Decisions · Technology
Where must LLMs not be used in loyalty
Never let an LLM decide, alter, or explain a member's points balance, accrual, expiry, redemption value, or tier status. Those actions create financial liability and must run on a deterministic ledger with human approval for any exception. LLMs may draft non-binding text or suggest segments, but not commit the programme.
The decision
A loyalty programme must decide where a language model may touch member value. The boundary is not the model's fluency; it is whether an incorrect output changes what the programme owes. Points, status, expiry, and redemption are liabilities. Those fields must remain on deterministic rules: the ledger calculates, a human confirms any exception, and the model may suggest elsewhere. The decision is to treat the points ledger and its term interpretation as a no-LLM zone.
The mechanism is ownership of the balance. An LLM produces probable text, not accountable computation. A member can rely on a written point total, so the programme needs a single source of truth that is auditable and replayable. Any model-generated statement about balances becomes a commitment the moment a member sees it, even if the programme later calls it a draft. Therefore the ledger itself cannot sit behind a model.
What it costs you
The cost of this boundary is engineering time and slower exception handling. Every balance, expiry, or redemption exception needs a deterministic code path plus a trained human check. An illustrative calculation: say a programme receives 500 balance disputes per month. If an LLM triaged them at 2 cents per call and automatically resolved 80 per cent, the 400 straightforward cases would cost 800 cents, or 8 dollars, and the remaining 100 would still need 4 dollars of staff time each, or 400 dollars. LLM-assisted handling would total 408 dollars. Without the model, all 500 disputes require staff at 4 dollars each, or 2,000 dollars. The no-LLM boundary therefore costs 1,592 dollars per month in extra handling for this volume. That figure excludes the larger risk of one incorrect 20,000-point extension at one cent per point, which is 200 dollars per member and scales without limit if the error repeats.
How to decide
Apply one test to any proposed LLM use. Write down the worst incorrect output the model could produce in that step. If that output changes a member's point balance, accrual rate, tier, expiry date, or redemption value, or if a member could reasonably quote it as a programme term, the answer is no. If the output only drafts a marketing subject line, suggests a segment for an offer, or summarises an internal report that no member sees, the model may be used. The division is reversibility and member reliance.
What would change our mind
Our recommendation flips only if a loyalty programme can demonstrate, before an external auditor, that a specific LLM call is deterministic, replayable from the same ledger state, and cannot alter a balance or entitlement without a separate human approval. If the model's output is bound to the same release process as code, including rollback and audit logs, then the no-LLM boundary would lift for that call. Until that condition is met, the risk of a plausible but wrong point total is not worth the saved handling cost.