the wider picture of where these tools fit.
Reading scopes with no text layer
Plenty of scope PDFs arrive as scans, faxes, or photos of a printout. A vision pass reads those pages the way your eyes do, before parsing starts. Skip that step and the extractor sees a blank page.
Extracting line items into a structure
Parsing pulls the code, description, quantity, and unit price off each line into a consistent shape. This is mechanical work, and a machine should do it instead of you.
Finding rooms on a drawing
When a scope includes a floor plan, room detection can find the rooms and export them as sketch data. Treat that as a starting point for a sketch, not a finished measurement.
Where AI must never decide: prices and codes
Two things AI must never invent in a claim estimate: a price and a product code. An estimate is evidence. At ESXPress the price on the PDF is the source of truth, every line keeps the unit price from the customer's own scope, and codes come from a lookup against a fixed reference rather than a model's best guess. A line we cannot match keeps its PDF price and gets flagged for review.
This is also why import behavior matters. When Xactimate reprices on import, the price list loaded on your machine sets the final numbers, and you want to understand that before you promise a number to a homeowner. The way a bid overlaps with overhead and profit deserves the same care, which is the subject of the overhead and profit trap.
How ESXPress uses AI, honestly
Our vision pass is DeepSeek-based. Vision reads the document, parsing pulls the line items, and a resolution engine matches product codes against a reference built from processed claim scopes. Vision reads scans and rescues hard pages. It is never the sole pricing judge.
Vision reads, parsing extracts, lookup resolves
You can follow the whole path on the PDF to ESX converter: a carrier scope goes in, a structured ESX comes out, and anything we could not resolve is marked rather than guessed. The Xactimate AI page walks the same pipeline in more depth, including the parts that stay human on purpose.
What the measured record shows
We ran a narrow study on 289 claim scopes. The question was simple: of the lines we extract, how many carry a code our lookup can resolve? The answer was 36.6%, which means those lines are eligible to reprice on import. The rest kept the price that came off the PDF. That ratio is worth knowing before you assume any tool, ours included, replaces a human read of the scope. About the data explains how we measure it.
What to expect from an AI estimating tool
Reasonable to expect: an image-only scope read correctly, lines pulled into a consistent structure, room detection from a drawing, an unknown code flagged instead of invented, and a total that sums the prices on your own document. Not reasonable: a tool that prices your work instead of using your price list, a guessed code, or a file that needs no review at all.
How to use AI on a claim without losing the review
- Convert the carrier scope so you are working from structured data instead of a re-typed PDF.
- Read the flags first. The unresolved lines are where a model would have guessed, and they deserve your attention.
- Check the completeness report against the job, not only the scope. Code-required items are the ones carriers leave out most often.
- Validate the file before import with the free ESX validator.
- Keep the human sign-off. AI reads. You decide.
The short version
AI is good at reading a carrier scope and bad at deciding what a claim is worth. Use it as a fast pair of eyes that hands you a structured list and a set of flags, then do the reviewing yourself. That is how it earns a place in an estimating workflow instead of eroding it.
The quickest way to see where this lands on your own files is to convert one carrier scope and read the completeness report on the output. The free plan is 5 conversions a month, no card required.