Security & data

Plain words about what the product does with your firm’s data — written from the code, updated when the code changes. Last reviewed September 12, 2026. The Privacy Policy and Terms are the contract; this page is the explanation.

Who can see a workspace

A workspace is one firm. Every company, signal, thread, packet, note, meeting record and decision belongs to exactly one workspace, and every page and API call is scoped to the signed-in member’s workspace. A person belongs to one workspace.
Members and admins. Admins invite and remove members, record decisions, approve firm memory and publish the board page. Removing a member deactivates them immediately and ends their sessions; their work stays attributed to them in the record.
Threads are shared on purpose. A research thread is private to its owner until it is shared with the workspace or with named teammates. Quick Ask conversations stay private to the person who asked; teammates with access see the agent’s deliverables, not the chat.
Nothing private is indexed or unfurled. Every workspace, account, API, board and intake route is excluded from crawling (robots.txt, noindex) and never appears in the public sitemap.
The board page is read-only, lives at an unguessable link, asks for a shared access code (stored only as a hash), and shows only what an admin has published — never the working draft.

Signing in

Email and password. Passwords must pass Django’s validators (minimum length, not too common, not all numbers, not similar to your name or email) on both sign-up paths. Verification email is sent; existing accounts are never locked out for lacking one.
Sessions. Cookies are Secure and HttpOnly in production behind HTTPS with HSTS. Settings → Security lists every browser and phone signed in as you and lets you sign the others out.
Two-step sign-in is not offered yet. We say so in the product rather than pretending; it is on the roadmap.
Sign-up does not reveal accounts. Creating a workspace with an address that already has an account shows the same “check your inbox” page as a new one and emails the account holder instead.

What leaves the workspace, and where it goes

To the AI model provider. A research run or a diligence-packet run sends the model the brief, the firm’s approved principles and lessons, the documents you attached to that run, and the public and provider data the run fetched. Company notes and deliverable comments are never part of it. Meeting audio recorded on the iPhone is transcribed by the provider’s transcription model; web meeting records never contain audio.
To market-data providers. When your firm connects PitchBook or CB Insights, runs query them by company name and identifiers. Your documents are not sent to them. DealCloud is linked read-only: companies and stages are read; the CRM is never written.
To the public web. Runs fetch public pages and cite them. They do not post anything.
By email. Packets and snapshots go only to the recipients a person chose; notifications carry a title, a line and a link — never the report text. Every delivery is recorded, and a failed delivery is shown as needing a check rather than assumed sent.

What is kept, and for how long

Research, packets, decisions and firm memory are kept as the record of the workspace until you delete them; deliverables keep their revisions so a decision can be re-read against the exact evidence it rested on.
Archived threads leave every surface at once and are purged seven days later; permanent deletion before that is a second, typed confirmation.
Meeting records. Web records hold notes or an imported transcript, never audio. On the iPhone, temporary audio is deleted within 24 hours by default and retained audio within 30 days; the transcript is never publishable, and every purge is logged. A consent attestation is recorded with every meeting record.
Documents you attach are stored with the thread they belong to and uploaded to the model provider for that run. Deleting the thread permanently removes our copy and asks the provider to delete its copies.
Product measurement stays in our own database: allowlisted event types with IDs (a version, a signal, a comparison) — never the text of a query, a document or a result — and no third-party analytics collector on any page.

What the agent never does

It never records a decision. Only a workspace admin records Advance, Hold or Pass, with a reason and a revisit condition.
It never scores a company, ranks the feed or moves events out of chronological order. Signals carry the plain reasons they surfaced; comparisons show sources and gaps, not a winner.
It never reads discussion. Notes and comments are for people; what shapes research is a principle a partner approved, and every time one is handed to a run is recorded.
It never writes to your CRM or to any provider.

Operations

Production runs behind HTTPS with HTTP Strict Transport Security, secure cookies, content-type sniffing disabled, and a signature-verified webhook for the model provider’s run notifications.
Every consequential action is a server-side state machine with an audit trail: memberships, packet delivery, publishing the board page, meeting purges, firm memory edits.
We have not yet completed a third-party security audit or a SOC 2 report. If your firm needs one, tell us and we will say plainly where we stand.

Questions

For a walkthrough of any of this on your own workspace, write to hello@signalvc.ai. Privacy requests go to privacy@signalvc.ai.