AI Shelf Audit for FMCG Modern Trade: Photo to Planogram Score

The traditional shelf audit is a lie that everyone agrees to tell.

Your field executive visits a modern trade store, eyeballs the shelf, ticks boxes on a form that asks whether the planogram is compliant, and moves on. The form says "compliant." The shelf has three SKUs out of position, two facings missing, and a competitor brand eating 30% of the space your brand plan allocated. The next morning, the ops deck says planogram compliance is 84%.

Nobody is cheating. Everyone is approximating. The human eye can't hold a planogram specification for 40 SKUs in short-term memory while also counting facings, checking positioning, and noting competitor intrusion. So field executives estimate. And estimates aggregate into a compliance metric that has no relationship to actual shelf reality.

Computer-vision shelf audits were supposed to fix this. They're now mature enough in the Indian modern trade context to actually work — but only if you understand what they do well, where they break down, and what "implementation ready" actually looks like for your team.

How an AI shelf audit works end-to-end

The process has four stages:

Stage 1: Capture

The field executive photographs the shelf section with their phone. Good capture systems guide the photo: the app prompts the executive to photograph the full shelf bay, flags blurry images immediately, and requests re-capture if coverage is insufficient. The whole capture process for a standard modern trade shelf bay (4–6 shelves, 3–4 bays wide) takes 45–90 seconds.

A key design choice here: the capture should work offline. The photo queues locally and uploads when connectivity resumes. If capture requires a live API call, you've reintroduced the network dependency problem that kills field execution in Indian semi-urban contexts.

Stage 2: Detection and segmentation

The uploaded image is processed by a computer vision model that:

1. Identifies shelf structures (shelves, bays, sections)
2. Detects individual products within those structures
3. Segments each product into a bounding box
4. Classifies each detected product (which SKU is this?)

SKU classification is where the model quality matters most. A model trained on 10,000 images of your brand's SKUs can achieve 94–97% accuracy under good conditions. A generic model trying to handle all FMCG brands simultaneously drops to 75–85% in realistic conditions — which is too low for meaningful compliance tracking.

The implication: the best shelf audit implementations use brand-specific fine-tuned models, not generic retail AI. If a vendor tells you their model works for all brands out of the box with no training data, that's a red flag.

Stage 3: Planogram scoring

Once products are detected and classified, the system compares the shelf state against the planogram specification:

• Facing count — does each SKU have the correct number of facings?
• Position compliance — is each SKU in the correct shelf position relative to the planogram?
• Share of shelf — what percentage of visible shelf space does the brand occupy?
• Out-of-stock gaps — are there empty facings where product should be?
• Competitor intrusion — are competitor SKUs present in space allocated to the brand?

The output is a compliance score (typically 0–100) with a breakdown by SKU and by planogram element. The whole scoring computation happens in under 4 seconds for a standard bay on modern cloud inference infrastructure.

Stage 4: Action assignment

The score without an action is analytics theatre. Good systems use the shelf audit output to auto-generate field tasks: "3 facings of SKU-X missing — restock from back-room stock" gets assigned to the executive at that outlet. The task stays open until re-audit confirms resolution.

Where AI shelf audit breaks down

Poor lighting

Indian general trade stores, and even some modern trade stores in Tier 2–3 cities, have inconsistent lighting. Halogen lighting from unusual angles, fluorescent glare on glossy packs, and dark spots in lower shelves all degrade model accuracy meaningfully.

Mitigation: enforce a minimum exposure in the capture app (reject images below a brightness threshold), and calibrate model performance benchmarks against images taken in realistic (not ideal) lighting.

Partial occlusion

Products that are partially hidden — by a promotional stand, by another product in front, by the shelf label holder — are the hardest classification problem. A model that achieves 95% accuracy on clear shelf images may drop to 80% on heavy-occlusion images.

This is a solvable problem with richer training data (occlusion-heavy images deliberately included in training sets), but it's not solved by generic models.

Similar pack designs

In FMCG, pack design convergence within categories is common. Two biscuit brands in the same sub-segment may have nearly identical visual signatures at shelf image resolution. Misclassification between visually similar SKUs is a systematic accuracy problem.

Mitigation: focus model training on the visual delta between your SKUs and your nearest competitor look-alikes. This is finer-grained than general category training.

Pack design changes

A new pack design invalidates the training data for that SKU. If you change the pack design for a key SKU and the model hasn't been updated, that SKU will score as "unrecognised" — which looks like an out-of-stock to the system.

Good shelf audit systems have a re-training pipeline that can incorporate new pack designs in 7–14 days. Ask vendors specifically how new pack designs are handled and how quickly the model is updated.

What the next 12 months look like

The shift happening now is from single-shot classification models to multimodal reasoning models that can:

• Describe shelf state in natural language ("the biscuit section has three brands competing for the same eye-level space — Brand A has 4 facings, Brand B has 6, your brand has 2")
• Answer questions about the shelf without pre-built planogram data ("does this shelf look like it's been recently restocked?")
• Compare current shelf state to last week's state automatically

For Indian FMCG ops teams, the practical implication is that shelf audits are moving from a compliance measurement exercise to a real-time competitive intelligence feed. The same photo that generates a planogram score today will, within 12 months, also generate a competitor pricing capture and a shelf trend analysis — automatically.

The data flywheel matters enormously here: the more outlet-shelf photos your field team captures, the better the models get, the more valuable the competitive intelligence becomes. Teams that start capturing systematically now will have a 12–18 month data advantage over teams that wait for the technology to "mature."

Practical readiness checklist

Before deploying AI shelf audit in your FMCG field operation:

• [ ] SKU catalogue complete and up to date — every active SKU with reference photos from multiple angles
• [ ] Planogram library digitised — current approved planograms for each channel (MT, GT, CSD, pharmacy)
• [ ] Offline-capable capture — shelf photos queue locally when connectivity fails
• [ ] Training data pipeline — process for capturing new pack designs and re-training within 14 days
• [ ] Action integration — shelf audit score triggers tasks in the same app the executive already uses
• [ ] Pilot scope defined — 10–20 stores across 2–3 cities before national rollout

How AI planogram compliance works in Kinematic

Kinematic Supply Chain includes AI planogram compliance as part of the retail execution layer. Field executives capture shelf images through the same app they use for check-ins and order booking — not a separate tool. Compliance scores generate tasks automatically and surface on the supervisor dashboard as an exception-driven view: which outlets need immediate attention, ranked by compliance gap.

The outlet data captured in Field Force, combined with shelf state from the AI audit, gives brand managers the first complete picture of in-store execution quality. See how it works for FMCG teams →