Loyalty Personalisation with AI: From Segments to Individuals

For most of the modern era of loyalty programs, personalisation meant segmentation: dividing the customer base into a manageable number of groups defined by demographics or behaviour, and communicating to each group with tailored content. This was an enormous improvement over one-size-fits-all messaging, and it drove real gains for the businesses that adopted it early.

But segmentation has ceilings. There are only so many segments a marketing team can manage. Customers in the same segment are still meaningfully different from each other. And segments are static — customers evolve, but the segment they belong to often does not update fast enough to keep up.

AI has quietly changed what is possible. Individual-level personalisation — not personalisation to a segment of one, but genuine, per-member decisions about what to communicate, when, and how — is now technically feasible in ways it was not a few years ago.

The most useful frame for AI-driven loyalty personalisation is next best action. For each member, given everything the program knows about them, what is the single action most likely to produce the desired outcome — an incremental visit, a larger basket, a resumed engagement, a completed milestone?

The action might be a specific message. Or the timing of a reminder. Or the choice of reward to feature. Or a decision not to communicate at all, because the member is currently over-messaged. The AI's job is to decide, based on the pattern of similar members, which of these actions is most likely to matter for this specific person right now.

This frame avoids the trap of treating personalisation as more messages sent. In many cases, better personalisation means fewer messages, timed better, with more relevant content.

There is a subtle but important line between personalisation that feels like a business remembering its customer well and personalisation that feels intrusive. The customer who is greeted by name and offered their favourite drink feels valued. The customer who is offered something based on a purchase they made at a competitor feels watched.

Staying on the right side of this line requires thoughtfulness about which signals are used, and honesty about which signals customers would reasonably expect the business to have. Loyalty program data — purchases, visits, redemptions, communication engagement — is well within the boundary. Data from outside the program, especially data acquired without explicit consent, usually is not.

Programs that respect this line tend to enjoy higher engagement, not lower. Customers reward the business that gets it right.

AI-driven personalisation, left unconstrained, will find edge cases that produce bad outcomes. It will send too many messages to a small subset of customers who happen to respond to the last message. It will offer overly aggressive discounts to customers who would have bought anyway. It will surface content that is technically correct but tonally off.

Guardrails matter. A cap on messages per member per week, regardless of what the model recommends. A rule that no member gets more than a certain discount per period. A tone-of-voice filter that catches messages that would embarrass the brand. A review queue for anomalies. These are not sophisticated interventions; they are hygiene.

The programs that get the most value from AI personalisation combine strong models with strong guardrails. The programs that trust the model without checks tend to have a bad month that requires them to walk everything back.

AI is very good at optimising within a defined strategy. It is not good at deciding what the strategy should be. The choice of what outcomes to pursue, what tone the program should have, what mechanics to introduce or retire, what edge cases to prioritise — these remain human decisions.

The best pattern we have seen is a small team of humans setting the strategy quarterly and the AI executing within it daily. The humans watch the aggregate outcomes, spot the anomalies, and adjust the strategy. The AI makes the millions of individual decisions the humans could never make manually.

This division of labour respects what each is good at. It also keeps the program aligned with business strategy in a way that fully automated systems drift from.

The right way to measure the value of AI personalisation is against a control. A share of members receives the AI-driven experience. A comparable share receives the baseline experience — segmented, but not individually personalised. The lift in engagement, redemption, and incremental revenue between the two is the honest measure of what the personalisation is worth.

This kind of test is uncomfortable because it requires accepting that some members will get a less-personalised experience for measurement purposes. But without it, the value of the personalisation is speculative. The businesses that run these tests consistently know exactly what their AI investment is producing; the ones that skip the tests are essentially trusting the vendor's claims.

Two directions are worth watching. The first is generative personalisation of the messages themselves — not just choosing from a library of templates, but producing message content tailored to the individual member. This is now feasible at reasonable cost and quality.

The second is cross-channel orchestration — decisions that span WhatsApp, wallet notifications, email, and in-store touchpoints as a coherent flow rather than isolated messages. This has been an aspiration for a decade and is now becoming operationally achievable for mid-market operators.

Both directions raise the ceiling on what personalisation can achieve. Both also raise the importance of guardrails and human oversight. The programs that get ahead of the operational discipline these directions demand will be well positioned when the ceiling raises again.