Community Of Bandits

A community of bandits is Braze's AI decisioning architecture for splitting a large marketing action space into collaborating contextual bandits, each responsible for a decision dimension such as channel, offer, timing, creative, or frequency.

Key points

• Braze frames the design as a response to the combinatorial problem in lifecycle marketing: a few plans, creatives, days, send times, and frequencies can already create more than a thousand possible customer-level actions ^[src-152].
• Instead of training one model over every possible combination, Braze partitions the action space by dimension. One bandit can choose a channel, then pass that decision as context to a second bandit choosing an offer, and so on ^[src-152].
• The source describes a bandit team containing a multi-armed bandit, an elastic-net bandit, a simple gradient-boosting bandit, and a complex gradient-boosting bandit, with the right bandit chosen based on use case and data richness ^[src-152].
• The stated operating goal is sample efficiency: learn quickly from limited campaign data without overfitting or exposing too many customers to bad options ^[src-152].
• The design makes AI Decisioning more modular. Marketers still define the success metric, decision dimensions, available options, and constraints; the bandit community learns the policy inside those boundaries ^[src-152].
• Braze reports a financial-services case study where the architecture varied message, creative, time of day, day of week, and frequency for business credit card referrals, producing a vendor-reported 92% lift in conversion rate and $16M in annualised value ^[src-152].

Related entities

• Braze

Related concepts

• AI Decisioning
• Contextual Bandits
• Multi Armed Bandits
• Reinforcement Learning For Marketing
• Marketing Bandit Optimisation
• Personalisation Gap
• Customer Feature Matrix

Source references

• ^[src-152] George Khachatryan, Nathaniel Rounds, Victor Kostyuk / Braze – "A community of bandits" (source page: "From multivariate testing to AI decisioning", 2025-09-30)