A budget is only useful if it looks like your company built it. When your Claude or ChatGPT turns a Biltum Project Packet into a draft, the first thing it does is not write. It reads: how your company structures, names, and prices work in JobTread. Only then does it build the job.
Two kinds of evidence, two jobs
The Project Packet carries the job: the measured floor plan and 3D scan, your photos with a map of when each was taken, the walkthrough video and timestamped frames from it, and a timestamped transcript of everything you said. Your JobTread account carries the company: your budgets, templates, catalog, units, formulas, and custom fields.
The packet’s instructions keep those roles apart. Your narration and the measurements decide what is in the job. Your company’s records decide how it’s built and priced. A general construction assumption never outranks either.
What your assistant reads first
- A company profile, if Biltum has one. When Biltum has already learned your company’s format, the packet includes it as a starting map: your usual structure, templates, and gaps. Your assistant still checks it against live JobTread records, and live records win.
- Your recent budgets and templates for comparable work — contractor-authored ones, not test jobs.
- The exact catalog and field records those budgets use: items, units, formulas, labor inputs, selections, and custom fields.
It reads only what it needs for this job, and expands where a convention that matters is still unclear. It records what it learned in a short company brief inside the chat, so the reasoning is visible.
How it builds the draft
The instructions set an order that mirrors how an estimator works: settle the structure and line items first, then measure, then calculate.
| Part of the draft | Where it comes from |
|---|---|
| Group structure and order | Your applicable budgets and templates — phase-first or area-first, at the depth you normally use |
| Line items | The work you narrated and photographed, matched to the exact catalog items your company already uses, with their native descriptions |
| Quantities | The scan’s measurements for the selected work, run through your own formulas and units |
| Prices | Your catalog and current rates — never invented, never a national average |
| Defaults you always include | Supported company defaults even when you didn’t say them on the walk; anything you explicitly excluded stays out |
What it does with uncertainty
Walkthroughs leave things open. The instructions don’t let your assistant stall on them or guess silently. If two materially different readings are still supported after it checks the evidence and your company’s practice, it writes the fuller requested scope with all its dependent work, and marks one row per decision REVIEW -. It never charges both sides of an either/or, and never adds rooms, trades, or upgrades you didn’t ask for. Ordinary defaults and defensible measurements don’t get flagged, so the REVIEW rows are the ones worth your time.
Writing into JobTread
Your assistant writes only because you asked it to in the chat; the packet’s instructions alone grant no write permission. Before writing, it rereads the job’s current budget so it doesn’t duplicate or overwrite existing work or your edits. It writes parent groups before their items, keeps your field values and formulas intact, and doesn’t create permanent catalog items or change company settings. Then it reads the written rows back to verify them and asks the few questions that matter.
What if you’re new to JobTread?
A company with little budget history has less to learn from, and the instructions don’t pretend otherwise. Your assistant starts from the templates and catalog you have, uses a sensible standard structure where yours is silent, and tells you what it couldn’t confirm. The draft sharpens as your history grows.
The practical payoff
The measure is what your estimator does when the draft arrives. If the answer is “rebuild it in our format,” it failed no matter how fast it was. If the answer is “check the REVIEW rows and refine the numbers,” the walkthrough just became the first draft of the budget. The cheapest way to find out is to install the app, walk one job, and read the draft that lands.
Source note: Behavior described here follows the budget-writing instructions shipped in every Biltum Project Packet (version 0.4.3), as verified on October 15, 2026. JobTread’s AI Connector is a JobTread product; see JobTread’s documentation.
Related: the step-by-step tutorial, how one walkthrough becomes a Scope of Work, and how accurate iPhone LiDAR measurements are.
