We have noticed an interesting pattern across retail chains trying to adopt AI. The CIO or AP team decides AI is coming, and the first move is to refresh the camera fleet across the chain to ensure that the infrastructure is "ready." New cameras, new cabling in some cases, new VMS. As soon as this is done, the AI evaluation begins.
And why not? Cameras are the foundation for visual AI for retail stores. If the foundation is old, the current approach is to fix that first and build everything on top of it to make the infrastructure work better.
The problem is that you don't know what "ready" actually means until you put real AI into a real store. And by the time you find out, you've already written the check.
What we are seeing in the field
We've been deploying visual AI in live retail for over five years. That means more than 2 million transactions a day, over 20,000 point-of-sale systems. Across that footprint, here's what's consistent:
- Most retailers have more usable camera coverage than they think. The cameras that came in for security investigations work fine for AI. We process video at 25 FPS on 640×480 at the checkout and 1280×720 in the aisles. Standard analog and 2MP IP cameras both work. Most stores already have this.
- The cameras that don't work for AI usually aren't the old ones. They're the ones that are wrongly placed. A 360° fisheye mounted dead center over a self-checkout pod is a great security camera and a poor AI camera, regardless of how new it is.
- Camera placement issues are use-case specific. The correct angle for queue monitoring is different from the correct angle for self-checkout missed scans. You can't design for both in a generic upgrade plan because the upgrade plan doesn't know what purpose the camera is going to serve.
If you upgrade cameras first and evaluate AI afterwards, it is most likely that you're making some guesses. Some of the new cameras will be in the right place while some will not be. Some areas you didn't consider for camera coverage will turn out to be significant. You'll eventually find out in a year's time. By then, the budget is already closed.
What actually works
Run the AI in a live store first, on your existing cameras. Let the AI identify and report the gaps. This way, you can upgrade with intent.
This is not a sales pitch dressed up as advice. It's how the largest retailer in the UK runs their program. They have the budget to do whatever they want. They still don't blanket-upgrade hardware ahead of AI evaluation, because they've learned that it costs more and produces a worse outcome than doing it the other way around.
Here is the other thing that is worth understanding: the AI itself shouldn't be demanding new hardware in the first place. A lot of vendors require GPU-heavy setups, dedicated servers in every store, extra cooling, and IT staff to keep it running. That cost gets folded into the "camera upgrade" budget and nobody notices what's actually driving the number.
SAI's visual AI platform runs on CPU for most use cases. A standard checkout deployment for a superstore is roughly £5,000 in hardware all in. Comparable GPU-heavy platforms run around £17,000. Same store, same outcomes, different architecture.
Where fixed cameras beat 360°
This comes up in almost every conversation that we have with retailers. Fixed cameras with defined viewing angles outperform 360° fisheye cameras for most AI use cases. Not because 360° cameras are bad, but because AI models are trained on consistent perspectives. A fisheye distorts geometry in ways the model has to compensate for, and you lose accuracy in the process.
This is the kind of thing you only learn by running models in real stores at scale. We learnt it the hard way. If your camera refresh plan is built around 360° coverage everywhere, you may be designing for the wrong outcome.
A different way to start
Before you commit to a chain-wide camera refresh, do a proof of concept in one or two stores using the cameras that are already in place. We'll deploy our platform, turn on the relevant use cases, and show you what works on the existing setup and what doesn't.
You'll discover three things:
- Real data from your own stores
- A clear picture of which camera positions are working for AI and which aren't.
- A hardware plan that's based on what you actually need, not what a generic upgrade spec suggested.
No commitment, no obligation to buy anything. If the answer at the end is "we still need to upgrade," at least it's an upgrade designed around real findings instead of assumptions.
If you're already in the middle of a camera refresh and starting to wonder whether the AI evaluation should have come first, that's definitely worth a conversation.
Pat O'Leary
patrick.oleary@saigroups.com
