The self-checkout shrink problem: what AI on CCTV can and can't fix

Self-checkout is the loss-prevention conversation of 2026. The BRC's 2025 Retail Crime Survey put the total cost of retail crime, including prevention spend, at £4.2bn, with £2.2bn of that direct customer theft, and unmanned lanes are squarely in the frame for a growing share of it.

The numbers that get quoted vary by source, but the direction is consistent: self-checkout lanes are widely estimated to see several times the shrink of a staffed till, and some grocers attribute a fifth or more of total loss to self-scan and checkout technology. Whatever the exact figure in your estate, the unmanned lane is leaking more than the staffed one.

Two very different problems wearing the same uniform

Lumping all self-checkout loss together is the first mistake. It is at least two problems:

• Honest mis-scans. A meaningful slice of self-checkout loss is genuine error, a shopper who misses an item under the trolley, double-bags by accident, or fumbles a barcode. This is a UX and process problem, not a crime problem.
• Deliberate theft. The rest is intentional, barcode switching (scanning a cheaper item), the partial non-scan (sliding items past unscanned), and the straight walk-away (bagging without scanning at all).

The two need different responses. You cannot deter an honest mis-scan, you have to design it out. And you cannot redesign your way out of deliberate theft, you have to detect and deter it. A single 'self-checkout loss' number hides which problem you actually have.

The first line of defence is not AI

This is the part vendors skip. The biggest levers on self-checkout shrink are physical and procedural, not algorithmic:

1. Lane and fixture design, sight-lines, the number of unmanned lanes per host colleague, and where that colleague actually stands.
2. Weight and scan checks at the point of bagging, the unglamorous controls that catch the unglamorous thefts.
3. Signage and friction, the well-evidenced deterrent effect of making it obvious the lane is monitored.
4. Staffing model, an attentive host at the bank of lanes is still the single most effective control most retailers have.

Where AI on existing CCTV genuinely helps

With that framing honest, here is where camera-based AI adds real value at and around the self-checkout, on the cameras you already have:

• Behaviour at the lane. Concealment and non-scan patterns at the bagging area are exactly the kind of camera-visible event a behaviour model is built to flag, for a human to review, never to action automatically.
• The walk-away, end to end. The highest-value self-checkout incident is the shopper who bags and leaves. Multi-camera tracking turns 'something happened at lane 4' into a forward-ready evidence pack with the subject followed to the door.
• Repeat offenders. The same faces work self-checkout across multiple sites. A cross-store watchlist is what turns a one-off loss into a recognised pattern, reviewed by a human every time.

What we would tell a loss-prevention lead to do

1. Split your self-checkout loss into mis-scan versus deliberate before you buy anything. The mix tells you whether your next pound goes on lane design or on detection.
2. Fix the physical and staffing controls first. They are cheaper, faster, and they move the honest-error slice that AI can never touch.
3. Then layer AI on the cameras you already own to catch the deliberate slice, with a human reviewing every flag, and a cross-store watchlist for the repeat names.