HealthcareIndustries and use cases
AI Chatbot for Mental Health Clinics: Where the Line Is
A chatbot on a mental health website has one job: remove the practical barriers to getting started, then get out of the way the instant a conversation turns clinical. This post starts with the rule that matters most.
Read this part first
Any message that suggests crisis, self-harm or thoughts of suicide must route to a human immediately, and the reply must show the local emergency and crisis numbers your clinic publishes. Not after a clarifying question. Not after the bot decides how serious it sounds. Immediately.
A bot must never be the only responder to a message like that. If your clinic cannot guarantee a person is reachable at that hour, the widget has to say so in plain words and point to crisis services that are staffed around the clock. Build this before you build anything else, then test it yourself.
Why detection alone is not a safety plan
Every vendor will tell you their model spots crisis language. Treat that as a helpful extra, not as the plan. People in distress do not always use the words a filter looks for. They write sideways, they understate, they ask about something else and mention the real thing in the last line.
So do not rely on detection as the only trigger. Put your crisis resources somewhere permanent: in the widget from the first message, on every page, and in the automated reply people get outside staffed hours. Nobody should have to phrase a message correctly to see that number.
Then set escalation to be generous. An unnecessary handover costs a staff member two minutes. A missed one is not a metric you want to learn from. The general pattern is in our handover protocols guide, but here err much further towards the human.
What the bot must never do
Write these rules into the bot instructions first, before any feature work:
It is tempting to let the bot stay warm and stay in the conversation, because that feels kind. It is not. Warmth that delays someone reaching a human is harm dressed as good manners. Acknowledge briefly, hand over, show the numbers. Our chatbot guardrails guide covers how to write and test rules that hold under pressure.
Say it is a bot, in the first sentence
This matters more here than anywhere else. The opening message should say it is an automated assistant, that it helps with fees, availability and getting started, that it is not a therapist or a crisis service, and when a human is available. People trust a clearly labelled bot. What breaks trust is an assistant that reads like a person until the moment someone needs one.
The questions it can genuinely help with
Once the boundaries are set, real value remains. Getting started with therapy is full of small practical unknowns, and those unknowns stop people. Many feel awkward to ask a receptionist by phone, which is why they get typed into a chat window at 9pm.
Language matters too. The AI runs GPT-4o and Claude on every plan and replies in many languages, detecting the visitor's language from what they type. For a clinic serving a mixed community, that removes a barrier no amount of front-desk goodwill can.
Booking capture without becoming a records system
Keep the request small. A name, a way to reach them, new or returning, in-person or online, and two or three times that suit. That is enough for your intake coordinator to confirm in one message.
Do not ask why they are seeking therapy, and do not ask about diagnosis, medication or history. Those answers are patient health information the moment they are typed. Collect what you need to make contact, and let the clinical conversation happen with a clinician.
PHI, the BAA, and being honest about plans
Keep patient health information out of the chat and hand clinical questions to your team. If your use involves client records, sort the agreement before you write a single bot instruction.
Realistic volume and cost
A single-site clinic sees a modest, evening-weighted trickle of chats rather than a flood. The free plan covers 500 messages a month, roughly three a day, with 2 seats, one chatbot, the mobile app. Paid plans are Starter at $25 a month and Professional at $95 a month, with 20% off on annual billing. Annual billing takes 20% off; the ladder is on the pricing page.
Frequently asked questions
What must happen when someone mentions self-harm or crisis?
It goes to a human immediately, and the local emergency and crisis numbers you publish appear in the reply straight away. The bot does not ask follow-up questions, does not assess risk, and does not try to keep the person talking. A bot must never be the only responder to a message like that. If your clinic cannot guarantee a human is reachable at that hour, the widget must say so plainly and point to crisis services staffed around the clock.
Can a chatbot do intake screening or a symptom questionnaire?
No. Screening tools produce scores people read as clinical findings, and interpreting them is a clinician’s job. The answers are also patient health information the moment they exist. Keep the widget to booking and practical questions, and run screening inside the clinical system your clinicians already use, with a person attached to the result.
Is it dishonest to run a bot on a mental health website?
Only if you hide it. Say in the first message that this is an automated assistant, that it answers questions about fees, availability and getting started, and that it is not a therapist and not a crisis service. State when a human is available and what to do outside those hours. People forgive a clearly labelled bot. They do not forgive one that pretends to be a person and then fails.
What can it genuinely help with?
The practical questions that stop people getting started. Fee ranges and whether you offer a sliding scale, which insurance you accept, whether there is a waitlist, whether you offer evening or telehealth appointments, what happens in a first session, how confidentiality works, parking, and what to bring. These are the questions people are too embarrassed to phone about, and answering them removes a real barrier to care.
What does it cost and how much volume should we expect?
A single-site clinic usually sees a modest number of chats a day, weighted to evenings. The free plan covers 500 messages a month, about three a day, with 2 seats, the mobile app. Teams and Slack work on every plan, Free included, which is how escalations reach staff fast. Keep patient health information out of the chat and hand clinical questions to your team.
Answer the practical questions. Escalate everything else.
EzyConn trains on your clinic website and pushes handovers to the mobile app on every plan, including free. 2 seats, 500 messages a month, no card.
Start freeLast updated . Nothing here is clinical, legal or compliance advice. If you or someone you know is in crisis, contact your local emergency number or a crisis line now. Follow your own clinic policy and your regulator's guidance, and sign a BAA before patient health information touches a chat widget. View more guides.