Semafor used AI embeddings and multi-agent reasoning to distill 4,900 claims from 300+ speakers at its World Economy conference into a structured editorial report.

Semafor Tech Editor Reed Albergotti built a prototype using OpenAI's Codex the Sunday after Semafor World Economy 2026 ended. Alastair Clements, Semafor's data lead, then spent 36 hours turning it into an analytical tool that parsed 4,900 distinct claims from more than 300 speakers, anchored to specific transcript quotes.

The system converted each claim into a numerical fingerprint (called embedding or vectorizing) to capture meaning rather than wording, then used multi-agent reasoning to surface supporting or opposing quotes tied to central themes. Journalists then reviewed every theme, stress-testing premises and editing down to the most clearly supported ones.

Current AI systems cannot generate insights reliably on their own compared to journalists, but they expanded the scope of what journalists could discover and analyze. The technology determined what was possible to surface; journalists determined framing and what was worth publishing.

The pipeline used Google BigQuery, embedding models from Voyage (owned by MongoDB), Anthropic's Haiku 4.5 and Opus 4.7 models, Cohere's ranker, and the open-source UMAP library. Total API calls and new database cost only a few hundred dollars.