Claim
AI products retain customers at rates between mobile games and social networks, with the stickiest cohorts reaching around 40% by month five while many products sit in the high single digits, because raw capability does not generate the habit and network effects that drive higher retention.
Mechanism
Social networks achieve retention through bidirectional relationships, content feeds, and network effects that make leaving costly. Mobile games use reward loops, streaks, and social competition. AI products have neither by default. Products that reach the 40% five-month ceiling do so through deep workflow embedding: the AI is inside the customer's daily operational sequence, and switching means retraining habits and re-integrating data. Products that sit in the high single digits have not achieved that integration and can be swapped for a competitor when the next benchmark reversal arrives. Capability is necessary but not sufficient; the colander retains what is sticky, not what is capable.
Conditions
Holds when: the AI product is evaluated primarily on capability and lacks meaningful workflow integration or switching costs; customer segments are not locked in by contract or embedded data.
Fails when: the AI product has deeply integrated into core customer workflows (ticketing, CRM, code review) and data gravity creates high switching costs; regulated industries with procurement constraints that override marginal capability differences.
Evidence
"AI is like a colander. Customers pour in, & many pour out."
Retention data from Tunguz's analysis: AI products retain "somewhere between high single digits & about 40%, with the stickiest foundational cohorts near the top" at month five. By comparison, SaaS companies retain around 90% in the first five months and Facebook/Instagram retain around 80%. Mobile games sit below AI on this spectrum. The 40% ceiling is not a ceiling of capability; it is the ceiling of habit and integration depth.
Signals
- Products at the 40% cohort cite frequent daily use and workflow integration rather than feature variety
- High-churn AI products have users who can describe a task but not a routine
- Retention curves flatten earlier for products that lack in-context data persistence
Counter-evidence
The benchmark period is still short. AI products as a category are newer than the retention patterns that defined social and gaming norms. A two-year cohort may look different from a five-month cohort. Products built on top of AI models inherit model switching costs only partially; application-layer switching costs (trained workflows, institutional memory in the product) can be substantial even when the model layer is commoditizing.
Cross-references
- The average frontier model holds its performance crown for 41 days, resetting customer negotiating power on the same cycle: model-layer pricing cycle applies to buyers; product-layer retention is a separate dynamic determined by integration depth, not benchmark timing.
- The AI deployment bottleneck has shifted from model capability to customer adoption, making embedded engineering teams the institutional moat: embedded engineering teams are one mechanism for moving products from the high-single-digit churn band into the 40% cohort by deepening workflow integration.