What AI Shelf Monitoring Looks Like After 2,000+ Real Store Visits

Most AI retail-execution demos are staged: a clean sample shelf, a scripted upload, a slide with round numbers. This isn’t that. It’s a walkthrough of a dashboard Interact runs today, live, for an active FMCG distribution client in Malaysia — real outlets, real shelf photos, real facing counts, updated daily. The client’s name is withheld at their request; everything else here reflects what the system is actually doing right now.

Interact Technology · Case Study · September 2026

1. The question every FMCG team asks and can’t answer fast

Brands and distributors spend heavily to get products onto shelves. What’s much harder to buy is an honest, current answer to a simpler question: what’s actually on the shelf right now, across every outlet that carries it? Historically that visibility has depended on manual audits — a merchandiser’s paper checklist, a monthly sampling round, a spreadsheet compiled days after the store visit happened. By the time a gap shows up in a report, the shelf has usually already moved on.

That lag is the real problem AI image recognition is built to close — not by replacing the merchandiser, but by turning the photo they already take at every visit into a structured, immediate measurement instead of a filed record nobody reopens.

2. Inside a live production dashboard

What follows is a screenshot-free walkthrough, out of respect for the client’s privacy — but every figure below is drawn from a real 30-day window of an Interact dashboard running today for a multi-brand FMCG distributor operating across Malaysia, rounded to protect specifics.

• Scale: over 2,000 outlet visits captured in the past month alone, each producing a geotagged, timestamped shelf photograph.
• Breadth: well over a dozen product categories tracked in the same view, spanning everything from beverages and biscuits to personal care and household goods.
• Depth: dozens of the distributor’s own brands measured facing-by-facing against well over a hundred competing products stocked in the same categories.
• Granularity: every number drills down to the individual visit — outlet, product, brand, category and facing count — with the source photo attached to the record.

What moved over the last 30 days: this distributor’s own-brand share of shelf climbed from the high-20s into the low-40s (percentage points) within a single month. That isn’t a claim that Interact’s AI caused the shift — the distributor runs its own trade activity, not us. It’s a demonstration of what changes when a swing like that is visible in near real time, instead of surfacing weeks later in a manual audit report that has already gone stale.

3. From shelf photo to dashboard row in under three seconds

At the outlet, a merchandiser opens the app and photographs the assigned shelf section — the image is automatically tagged with outlet, category, GPS position and timestamp. Interact’s AI image recognition model then classifies every visible SKU in that frame and counts facings, cross-referencing the result against the brand’s must-stock list and the competitor set defined for that category. The reading lands in the dashboard within seconds — no batch upload, no manual tagging, no waiting for an overnight job.

The same capture drives more than one measurement from a single photo: Share of Shelf (own facings versus competitors’), On-Shelf Availability (a must-stock SKU that should be present but isn’t) and Pricing Compliance (shelf price against the agreed price) are all produced from the one shot, not three separate visits.

Why the speed matters: a gap report that lands three seconds after the photo is taken changes what a field team can actually do about it. A report that lands after the monthly review meeting is a record of what already happened.

4. What this proves for anyone evaluating the same thing

If you’re weighing whether AI shelf recognition is a real operating capability or still mostly roadmap slideware elsewhere, this is what a production system running at this scale actually looks like:

• It runs on outlets nobody controls: general trade and modern trade stores a merchandiser only visits — not a staged shelf under studio lighting.
• It measures against a real must-stock list: the specific SKUs and facing targets a brand has actually agreed with the retailer, not a generic catalogue.
• It reads competitors automatically: the same photo that measures your facings measures theirs, with no separate data-entry step.
• It’s a system of record, not a one-off report: every visit stays queryable months later, by outlet, product, brand or category.

5. See it for yourself

Everything above has been described in aggregate. Here are two of the actual shelf photos behind those numbers, exactly as the system produced them — the only edit is cropping out the outlet name and visit timestamp that the source system overlays, which identify a real store and don’t belong in a public example. Every bounding box, SKU label and price tag below is the AI’s own output, not a mockup of one.

What’s not edited here: the bounding boxes, labels and price tags are the AI’s actual output on these actual shelves, on an ordinary Tuesday, at outlets Interact doesn’t operate. The only thing removed is the outlet name and timestamp the source system attaches to every capture — identifying detail a real store doesn’t need to appear in a public example.

What’s available now

This case study describes a live Interact deployment for an active FMCG distribution client in Malaysia. The client’s name and any store-level identifying detail have been withheld at their request. Scale figures (visits, categories, brands, share-of-shelf percentages) are rounded from the underlying production data to preserve confidentiality while remaining directionally accurate. Figures 1–3 are redrawn illustrations of the live dashboard’s structure, not screenshots; category names are real, values are rounded. Figures 4–5 are real, unedited AI-annotated shelf photos from the live system, with only the outlet name and timestamp cropped out.