How to catch self-checkout theft

Self-checkout tills were built to save on staff, and they do, but the trade-off is a bank of machines with one colleague watching six screens and none of their hands. Scan avoidance, the banana-price trick on something heavier, a barcode covered by a thumb: none of it trips a scale or an alarm, because the machine only knows what got scanned. US retailers lost $112 billion to shrink in a single year, most of it preventable theft that cameras recorded but nobody watched (National Retail Federation, 2023), and the self-checkout bank is where a disproportionate amount of that recording goes unwatched, because there was never anyone free to watch it live.

What actually happens at the till

Most self-checkout loss isn't a criminal plan. It's an item that slides past the scanner in the general shuffle of bagging, a trolley base nobody thinks to unload, a barcode swapped for a cheaper one on something that looks close enough. The pattern that costs the most is rarely dramatic: it's ordinary and it repeats, till after till, shift after shift, and it's exactly the kind of thing a person glancing over six screens will miss nine times out of ten.

Why the till's own hardware can't catch it

The scale catches a weight mismatch. It has no idea an item never approached the scanner in the first place, and it can't tell a swapped barcode from a correct one, because both register as a successful scan. The gap sits entirely in what the camera sees and nobody watches: the hand movement, the bag filling before the total's been paid, the item that goes straight from basket to carrier bag.

Reading the till bank as it happens

Self-Checkout Fraud is a Svid check built for exactly this: it watches the self-checkout cameras and judges a window of recent frames as they arrive, so a scan-avoidance move reads as a sequence rather than one blurry still. An item passed straight from basket to bag without the scanner beep's worth of motion, a barcode covered mid-scan, a trolley base that never gets unloaded: it flags the moment for a colleague to glance at while the customer is still standing at the till, not an hour later when the drawer's already short. Delivery is on screen, by email, or as a generic JSON webhook into systems you already run, so the alert can land wherever the supervisor's desk already looks.

1. 01

   Point a camera at the self-checkout bank

   An existing RTSP feed, an upload, or a phone signed in at app.svid.ai/live.html all work. It only needs a clear view of the tills and the bagging area.

2. 02

   Turn on Self-Checkout Fraud for that camera

   Checks are enabled per camera, so it can run on the self-checkout bank alone if that's all you want watched.

3. 03

   Try it over an existing export first

   Recorded footage runs through the same pipeline as live, so you can point it at last week's till footage and see what it would have flagged before putting a live camera on it.

4. 04

   Act on the flag, not the assumption

   An alert is a prompt to look at a clip, not a verdict. The usual outcome is a colleague asking if they can help with the bagging, which resolves most of it without anyone being accused of anything.

What it doesn't do

It doesn't plug into the till software or the POS system beyond a generic JSON webhook into whatever you already run there; there's no claimed integration deeper than that. And it doesn't recognise a specific customer from one visit to the next. There's deliberately no facial recognition and no re-identification anywhere in Svid: the check describes behaviour in the frame, a barcode covered, an item unscanned, and leaves who someone is to a human looking at the clip.

What it costs

The Store prices Self-Checkout Fraud from your own usage before you buy, shown as the extra it would add per month at your camera count and hours watched. It sits on a subscription that's also priced from your own usage, floored at $9 a month, with the whole derivation on screen at signup. A shop running one or two self-checkout cameras sits well toward the cheap end. The checks page has the full list if you're weighing it against the shop-floor Theft Watch check too, which most retailers running self-checkout end up pairing it with.

Can a camera really tell if someone skipped scanning an item?

It reads the behaviour, not the barcode. An item that moves from basket to bag without the motion a scan would involve, or a barcode covered mid-pass, is exactly the pattern the check is built to flag, judged over a short window of frames rather than a single image.

Does it need to connect to the till or POS system?

No. It reads the camera feed over the till bank; there's no integration with the till's own software beyond a generic JSON webhook you can point at whatever system you already run.

Will it recognise a repeat offender?

No. Svid has no facial recognition and no way of matching a person across separate visits. Every flag is judged on what's visible in that moment, nothing more.

What happens after an alert fires?

It lands on screen, by email, or by webhook with a description and a link to the clip. A colleague reviews it and decides what to do; nothing is automatic beyond the flag itself.