Journal / Kirill Naumov
AI personalisation in retail: useful, not unsettling
A practical framework for retail personalisation with AI: relevant signals, visible customer value, consent and meaningful measures.
By Kirill Naumov ·
Personalisation often starts with the question, “What else do we know about this customer?” Start somewhere more useful: “Which customer task can we make easier?” A large volume of data does not automatically make an offer relevant. It may simply help a business make a confident mistake.
Choose a moment before choosing a model
Pick one moment: a shopper is finding a suitable product, comparing options or deciding what to do after a purchase. Describe what the person is trying to accomplish, which information is genuinely needed and what a useful outcome looks like. If the team cannot explain the value in plain language, it is too early to begin with an algorithm.
Use signals that belong to the task
A preference a customer has chosen to share may help identify the right size or format. A recent visit to a category may make a comparison more useful than a random discount. But a signal should be current, understandable and connected to the customer’s task. A click is not permission to infer someone’s entire life.
Make the recommendation visible and controllable
Explain why a recommendation appeared, let the customer correct it and make it easy to change data preferences. Store colleagues need the same clarity: where a suggestion came from, which approved information supports it and when to verify it or ask a specialist. An unexplained recommendation is difficult to defend in a customer conversation.
Compare AI with the simpler option
Personal service does not always require machine learning. Clear navigation, useful filters, a customer-approved purchase history or quick answers to common questions may solve the problem. Compare the AI approach with the simplest alternative the team can actually maintain. If the customer sees no difference, the added complexity has not earned its place.
Measure more than clicks
Before a test, decide what to observe: task completion, search time, recommendation dismissals, returns, complaints or repeat use. Include answer quality and the cost of maintaining the experience. More clicks on a banner do not prove that choosing became easier or that the offer created an incremental sale.
Start small and keep a human way out
Limit the test to one category or channel, name the owners of the data and process, and define the errors that stop the experiment. Give customers a clear choice and colleagues permission to reject a questionable suggestion. Good personalisation does not show off how much a business knows. It saves people effort without asking them to trust blindly.
A first experiment needs one customer moment, available baseline measures and a team accountable for the outcome. At the four-hour AI for Retail workshop on 7 November, we will work through personalisation alongside sales, learning and loyalty.