AI data handling
Last updated
This page states precisely when your data is transmitted to a third-party AI model and when it is not. The rule in one line: document content reaches an AI provider only as the direct result of an action you took, never as a side effect.
Principles
- Extraction is local. Reading text out of your uploaded files (PDF, DOCX, PPTX, XLSX, CSV) happens in your browser, client-side. Uploading a file transmits it to storage — not to any AI model.
- Transmission is explicit. Content goes to an AI provider only when you trigger an AI action on it: generate a summary, a brief, a memo; run a review; submit verification evidence for classification.
- Scope is per-user. AI requests are built from data your account can read under Row Level Security. There is no cross-tenant context, ever.
When data reaches an AI provider
| Action | What is sent |
|---|---|
| AI panel conversation | Your message plus page context from your own data |
| Document review / summary | The text of the specific document you requested analysis on |
| Deal brief / investment memo | The company's profile or watchlist data you can already see |
| Intake parsing | The text you pasted or files you uploaded to the parser |
| Verification classification | The evidence document you submitted to the slot |
Requests are processed through the provider's API under API data-use terms — API inputs are not used to train the provider's models.
What never happens
- No document is sent to an AI provider because you uploaded it.
- No background job reads your documents into a model without an action from you.
- No AI request ever includes another user's private data.
- Nothing the AI produces is shown to a counterparty without your confirmation.
Technical safeguards
- AI provider keys live server-side as encrypted secrets — never in the browser bundle.
- All AI calls route through server functions that enforce authentication, rate limits, and timeouts.
- Usage is metered per user per day, bounded by plan limits.