Start with why they work on humans, because every architecture is applied psychology. Points exploit the endowment of progress: Nunes and Drèze's car-wash experiment showed customers given a 10-stamp card with 2 'free' stamps completed it far more often than customers given an 8-stamp card, though the task was identical. Visible progress toward a goal accelerates behaviour. Expiring points weaponise loss aversion (spending to save what you 'own'), tiers convert it into altitude anxiety (the gold member flying in December to keep gold), and paid memberships flip the psychology entirely: sunk cost makes the customer their own retention department, using you to justify the fee they paid.
Architecture selection is a business-model question, not a preference. Points suit high-frequency, thin-margin retail where a deferred 2–5% works as a discount you might never pay (breakage) and the data is worth more than the liability. Tiers suit categories with steep heavy-user concentration and real service levers (airlines, hotels, where upgrades cost little and mean much). Perks-only suits brands where the maths would insult the relationship (a café remembering your order beats 2% back). Paid membership suits ecosystems with enough benefit surface to justify the fee, and it's the only architecture where the program itself is a P&L line rather than a liability. The wrong pairing fails predictably: tiers without service levers are wallpaper, and points in a twice-a-year category expire before the second purchase.
The economics must be designed before the enthusiasm. Earn rate sets the visible generosity (1 point per ₹100). Redemption value sets the real one (a point worth ₹0.25 makes the program a 0.25% rebate wearing a 1% costume). Breakage, typically substantial, funds the gap between promised and delivered value, but is also a measure of how boring your rewards are. And the accounting is real: unredeemed points sit as deferred-revenue liability, which is why finance eventually audits what marketing launched in a fever. The design discipline: model the program as a pricing change (because it is one), simulate redemption behaviour at maturity (year three, not launch quarter), and decide upfront what breakage rate you're targeting, since too low bleeds margin and too high means nobody values the currency.
Then judge it like an economist, not a romantic. The program's return has exactly three legitimate sources: incremental behaviour (purchases that wouldn't have happened, measured against holdouts, not against members' own pre-enrolment habits, since joiners are self-selected heavies), data value (item-level purchase histories powering assortment, pricing, and targeting, often the biggest prize and the least celebrated), and switching-cost drag (defection now has an exit fee of forfeited status and points). Against these, the double-jeopardy caution: heavy buyers were staying anyway, light buyers rarely enrol, so a program graded on members' loyalty is grading water on wetness. The question that survives audit: what did behaviour do net of who enrolled?