A Checklist for Choosing a Computer Vision Vendor for Retail

Score vendors on your store reality, not their demo

Shelf-monitoring demos run on clean planogram photos in good light. Your stores aren't that. This computer vision vendor checklist for retail is built around whether a vendor survives real fixtures, real connectivity, and real packaging churn — the things that decide whether the system gets used after the pilot.

1. Capture and hardware fit

• Fixed cameras, staff phones, or shelf-scanning robots — which, and does it physically fit your fixtures and aisle widths?
• What gets installed per store, who maintains it, and what's the power and network footprint?
• Who owns hardware failure and replacement across the estate, and what's the mean time to get a dead store back online?

2. Connectivity model

• Does it upload full-resolution photos to the vendor's cloud, or run detection on-device and send only results?
• What happens in a store on 4G failover or business broadband? A multi-megabyte shelf photo takes seconds to move on a slow uplink, multiplied across every photo and every store.
• Are events retained or queued for a store that drops offline and reconnects?

3. Catalogue and packaging coverage

• Does it already recognise your SKUs, including private label?
• How fast does it retrain for new, seasonal, and promotional packaging, and who does that work — you or the vendor?

4. Accuracy and false-alert handling

• Is there confidence-based routing, or a single fixed threshold?
• Where do ambiguous detections go — a review queue, or straight onto a store associate's task list?
• Can you see precision and recall on your photos from your stores, not their benchmark numbers?
• When it meets packaging it hasn't seen, does it degrade quietly into wrong-but-confident alerts, or flag low confidence and route for review?

5. Integration

• Do detections land in your store-operations dashboards and data model, or only in a vendor portal?
• Is there a real API or event stream out — webhook, message queue, batch export?

6. Data handling

• Photos of your stores and staff: where are they stored, for how long, and which sub-processors touch them?
• Can you require on-device processing with no image retention?

7. Pricing and lock-in

• Per-store, per-scan, or per-detection — and how does the bill behave across the whole estate and at audit peaks?
• If you leave, do you keep the trained model, the historical detections, and the hardware?

How to weight it

Standard fixtures, good connectivity, and a catalogue the vendor already covers push the decision onto sections 3 and 4. Unusual store formats or thin connectivity make sections 1 and 2 the whole decision — a vendor that needs photo upload to its cloud is out regardless of how well it scores on accuracy. Our computer vision project was shaped by exactly that constraint: on-device inference sending only a small structured result, because the store connections couldn't reliably do anything else.

Red flags in the sales process

• The demo is only on their photos. Insist on a pilot in your worst-lit, most-cluttered store.
• "It detects everything." Real systems are honest about occlusion, glare, promo packaging, and depth of stock.
• Vague on the offline-store case. "What happens when a store loses connectivity mid-shift" should get a specific answer.
• Pricing that won't be quoted at full estate size plus audit-peak volume.
• No way to export detections or the trained model.

Where this checklist misleads

• Feature-counting. A vendor can tick every box with capabilities you'll never switch on. Weight by your store reality.
• Demo accuracy. A curated planogram photo tells you nothing about aisle nine at 6pm on a Saturday.
• Smallest estates. For a handful of stores, manual audits may still beat any vendor rollout.

FAQ

Build or buy — which does this checklist point to? Either. Use it to score SaaS vendors, or use the same questions as requirements if you build.

What's the single most important item? Whether detection runs on-device or needs photo upload. On a real store network, that one answer gates most of the others.

How long should a pilot run? Long enough to cover restock cycles and a promotional packaging changeover in a genuinely hard store — weeks, not a one-day walkthrough.

ISTRALLEN builds retail computer vision to this bar — on-device, confidence-routed, integrated with your operations stack; see AI for Retail.