AI in retail: what it actually does on the floor

AI arrived in retail from the back door. It went into the warehouse, the pricing engine, and the website, because that is where the data already was. The sales floor got cameras that count people. The conversation at the counter stayed exactly where it was ten years ago: unheard by anyone except the two people having it.

Where AI already works in retail

• Demand forecasting and replenishment. Sales history predicts what to order and when. The oldest and the most settled use case.
• Pricing. Competitor prices, margins, and stock levels set the shelf price without a weekly meeting.
• Online recommendations. What a shopper clicked predicts what they might click next. This is the use case most people picture when they hear retail AI.
• Cameras. Footfall, queue length, dwell time in a zone. Cameras answer how many and how long.
• Chatbots. The website conversation is already text, so it has been analyzed for years.

What all of that leaves out

Every item on that list reads numbers or pixels. None of them hears a word. The till records that a customer walked out without buying; it does not record that they asked about a payment plan and nobody offered one. The camera records a queue of six; it does not record that the person at the counter skipped the question that usually leads to a second item.

The result lives in the numbers. The reason lives in the conversation, and the conversation has never been written down.

One shift on one floor holds hundreds of conversations. A mystery shopper reports on one of them, a supervisor overhears a handful, and the rest were never evidence of anything.

What changes when the conversation becomes data

• Lost sales get a reason. Not a conversion rate, but the sentence where the customer said why.
• Objections stop being anecdotes. The same hesitation heard across twenty conversations is a pricing or a messaging problem, and it reads differently once you can see it repeat.
• Missed offers become visible. The upsell nobody made, the question nobody asked, the alternative nobody suggested.
• Training gets an address. The same skipped step across different people is a training problem, not a personal one.

What it does not do

• It does not see the store. Shelf condition, cleanliness, and queue length are seen by eyes and cameras, not heard by a microphone.
• It does not replace footfall counting. Those are different questions and different tools.
• It does not hold customer personal data. It is masked in the transcript before any analysis runs.
• It does not deliver a verdict about a person. The system scores against your criteria and shows the evidence; the decision stays with a manager.

Where to start

1. 1Write down the standard. The greeting, the questions your staff should ask, the objections they should handle. Without it there is nothing to score against, and a generic template scores every business the same way.
2. 2Record on what the team already carries. Nothing gets installed on the floor, and nobody is taken off a shift.
3. 3Read the first shifts before changing anything. The first week is for comparing what the standard says with what the floor actually sounds like.

This is the part of retail AI that is still mostly empty, and it is the part Lansy AI works on: the conversation itself, scored against the standard you set, with feedback reaching the employee while they still remember the shift.