Crisis Intervention Protocol (AI Companion Service)

This protocol describes how our AI-assisted chat service responds when a user may be at risk of self-harm or another mental-health crisis.

This page is a deployer-facing template: items marked [to be completed by the operator] must be replaced by the service operator according to their jurisdiction and actual resources.

Conversations in this service are assisted by an AI assistant; a human team can step in at any time. The first reply of a new conversation discloses this to the user.

Every user message runs through deterministic crisis detection (multi-language coverage of self-harm expressions), guarded by permanent automated tests (100% recall required on severe cases, no false alarms on idioms).

Signals are tiered: severe triggers the full intervention flow; elevated injects preventive safety instructions.

Any AI reply that could encourage self-harm is fully overridden by a safe reply before sending (a hard outbound red line, pinned by permanent tests: final output must contain zero self-harm-encouraging content).

On severe signals, the reply includes crisis-support resources (hotlines, next section) and responds with an empathetic, non-judgmental tone.

[To be completed by the operator: human escalation path — when the human team takes over, response SLA, on-call arrangements.]

[To be completed by the operator: list crisis hotline names and numbers for the regions you serve — e.g. national suicide-prevention or mental-health support lines; cover at least your primary service regions.]

The same resources are configured inside the system and are provided to users automatically during severe conversations.

The system keeps persistent counters of crisis handling (severe detections / safe-reply overrides / resource referrals, daily buckets, exportable) so operators can meet annual-reporting duties (e.g. California SB 243 reports starting 2027).

Counters contain aggregate numbers only — never conversation content.

Compliance contact for this site: BOUNDLESS (service provider), messaging @WJKJ2026, email hello@bd2026.cc.

If you deploy this toolkit for your own customers, replace the line above with **your own** compliance contact — the legal duty for disclosure, public notice and annual reporting sits with the service operator (deployer), not the tool vendor.

This template is not legal advice; operators should finalize the content per their local regulations (e.g. EU AI Act Art. 50, California SB 243, New York GBL §1700).