AI features make people pay faster. They also make people leave faster. RevenueCat's latest data shows AI subscription apps earn 41% more per paying user in year one, but only about one in five annual subscribers renews. For other apps it's closer to one in three. I build iOS apps with subscriptions, so I went through the numbers and wrote down what I'd actually change.
Short version: subscriptions are paid for by habit, not by the model inside the app. Judge an AI app by its second year, not its first month.
Where the numbers come from
State of Subscription Apps 2026 by RevenueCat, the company whose SDK handles roughly 20% of global subscription app revenue. The data covers 2025: 115,000+ apps, $16B+ in revenue, 1B+ transactions. An "AI app" here means one where AI is part of the core experience, which is 27% of the sample.
AI vs non-AI apps
| Metric | AI apps | Non-AI apps | Difference |
|---|---|---|---|
| Revenue per payer, first month | $18.92 | $13.59 | +39% |
| Revenue per payer, first year | $30.16 | $21.37 | +41% |
| Trial start rate (of downloads) | 8.5% | 5.6% | +52% |
| Paid by day 35 (of downloads) | 2.4% | 2.0% | +20% |
| Annual plan renewal | 21.1% | 30.7% | −9.6 pts |
| Monthly plan kept for 12 months | 6.1% | 9.5% | −3.4 pts |
| Weekly plan kept for a year | 1.2% | 1.7% | −0.5 pts |
| Refund rate (median) | 4.2% | 3.5% | +0.7 pts |
| Share of monthly plans | 59.8% | 26.2% | ×2.3 |
The top half of the table is about how well AI sells. People start trials more often, convert more often and pay more. The bottom half is about how well AI retains people, and it's worse on every line. Out of 100 annual subscribers, 21 come back for year two in an AI app and 31 in a regular one.
Why they leave
The report gives the numbers, not the reasons. What follows is my own reading of them.
- People buy an effect, not a habit. Generating an avatar or reading calories off a food photo is impressive the first time. If the job is one-off, there's no reason to come back. Across categories, 26–40% of voluntary cancellations cite "not using it enough".
- Monthly plans. 60% of AI subscriptions are monthly, compared with 26% for other apps. A monthly plan asks "still worth it?" every 30 days.
- Expectations run ahead of results. Refunds are higher, and the ceiling is higher too: 16% vs 12.5%. The ad promises magic, the model sometimes gets it wrong, and the user notices while a refund is still easy.
- Price. Model costs push AI apps to higher prices, and "too expensive" is the top cancellation reason, at 25–45% depending on category.
Five things I'd do with an AI app
Examples from my own apps are illustrations, not measurements.
1. Sell annual, not monthly. Annual plans renew less often than you'd like, but you collect twelve months up front. If the user cools off in month three, a monthly plan just ends. The median annual price in the report is $34.80, up from $31.60.
2. Use long trials instead of 3-day ones. Trials of 17+ days convert at 42.5% vs 25.5% for trials of four days or less. 55% of 3-day trial cancellations happen on day zero, so a short trial tests the wow moment, not the habit. In Clay Sniper, a shooting journal, I moved the Pro trial from three days to two weeks so a shooter gets at least a couple of range sessions in.
3. Put AI inside a recurring job. AI keeps users when it removes friction from something they do every week anyway, not when it's a standalone party trick. In Lunar Fasting, calories from a photo are one line in a daily log next to water, fasting and weight, not a separate "scan your food" screen.
4. Push marginal cost toward zero. Every cloud request costs money, which is why AI apps tighten freemium, shorten trials and add expensive AI tiers. That fights tactic 2. In Lunar Fasting, photos can be analyzed on-device with Apple Foundation Models, offline and free. The cloud is only used with the user's own API key. If a request costs nothing, a long trial doesn't lose money.
5. Don't give away a free tier you don't need. Hard paywalls convert 10.7% of downloads by day 35, freemium converts 2.1%, and annual retention is nearly identical (27% vs 28%). The paywall model decides how many people pay, not how many stay.
Modeling the economics
An AI app takes about 63% of its first-year revenue per payer in month one, roughly the same share as a non-AI app. The difference shows up only at renewal, outside a first-year window. If your ad bidding optimizes for short-horizon value, it's easy to overpay for AI users. Plan for a lower lifetime value than market averages suggest.
One more detail from the report: on Google Play, 32% of cancellations are failed payments rather than a user decision. On the App Store it's 15%. Part of your "churn" may be fixed by a card-update reminder, not by the product.
Bottom line
I wouldn't build the product around the model. Models get cheaper and more alike, and in a year your competitor will have the same one. I'd build around a job people come back to every week and let AI take the tedium out of it. Sell an annual plan with a long trial, run inference on the device where possible, and look at year-two revenue.
Originally published at klukyanov.ru.
Shorter weekly write-ups (in Russian) — on Telegram.
