AI in retail loyalty: measure incremental profit, not redemptions

By Kirill Naumov · 2026-10-24

Eight thousand people open a message. Two thousand redeem the offer. The campaign dashboard is happy; the commercial question is still waiting: how many of those purchases happened because of the offer? A redemption count cannot tell you who would have bought anyway, who added to the basket, or who used a discount that quietly reduced margin.

Decide what should change before asking AI to optimise

Name one outcome: an additional visit, a second purchase, a lapsed customer returning, a larger basket or stronger contribution after reward costs. Set the measurement window before launch. Opens, clicks and redemptions remain useful delivery signals, but they are not proof of incremental value.

A model may predict who is likely to redeem. That is not the same as predicting whose behaviour can change profitably. If a programme optimises only for response, it can become remarkably efficient at discounting customers who were already ready to buy.

Give the offer a fair comparison

Where permissions and data quality allow, randomly hold back a suitable group of eligible members and compare it with the AI-selected offer group over the same period. If withholding a message is inappropriate, compare the new offer with the existing business-as-usual experience. Decide in advance how to handle overlapping campaigns, stock-outs, returns and channel differences.

Digital offers may support member-level tests; where customer identity is incomplete, store or regional comparisons may be more practical. The method matters less than agreeing beforehand what the comparison represents. A control group invented after the campaign may tell a persuasive story, but it cannot reliably establish what caused the result.

Count the economics the model does not see by default

Measure incremental contribution, not only revenue. Include discounts and reward costs, fulfilment or service expense, returns and material campaign costs. Check for cannibalisation: did the offer move a planned purchase forward, shift spend away from a higher-margin option or reward someone who would have paid full price?

For a deliberately simple illustration, imagine the offer group spends $10,000 more than a comparable holdout. If rewards cost $4,000 and additional service and returns cost $2,000, the apparent sales lift is not $10,000 of new profit. The illustrative incremental contribution is $4,000 before other overheads. A real calculation must use the retailer’s own margin and cost definitions.

AI can help form testable hypotheses, select from an approved offer catalogue and summarise results. It should not change eligibility rules, invent offer terms or declare a winner based on redemptions alone. A person should review unusual cases and the economics before the next campaign goes out.

Run a compact pilot that answers a real question

Choose one customer moment, such as encouraging a useful second visit. Define the eligible group and primary outcome; compare the AI-selected offer with a holdout or the current experience; agree the time window, costs and guardrails; then track total spend and contribution alongside complaints, opt-outs and operational exceptions. Scale only when the result is commercially credible and acceptable to customers.

The goal is not to build a laboratory with a loyalty card attached. It is to avoid spending the reward budget on activity that looks busy but changes very little. If a simpler message works just as well, that is a useful finding too. We explore practical AI across retail, sales, employee learning and loyalty at the AI for Retail workshop on 7 November.

Explore the AI for Retail workshop → · Start a retail AI diagnostic

Further reading: NIQ on measuring retail personalisation · Google Meridian on experiment design.

Start a conversation →