Positions
LLMs do not belong in the points ledger
LLMs do not belong in the points ledger because the ledger must be deterministic and auditable, and a language model's output is neither. Across 217 profiled programmes, the points ledger's role is to apply exact earn, burn, and tier rules; an LLM introduces unverifiable judgement into a record that must balance to the unit.
The argument
The claim that LLMs do not belong in the points ledger starts from a definition of what a ledger is. A ledger is an append-only record of debits and credits where every entry must be exactly recoverable, every balance must reconcile to the unit, and the same inputs must always produce the same state transition. An LLM is a probabilistic model that generates a likely next token. It has no internal arithmetic, no guaranteed referential integrity, and no mechanism for replaying a transaction and arriving at the same result. These two objects solve different problems. Putting an LLM inside the ledger path replaces a deterministic state machine with a sequence predictor, and that is not a small implementation detail; it changes the nature of the system from auditable to merely plausible.
The register's own counts show how much of loyalty operation is already expressed as explicit, deterministic rules. It profiles 217 programmes total. Of the 214 programmes whose tier status is known, 112 run tiers. That means more than half of known programmes have a published or discoverable tier structure, with thresholds for earn, spend, or flight segments. The sector detail makes the point sharper: 61 of 61 airlines with known status run tiers, and 18 of 18 hotels with known status run tiers, while 0 of 18 grocery programmes and 0 of 12 fuel programmes run tiers. These are not natural facts; they are choices a programme makes and then encodes as rules. A points ledger must apply those rules exactly. An LLM that paraphrases the rule or infers it from examples will eventually misclassify a member at a threshold, because threshold enforcement requires exact comparison, not semantic similarity.
The counter is sometimes made that LLMs are only needed for the messy edges: merchant category codes, partner feeds, and promotional language. But that argument concedes the main point. Those messy edges are classification problems that happen before the ledger, not inside it. Once a transaction is classified, the points ledger must perform arithmetic and update state. The register profiles 87 vendors, and 48 of them publish full or partial pricing. That transparency expectation is incompatible with a model that cannot produce the exact rule by which it arrived at a balance. An operator who cannot explain a point deduction in terms of a deterministic rule has lost the audit trail that members and regulators expect.
The strongest counter-argument
The best case for putting an LLM in the points ledger is that the ledger is not only arithmetic; it is also classification under ambiguity. A member buys from an online shop whose merchant category is unclear, or a partner sends a feed with a narrative description of a bonus, and a rigid rule engine fails. An LLM can read the text and assign points consistently enough, replacing thousands of human judgement calls. In that view, every system has an error rate, and an LLM trained on the programme's own history may make fewer errors than brittle code or slow manual review. This is a real argument, not a straw man, because loyalty operations are full of exceptions and an LLM is genuinely good at extracting structured meaning from prose.
What would change our mind
We would change this position if any one of the 217 profiled programmes published an independently audited comparison of its existing deterministic ledger against an LLM-generated ledger over a full earning and redemption cycle, and the two matched exactly, transaction for transaction, with no post-hoc corrections, no hidden repair layer, and no manual overrides. That would show that an LLM can in practice reproduce the behaviour of a deterministic ledger in a live environment. Until that evidence exists, the difference between usually right and auditably right remains the whole point of a ledger.
What follows if we are right
If we are right, the person running a loyalty programme should keep LLMs out of the transaction path entirely. Use them for member-facing chat, call summarisation, fraud investigation leads, or offer copy, but never to compute balances, accrue points, expire points, or move a member between tiers. The ledger itself must remain a deterministic system whose every state change can be replayed and audited. A programme that is tempted to insert an LLM because the rules are too complex should first simplify the rules, not surrender the ledger's auditability.