The strongest replenishment program you can build is the one that never charges anybody by surprise.
Amazon's Auto-Buy settled the demand question. People will let software buy for them when the purchase is predictable and boring. The interesting move is not to copy it. It is to keep the prediction and hand the decision back.
You already hold the signals: bottle size, average days between orders, seasonality, household size, the last time support heard from them. Flow watches those and fires 1 message when the model says the customer is roughly 5 days from empty. SMS or WhatsApp, a single line, a single tap to confirm, shipped that day. No dashboard, no plan, no cancellation flow, no card charged in the dark.
Compare that to a standard subscription on the metrics that actually matter. Most subscription churn comes from shipments people did not want yet, and this model cannot produce that failure. You trade a lower conversion rate per prompt for a much lower cancel rate, near zero disputed charges, and support tickets that stop being about refunds. Repeat rate holds because nobody is escaping a plan.
The honest catch: you lose the forecast. Finance likes subscriptions precisely because the revenue is bookable, and asking permission every cycle removes that certainty. You also need message discipline. Prompt too early and you are spam. Prompt on the wrong channel and you are ignored. Cap the frequency hard, make the tap genuinely 1 tap, and let people mute the agent without leaving the brand.
Subscriptions were designed around the merchant's forecast. This one is designed around the customer's shelf. Only 1 of those gets cancelled on a Sunday night.
