A concealment happens in aisle four. The cloud pipeline is still uploading when the person turns toward the exit. By the time the alert arrives, staff are reading a notification for an event that already left the building.
That is the market challenge for Australian retail CCTV AI: not "do we have cameras?" but "can we act while they are still on the floor?"
> Quick answer: Edge AI processes video on-site with a typical alert path of about two seconds. Cloud CCTV analytics round-trip frames off-site and often add 5 to 15 seconds of delay.
At a glance
Both use AI. The difference is where inference runs.
Cloud path: camera → upload → queue → remote inference → result back → alert.
Edge path: camera → local inference → alert.
For loss prevention, the cloud middle steps are the gap between intervention and walk-out. Economics: cost of retail shrinkage.
| Factor | Cloud analytics | On-premises edge |
|---|---|---|
| Typical alert latency | 5 to 15 seconds | ~2 seconds |
| Internet required for detection | Yes | No |
| Peak-hour NBN contention | Degrades upload | Local LAN only |
| Floor display during outage | Often fails | Continues |
| Bandwidth cost | High continuous upload | Low (updates only) |
Edge wins for concealment. Typical end-to-end staff notification is about two seconds on compact edge hardware - person still in the aisle.
Cloud round-trips typically take five to fifteen seconds depending on NBN upload, encoders, server load, and queue depth. Fifteen seconds is enough to pocket a high-value item, move two aisles, and blend into foot traffic.
Buyer overview: how AI detects shoplifting in real time.
With cloud processing, frames leave the building - often to overseas data centres. Edge keeps processing local for routine inference.
Under the Privacy Act 1988, identifiable images are generally personal information. Overseas processing can raise APP 8 cross-border questions. This is not legal advice - confirm with your adviser. Deeper custody piece: on-premises AI for retail CCTV privacy.
Cloud systems usually fail when upload fails. Contended business NBN links show this during peaks, not only during outages.
Edge systems can keep detecting, driving local displays and sirens offline. Internet is for optional remote monitoring and small model-weight updates - not live video streams for inference.
| Factor | Cloud | Edge |
|---|---|---|
| Monthly bandwidth | High (uploading video 24/7) | None for inference |
| Cloud compute | Per-camera or per-event fees | None (local hardware) |
| Hardware | Cameras only | Cameras + edge appliance |
| Ongoing model updates | Often included | Included in managed services |
| Total cost of ownership | Scales with cameras and usage | Often flat per-store fee |
For 10 to 15 cameras, continuous upload plus per-camera AI fees can exceed a flat per-store edge subscription. Always model your camera count and NBN plan. Cloud quotes that ignore bandwidth are incomplete.
For stop-theft-before-exit, edge matches the job. Alert quality after week two still decides staff trust: false positives explained.
IntelliGuard uses on-premises edge AI for Australian retail and pharmacy loss prevention. Detection on the NeuraIQ appliance. No video upload for inference. Alerts in about two seconds with who-to-approach context and aisle location.
$299 to $369 per store per month (excl. GST). 30-day money-back guarantee. Pilot the highest-shrink site first.
Cloud and edge both use AI. They optimise for different constraints. If the job is stop theft before exit on an Australian retail floor, latency, privacy, and offline reliability point the same way: process where the cameras live.