Stock counted from a photo of the shelf
A photo of the shelf becomes next week's order, waste log and stock count before you finish your coffee.
The problem
Counts happen on clipboards at 6am, often by the person who also needs to open the kitchen. Numbers get guessed, waste is scribbled on a napkin and pars live in someone's head. By Wednesday you are out of basil, holding thirty too many burger buns and nobody knows why.
What AI makes possible now
A kitchen hand takes three photos of the dry store, walk-in and prep bench with a phone. The system identifies each SKU from its label, can or packaging shape, tallies visible units and compares the total against the stored par for that day of the week. It cross-references yesterday's sales from the POS and writes a draft order for the supplier, flagging any waste where the count went down by more than usage can explain. A manager reviews the flagged lines, taps approve and the order sends.
How it works
- Staff photograph shelves and prep areas using a dedicated camera app that timestamps and geotags each capture.
- Object detection identifies products via label text, barcode, can shape or consistent shelf position and counts visible units.
- Counts are compared against stored weekday-specific pars and recent POS depletion rates to generate a draft supplier order.
- Discrepancies where stock fell faster than sales suggest are logged as probable waste and surfaced for manager review with the original photo.
The first thirty days
In the first thirty days three shelf zones are mapped, recognition works for the top forty SKUs by volume and a daily email lands in the manager's inbox with a draft order and a short waste-flag list for approval. No POS integration yet; depletion is estimated from a flat average.
How it earns
The value sits in three pockets: labour hours clawed back from manual counts, waste identified quickly enough to fix a portioning error tomorrow instead of next month, and stockouts avoided because the reorder suggestion knows you sell more ciabatta on Friday than Tuesday. A single site breaks even when it prevents one missed delivery run per fortnight.
Why now
Off-the-shelf vision models can distinguish fifty common food-service pack formats without custom training, and supplier catalogues are finally available as structured data feeds rather than PDFs.
First customers
Multi-site cafe groups and small supermarket chains where a district manager currently drives between locations just to check the cold room.
The hard part
Packaging changes break recognition until the model is retrained, and a badly lit photo of identical white sauce tubs can still confuse count tallies. This needs a short human verification loop baked in, not removed.
Build this with us
We map your shelf layout, train recognition on your actual SKU range and build the photo-to-order pipeline with the supplier feed wired in. The manager still signs off the order each morning, but the clipboard count is gone. If you want to stop guessing stock at dawn, apply to build it with us.