Multi-armed, contextual, and combinatorial bandit algorithms are existing experimentation frameworks that find an optimal variation — e.g. which onboarding flow converts best — in real time, by continuously reallocating traffic toward better-performing arms rather than waiting for a fixed-horizon A/B test to conclude.
Today these pair with human-authored variants: a person designs the candidate variations, and the bandit framework decides which wins. The speculation is that AI-generated variations could plug into the same bandit frameworks, turning them into open-ended, higher-volume experimentation engines that don't require a human to hand-author every candidate.
Relates to Experimentation Velocity Formula: (Volume × Quality) ÷ Time-to-Launch.