Two very different businesses — grocery retail and auto parts — ended up with the same system shape: a person points a phone at the real world, and structured data comes out the other end. The photo replaced the form.
Forms are where field data goes to die
Field work generates the most valuable operational data a company has, and captures it in the worst possible way: a tired person, standing up, thumb-typing into dropdowns designed at a desk. The result is late, thin and wrong. Merchandisers audited shelves store by store, by hand; parts staff spent twenty minutes typing per listing. Both teams already photographed everything anyway — the photos just went nowhere.
What happens after the shutter
The pattern that works is a pipeline, not a chatbot:
- Extract — a vision model reads the photo: which products are on the shelf, what part is on the bench, what the label says.
- Structure — the reading is forced into a schema: SKUs, facings, part numbers, condition. No free text.
- Validate — the structured result is checked against what is known: the planogram, the parts catalog, price bounds. Anomalies go to a person.
- Deliver — the output lands where work happens: a compliance report, a ready-to-publish listing.
What it changed
The retail chain saw +11% on-shelf availability — audits that happen every visit catch gaps that quarterly audits never did. The parts seller cut listing time by 65%, which in practice meant inventory got listed at all instead of waiting for a quiet afternoon.
Where the pattern fits — and where it doesn’t
It fits wherever the truth is visible: shelves, vehicles, meters, sites, damage, documents. It does not fit where the truth is a judgment call a camera cannot see, or where a wrong reading is expensive and hard to catch — the same readiness rules as any automation. Start where a person already takes the photo.
