Where AI speaks — and where it deliberately does not

The AI does not ask your scored questions

When your clinic assigns forms or instruments — your own intake paperwork, PHQ-9, GAD-7, AUDIT, or any of your templates — a deterministic form engine runs the conversation. Validated instruments are administered item by item in their published wording and order, the way they are meant to be administered. A language model is used for exactly three things: mapping the patient's natural answer onto your response options (“pretty much every day” becomes Nearly every day), explaining a question in plain words when the patient asks, and answering the patient's own questions — about confidentiality, for example.

• Every answer carries the patient's exact words. The chart receives the clean structured value and a verbatim quote — validated in code so a paraphrase can never masquerade as a quote. The full transcript is preserved besides.
• Answers are validated before they can reach the chart. A choice answer must be one of your actual options; dates and numbers must come from the patient's words — never inferred; a denial is never recorded as the thing denied (“no allergies” can never become an allergy list); information about another person never fills the patient's own field.
• Declining is respected and recorded. “I'd rather not say” marks the item declined and moves on — you see exactly which items were declined.
• Progress honesty. A patient who asks how much is left gets the real count; a tired patient is offered — once, not repeatedly — the option to finish another time, with everything saved.
• Identity data is never re-asked. What your chart already holds (name, date of birth) is pre-filled; contact and insurance on file are read back once for confirmation rather than re-typed through chat.

Interview & daily sessions: AI phrases, deterministic logic governs

In the pre-session interview, a conversation model is appropriate — a patient disclosing something difficult needs a human-feeling follow-up, not a template. Its role is strictly bounded: coverage of the 13 clinical areas is tracked deterministically, a separate extraction step records what was actually answered, and plain code decides what to cover next and when the interview is complete. The closing of every session is a fixed, composed message — the system never generates a free-form recap of what the patient said, because a generated recap could misstate it. You read the actual record, not an AI retelling. Between visits, provider-directed sessions produce the mood trajectory and daily-life picture that, with the intake, give you a two-time-scale view before each appointment. Crisis signals in any session reach your EMR immediately.

Safety runs on rules, not model judgment

A deterministic, category-aware risk gate screens every patient message in every mode — separate from any language model — distinguishing self-harm risk, harm to others, abuse, and medical emergencies, each with its own protocol. Once flagged, a safety posture persists for the rest of the session. A standing capability-honesty rule prevents the system from claiming abilities it does not have — even to comfort. Full protocol: Safety & Crisis Protocol.

Chart safety: six layers between a conversation and the record

Each layer is enforced in code, not entrusted to model behavior. An answer passes through all of them, in order:

1. 01Extraction rules that forbid guessing, inference, paraphrase, and stretching one answer across multiple fields.
2. 02Deterministic value validation — option membership, type and format checks, contact-field shape checks.
3. 03The patient’s verbatim words attached to every recorded answer.
4. 04A question guard: no off-script question can reach a patient during forms.
5. 05Composed, never generated, session closings — no AI retelling of clinical content.
6. 06Scores computed arithmetically and flagged for provider review.

The standing quality bar, verified by an adversarial test suite before any change ships: zero misdirected questions and zero wrongly-recorded answers — including under typos, out-of-order answers, declines, and interruptions. Each new form family your clinic introduces passes the same gate before going live for your patients.

Your data stays on dedicated infrastructure

Patient-facing conversational AI runs entirely on ReasonCare-operated, self-hosted models — clinical conversations are not sent to third-party AI services. Integration is HIPAA compliant, built on SMART on FHIR, and native to ReasonEMR: enrollment, form assignment, and results all flow through your existing chart, and you remain the decision-maker throughout — the system collects, structures, and faithfully transmits; it does not diagnose, treat, or interpret.