For your InfoSec team
Our answers to the questionnaire you were going to send us
The security questionnaire is usually the slowest step in a healthcare purchase, and our answers do not change depending on who is asking. So they are here, including the ones that are no. A vendor response with nothing but green ticks is not reassuring to anybody who reads these for a living.
Where the AI runs
Does our data go to OpenAI, Anthropic, Google or any other AI provider?
No. Every model call — the language model, the embedding model and the reranker — runs on hardware we operate. There are no third-party AI provider credentials in the environment at all, which is a stronger statement than a policy: there is nothing configured that could send your content to one, even by mistake.
Do you train or fine-tune models on our content?
No. The models are run, not trained. Your documents are indexed for retrieval within your own organization and are never used to adjust model weights, and never leave your organization's boundary to inform anyone else's answers.
Is the product multi-tenant, and how is our data separated?
Multi-tenant, with separation enforced in the database by row-level security rather than by application code remembering to filter. Retrieval is additionally scoped by role, so a document class can be restricted to the staff entitled to see it.
Subprocessors and data flow
Who are your subprocessors?
Two, plus one that is optional. Resend handles transactional email. Cloudflare provides DNS, CDN and WAF in front of the application. Microsoft Graph is involved only if you switch on SharePoint synchronisation, which runs against your own tenant with your own consent and can be left off. Hosting is our own infrastructure, not a hyperscaler, so there is no cloud provider in the list.
Does any content leave your infrastructure?
For a narrow set of medical-coding questions the assistant can attach links to public reference sites such as AAPC or the AMA. Where a lookup is performed it goes through a gateway on our own infrastructure — there is no third-party scraping or search service in the path. Your documents are never sent anywhere.
Will you sign a BAA?
Yes, and BAA execution and tracking is a feature of the product rather than a filing cabinet exercise. Ask the harder questions too, of us and of everyone else: where inference runs, what is retained, whether prompts are logged, and what happens on termination.
What we retain, and why
Do you log the questions our staff ask?
Yes, deliberately, and this is a feature rather than a side effect. A question the knowledge base could not answer is a precisely located gap in your documentation, described in the words one of your own staff actually used. Your administrators see that list and close the gaps, and the corpus gets better because of it. A system that discards its failures throws away the most useful thing it produces.
Could a question contain PHI?
It could — somebody can type anything into a box. That is why question logs are treated as a protected data class in their own right, visible only to your own administrators, and never pooled across organizations or used to inform anyone else's answers.
What does the audit trail cover?
Identity, the query, the documents returned, and the timestamp — the query itself being the part most audit logs omit, which is what makes them impossible to reconstruct from later. Role changes and document access are recorded alongside it.
The things we do not have
Do you hold a SOC 2 Type II attestation?
No. We are self-hosted, so there is no host attestation to inherit either, and any vendor telling you they are "built on SOC 2 infrastructure" is describing their landlord rather than themselves. We would rather say this on a public page than have you find it in week six of an evaluation.
Do you hold HITRUST certification?
No.
Does the product make clinical decisions?
No, and it is scoped so that it cannot drift into doing so. It answers operational questions — which form, which policy, which escalation path, what the documented procedure is. Anything touching diagnosis or treatment selection is a different product category carrying a different regulatory burden, and a vendor blurring that line should worry you.
Access and identity
How is access controlled?
Role-based, with organization administrators managing their own users. Multi-factor authentication is available and can be made mandatory for an organization.
Can we restrict which documents a role can retrieve?
Yes. Retrieval is role-scoped, not just the document list — a role that should not see a class of material does not get it in an answer either, which is the route most systems leave open.
Anything not covered here
Business continuity, retention periods, termination and data export, and penetration testing are handled per engagement and provided on request — they depend on how you deploy and we would rather give you the real answer for your situation than a generic one on a web page.
Ask us the ones we have not answered