AI Safety Policy.
Last reviewed · October 17, 2026
Summary
Behavioral health AI carries safety obligations that horizontal AI does not. DIAL3D's safety architecture addresses: AI disclosure to patients, hallucination prevention and mitigation, crisis-line protocol activation, human-in-the-loop guardrails on high-stakes decisions, and ongoing quality monitoring.
This page describes our standing safety commitments. We update it when our practice changes; the last update date is at the top.
AI disclosure
DIAL3D's voice and chat agents identify themselves as AI in the first ten seconds of every interaction. Patients can ask to speak with a human at any point during the conversation, and the agent connects them or takes a message, following the routing your facility sets.
The agent does not impersonate a human under any circumstance. If a patient explicitly asks whether they are talking to a human or an AI, the agent confirms it is an AI.
Hallucination prevention
DIAL3D's clinical and operational responses are grounded in a structured knowledge base configured during onboarding. The agent does not generate free-form clinical advice or diagnostic statements.
When the agent does not know an answer, it says so and warm-transfers to a human. We measure 'do not know' rates as a quality metric — too low means the model is confabulating, too high means the knowledge base needs expansion.
Crisis-line protocol
Active-risk language (suicidal ideation, overdose risk, immediate harm) triggers DIAL3D's crisis-line protocol. The agent stays on the line, drops to a slower trauma-informed cadence, and connects the caller with help: your designated team where your line transfers calls, and 988 otherwise.
Crisis detection uses patterns written for substance use presentations, not generic risk language, and runs on every turn. We treat a missed crisis call as the most serious quality failure.
Human-in-the-loop guardrails
Several decision categories require human handoff regardless of AI confidence: clinical fit determination for complex cases, single-case agreement quotes, custody/family-law-adjacent intake, and any explicit request for a human.
Counselors remain the decision-makers on whether to admit a specific patient. DIAL3D's role is qualification, intake, and warm transfer — not admission decisions.
Model monitoring
We test changes against a fixed set of known scenarios before release, and we can roll back to the prior version.
Customer-specific scripts and knowledge bases are versioned. Customers can review and revert changes.
Bias + fairness
Conversational AI can encode bias from training data.
We do not screen out patients based on protected-class attributes. Qualification floors are based on payor, level of care, and clinical fit, not demographic factors.
Voice cloning consent
DIAL3D offers one-time voice cloning of an existing team member with their written consent. We never clone a voice without explicit, documented consent from the individual.
Cloned voices are limited to the customer who provided the source recording. We do not reuse cloned voices across customers.
Need this document, a security questionnaire, or a deeper conversation?
Need our AI safety documentation, crisis-protocol specifics, or to discuss a specific safety question? Our safety team responds within one business day.
Contact our compliance team