Perfect store programs die in the last mile of data

Perfect store programs rarely fail at the camera. Shelf recognition now reads a photo in under 90 seconds at better than 95% accuracy. They fail in the pipeline behind it. The score crawls through a vendor, a warehouse and a dashboard before anyone acts, so the average shelf gap is fixed four to seven days later on the next visit. By then the promotion has ended and the rep is three routes away.

Why does a shelf photo take a week to change anything?

The standard perfect store stack is a relay race with a slow baton. The rep app captures the shelf, uploads it to an image recognition vendor, which returns SKU-level detections on a batch service level. The warehouse ingests those detections overnight, the BI layer refreshes them into a dashboard, and the numbers surface at the weekly or monthly commercial review. Every hop is engineered for reporting, not intervention. A vendor batch that returns overnight is fast by reporting standards and useless by shelf standards. Nobody in that chain is holding a task that reads go back to outlet 4471 and face the cooler before the weekend.

The first hop already leaks. The rep app assumes the photo uploads while the rep is still standing in front of the shelf, but across much of the fragmented trade the connection is simply not there, the problem we picked apart in if the app needs 4G, the fragmented trade doesn't have it. So the image sits in an outbox, syncs hours later from a depot, and the clock on the entire pipeline only starts once the rep has already driven away.

The recognition isn't the bottleneck

It is tempting to blame the model, but the accuracy question was settled a while ago. Vendor and independent benchmarks now put shelf recognition above 95% accuracy in real store conditions, with reported planogram-placement gains around 40% and 15 to 25% higher sales per square foot where compliance holds. The camera does its job. The failure is that a correct answer delivered late is operationally worthless. A perfectly scored cooler is still warm, still half empty, still missing the promotional facing the moment the dashboard finally lights up.

The execution numbers make the point in a different register. Roughly 90% of companies fail to deliver their in-store promotional plan, and fewer than half of consumer-goods leaders say their merchandising plans are executed as intended. Those are not detection failures. Recognition tells you the plan slipped; it does nothing about the slippage. You cannot photograph your way out of a latency problem, and the model is now the cheapest, most reliable component in the whole chain.

Where do the days actually go?

Detection speed and correction speed are different clocks, and the perfect store stack optimises the wrong one. Vision Group Retail's own breakdown of correction windows is worth reading literally:

The conventional photo-recognition programme, for all its sophistication, lives in the top row. It routes a machine-precise score through the slowest correction path on the board, four to seven days to reach the same shelf. The bottom row is the identical sensor wired to a different clock. Nothing about the hardware or the model changes between them; only the plumbing does.

Why the dashboard measures a store that no longer exists

By the time the commercial review convenes, the shelf has moved on. Global grocery runs an 8.3% out-of-stock rate, and between a quarter and 60% of those gaps are product that is already in the building, in the backroom or on a pallet, that never makes it to the shelf. That is precisely the failure a rep standing in the aisle can fix inside the same visit and a dashboard, three days later, cannot. The stakes are not marginal: roughly a three-point gain in on-shelf availability tracks to a one-point gain in sales for a manufacturer, so every day of lag is measurable revenue left on the pallet.

There is a second failure hiding under the timing one: the score, the order and the delivery live in three different systems. The recognition result sits in the vendor's portal, the reorder in the distributor's DMS, the physical fix on a truck governed by the ERP. That fragmentation is the same one we traced in your distributors know your customers better than you do, and it is why a corrective task cannot simply ride the next drop today. When the fix does travel physically, it belongs on the vehicle already heading there, the same logic behind the truck that delivers full and returns empty-handed.

So where does the fix actually live?

The latency is architectural, not analytical. Score the photo where it is taken, on the same event bus that already carries the order and the delivery, and emit a corrective task before the visit closes: a re-face for this rep now, a missing SKU appended to the next drop. That is a plumbing decision far more than a data-science one, and it is the shape of the single event bus we build these suites on, where a shelf score is an event that triggers work rather than a row that waits for a meeting.

The camera race is effectively over; the next few years of retail execution will be won on loop-closure time, not model accuracy. As recognition moves onto the device and scores land in the same second the shutter clicks, the meaningful metric stops being how well you saw the shelf and becomes how much of it you fixed before the rep pulled out of the car park. Programmes that keep routing that answer through a warehouse will keep paying for a high-resolution photograph of a problem they could have solved on the spot.