How AI Shelf Photo Analysis Actually Works in FMCG Merchandising (And Where It Still Fails)

By Sufyan · 2026-07-28 · 5 min read

Last month a merchandising head at a dairy company in Dubai showed me his phone. 847 shelf photos taken by his team in a single week. He'd reviewed maybe 40 of them. The rest just sat there, taking up storage, proving nothing.

That's the dirty secret of shelf photos in FMCG. Everyone takes them. Almost nobody looks at them.

And this is exactly why AI shelf photo analysis stopped being a nice-to-have around 2023 and became something sales ops leaders are actively budgeting for in 2025. But there's a lot of nonsense being sold in this category, so let me walk through what's actually happening — the good parts, the parts vendors don't tell you, and how to think about it if you're evaluating tools this quarter.

What AI shelf photo analysis actually does

At its simplest: a field rep snaps a photo of a shelf. The AI reads the image and tells you what SKUs are on the shelf, how much space each brand is taking, whether your planogram is being followed, and whether competitors have stolen your facings.

That's the pitch. The reality is more interesting.

Good systems (and honestly, there are only about four or five worth naming globally) can identify individual SKUs at roughly 92-96% accuracy in decent lighting. They calculate share of shelf, spot out-of-stock situations, catch price tag mismatches, and flag when your premium SKU has been shoved to the bottom shelf behind a competitor's promo display.

The rep doesn't have to do anything except take the photo. No manual counting. No filling out a 40-field form. Just click, upload, done.

And this is where the ROI gets interesting. A merchandiser in Riyadh visiting 25 outlets a day used to spend 6-8 minutes per outlet doing manual compliance checks. With AI analysis running in the background, that drops to under 2 minutes. Multiply that across a 60-person team and you've bought back about 180 field hours a week.

Where it still breaks (nobody tells you this part)

Honestly? I used to oversell this capability in early Zivni demos. Then customers pushed back and I had to get more careful.

Here's what still trips up even the best AI shelf recognition:

Cluttered traditional trade shelves. A modern trade shelf in Carrefour or Tesco is relatively clean. An FMCG shelf in a kiryana store in Karachi or a baqala in Muscat? Products stacked three-deep, half the labels facing sideways, cardboard displays blocking the view. Accuracy drops to maybe 70-78% here. Still useful, but not magic.

New SKU launches. Every time you launch a new variant, the AI has to be retrained or at least fed reference images. Some vendors do this in 48 hours. Others take three weeks. Ask this question specifically during demos.

Lighting and angles. Reps in a hurry take bad photos. A rep in Sharjah in a poorly-lit grocery at 8pm is going to produce images the AI can barely parse. Look, no vendor will admit this openly, but roughly 8-12% of real-world field photos need re-shoots.

Beverage bottles that look identical. Two SKUs of the same drink in 500ml vs 600ml? The AI will confuse them until you specifically train it. Same with flavor variants that only differ by a small color band on the label.

I'd rather you know this going in than get sold a fantasy.

What sales ops leaders should actually measure

If you're rolling out retail execution software with AI shelf analysis in 2025, don't just measure photo counts. That's a vanity metric.

Track these instead:

One of our customers in Pakistan — a snacks distributor with about 340 outlets across Lahore and Faisalabad — went from a 61% planogram compliance rate to 84% in four months. Not because the AI was magical. Because their supervisors finally had objective data to have real conversations with reps. Before, it was "you're not doing your job." After, it was "here are the 7 outlets where your compliance is under 50% — walk me through what's happening."

That's the actual transformation. Not the AI. The conversations the AI enables.

How to buy it without getting burned

A few practical things I'd tell any sales ops leader evaluating AI-powered field sales tools right now:

Ask to see the AI running on YOUR shelf photos, not the vendor's polished demo images. Send them 30 real photos from your reps. See what comes back.

Ask how new SKUs get added. If the answer involves "contact support and wait," that's a problem.

Ask about offline mode. Reps in Oman driving between towns lose signal. The photo analysis should queue and sync, not fail.

Ask what happens when accuracy is low. Does the system flag uncertain results for human review, or does it just guess and pretend it knows?

And finally — and this is the one people forget — ask whether the AI shelf photo analysis is bundled or a paid add-on. At Zivni we keep the base field sales automation at $5/user and layer the AI as a modular add-on because not every customer needs it on day one. Some vendors force it into the base package and inflate the whole license. Fine if you'll use it. Wasteful if you won't.

So where does that leave you if you're planning your 2025 field sales tech stack? Probably somewhere between excited and skeptical, which is honestly the right place to be.

What's the compliance rate you're actually seeing in your general trade channel right now — and do you trust the number?

© 2026 Zivni. All rights reserved.