Top-3 European grocer: 400 stores, no new hardware

The retailer is anonymised at their request pending a named-reference agreement.

The situation

A top-3 European grocery retailer running roughly 400 stores had two separate problems that turned out to share a root cause. Fresh produce was slowing down every checkout, because loose items needed a shopper or colleague to find the right code. And checkout shrink was rising with self-checkout adoption, visible only in the stock count weeks later.

Both problems came from the same gap: the terminals could see what was happening and could not interpret it.

What was ruled out

• New smart terminals: a multi-year capital programme across 400 stores, rejected on payback.
• Cloud vision: per-lane bandwidth and a round-trip latency budget that could not meet checkout timing, plus a data-protection review nobody wanted.
• In-store servers: a new estate of hardware to manage in order to avoid buying an estate of hardware.

What was deployed

Results

Accuracy improved after go-live rather than degrading, because each store's model kept training on that store's own produce. Sites that joined the rollout later reached target accuracy faster, benefiting from learning shared across the estate without any store's data being shared.

What we would flag to the next retailer

Two things consistently take longer than teams expect, and it is better to plan for them:

• Device inventory quality. Most estates do not have an accurate list of what is actually deployed, by model and firmware. The first fortnight is usually discovery.
• Colleague comms. Intervention prompts change what colleagues see at the lane. Stores briefed properly reached steady state noticeably faster than stores that were not.