Assistive AI, with the clinician in control.

Server-side only — keys never in the browser

AI runs behind the server, so model credentials are never exposed to the client or the patient's device.

• Model API keys remain server-side, never in the browser
• AI calls are brokered by the server, not the client
• PHI sent to models is governed by the same access controls as the record
• No third-party model key ever ships to a device

Assistive drafts & flags for clinician review

AI proposes; the licensed professional disposes — nothing AI writes becomes part of the record without human review.

• AI-assisted summaries, drafts, alerts, and suggestions for clinician review
• Clinical decisions remain with licensed professionals
• Output is a starting point a clinician edits, accepts, or discards
• AI is never an autonomous decision-maker

Patient-facing AI is non-diagnostic & non-prescriptive

Companion's patient-facing assistant supports recovery conversations without diagnosing or prescribing.

• Non-diagnostic and non-prescriptive by design
• Grounded in the patient's care plan, not open-ended medical advice
• Routes questions and concerns to the care team
• Directs emergencies to local emergency services

Human-in-the-loop coding

AI helps capture the documentation elements relevant to downstream coding review — it does not code the visit for you.

• Captures documentation elements relevant to downstream coding review
• The E/M level is not surfaced to the provider during documentation
• Coding decisions stay with qualified billing and coding staff
• Supports revenue-cycle workflows; does not guarantee coding outcomes

Audit of AI access

Every AI touch of protected health information is recorded on the same tamper-evident trail as the rest of the record.

• AI access to PHI is written to a hash-chained audit trail
• Who, what, and when are captured for review
• The same tamper-evident audit layer as all PHI access
• Row-Level Security governs what any AI process can read

Server-side call

The request is brokered by the server, so the model API key never reaches the browser or the patient's device. What any AI process may read is governed by Row-Level Security.

Assistive result

The model returns a draft, summary, suggestion, or flag — a starting point for a person, never a clinical decision.

Clinician review

A licensed professional edits, accepts, or discards the output. Nothing the AI produces enters the record unreviewed.

Audited access

The AI's access to PHI is written to the hash-chained audit trail — who, what, and when — on the same tamper-evident layer as every other PHI access.

Forge — scheduling & revenue cycle

• AI pre-visit intake returns a clinician summary, suggested ICD-10, and red-flag triage that pre-fills the encounter for the clinician to confirm
• No-show risk and smart-scheduling slots are suggestions the scheduler weighs, not automatic actions
• AI-drafted appeals and patient-proposal narratives are reviewed by billing staff before anything is sent

Revela — surgical safety

• The AI surgical-safety screen is built on validated risk scores — Caprini, STOP-Bang, RCRI, ASA — surfacing graded safety flags for the surgeon to review
• Contraindications, drug interactions, and missing pre-op items are flagged; the surgeon is the decision-maker and the system is the safety net
• Server-side AI drafts notes and PROMs insights; PHI is scrubbed server-side before it reaches a model

Mind — rating-scale alerts

• The longitudinal rating-scale engine scores validated instruments and fires safety alerts — for example, PHQ-9 item-9 to a care-team alert
• Scale interpretation and clinical decisions stay with the clinician; the engine surfaces, it does not decide
• The E/M level is never displayed during documentation — it is computed downstream at the billing layer

Continuum — perioperative drafting

• A server-side Edge function drafts the operative note in real time as documentation progresses, with dictation cleanup
• ICD-10-CM and CPT coding suggestions and drug-interaction checks are presented for review, not applied automatically
• AI surgical flags augment standard calculators; the Universal Protocol time-out and clinical judgment remain the surgeon's

Companion — patient-facing assistant

• Non-diagnostic and non-prescriptive: it never diagnoses and never changes medications or doses, offering general education grounded in the patient's own care plan
• Patient-reported check-ins surface red-flag alerts to the care team when thresholds are crossed; emergencies are directed to local emergency services
• AI document extraction is gated by patient review — the patient edits or soft-voids with a required reason before anything is accepted

Profiler — intake summaries

• Adaptive AI-branching intake is condensed into a draft summary the clinician reviews
• Suggested codes and triage flags pre-fill the encounter, but the clinician confirms before anything is relied upon
• Runs on the shared record, so the same server-side and audit guardrails apply to intake