Step-by-step: deflect crisis routing widget questions for…

Before you touch a setting, list out the exact crisis scenarios your service handles — suicidal ideation, self-harm, panic attacks, substance crisis, or immediate safety concerns. For each, define:

• What the user says. Actual phrases people type, not theoretical keywords. Pull them from past intake forms, support tickets, or chat logs.
• What the immediate response should be. Is it a calm reassurance, a hotline number, an instruction to call 911, or a direct handoff to a crisis counselor?
• What information you need to collect. Location? Urgency? Name? If a custom action will trigger your internal on-call system, decide what fields are mandatory.
• Who gets alerted and how. Does a Slack alert go to the crisis team? Should the conversation automatically hand off to a human in the shared inbox? Write a one-paragraph escalation protocol.

Then gather the source material your agent will use: your crisis response policy, a list of local and national hotlines, a self-care guide, and any pre-written scripts for de-escalation. Organize these into documents (PDFs or plain text) that are clear enough for an AI to parse. For Mental Health Services, every document must include explicit safety language, never advise a user to wait, and always default to a human when risk is mentioned.

The output of this planning step is a simple table that maps user intent → response + action + escalation. That table becomes your blueprint.

Build the knowledge base

Create a new Chatref agent and upload those crisis documents. Add your website’s crisis resource pages as URLs if they contain the same protocols. Chatref will learn the material so answers stay grounded in your own content — not generic web search.

Test the agent in the playground with example crisis statements from your planning table. Correct any responses that are vague or that don’t prompt the custom action. Refine the documents if needed (shorter sentences, clear headings, numbered steps). You’re aiming for the agent to consistently recognize a crisis signal and respond with your exact protocol.

Create the custom actions

Under the agent’s Custom Actions settings, add actions that mirror your escalation plan:

• Crisis triage — define fields: crisis_type (dropdown of your defined categories), severity (high/medium), user_location (free text). Set the action to collect these and then hand off the conversation to the shared inbox. Optionally, configure a webhook to post a message to your team’s Slack or trigger an SMS via Twilio (if you have that integration set up).
• Hotline display — a simple action that, when the agent detects a crisis but the user is not at immediate risk, replies with relevant hotline numbers and next steps from your knowledge base, then offers to connect a human.

Make sure the action’s description clearly tells the agent when to trigger it — for example: “Trigger this when the user mentions self-harm, suicidal thoughts, or asks to speak to a crisis counselor immediately.” The agent will use that instruction to decide.

Embed the website widget

From the agent’s Widget tab, copy the embed snippet. Paste it into the HTML of every page where someone might trigger a crisis — your homepage, contact page, resource hub, and any landing pages for at-risk populations. Set the widget’s allowed origin to your domain (e.g., yourpractice.com) so it only loads on your site.

Customize the widget launcher with a calm, accessible label like “Need immediate help?” and use a color that signals safety. In the agent’s settings, enable the shared inbox so your crisis team can join any conversation directly from the dashboard.

Internal dry run

Before going live to patients, put the widget on a staging site or a password-protected page and have your crisis team test it. Run through at least 20 scenarios from your planning table — typed exactly as a person might. Check that:

• The agent recognizes a crisis and triggers the correct custom action (crisis triage vs hotline display).
• The shared inbox receives the handoff with full chat context and the collected fields.
• The team can respond inline without losing the thread, and their replies show up in the widget in near-real time.

Adjust agent instructions and document phrasing until the team agrees it’s safe.

Soft launch

Turn on the widget for a limited window — say, your 24-hour crisis line’s quietest hour — and monitor the inbox. Let the team handle handoffs with a buddy system: one person watches the dashboard, one is on standby to answer. After a few consecutive sessions, expand to normal hours.

Keep an internal checklist: if the agent ever fails to hand off when it should, pause the widget and diagnose. Common causes are unclear document structure, missing crisis phrases, or the custom action’s trigger description not matching the user’s wording. Iterate on those, then re–test.

Team readiness

Schedule a 20-minute walkthrough for every staff member who will use the shared inbox. Cover how to see ongoing conversations, join a thread, see the collected action fields, and escalate to another team member. Stress that they’re always in control: the agent defers to a human the moment they join.

What to track

• Deflection rate: what percentage of crisis-intent conversations are handled end-to-end by the agent (i.e., answered with a hotline or resource) vs. handed off to a human? A low deflection rate on deadly-serious issues is fine; the goal is to never miss a handoff, not to automate it completely.
• Handoff accuracy: for every escalated conversation, did the agent collect the fields you needed and pass them to the inbox without losing info? Measure the percentage where the counselor did not have to re-ask for location or crisis type.
• Response time: the time from crisis statement to agent reply (or handoff). Under 15 seconds is the target. If the handoff takes longer, check the shared inbox notification settings for your team.
• User feedback: add a simple post-chat rating (“Did you get the help you needed?”) to spot conversations that felt unsafe or confusing.

Use Chatref’s insights to surface the top crisis phrases and see which ones the agent missed. Update your knowledge base documents accordingly, and refine the custom action triggers.

Adjusting the playbook

After the first month, compare the team’s workload before and after. Did the number of late-night crisis calls that converted to a human decrease because the agent provided resources directly? Or did more people request a counselor because they now trusted the widget? Either outcome informs your resourcing.

Revisit your planning table quarterly, especially when protocols change (new hotline numbers, updated intake forms). Sync the documents, re-test the agent, and retrain the team. That closed loop is what keeps a crisis routing widget trustworthy over time.

What causes crisis routing widget problems for Mental Health Services?

Problems usually come from three places: the knowledge base lacks precise crisis language (e.g., it uses clinical terms when users type “I don’t want to be here”), the custom action triggers are too narrow (missing variations of a plea for help), or the handoff pipeline isn’t tested regularly with real user phrasing. Additionally, if the widget is not available on every critical page — resources, contact, intake forms — someone in crisis won’t find it.

How do I improve crisis routing widget for Mental Health Services?

Start by reviewing failed handoffs in Chatref insights and expanding your documents with the exact language users employed. Widen your custom action triggers to cover more phrasing. Add a fallback: if the agent is unsure of severity, it should always escalate to the shared inbox. Run monthly tabletop tests with your team using fresh user language, and update the escalation protocol whenever your on-call process changes.