One comparison decides it
The decision reduces to one comparison. What does a stale or conflicting answer cost, against what an error page costs? Run it per feature, not per company, because most products hold both kinds of data.
CP fits where a wrong answer creates work or losses that dwarf being down. Account balances, the last item in stock, distributed locks, leader election: anywhere two nodes disagreeing means money moves twice or two schedulers both believe they are in charge. Coordination data that lies is worse than coordination data that pauses.
AP fits where staleness is invisible or cheap. A like count reading 41 instead of 42. A feed missing the newest post for five seconds. A product page showing a review to some people and not others for a minute. Nobody notices, and an error page would be strictly worse.
Where these came from
These grew out of real workloads. Amazon's original Dynamo paper says plainly that a shopping cart should accept writes during failures and merge conflicts afterwards, because a customer who cannot add to their cart is lost revenue.
Two practical things follow. Many stores let you pick per operation: Cassandra sets consistency per query, and DynamoDB serves eventually consistent reads by default with strongly consistent ones on request. The letters are dials, not tattoos.
Announcing that you choose AP for a whole design. Say the payment ledger needs CP behaviour and the activity feed tolerates AP, and you sound like somebody who has shipped both.
Worked example
A ticketing startup sells the final 200 seats for a concert while a partition splits its two regions for 90 seconds. The seat inventory lives in a CP configuration: writes require a cross-region majority, so during the split, the minority region shows a spinner and about 300 buyers see "try again shortly." Annoying, but no seat sells twice. The same product's "people are viewing this event" counter and recommendation carousel run on an AP store; both regions keep serving slightly stale numbers and nobody can tell. Marta, the architect, gets asked afterward why buyers hit errors at all. Her answer: the alternative was overselling roughly 40 seats and refunding angry customers at the venue door. For inventory, 90 seconds of refusal was the cheap failure.