True AI, securely and compliantly,
running in your own cloud.
Every practice and hospital sits on decades of records an AI could read in seconds — and almost none can safely let it. The lock is there for good reason; it is also why care still runs on typing, faxes and hold music. PHI AI opens it without breaking it: a frontier model working on your real charts, in your own cloud, on your own keys. Notes written. Scans read. Denials fought. Questions answered. It sees only what your role allows. Compliance at its core, and purpose-built for AI. Suitable for family practices or large health systems.
What it does, out of the box
Every feature is AI-native. Everything below is running on this site right now, each powered by a managed model in the platform's registry — with live analytics, drift recognition and alerting behind every one.
Retrieval-tuned AI over the record with citations per claim, role-scoped tools, and refusals that name their rule — on the foundation model you bring.
Every pending claim scored before submission, from your own adjudication history — each factor cited to real counts.
Criteria checked against the chart, evidence gathered, and the appeal drafted — into the signature queue, never past it.
Two-directional review: findings the record supports but the codes miss, and billed codes the documentation does not support.
Every release checked against the requesting jurisdiction's requirements; hard failures block with the requirement named.
Records classified at the door — Part 2, mental health, HIV, reproductive, genetic — with low-confidence calls queued for humans.
Any model, any kind: registered, enabled and activated into capability slots through one audited control panel.
Live performance metrics, population-stability drift recognition, and alerts that never close themselves.
Six vendors, bulk and streaming, in and out — each vendor's real seams honored and refusals recorded.
Cohorts, quality measures, no-show risk, trial pre-screening — on a plane that returns counts, never patient lists.
Record a visit from the browser — behind a state-law consent gate — and the capture lands in the Data Store, on the chart, audited.
A genuine vision inference over the pixels: technique, findings, measurements from the acquisition scale — drafted to the queue.
Plain-language instructions written from the chart — then a second model pass fails the draft on a single invented assertion.
A transparent risk score — every factor and its points printed — and an AI call script per patient, staged for a human to dial.
Protocols screened against the whole corpus in SQL; the model reads the charts as a recall-biased second reader that never rules out.
Real numerators over real denominators, computed from the corpus on page load — arithmetic in the open, counts never lists.
The claims model examined across sex, age and payer subgroups — calibration gaps, printed thresholds, an AI reading of the results.
Integrity rules sweep every record in the store; the model argues root causes and fix order — it never quarantines anything itself.
AI output is a draft until a licensed person signs it. Only the signed artifact acquires a write path to the EMR.
The hash chain is re-verified on screen, in front of you — including the refusals, which is the point.
Works with what you already run
Out of the box: the major EMRs in both directions, the three major clouds as your own account, and the frontier models — plus anything you bring.
All product names are trademarks of their respective owners; a listing means PHI AI speaks that vendor's published interfaces, not affiliation or endorsement. Vendor capabilities per their documentation — the seams, honestly, in EMR connections.
Built for PHI. Not adapted to it.
The model reaches data only through role-scoped, audited retrieval tools — the same gates people go through. It receives excerpts, never a database connection, and every answer cites the records it came from.
A hash-chained, append-only audit trail carries every question, every record read, and every refusal. Withheld records are counted and stated, never silently missing.
Population tools return counts and cannot enumerate patients. Unsigned AI drafts have no write path. Psychotherapy notes never export. The refusal is in the shape of the API — not in a policy document.
Deploys into infrastructure you own: envelope encryption under your KMS, and a web tier that makes no third-party calls. You bring your own LLM — any foundation model, under your own BAA with your AI provider — and your organization's identity system governs who sees what.
Any model of any kind joins the registry — foundation, predictive, classifier, imaging, or yours — behind HTTPS and an audited lifecycle.
Two deployment shapes, chosen deliberately: a clinic on one profile, a health system at terabyte–petabyte scale on the other.
Every switch, model, feed and threshold governed from the Control panel — and every change lands on the audit trail under a name.
HIPAA minimum necessary, 42 CFR Part 2, psychotherapy separation — enforced in the architecture, verified live, documented in full.
See it working
Pick a role, ask the assistant a question, try to reach a screen your role cannot see — and then read the audit trail that watched you do it.