What a party's IT person needs to know.
The short version: your material sits in our own database on our own servers in the EU, one workspace cannot read another's rows, and every AI provider we send work to is named below. If something here is not enough for a procurement review, ask — we would rather answer the question than be found out by it.
Tenant isolation
Every table that holds a workspace's material carries that workspace's id, and the isolation is enforced by the database itself through row-level security, not only by application code. The policies are forced, so they also apply to the account that owns the tables, and they have a test suite of their own. Personas, documents, preparation, transcripts and reports are visible inside one workspace and nowhere else.
Self-hosted in the EU
Postgres with pgvector, the sign-in service and the application all run on our own machines in the EU. There is no third-party authentication vendor and no managed-database vendor in the path, and nothing you put in is used to train a model.
Where data leaves the EU
Language and voice models are the one thing we do not run ourselves. These four flows are all of it.
- Live voice debate — the candidate's microphone is streamed to OpenAI in the United States as they speak: one stream for every AI participant, plus one that turns speech into text. The debate streams are sent with storage switched off.
- Preparation, analysis and the brief — knowledge-base text, uploaded documents and the transcript being analysed go to Anthropic and OpenAI under their API terms. Neither provider trains on it.
- Real-debate debrief — a recording you upload is sent whole to ElevenLabs in the United States, which writes the transcript.
- E-mail — the names and addresses of the people you invite go to Resend, which delivers the mail.
Who gets in
Access is by invitation only, and the door is closed at the API rather than only in the interface — there is no public sign-up to find. An invitation carries one of four roles: owner, admin, member, viewer. Imported opposition research is split into verified material and interpretation that needs review, and unapproved interpretation reaches no AI persona until somebody on your team approves it.
What we have not built yet
Single sign-on, an exportable audit trail and customer-specific data agreements are not in the product today. Neither is a certification against any security standard: we are a beta with design partners, and the accurate answer is that the isolation above is tested and the paperwork is not yet done. We would rather write that here than in an answer to a questionnaire.
Still have a question
Send the questionnaire. We answer the awkward rows too.
Request access →