Build vs Buy: Off-the-Shelf Merchandising Tools vs Custom Computer Vision

Not a two-way choice

Build vs buy for computer vision in retail usually gets framed as "licence a shelf-monitoring platform or build our own model." There's a middle option most chains with unusual store formats or thin connectivity end up at: a custom pipeline assembled from proven components — a standard detection model, a portable inference runtime, on-device hardware, a lightweight messaging layer — sized to your stores rather than started from a blank repo.

Option 1: buy a packaged shelf-monitoring platform

SaaS with fixed cameras, or shelf-scanning robots. Fast to pilot, and the vendor owns the model and its upkeep. Comparable published outcomes exist — vendors in this space market out-of-stock reductions in the 20-30%+ range — so the category works.

What it costs you:

• Per-store or per-scan pricing that scales with the estate and spikes around audit cycles.
• Their capture model. Their cameras or robots have to physically fit your fixtures and aisles. Non-standard formats are where this breaks.
• Their catalogue coverage and retrain cadence. New, seasonal, and private-label packaging is recognised when the vendor gets to it, not when you need it.
• Data leaves. Photos of your stores and staff go to their cloud under their retention terms.

Option 2: build in-house from scratch

Full ownership: model training, edge deployment across every location, a messaging layer, a review console, an operations pipeline. It's a multi-quarter project with a long maintenance tail, and it only makes sense if shelf intelligence is close to core business and you have computer-vision and MLOps capacity to own it indefinitely.

Option 3: a custom pipeline on proven components

This is the shape of our computer vision project: a standard single-shot detector for the model, a portable inference runtime so one trained model runs across whatever edge hardware ends up installed, on-device inference so only a small structured detection result leaves the store rather than the raw photo, a lightweight publish-subscribe protocol carrying events from 150+ stores to one dashboard, and confidence-based routing into a review queue. Proven parts, assembled for the client's actual connectivity and fixtures, built inside a 12-week window.

Signals you've outgrown "buy"

• Non-standard fixtures or store formats a vendor's cameras or robots don't handle.
• Connectivity that can't stream photos to a vendor cloud reliably.
• Packaging or private-label churn faster than a vendor retrains.
• Per-scan pricing across a large estate is a big, forecastable line item.
• You want detections in your own operations dashboards and data model, not a vendor portal.

The middle path in practice

These aren't mutually exclusive across an estate. A chain with a mix of store formats can pilot a packaged vendor in the standard-layout locations where its cameras fit and its catalogue coverage is good, and run a custom pipeline in the older or oddly-shaped stores the vendor can't serve. The vendor data also does double duty as a scoping input — it tells you what detection accuracy and alert volume to design the custom side around before you build it.

Where "buy" stays the right answer

• Standard fixtures, decent connectivity, catalogue the vendor already covers. The integration work you'd save is the whole project.
• Small estate. Build and run don't pay back against a handful of stores.
• No computer-vision or MLOps capacity to own a pipeline after it ships.
• You need a result in weeks with zero engineering.

A quick way to place yourself

1. Can a vendor's capture hardware physically fit your fixtures and connectivity? No pushes toward custom.
2. Does packaging or private-label change faster than a vendor would retrain? Yes pushes toward custom.
3. Do you have computer-vision and MLOps capacity to own a pipeline indefinitely? Only "yes" justifies a full from-scratch build over custom-on-components.
4. Is a vendor portal an acceptable home for the data, or does it need to be in your operations stack?

FAQ

Is a shelf-scanning robot "build" or "buy"? Buy — it's vendor hardware plus a vendor model. The real question is whether it fits your floor plan and whether the estate economics work.

Can we start with buy and move later? Yes, and it's a reasonable path. Pilot a vendor in one region, use the data to scope a custom build for the rest.

What's the long pole in a custom build? Not the model. Edge deployment across many stores, the messaging layer, and keeping training data current with packaging changes.

ISTRALLEN builds custom retail computer vision on proven components, sized to real stores and connectivity; see AI for Retail.