How we keep clinical answers safe
Trust is the whole product. Below is exactly what Primary AI does to make answers safe to act on — from verifying that every citation is real to protecting the data that flows through the Service. We only claim what is actually true; we do not hold certifications we have not earned.
Our commitments
Private Assistant history
Clinical questions must be de-identified by default. Authorized clinicians can use patient information in Primary AI Assistant. Approved Canadian Business API customers can use a separate fail-closed PHI path. Assistant transcripts and notes are private to the user and deleted after 30 days.
Citations verified against live lookups
Every reference is checked against a live PubMed and DOI lookup so the paper actually exists before it is shown to you — the single biggest failure mode of medical AI, engineered against directly.
No-fabrication guardrail
A hard guardrail instructs the model to refuse rather than invent a citation when it cannot find a real one. We would rather say 'no source' than show you a fake one.
Evidence grading
Recommendations flag the type of evidence behind them — RCT, meta-analysis, guideline, or expert consensus — so you can weigh them at a glance.
Safety highlights surfaced, not buried
Life-threatening risks, contraindications, and black-box warnings are surfaced explicitly rather than hidden in prose.
Encrypted in transit
The app uses HTTPS for service connections. Access controls restrict access to saved records.
Private Assistant records
Database access rules restrict Assistant records to their owner. Service systems also process records to provide the requested features.
Named AI providers
OpenAI generates responses, Google Gemini interprets clinical questions, and Voyage AI creates text embeddings for retrieval. The Privacy Policy describes additional transcription and search providers.
Data handling in plain language
Primary AI follows a principle of data minimization: we collect only what is needed to operate and improve the Service safely. Our Privacy Policy describes retention, provider processing, advertising uses, and your choices.
Clinical questions must be de-identified unless they use a separately approved Canadian Business API PHI path. Approved PHI requests fail closed unless the customer agreement and every provider control are active. Their question and answer are not written to application query logs. Assistant database access is restricted to the record owner. Records expire after 30 days and are removed by an hourly job. Backup and provider retention are separate. Full detail is in our Privacy Policy and Terms of Service.
Reporting a vulnerability
We welcome responsible disclosure. Email rishad@primaryai.world or see our security.txt. We acknowledge valid reports.