Thinking Fast & Slow for a Personalized Notification System

2026-06-19 · Source: Netflix TechBlog - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Intermediate, medium

Summary

Netflix has implemented a hierarchical personalized notification system, inspired by Daniel Kahneman's "Thinking, Fast and Slow," to manage its hundreds of millions of daily messages. This framework addresses the inherent conflict between optimizing for immediate engagement and mitigating long-term member fatigue or opt-out risk. The system comprises a "slow" policy that makes strategic, personalized decisions about a member's weekly messaging plan, including frequency per channel, offering approximately O(100) distinct cross-channel pacing strategies. Concurrently, a "fast" policy handles tactical, real-time message selection when send opportunities arise. This architecture, which uses a feature store to bridge communication, explicitly incorporates a utility function balancing positive engagement signals with a universal message cost to prevent over-messaging. The new system resulted in significant production metric lifts, particularly benefiting casual viewers, by enabling independent evolution of pacing strategies and content ranking.

Key takeaway

For AI Engineers or Product Managers designing personalized communication systems, adopting a hierarchical "slow" and "fast" policy architecture is crucial. This approach allows you to decouple long-term strategic pacing from real-time message selection, explicitly managing user fatigue and optimizing for sustained engagement. You should implement a feature store to enable asynchronous communication between policies, facilitating independent iteration and A/B testing of pacing strategies and content ranking models for significant metric lifts.

Key insights

A hierarchical "slow" and "fast" policy system effectively balances short-term engagement with long-term user experience in personalized notification delivery.

Principles

• Decouple strategic pacing from tactical message selection.
• Explicitly model long-term costs like user fatigue.
• Use a universal message cost to prevent over-messaging.

Method

The system uses a "slow" policy to define personalized weekly message pacing (O(100) combinations) by maximizing a utility function balancing engagement and message cost. A "fast" policy then executes real-time message selection within these strategic guardrails, communicating via a feature store.

In practice

• Implement a feature store for policy-to-policy communication.
• Tune a universal message cost parameter empirically.
• A/B test pacing strategies independently of ranking logic.

Topics

• Personalized Notifications
• Hierarchical AI Systems
• Dual-Process Theory
• Feature Stores
• User Engagement
• Notification Fatigue

Best for: AI Architect, AI Product Manager, Product Manager, Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• I built a custom Slack inbox. It was easier than you’d think. | Yash Tekriwal (Clay) — AI agents and platforms can transform overwhelming digital communication into actionable, prioritized workflows.
• Building My Own Personal AI Assistant: A Chronicle, Part 2 — API-driven data access and AI-powered task decomposition significantly enhance personal agent efficiency and utility.
• Dreaming: Better memory for a more helpful ChatGPT — ChatGPT's "dreaming" memory system automatically curates and updates user context for personalized, current, and relevant interactions.
• From Chat Fatigue to Instant Action // Donné Stevenson — Dynamic, action-oriented UIs overcome "chat fatigue" and boost agent engagement more effectively than open-ended chat.
• Slow AI Is Dead — For AI tools to be useful in daily work, rapid inference speed is paramount, rendering "slow AI" obsolete.
• How Snapchat Serves a Billion Predictions Per Second — Snapchat's Bento platform scales ML ranking to a billion predictions/second by optimizing for asymmetric workloads and continuous iteration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Netflix TechBlog - Medium.