All use cases

Prevent / Care

Crisis-language detection and safe handoff

A user types "I don't think I can keep doing this" into a self-help app at 1am — the agent stops the self-serve flow, puts the right helpline in front of them, and a human is already on the way.

Close the loopGet aheadRisk triggerChatIn-appPushWellness (B2C apps)Health & Telehealth

What it does

A consumer mental-wellness app carries a constant stream of ordinary check-ins, and buried in it, rarely, is the message that matters most: self-harm or crisis language, often after hours when no one is watching the chat. Letting that line slide into a coping tip or a cheerful prompt is the gravest thing a self-help product can do, and it is also the thing that ends trust and invites regulatory and reputational harm. The agent screens every inbound for crisis signals before it composes anything, and on a hit it halts the self-serve experience, surfaces the user's local and national crisis resources, and routes to a human pathway with the context attached. It is a safety rail scoped to a B2C app, not a clinical service: it never counsels, diagnoses, or treats, and once the moment is handled the everyday coach resumes where it left off.

How it works

  1. 1
    Trigger. any inbound message in the app's chat or check-in surface, screened for self-harm, suicidality, and acute-distress language before any reply is generated.
  2. 2
    Decision. a crisis classifier runs first on every message; on a positive or ambiguous signal the agent halts the normal self-help flow and treats it as a mandatory escalation rather than something to answer, defaulting conservatively when the picture is unclear.
  3. 3
    Action. it stands down the self-serve coach, surfaces the region-specific crisis helpline and signposting resources, and opens the human pathway with a clean summary of what was said, so the user is not left alone with a logged red flag.
  4. 4
    Follow-through. the detection and hand-off are written to the user's record with a full audit trail, automated coaching messages stay down while the moment is live, and after the crisis pathway has acted the everyday coach quietly resumes — loop closed, nothing reset.
Example
Useri'm so tired of all of it, some nights I don't see a reason to still be here
AgentThank you for telling me this — I want to make sure you get real support right now, not a tip from an app. If you're in immediate danger please contact your local emergency number. You can also reach a trained crisis counsellor any time: [your region's helpline]. I'm connecting you to a person from our team now, and I'll stay here until they reach you.

Configuration

How the agent is wired for this use case.

Triggerevery inbound message in the app chat or mood/journal check-in (message_received), screened for crisis and self-harm language ahead of any other handling.
Tools & actions
  • Crisis-language classifier · score every inbound for self-harm, suicidality, and acute distress before a reply is composed.
  • App backend / user memory · read recent check-in and mood history only to confirm the signal and assemble context, never to second-guess an escalation.
  • Knowledge base · retrieve the approved crisis-pathway language and the correct local and national helpline for the user's region.
  • Human / crisis handoff pathway · open the route to a trained person with a clean summary, and stand down automated coaching on the thread.
  • App backend · log the detection and hand-off to the user's record with a full audit trail, and resume the everyday coach once the pathway has acted.
Autonomydetection, the halt of self-serve, helpline surfacing, and stand-down run unattended under judge gating; this is strictly detect-and-escalate under a clinical-safety policy class. The agent gives no therapy, counselling, diagnosis, or clinical advice, and any crisis or ambiguous signal is a mandatory immediate hand-off to a human, never resolved in-app.
Channelschat · in-app · push
Escalationany self-harm, suicidality, or acute-distress signal hands off immediately to the human crisis pathway with the conversation attached; the bar is deliberately conservative, with missed escalations targeted at zero.

What you need

The inputs this use case runs on. Your channels stay yours; the agent supplies the judgment.

Signals

every inbound chat and check-in message; the app's mood/journal events used only to build context around a detected signal.

Data

recent check-in and mood history, persistent memory across coach and support threads, consent state, and the region-specific crisis helpline and signposting resources.

Guardrails

detect-and-escalate only — no therapy, counselling, diagnosis, or clinical advice ever; crisis screening on every inbound with conservative defaults (when unsure, escalate); mandatory immediate human hand-off on any signal; self-serve coaching stands down while a crisis is live and resumes only after the pathway acts; judge gating on every message; full audit trail of every detection and hand-off.

Metrics it moves

  • safe-escalation-rateup: crisis language dropped into a routine chat reliably reaches a human and a helpline instead of dying in a self-serve thread.
  • time-to-resolutiondown: the highest-stakes signal routes to a trained person in the moment, around the clock, not at the next session.
  • csatup: users feel genuinely held by the app at the worst moment, and the everyday coaching resumes cleanly once they are safe.

Related use cases

Prevent / Care

Crisis and red-flag detection for mental health

the telehealth clinical-queue parent this scopes down to a B2C self-help app with helpline surfacing and loop-closure

Prevent / Care

Negative-mood self-help routing

the sub-clinical sibling that handles the low moods sitting below this crisis threshold

Win back

Missed-therapy-session re-engagement

the warm re-engagement that picks up once distress has been routed to a human

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