Building audiences with synthetic data in seconds

Overview

Audience planning has always had a bottleneck: turning data into clear, consistent audience descriptions that match what a campaign actually needs. It was slow, manual, and the results rarely lined up.

A global media and technology group set out to fix that with AI, working with Elsewhen to put large language models (LLMs) to work without losing precision or control. The system pairs retrieval-augmented generation (RAG) with synthetic data to write rich audience descriptions on demand, structured and explainable, that planners shape in their own words and refine before use.

62

countries served globally

5 min → 2 sec

audience query latency

£17m

in planner time saved per year

Before

Stuck maintaining segments manually

Audience segmentation absorbed huge effort for little return, planners maintained static segments and reconciled inconsistent vendor metadata that didn't reflect how campaigns were actually run. At scale that friction cost real money: slower planning cycles, wasted senior time, and missed windows in fast-moving campaigns.

Left unchanged, it would keep pulling skilled planners into data hygiene and away from strategy.

After

Creating audiences in moments

Planners now generate and interrogate audience descriptions on demand, each with a clear account of how and why it was built. Rather than hand-curating segments, they describe what they want and refine what comes back, faster and with more confidence.

The system runs as a co-pilot, built on RAG and synthetic data that together produce rich, structured audience descriptions on demand. Reliability was engineered in: model temperature tuned down to cut variation, a second model to flag errors, JSON validation to enforce structure, and humans-in-the-loop approving outputs through an admin panel before anything ships. Planners now build in seconds what used to take an afternoon, and can explain every call.