AI Shelf Photo Analysis in FMCG: What Actually Works (and What's Marketing Fluff)
A merchandiser in Riyadh once sent me 47 photos of a single Panda Retail aisle in a WhatsApp group. The regional manager wanted "proof of execution." Nobody looked at them. Nobody had time.
That's the problem AI shelf photo analysis is supposed to solve. And honestly? It does — most of the time. But there's a lot of nonsense being sold right now, so let me walk you through what's real and what isn't.
What AI shelf photo analysis actually does
A rep walks into a store. Takes a photo of the shelf. The software then identifies every SKU, checks share of shelf, flags out-of-stocks, compares against the planogram, and scores the execution. All in about 8–12 seconds on a decent 4G connection.
That's the pitch. Here's what it looks like in practice at a real distributor we work with in Karachi: their reps used to spend 4 minutes per outlet manually ticking SKU availability on a checklist. Now it's one photo. The AI does the counting. The rep moves on.
Multiply that by 35 outlets a day. That's roughly 2 hours of admin time recovered per rep, per day. For a 60-person field team, you're looking back at 120 hours a day you didn't have before.
But — and this is the part nobody tells you — the AI is only as good as the training data behind it. If your product SKU changed packaging three months ago and the model wasn't retrained, it'll misidentify things. We had this happen with a laundry detergent client in Sharjah. New pack design, 30% error rate for two weeks until we retrained. Honest truth: any vendor telling you their AI is "99% accurate out of the box" for your specific SKUs is lying, or hasn't actually deployed at scale.
The four things it's genuinely good at
Share of shelf. This one's a slam dunk. Instead of a rep guessing "about 40%," the AI measures actual linear space per brand. When a modern trade key account manager is negotiating with LuLu or Carrefour, having hard numbers changes the conversation.
Out-of-stock detection. Empty facings, gaps, missing SKUs. The AI catches these fast. We've seen clients cut OOS incidents by 22–34% within the first quarter just because gaps get flagged in real time and reps can't "forget" to report them.
Planogram compliance. This is where planogram compliance AI earns its keep. You upload the agreed planogram, the rep takes a photo, and the system scores compliance percentage. No debate. No "the store manager rearranged it." Just a score, a photo, and a timestamp.
Competitor tracking. Look, this is quietly the most valuable feature for most brand teams. Your reps are already in the store. They might as well capture what Unilever or P&G is doing next to your product. New SKU launch? Price change? Promo? The photos build a competitive intel database without anyone doing extra work.
Where it falls short (the honest bit)
I got this wrong at first. Early on at Zivni I assumed AI shelf photo analysis would replace merchandiser judgment. It doesn't. It replaces the counting, not the thinking.
A few things the AI still can't do well:
- Poor lighting in traditional trade kiryana stores messes with recognition. A dim shop in Old Lahore is not the same input as a Waitrose aisle in Manchester.
- Products stacked awkwardly (chips bags, for example) confuse the model.
- If your SKUs look almost identical — same brand, different flavors, similar packaging — accuracy drops. Think shampoo variants or spice packets.
- It won't tell you why the shelf looks the way it does. That's still the rep's job.
So the pitch of "replace your merchandisers with AI" is nonsense. What actually works is giving your existing merchandisers superpowers so they spend time on relationships and problem-solving instead of clipboards.
What to actually look for when buying
If you're evaluating an FMCG merchandising software with shelf AI baked in — whether it's Zivni, FieldAssist, BeatRoute, or Repsly — ask these five questions, in this order:
- How long does it take to train the model on my SKUs, and who pays for retraining when packaging changes?
- What's the offline mode like? Because half of Oman's traditional trade has patchy signal, and I don't want photos lost.
- Can I export raw compliance data into my BI tool or ERP? (If the answer is "through our dashboard only," walk away.)
- What happens when the AI is unsure — does it flag for human review, or just guess?
- Show me a live customer in my category. Not a case study PDF. An actual reference call.
The fifth one filters out 70% of vendors immediately.
A quick note on cost vs. value
A lot of sales ops leaders get stuck on per-user pricing. At $5/user/month base with modular add-ons like shelf AI, the math is easy for a 20+ rep team. What's harder to price is the decision quality improvement. When your national sales manager can see planogram compliance across 4,000 outlets in Saudi Arabia on a Monday morning instead of waiting for Friday's spreadsheet — that changes how fast the business can react.
And reacting fast in FMCG is basically the whole game.
One last thing before I close this laptop — if you're piloting shelf AI, don't roll it out to your whole field team on day one. Pick 8–10 reps in one city. Run it for six weeks. Fix the SKU library. Then scale. Every client who's tried the big-bang approach has regretted it, including two of ours who came back sheepishly after trying to skip the pilot phase.
What questions are you stuck on? Happy to talk through specifics if you're mid-evaluation.