This is the first thing every behavioral health buyer asks, and the answer is architectural rather than a promise. The agent stays on the operational side of the line. It does not diagnose, assess, or advise on treatment, and it will not imply that a visitor can request a specific provider when your process is to match them. It answers only from content you approve, and when a question falls outside that content it says so and hands off instead of improvising. You control that content line by line, including removing individual pages you do not want it drawing on. Three moments carry the real risk: crisis response, coverage qualification, and identity checks. All three run as deterministic flows you define once and review as written. Before launch you test the agent against your own worst-case inputs and turn those into standing evaluators you rerun on every change.

A deterministic crisis protocol runs. You define the trigger language and the exact response: the agent surfaces the 988 Suicide and Crisis Lifeline and any crisis resources you specify, stops the intake questions, and escalates to a live person immediately with the full conversation attached. It runs the same sequence on web, WhatsApp, SMS, and email, and at 3am exactly as at 3pm. Your clinical team signs off on one protocol and knows what a visitor will be told.

Yes, with a BAA available, alongside SOC 2 Type 2, ISO 27001, and GDPR. Because someone can disclose a condition or an intention in a chat on a public page, you also get role-based access, retention periods you control, a full audit trail per conversation, and explicit control over what the agent can access and store.

It books. A branded booking card appears inside the conversation with live availability from your hosts' Google Calendar or Outlook Calendar accounts, and the person picks a time without leaving the thread. You set duration, buffers, minimum notice, booking window, working hours, and time zones, and distribute across your team by round robin, least busy, or priority order. The agent collects the qualifying details you need during the booking itself and does not re-ask for anything already given. The attendee gets a calendar invite and a confirmation email, and the host gets a notification. Plenty of teams start with routing to a phone number or a form and switch the calendar on later, which is a perfectly normal phase one.

That is the right way to judge this. Conversion funnels run from visitors through conversations to qualified completions, lead sources break out by UTM source, campaign, and landing page with a conversion rate on each, and every trend compares to the previous period. A goal only counts when its trigger actually succeeded, so a booking is recorded when the record wrote through rather than when the agent claimed it. Every number opens down to the individual conversation behind it, and anything on screen exports to CSV. You can take conversion by campaign, service line, and location to a budget review and defend the spend with appointments.

Yes. Qualified inquiries and bookings write into Salesforce or HubSpot as structured records with the transcript, the originating page and campaign, and a distinct lead source, so they stop collapsing into a generic form fill. Qualification outcome lands in real fields your admins can report and route on rather than in a free-text notes box. Partial captures count too: if someone shares a name and number and then leaves, that record still exists, where a form would have kept nothing. Anything beyond Salesforce and HubSpot, including your EHR or practice management stack, connects through APIs and webhooks scoped to your account. And if you have not chosen a CRM yet, the agent works before and after that decision.

Standing up a chat box on a model is the easy part, and it is why so many teams already have one nobody uses. The hard part is everything after: deterministic crisis and qualification flows your clinical and legal teams will sign, evaluation that proves how the agent behaves on the worst inputs before a patient sees it, one conversation that follows a person across web, email, WhatsApp, and SMS, audit trails and HIPAA posture, booking against real calendars, attribution into your CRM, and human takeover in the same thread. That is what Tars has spent 8+ years on, and you get implementation and guardrail setup from people who have done it in behavioral health. You go live in weeks, not over several quarters of your own roadmap.