The problem is review capacity, not coverage
Most stores are already well covered. A mid-size branch may run twenty or thirty cameras recording continuously, and almost none of that footage is ever watched. Shrinkage is usually discovered weeks later during a stock count, by which point the recording has rolled over and the investigation has nothing to work with.
AI-assisted detection does not add cameras. It adds a shortlist. Instead of asking a loss prevention lead to scrub through hours of video, the system surfaces a handful of timestamped moments that a person can review in a few minutes.
What the system reports
The events are deliberately descriptive. They state what was observable in frame and nothing more:
- Shelf Interaction Without Checkout — an item was taken from a shelf and the exit was crossed with no corresponding till transaction
- Handled Item Out Of View — an item stopped being visible while still held, before reaching a checkout
- Suspicious Behaviour — dwell, repeated approach or movement patterns that fall outside the zone baseline
- After-hours presence — activity in a stockroom or sales floor outside trading windows
None of these is a finding of theft. A customer may put an item back out of frame, hand it to a colleague, or use a click-and-collect flow the camera does not see. Every event is raised as Manual Review Required, and the reviewer decides what actually happened. Detect Suspicious Behaviour with AI Vision goes further into how these patterns are defined and tuned.
Working with the cameras you have
Existing estates vary enormously, and the honest answer is that some views will work and some will not. Before committing to a rollout, check three things per camera:
- Angle — a shelf-facing or three-quarter view supports interaction detection; a ceiling dome pointed straight down usually does not.
- Resolution at distance — the hand-and-item region needs enough pixels, which in practice means the shelf is within roughly eight to ten metres of a 1080p camera.
- Lighting stability — shopfront glare and overnight lighting changes cause more misses than any model limitation.
Cameras that fail these checks remain useful for context and timeline reconstruction, just not for interaction-level events. Scoping this openly avoids the familiar pattern where a pilot is judged against views that were never going to work.
Tying events to transactions
The strongest signal comes from correlating exits with till data. When a checkout transaction is available, an exit event can be matched and closed automatically, which removes the large majority of candidate events before anyone looks at them. Where a point-of-sale integration is not available, the same events can be grouped by time window and reviewed as a batch, at the cost of a longer queue.
What to measure
Individual events are evidence; patterns are the business case. Track events per trading hour by zone, share closed as explained after review, median review time, and shrinkage variance between instrumented and non-instrumented departments over a full stock-count cycle.
Governance and staff communication
Two rules keep this workable. Keep the language factual in the interface, in reports and in any conversation with a customer or colleague: the system observed a sequence, a person reviewed it. And tell staff what is detected and why before switching it on, because loss prevention tooling that arrives unannounced creates more internal friction than it saves.
Rollout note
Start with one department that has a known shrinkage problem, review every event for two weeks, and expand only once the queue is small enough to be genuinely worked through each shift. The same camera connection also supports Possible Smoke events in stockrooms and back-of-house areas, which is covered in Early Smoke Detection Using Existing IP Cameras.
