Journal / Kirill Naumov
AI in Retail Pricing: Optimise the Decision, Not Just the Number
A practical guide to AI-assisted retail pricing: separate price response from demand noise, protect margin and trust, and test changes before scaling.
By Kirill Naumov ·
“The competitor dropped the price. Should we follow?” It sounds like a quick question. In reality, the answer depends on whether customers compare these products, whether the item is available, what the promotion is doing, and what the retailer is trying to protect. AI can help analyse that decision. It should not turn every price movement into an automatic race to the bottom.
First decide what the price is meant to do
A pricing model needs a business objective and guardrails. Is the task to clear seasonal stock, preserve a margin floor, respond to a meaningful competitor, maintain a known price image, or test a value proposition? Those are different decisions. Define the objective, eligible products, minimum margin, maximum change, review owner and customer-facing rules before asking a model for a recommendation.
A price change is not proof of a price effect
Sales move for many reasons: availability, season, placement, promotion, competitor activity and customer mix. If a price falls just as a display improves, the resulting sales change cannot be cleanly attributed to price. Keep a record of concurrent changes and compare similar products, stores or periods. Where feasible, use a controlled test with a pre-agreed stopping rule rather than interpreting a before-and-after chart as a verdict.
A [Management Science field experiment on competition-based online retail pricing](https://pubsonline.informs.org/doi/10.1287/mnsc.2017.2753) explicitly asks whether a retailer should respond to a competitor, to whom, and by how much. The researchers tested randomized prices and a pricing strategy in a specific retail setting. Its reported result is evidence from that setting—not a forecast that another retailer will achieve the same outcome.
Keep the offer consistent wherever it appears
Before changing a price, check the full chain: product and variant, store or delivery area, promotion eligibility, start and end time, shelf label, product page, checkout and customer-service guidance. A correct model recommendation can still create a poor customer experience if one channel shows yesterday’s price or the terms are hidden in a footnote.
Treat price, promotion and availability as related but separate facts. A personalised product recommendation does not require a personalised price. Do not feed individual customer profiles into pricing just because the technology can accept them. The [US Federal Trade Commission’s January 2025 staff study](https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-surveillance-pricing-study-indicates-wide-range-personal-data-used-set-individualized-consumer) describes how consumer-related data may be used in individualized pricing; it is a US study, not a complete account of rules in every market. The practical lesson is to make data use, price logic and customer communication deliberate, reviewed decisions.
Show the human the reason and the uncertainty
A pricing or category manager should see what drove the suggestion: stock position, competitor match quality, demand pattern, promotion status and confidence. If an input is stale, a competitor comparison is weak or the recommendation crosses a guardrail, route it for review. Keep an audit trail of the recommended price, the price actually applied, any override and the reason for it.
Measure more than revenue
Agree on a balanced scorecard: gross margin, units, sell-through, markdown exposure, stock availability, basket effects, customer complaints and corrections across channels. Compare against a credible control and account for products that move together. If the model lifts units by discounting too deeply, it may have made the dashboard look busy while weakening the economics.
Start with one category, a limited set of eligible decisions and a human approval step. Expand only when the customer offer remains accurate, the commercial objective is met and the team can explain when the model should be ignored. For practical AI applications in retail, sales, learning and loyalty, see the [AI for Retail workshop](https://www.kirillnaumovai.com/workshops/ai-retail-november).