Loss Prevention Was Never the Destination: How AI Is Transforming Retail Operations

I came to retail technology sideways. I raced karts competitively in Britain as a teenager, ran a thirty-person business at sixteen, then spent seven years in tier-one investment banking before concluding I would rather build companies than finance them. Three venture-backed startups followed. None of that is an obvious route into supermarket operations.

But karting taught me something that has shaped every operating role since. Everyone on the grid has roughly the same engine. The race is won on execution: the line you take, how fast you get the car set up, the tenth of a second nobody watching ever notices. The machine is rarely the differentiator. What you do with it is. That is exactly where retail AI now finds itself.

What pulled me into this was not a technology demo. It was watching a cashier work.

In most grocery stores, a cashier is expected to memorise an encyclopedia. A banana is one code. An organic banana is a different code. Loose kale, vine tomatoes, a bag of shallots: all different, all committed to human memory, all keyed in by hand thousands of times a day. When the code will not come, the cashier picks the nearest thing they can remember. At a self-checkout, the customer does the same, only faster and with less patience.

None of that is theft. It is friction, and it produces errors at an enormous scale. The largest addressable pool of retail loss, it turns out, is not adversarial. It is operational.

Removing the False Trade-Off Between Security and Customer Experience

Retailers are often told they must choose between controlling loss and keeping the experience smooth. Every anti-theft measure adds a gate, a guard, a receipt check, a lock on the deodorant. Every convenience measure opens a door.

That framing is wrong, and it is the biggest obstacle in the industry today. Self-checkout runs shrink several times higher than staffed lanes, and the instinctive response has been to add friction back in. But the customer at a self-checkout does not want to skip the interaction. They want certainty that they will not be suspected of taking something they did not pay for.

Remove the guesswork and you remove both problems at once. Faster transactions and lower loss are the same intervention, not competing ones.

What matters is the difference between detection and intervention. An alert reaching a loss prevention manager two weeks later changes nothing. A prompt at the lane, while the basket is still open, changes the outcome.
Building AI Solutions That Work in Real Retail Environments

Three principles have held up throughout my leadership journey.

Deployability is a product decision, not an implementation detail. If your model needs a $50,000 server in every store, you have not built a product, you have built a science project with a procurement problem. We deployed across all 520 locations of one US retailer, Hawaii included, without an engineer leaving the office, because the software runs on hardware the stores already own.

Be honest about what is actually hard. The models are largely solved. Deployment is not. Most of the industry is still talking about AI while a much smaller group is quietly getting it into stores.

Protect the humans in the loop. Real-time correction turns loss prevention from a punitive function into a coaching one. Cashiers stop being audited and start being helped, and retention improves alongside the shrink numbers.

Turning Existing Retail Infrastructure into an Intelligent Network

Loss prevention is the beachhead, not the destination.

Every scale, scanner and checkout terminal in a store is already a sensor. It sees, it weighs, it reads barcodes and RFID tags, it knows what passed through it and when. We have spent a decade treating this equipment as cash registers. Over the next five years, I expect retailers to recognise it for what it is: a distributed sensory layer covering every square metre of the estate, running on infrastructure they have already paid for.

This matters because of where AI is heading. Agentic systems are only as useful as their grounding in the real world, and the physical world has no data layer. The models reason superbly over text. They are blind to whether the pallet arrived, whether the shelf is full, whether the item on the belt is what the label claims.

The devices that could answer those questions are already installed in millions of stores, factories and warehouses. They simply are not being asked.

The retailers who move first will extend this backwards through the supply chain, from the manufacturing line to the shelf edge, on the same devices running the same models. The question stops being, “Did we catch that loss?” and becomes, “Do we know, continuously and without asking anyone, what is actually happening across our operation?”

Learning the Operation Before the Algorithm

Learn the operation before you learn the algorithm. The people who succeed at the intersection of AI, retail and business leadership are almost never the ones with the best model.

They are the ones who have stood behind a checkout during a Saturday rush, who know what a store manager is measured on, and who understand that a solution needing three minutes of staff training per shift will not survive contact with a real store.

Be suspicious of anything that only works in a demo, and get comfortable with unglamorous problems. There is a great deal of prestige in frontier research and very little in produce codes.

The produce codes are where the value is.