Skip to content

Keeping PHI on-prem

Faxart runs OCR and language-model inference on hardware you own. Cloud LLM APIs are not used for fax content, even with a business-associate agreement in place. This page explains why that rule is stricter than HIPAA requires, on purpose.

To read a page, the page image has to reach the compute node. A fax archive is PHI. So a rented cloud GPU would put PHI off-premises. HIPAA would permit that under an agreement; Faxart does not. The cheapest, most durable way to keep PHI on-prem is to never let it leave in the first place, rather than to rely on a contract and a vendor’s controls.

There is a practical corollary: for the kind of degraded fax this system reads, OCR is bound by the page detector, not by raw compute, so a rented high-end GPU buys almost nothing. The biggest compliance cost would have purchased one of the smallest speedups. On-prem is both the safer and the cheaper choice here.

HIPAA wants PHI reads logged, not just writes. An application-layer audit log captures who clicked “download” with useful context, but it can be bypassed by anything that talks to the database directly. So Faxart uses two layers:

  • A database-level audit floor that logs reads even if the app is bypassed. This is the boundary that cannot lie.
  • Application-layer audit for the human-readable who and why overlay.

Together they give you a record that is both complete and legible.

A page that cannot be classified, a delivery that fails, a print that jams: each becomes a visible, auditable state (held, quarantined, failed), never a silent loss. For a paper-only workflow a lost print is a lost fax, so it surfaces like any other failed delivery.