Guide 35 · Technology
AI in loyalty programs: rules, machine learning or LLMs?
Use deterministic rules for anything that changes what a member holds: points earned, redeemed or expired, and tier status. Use machine learning for predictions a person or a rule acts on, such as churn risk, fraud scores and offer ranking. Use large language models (LLMs) for drafting, summarizing and answering questions, but never to calculate or change a balance, and have people check anything a member could treat as a promise.
AI in loyalty programs works best when each kind of system has its own job: deterministic rules for points balances, machine learning for predictions, and large language models for language. AI is now part of most loyalty platform pitches, from churn prediction to chatbots that answer member questions. Some of these uses are valuable; others put the program’s most sensitive asset, the points balance, at risk. The useful question is not whether to use AI, but which part of the program each kind of system should touch. This guide sets out a simple division and the tests that keep it in place.
Three kinds of system
What is the difference between rules, machine learning and LLMs?
They do different jobs and fail in different ways.
- Rules are explicit conditions anyone can read: “1 point per dollar; double points on Tuesdays; points expire after 18 months without activity.” The same input always gives the same output.
- Machine learning fits a model to past data to predict something, such as the chance a member will lapse. Its output is a probability or a score, and it can drift as behavior changes.
- Large language models (LLMs) generate text. They are good at drafting, summarizing and understanding questions, but they predict plausible words, not exact calculations, and can state wrong things confidently.
A loyalty program has two layers. The balance layer records what members have earned, redeemed and expired, and their status. The decision layer predicts and suggests: who is at risk, which offer to send, which transaction looks suspicious. The right tool depends on the layer.
Rules for balances
Which parts of a loyalty program should run on rules?
Everything that changes what a member holds. Points are a promise, and many companies carry them as a liability; Starbucks, for example, defers revenue for the Stars its members earn and records a liability for them. A balance therefore has to be calculated exactly, the same way every time, and be possible to replay and audit.
That means earning, redemption, expiry, transfers, tier qualification and reversals should run on a rules engine writing to a ledger. Rules have costs: every promotion or exception is a change request, and a mistyped rule can over-issue points. But a rule error is visible in the configuration and has a known size. An error made by a model that invents a bonus may only become visible after members have redeemed it.
Machine learning
Where does machine learning help?
In the decision layer, where the output is a prediction that a rule or a person then acts on. Good uses include:
- Churn and reactivation scoring, to decide whom to contact before they lapse.
- Fraud scoring, to flag accounts and transactions for review.
- Offer ranking and personalization, choosing which of several approved offers to show.
- Redemption and breakage forecasting for finance.
Machine learning needs labeled outcomes, regular retraining and monitoring. Without them, performance declines quietly. Keep a person or a clear rule between the score and any action that affects a member’s account, and test models against a control group so you know they add value.
A fraud model is a good example of the split. The model scores a transaction as suspicious; a rule then holds the redemption for review, and an analyst decides whether to release it or reverse the points. The model never changes a balance itself. The same pattern works for churn: the model picks the members at risk, and an approved offer with fixed terms goes to them.
LLMs: yes and no
Where can LLMs be used, and where not?
The dividing line is whether a wrong output could change what the program owes or be taken by a member as a promise.
| Use | Layer | Use an LLM? |
|---|---|---|
| Drafting campaign copy and subject lines | Content | Yes, with human review |
| Summarizing service calls or internal reports | Internal | Yes |
| Suggesting segments or offer ideas | Decision | Yes, a person approves |
| Answering general program questions in chat | Member-facing | Yes, if answers come from approved content |
| Stating a member’s balance | Member-facing | No, read it from the ledger |
| Calculating earn, redemption, expiry or tier | Balance | No |
| Approving point adjustments or exceptions | Balance | No |
| Interpreting the terms for a specific dispute | Member-facing | No, a trained person decides |
Even where an LLM is allowed, make the facts come from the system of record. A chatbot that reads the balance from the points ledger and quotes the approved terms is safe in a way that one generating answers from its own training is not.
There is legal precedent for treating a chatbot’s answer as the company’s own. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline responsible after the chatbot on its website gave a customer wrong information about bereavement fares, said it makes no difference whether information comes from a static page or a chatbot, and ordered Air Canada to pay $812.02. A loyalty chatbot that invents a bonus or an expiry extension creates the same exposure, so plan to honor or correct what it says, and log every answer.
Explain decisions
Should AI decisions be explainable to members?
Yes, when the decision changes what a member earns, keeps or redeems: a declined redemption, a reversed credit, a status downgrade or an account frozen for suspected fraud. The member should be able to see the rule that applied, the information used and how to challenge it. That does not mean revealing the model or the fraud signals; it means giving the reason and the remedy.
Explanations also reduce support contacts, because members who understand a decision call less often. Privacy law points the same way: the California Consumer Privacy Act, for example, gives California consumers the right to know what personal information a business collects about them and how it is used. Decisions that only rank offers or personalize a message generally need no individual explanation.
Build the explanation into the decision record. When a rule acts on a model’s score, store the rule, the inputs and the outcome together, so customer service can read the reason back to the member and correct it quickly if it is wrong.
Test a feature
How do you test a proposed AI feature?
Run three checks before building it.
- The balance test. Trace the feature from input to output. If the output credits, debits, expires or changes a member’s tier, it must be a rule, not a model.
- The worst-output test. Write down the worst wrong answer the system could give. If a member could quote it as a program term or a promise, keep the LLM out or put a person in between.
- The screenshot test. Imagine a member screenshots the output and posts it. If correcting it would mean apologizing for points already spent, the feature sits in the wrong layer.
When a vendor pitches an AI feature, ask the same three questions in the demo, and use how to evaluate loyalty platform vendors for the rest of the review. If a program is tempted to use AI because its rules are too complex to manage, simplify the rules first. Clear rules make every other layer, including the AI, easier to trust.
More answers
Frequently asked questions
Can AI calculate loyalty points?
It should not. Points are a promise to members and often a liability in the accounts, so every balance must be calculated the same way every time and be possible to audit. That is a job for a rules engine and a ledger. AI can help decide which offer to send, but the points themselves should follow written rules.
Where does machine learning help most in loyalty?
In predictions: which members are likely to lapse, which transactions look fraudulent, which offer a member is most likely to use, and how many points will eventually be redeemed. These outputs are probabilities, so they work best feeding a decision that a rule or a person makes.
Can an LLM run a loyalty program chatbot?
It can answer general questions and help members find information, if balances and program rules come from the system of record rather than from the model. The chatbot should read a member's balance from the ledger, not work it out, and it should not be able to promise points, extensions or exceptions.
What is an explainable AI decision in loyalty?
One where the member can see why it happened: the rule applied, the information used and how to challenge it if it is wrong. It does not mean revealing the model or fraud signals. It matters for decisions that change what a member earns, keeps or redeems.
Do privacy laws affect AI in loyalty programs?
Yes. AI features use member data, so consent, transparency and members' rights apply. In California, for example, consumers have the right to know what personal information a business collects about them and how it is used. Check the rules in each region with counsel before launching automated decisions.
What is the risk of an LLM making a mistake in a loyalty program?
An LLM can produce confident but wrong statements. If a member sees a wrong balance, a bonus that does not exist or an invented rule, they may act on it and expect you to honor it. In Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for wrong fare information from its website chatbot. Keep LLM output away from balances and terms, and review anything member-facing that states what members are owed.
What about AI agents that act for members or for the program?
Apply the same layers. An agent that books, redeems or adjusts points is changing the balance, so every action it takes should go through the same rules, limits and sign-in checks as a member or staff action, with a log of who or what acted. Start agents on read-only and drafting tasks, and add actions one at a time with limits on value and frequency.