AI Audience Segmentation

Context

My product area covered how marketers built cohorts for downstream personalization — targeted email, push, and in-app campaigns. Segments were built manually using filters and SQL, which meant they were often stale, overly broad, and dependent on the marketer's technical skill. Conversion rates were inconsistent across campaigns as a result.

Problem

Targeting quality was a function of who built the segment. Advanced marketers built precise cohorts; less technical ones built blunt ones. The floor was too low and the ceiling too hard to reach consistently.

Approach

1. 1Instead of building a smarter segmentation tool, we generated high-intent cohorts automatically. We clustered behavioral event data to surface natural user groups based on recency, frequency, engagement depth, and product interactions.
2. 2Used an LLM to interpret and label those clusters, producing activation-ready cohorts with human-readable names like 'high intent trial users' or 'churn risk accounts'. The LLM handled interpretation; the clustering handled the actual grouping logic.
3. 3Surfaced the generated cohorts directly inside the activation workflow so marketers could use them in one click, without needing to understand how they were built. Cohorts refreshed dynamically. Validated impact with controlled experiments comparing AI-generated cohorts against manually defined segments.

What I shipped

• ✓Behavioral clustering pipeline grouping users by recency, frequency, engagement depth, and product interactions
• ✓LLM labeling layer interpreting clusters into activation-ready cohort names and descriptions
• ✓Cohort surface integrated directly into activation workflows with one-click use
• ✓Dynamic cohort refresh so segments stayed current without manual maintenance
• ✓A/B experiment framework validating AI cohorts against manually built segments

Impact

• ·AI-generated cohorts improved activation targeting performance by 15% compared to manually defined segments.
• ·Targeting quality became consistent and scalable — no longer dependent on how advanced a marketer was at building filters.
• ·Reduced the time marketers spent on segment creation, freeing them to focus on campaign strategy.

Learnings

• Automation works best when it raises the floor, not just the ceiling. The goal wasn't to replace expert marketers — it was to make every marketer perform like an expert. Consistent quality at scale was the metric that mattered.
• LLM labeling made the output usable. Clusters without labels are just numbers. The labeling step was what turned a data science artifact into a product feature marketers could actually trust and act on.
• Controlled experiments are non-negotiable for AI features. Self-reported improvement isn't enough — we needed a clean comparison against the baseline to prove the cohorts were actually better, not just different.

One key tradeoff

Metric definitions