AI-powered voice extraction
Automatically surface the most relevant voices from social conversations for every poll.
Until this release, every Saut poll was anchored to a fixed pipeline of platforms and a fixed query set. That made the system reliable, but it left a lot of the signal on the table — the most articulate voices on any given topic are rarely sitting where the algorithm expects them to be.
Voice Extraction is the first model in the Saut stack whose only job is to find the people whose opinions actually move a conversation. It reads every post the scrapers return, scores each one for substance, originality, and reach, and elevates the top decile into the poll feed before anything else is shown.
What changed under the hood
We replaced the legacy keyword-density filter with a two-stage model. The first stage is a fast embedding classifier that drops the obvious noise — bots, ad copy, off-topic replies. The second stage is a tuned Claude pass that reads the remaining posts in context and assigns each one a stance, a confidence score, and an extracted quote that represents the post in roughly twenty words.
The extracted quote is what shows up in the Voices panel. It is never the full post; it is the most defensible single line from the post that a human briefing would actually use.
What it means for users
- Every poll now has a Voices tab populated within minutes of going live.
- You can sort voices by stance, platform, or recency without losing context.
- The original post is always one click away — Saut never paraphrases without attribution.
We are rolling this out to every workspace this week. Existing polls have already been backfilled — open any poll from the last 90 days and the Voices tab will reflect the new extraction.