12 AI Growth Lessons for Subscription Apps – Phil Carter, Elemental Growth

In the first minute or two, an 'oracle' asks the user questions about themselves; answers are used to match them with a Tolan that has its own personality, and the app explains why they were matched. The user is then quickly dropped into a voice-to-voice chat with the AI creature, which reacts to their answers in near real time. The match result is shown as a shareable personality-profile infographic.

After the onboarding quiz matches the user with a Tolan, the app shows an infographic of the Tolan's personality profile and how it maps to the user's, designed to be shared.

Runna skipped the casual tracking market Strava owned and targeted runners training for a specific race (marathon, 10K) with a customized plan that adapts to logged runs. Ladder similarly mapped segments (home vs gym, men vs women), sized each pocket, recruited coaches/personalities to attract each one, and adjusts sets and weights from logged workouts (progressive overload). Onboarding quiz answers seed the personalization.

Granola links to the user's calendar and sends a desktop/mobile notification before each meeting, which is useful on its own and reminds them to turn on AI notes. Wispr Flow keeps a small black sliver widget at the bottom of the laptop screen (press-and-hold Fn to dictate) that later became a Dynamic Island-like element popping up tips to deepen usage; on iOS it also ships a keyboard with a quick dictation button.

StudyFetch ships many AI features for college students (lecture recording, turning handwritten or digital notes into outlines/takeaways, extracting key terms from an uploaded diagram). Phil recommends shipping monthly to daily and announcing each feature on TikTok, Instagram and Discord so shipping doubles as acquisition.

Marketing team used AI tools (e.g. ElevenLabs for voice, video generators) to increase ad creative concepts from tens to 400+ per month, testing them on Meta and TikTok to find and double down on winners faster; learnings also feed product roadmap decisions.

Avoid fully AI-generated ads (e.g. Veo 3 generating person, voice and music end to end) because quality isn't fully convincing, fake testimonials break FTC rules, and consumers react badly when they detect AI. Instead shoot real people, then use ElevenLabs to translate one spokesperson into ~30 languages and Suno to layer multiple background tracks on one 30-second YouTube/TikTok ad to test which performs best.

After launching AI background removal/replacement, Photoroom led its Meta and TikTok ad concepts with the AI feature, placed within the first six seconds of the ad. This lifted conversion enough to make paid acquisition work in cheaper-CPM emerging markets (Mexico, Brazil, Indonesia) that previously didn't pencil out.

Every presentation/website made in Gamma is shared inside companies and at conferences (personal/social viral loops), and published 'gammas' get indexed by search engines and cited in ChatGPT/Claude answers (SEO/AEO), bringing in new users without ad spend.

In early versions of AI presentations, Gamma picked long-tail models (not always frontier OpenAI/Anthropic ones) where quality was good enough, gaining speed and much lower compute cost. Phil's rule: frontier models for bleeding-edge prosumer use where cost is no object; cheaper models for price-sensitive consumers, factoring in fast-falling model prices over 6-12 months.

When launching Notion AI in 2023, Notion reused its early community playbook: early AI adopters on YouTube, TikTok, Instagram and Discord posted how they used Notion AI, shared templates and databases, and tips to get value out of it.

Replaced a single hard paywall with a multi-step paywall: the product is free forever, but users are invited to try the best version free for 7 days, then asked to keep paying for it. Rolled out with other pricing and packaging optimizations. Phil warns it only suits highly retentive products; hard paywall is right for most apps.

Client attempted a shift from hard paywall to freemium; pulled back after a couple of weeks. Learnings are informing a planned second attempt.

Basic plan gets non-AI features and a limited taste of AI; full/unlimited AI (e.g. Duolingo Max's voice role-play with Lily) requires a higher Pro/Max tier, to cover compute costs. Also seen at Perplexity, ChatGPT, Claude, ElevenLabs, Tolan.

Tolan launched with a 7-day free trial and has adjusted trial length and its premium/free offering over time, to avoid free users running up compute costs and churning.