It is off until you turn it on
There is a switch in Settings → Security, for owners and admins only. A workspace that never turns it on has never sent a respondent's words anywhere.
- Answer text goes to Anthropic's Claude API. Anthropic does not train on it, and the prompt is not kept.
- Nothing is ever sent automatically — only when you run an analysis.
- Verified email addresses and the technical details recorded with a response are never fetched at all.
- Email addresses and phone numbers typed inside a written answer are stripped before anything is sent.
Response insights
On a form's Insights tab, you get the recurring themes across written answers, how many people expressed each, and the minority view a list of themes would otherwise hide.
- Every theme carries verbatim quotes, checked against the actual answers — anything not genuinely there is dropped.
- You see the cost before anything is sent, as tokens against your monthly allowance. A run you cannot afford says so up front rather than failing after the wait.
- Coverage is always stated: how many responses were read out of how many exist. Past a cap it reads an even spread across the whole set, never just the first few hundred.
Scoring assist
On a scored form, this shows the real distribution of scores from your own responses, and suggests bands placed where the answers cluster — with the number of responses that would land in each.
Nothing is applied for you and your own bands are kept. The boundaries are computed, not generated, so running it twice on the same responses gives you the same answer both times.