> ## Documentation Index > Fetch the complete documentation index at: https://docs.x.build/llms.txt > Use this file to discover all available pages before exploring further. # How the AI works > What happens between your first message and a finished, priced proposal. ## What the AI does with what you give it Every XBuild project is a chat: describe the job as you would to a coworker, hand over photos, measurement reports, or a carrier estimate, and the AI turns them into a scoped, priced proposal. This page covers what it reads, what it asks, what steers it, and why the result is yours to review before it goes out. ### 1. You describe the job Start in plain language: the trade, the property, what's damaged or being sold, and any quantities or materials you already know. The AI works in two modes — **retail** (a homeowner proposal priced from your packages and rates) is the default, and it switches to **insurance** (a scope in Xactimate or Symbility format) when you mention a claim or a carrier, or upload a carrier estimate. An explicit statement from you — a measurement, a material, a price — overrides every other source the AI has. ### 2. Photos, documents, and voice notes | Input | What the AI does with it | | --------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Photos** | Scanned on arrival; each gets an AI title and [description](/features/photo-descriptions) and is read for scope — an extra shingle layer, a chimney, skylights, a second structure. Treated as the weakest evidence (unreliable for counting and for identifying materials at a distance), so the AI asks when a photo is the only basis for a call. | | **Documents** | Read automatically. Measurement-report values flow into the scope exactly as reported, including the report's own suggested waste factor. A carrier estimate is digitized in full: claim header, folder structure, and every line item with its pricing. | | **Voice notes** | Transcribed, so talking through a job on site works like typing it. | | **Videos** | Never scanned. They play on the digital proposal but don't feed the estimate. | Chat accepts PDFs, images, and measurement or estimate files — Xactimate `.esx`, roof measurement `.xml` from EagleView and similar vendors, and `.skx` and `.fml` sketches — whose measurements and line items flow into the scope like a report's. Accepted types and the per-upload limit are in [Creating an estimate](/features/creating-an-estimate). ### 3. It asks before it guesses Every quantity traces to a source: a measurement report, a carrier estimate's line, or a number you gave. When the AI can't name the source, it asks a clarifying question instead of inventing one — expect a short back-and-forth when information is thin. Prices likewise: it never fabricates one, and it labels where each came from. ### 4. How the scope gets built **Retail** — the AI selects from your [component packages](/features/component-packages), your reusable assemblies of material, labor, and other items drawn from [Your Catalog](/features/your-catalog). It applies each item's quantity formula to the measurements, keeps your prices exactly as defined, and applies your margin. For trades you haven't packaged, it builds the items from trade knowledge and prices them from market data, flagged as such. **Insurance** — the AI builds or supplements a scope in the carrier's format ([Xactimate](/features/publishing-xactimate) or [Symbility](/features/publishing-symbility)). Carrier lines are copied verbatim — codes, descriptions, and quantities — and anything new or adjusted is flagged so an adjuster can see exactly where the changes are. See the [insurance supplement flow](/features/insurance-supplement-flow). ### 5. What steers the estimate | Control | What it sets | | ---------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Construction logic** on packages and items | Plain-language business rules — waste factors, labor ratios, conditional inclusions. Item-level instructions beat package-level ones. See [Component packages](/features/component-packages). | | **[Your Catalog](/features/your-catalog)** | Every item the AI can put on a retail estimate, with its unit, price, and price source. | | **[Estimate preferences](/features/estimate-preferences)** | Standing company-wide rules — what you always include, brands you install, how you treat carrier quantities — applied wherever a package, item, or layout doesn't say otherwise. | | **[Pricing & Profit](/features/pricing-and-profit)** | Labor rates, default margins, default pricing, and taxes; every estimate is priced from them. | | **[Proposal layouts](/features/proposal-layouts)** | The document's structure. A layout's instructions are followed strictly. | | **Chat** | Preferences you state and facts you provide apply immediately, and an explicit request in the conversation overrides everything else. | The better tuned these are, the less correcting you do per job. ### 6. A versioned proposal you can edit The proposal renders live in the preview. Every save creates a new **version** — versions are never overwritten. Ask for changes in chat, or click the pencil icon to edit fields directly; either way, saving cuts a new draft version. When it's right, **Publish** generates the link and PDF. See [Editing estimates and proposals](/features/editing-estimates-and-proposals) and [Publishing your proposal](/features/publishing-your-proposal). ### 7. Review before you send The AI is a fast, well-briefed estimator, not a replacement for your judgment. Photos can hide layers, market prices are estimates, and only you know what you agreed with the customer — check the scope and quantities against what you saw on site before you publish. The boundaries are listed in [What XBuild can and can't do](/features/what-xbuild-can-and-cant-do). AI work runs on tokens: every turn consumes some, and more files and revision rounds cost more, so a tight job description up front is the cheapest path to a finished proposal. See [How tokens work](/plans-billing/how-tokens-work). For further assistance, contact XBuild Support at [support@x.build](mailto:support@x.build).