A custom cabinetry company was reading dense supplier order PDFs and hand-entering the real line items before a job could even start, with no fast way to see where anything was in the warehouse. We built a job app that turns an order PDF into a clean item list with AI and tracks every piece by QR scan from received to delivered.
The problem
A regional custom cabinetry company runs dozens of jobs at a time, each moving from order, to the cabinet supplier, into the warehouse, onto a truck, and out to the customer. The work got done, but two parts of it ran on manual effort and memory: getting supplier orders into a usable form, and knowing where every item was at any given moment.
Cabinet suppliers send order confirmations as dense, multi-page PDFs. A single order can run 40 to 60 lines, and only some of those lines are actual cabinets. Mixed in are modifications attached to a cabinet (increase depth, finish a side, remove the toe kick), reference lines for door style and wood species that aren't separate items, and non-plan or $0 lines that exist for the supplier's own accounting. Before this project, a person read through all of it and hand-entered the real items to build the job, sorting signal from noise on every order. It was slow, easy to get wrong, and it had to happen before warehouse tracking could even start. On the floor, basic questions, what arrived today, what's staged, what got loaded, what's out for delivery, lived on spreadsheets, sticky notes, and in people's heads.
What we built
One app with two connected halves: AI order intake, and QR-based warehouse tracking. The financial side, quotes, invoices, and purchase orders, links in from their existing accounting system, so the job record shows the money and the physical items in one place.
For order intake, you upload the supplier PDF and the system builds the job's item list:
- It reads the whole document, including the messy multi-line formatting that varies from one supplier to the next.
- It pulls out the real cabinet line items and lists them as job items.
- It attaches modifications to the cabinet they belong to instead of cluttering the list with them as standalone entries.
- It skips the reference and $0 lines that aren't physical items.
- Anything it can't confidently sort gets set aside and flagged, so a person glances at a short exception list rather than re-reading the whole order.
The person reviews the result, makes any edits, and owns the list from there. What used to be 45 minutes of reading and typing becomes a few minutes of checking. The design choice that made it stick: it fits how the team already works. Nobody had to change how they place orders or what suppliers send. The automation absorbs the existing format instead of asking people to adopt a new one, which is usually where this kind of tool stalls.
Once a job's items exist, each one gets a QR label staff print from the browser. Warehouse staff scan an item to update its status as it moves: ordered, received, ready to ship, on truck, delivered. Scanning any item pulls up the full job and every other item on it with its current status, all from a phone or tablet on the floor. The result is a live picture of where every piece of a job is, without anyone walking the warehouse to find out.
Where AI earned its place
Reading a dense, inconsistent order form and pulling the real items out of it is exactly what AI is good at, and exactly what breaks rigid templates. We used it there and nowhere it wasn't needed. The parsing itself costs a few cents per order, which is nothing against the labor it replaces.
What it's worth
These are estimates based on a mid-volume shop. Real numbers vary with order volume and labor rates, but the model is conservative.
At 6 to 10 jobs a week, manual entry and verification ran 45 to 60 minutes a job. Reviewing the AI-generated list runs 5 to 10. That nets roughly 35 to 50 minutes saved per job, or about 250 to 320 hours a year, worth somewhere around $7,000 to $10,000 in recovered labor at a mid-volume shop, and more as volume climbs. Catching even one order problem a month before it reaches a truck avoids a re-delivery or an install-day scramble, which adds a few thousand more a year in avoided rework. On the intake savings alone, a fixed-bid build in the low single thousands pays for itself in three to five months.
It worked because it removed the worst manual task without asking anyone to change their habits, gave the warehouse a live view of every job for the first time, and used AI where AI is genuinely good, reading messy documents and pulling structure out of them, rather than bolting it on for show.
Drawn from real engagements. Details changed to protect client identity.