Prevent / Care
Crisis and red-flag detection for mental health
Someone types "I don't see the point in any of it anymore" into a wellbeing check-in at midnight — the agent stops, and a trained human is reaching out within minutes, not at the next appointment.
What it does
In a mental-health conversation, the signals that matter most are the ones a person drops in passing, often at the hours when no clinician is online — a line of hopelessness in a journaling prompt, a mention of self-harm inside a routine check-in, a mood that has been quietly sliding worse across weeks. Missing one of those is the gravest failure a wellbeing service can make, yet the same surface also carries a constant stream of ordinary, low-risk messages that don't need a clinician at all. The agent listens across the conversation and the member's check-in history, recognizes crisis language or a deteriorating trajectory, and escalates immediately to a human and the crisis pathway with a clean summary, while routine concerns stay in the everyday coaching conversation. It is a safety rail, not a therapist: it never counsels, advises, or attempts to treat — it detects and hands off.
How it works
- 1Trigger. a message in any wellbeing conversation, or a risk pattern surfacing across the member's logged check-ins — a worsening mood trajectory, a sudden drop, repeated distress cues — with no separate screening step required.
- 2Decision. crisis- and self-harm-language detection runs on every inbound before anything else, and the agent weighs the single message against the member's history: an acute crisis cue, a worsening trajectory, or an ordinary low-risk concern. Defaults are conservative — when the picture is ambiguous, it escalates.
- 3Action. an acute risk signal skips all self-serve, surfaces the crisis-pathway resources, and routes immediately to a human with a clean, sourced summary; a worsening-trajectory pattern is handed to the clinical team for review; ordinary concerns are acknowledged and stay in the everyday coaching conversation. Automated replies stand down on any thread a human now owns.
- 4Follow-through. every escalation carries the full picture so the member never has to repeat themselves in a fragile moment; the clinician's outcome writes back to memory; the boundary holds throughout — the agent offers no diagnosis, no counselling, and no clinical advice at any point.
Configuration
How the agent is wired for this use case.
message_received) screened for crisis and self-harm language, plus mood_trajectory and check-in signals from the wellbeing app that flag a deteriorating pattern.- Crisis-language classifier · score every inbound for self-harm, suicidality, and acute-distress cues before any reply is composed.
- Member memory / check-in history · read the mood and check-in trajectory to distinguish a worsening pattern from a one-off low.
- Clinician / crisis escalation queue · route an acute or worsening signal to a human immediately with a clean, sourced summary and surface the crisis-pathway resources.
- Knowledge base · retrieve the approved crisis-pathway and signposting language for the member's region.
- Care-record system · log every detection and escalation decision to the member's record with a full audit trail.
What you need
The inputs this use case runs on. Your channels stay yours; the agent supplies the judgment.
Signals
every inbound wellbeing message, mood and check-in events, and a trajectory flag when a member's pattern deteriorates over time.
Data
the member's check-in and mood history, prior risk signals, persistent memory across coaching and support threads, consent state, the region-specific crisis-pathway and signposting resources.
Guardrails
the agent detects and escalates only — no therapy, counselling, diagnosis, or clinical advice ever; crisis- and self-harm-language detection on every inbound with conservative defaults (when unsure, escalate); mandatory immediate human hand-off on any acute or worsening signal; automated replies stand down once a clinician owns the thread; judge gating on every message; full audit trail of every detection and escalation decision.
Metrics it moves
- safe-escalation-rateup: crisis cues and worsening trajectories dropped into ordinary conversations reliably reach a human instead of dying in a chat log.
- time-to-resolutiondown: an acute signal routes to a trained human in the moment, around the clock, rather than waiting for the next appointment or office hours.
- csatup: members feel genuinely watched over by the service, and routine concerns are handled without friction while the highest-stakes ones get a person fast.
Related use cases
Clinical red-flag detection and safe escalation
the medication-program parent this adapts to mental-health crisis signals
Between-session engagement for therapy and mental health
the engagement layer whose every distress signal hands off to this rail
Clinician copilot — approve instead of writing
what the care team works from once a case is escalated
See it on your own customer journey
Bring one drop-off, one churn cliff, or one silent segment. We will show you what a proactive agent with memory and judgment does with it.
Book a demo