Learning from Millions of Unique Messages
Imagine trying to personalize messaging when you have thousands of different message variants, each with unique copy, images, offers, and calls-to-action. How can an AI agent learn what works for each user when every message is completely different? Labels solve this fundamental challenge by abstracting millions of unique messages into a small set of learnable concepts that agents can experiment with and optimize.What Are Labels?
Labels are conceptual tags that describe the key characteristics of your messages. Instead of trying to learn from raw message content, agents learn which types of message concepts resonate with each user. Examples include tone (casual, urgent, friendly), value proposition (discount, new feature, social proof), call-to-action (shop now, learn more, try free), and urgency level (high, medium, low).How Labels Enable Learning
Without labels, an agent sees hundreds of completely different messages about sneaker sales. With labels, the agent just sees [value, fomo, performance, fashion]. This dramatically simplifies the learning problem, and the agent can work to identify the label that resonate with its user. Labels allow agents to identify successful patterns across different message variations, generalize learnings to new messages with similar label combinations, make predictions about which label combinations will work best, and optimize systematically rather than randomly.The Label-Agent Learning Loop
Every message gets tagged with relevant labels when created through automated content analysis or manual tagging. Each agent then tests different label combinations, observes user responses, and builds personal preference profiles. Agents identify which labels drive engagement, learning patterns like “user responds to urgent plus discount combinations” or “user prefers educational content over promotional.” When sending new messages, agents choose based on learned label preferences, prioritizing messages with high-performing label combinations while avoiding patterns that historically underperformed and balancing exploration of new patterns with exploitation of known winners.The Power of Label Combinations
Agents don’t just learn individual labels—they learn combinations that work. One user might respond to [urgent + discount + shop_now + evening] while another prefers [educational + new_feature + learn_more + morning]. As user preferences evolve, agents dynamically adjust their label preferences, perhaps shifting from promotional to educational content or changing timing preferences over time.Business Benefits of Labels
Labels provide a systematic approach to message creation and testing with reusable insights across campaigns. They enable deeper user understanding through behavioral profiles and drive efficient optimization with:- Faster learning through concept abstraction
- Better predictions for new message performance
- Reduced testing time through intelligent combinations
Labels in Action: A Real Example
An e-commerce company sends hundreds of product promotion messages. Different products, discounts, urgency levels, and tones. Instead of treating each message as unique, they label them. A message like “Weekend Flash Sale! 25% off designer handbags - Limited stock!” gets labeled as [weekend, flash_sale, discount_25, accessories, limited_stock, urgent]. Agents then learn individual preferences:- User 1: Responds to [weekend + discount_25 + urgent]
- User 2: Prefers [accessories + limited_stock] regardless of discount
- User 3: Ignores anything with [urgent] but loves [weekend] timing