An AI phone agent for the therapy practice's front office — never for the therapy

A counseling practice's phone rings while every clinician in the building is in session. New clients call to ask whether the practice takes their insurance and has openings; current clients call to reschedule, ask about a superbill, or cancel inside the fee window; and everything lands in a voicemail box that gets cleared between sessions. The prospective client who reaches voicemail often just calls the next practice on their list.

ThunderPhone gives an outpatient mental-health or counseling practice an AI phone agent for the administrative side of that phone line: intake inquiries, waitlists, scheduling, paperwork reminders, and billing routing, following scripts the practice writes. It is a front-office worker and nothing more. It never counsels, never assesses a caller's state, and never discusses anyone's treatment — and before any of that matters, it carries the practice's crisis script at the top of every call.

This is not a crisis line — and the configuration must say so first

An AI phone agent is not a crisis service, and no therapy practice should deploy one as if it were. The first rule written into a therapy-practice agent — before any scheduling skill, before any greeting polish — is the crisis path: if a caller expresses crisis, thoughts of self-harm or harm to others, or any safety concern, the agent immediately stops the administrative conversation and delivers the practice's configured crisis response. For many practices in the United States, that configured script directs the caller to call or text 988 or to call 911 or go to the nearest emergency room — but that wording belongs to the practice, not to the platform; the practice's clinicians decide what their callers should hear, exactly as they already decide what their after-hours voicemail greeting says. Alongside the script, the practice configures an immediate escalation path: a transfer to the on-call clinician or answering line where one exists, or whatever handoff the practice's protocol specifies.

The division of labor is absolute. The practice defines the trigger phrases and the response; the agent matches and follows. It never counsels the caller, never asks assessment questions, never weighs how serious the situation sounds, and never continues booking an appointment first. There is no judgment call for the software to make — a trigger fires, the script runs, the escalation happens, immediately. This boundary is repeated in the guardrails and FAQs below because it governs everything else on this page: a practice that is not prepared to write and rehearse its crisis script should not put an AI agent on its phone line.

What a therapy-practice agent can be configured to do

Every workflow below is administrative. The agent follows practice-written scripts, collects the fields the practice requires, and escalates on practice-defined triggers — with the crisis path active in every one of them.

New-client inquiry intake

The call that grows a practice is the one that usually hits voicemail. The agent can answer it immediately, walk through the practice's own inquiry form — name and contact details, insurance carrier or self-pay, the availability windows that work for the caller, preferred modality or clinician type where the practice's form asks (in-person or telehealth, individual or couples, and so on), and how the caller found the practice — and deliver a structured record to the intake coordinator. It states plainly that a clinician or coordinator will follow up to confirm fit; matching a client to a therapist is the practice's decision, never the agent's.

Waitlist management

When the calendar is full, the agent can offer the waitlist instead of a dead end: collect the caller's details and availability windows, set expectations using the practice's approved wording, and log the entry. Outbound campaigns can work the same list in reverse — when a slot opens from a cancellation, the agent can call waitlisted clients in the order the practice specifies and book the first taker, so cancelled hours get refilled instead of sitting empty.

Appointment scheduling, rescheduling, and reminders

Connected to a calendar or practice-management system that permits it, the agent can book, move, or cancel sessions during the day or after hours, and confirm the result back to the caller. Outbound reminder campaigns can call ahead of upcoming sessions, confirm attendance, and capture a reschedule request on the spot rather than discovering it as a no-show. Reminder wording is the practice's — including how much, or how little, the reminder says about the nature of the appointment.

Cancellation handling per practice policy

Most practices have a written cancellation policy — a notice window, a late-cancellation or no-show fee, exceptions the practice honors. The agent can process a cancellation, read back the relevant policy exactly as the practice wrote it, record the reason the caller gives, and route disputed or exceptional cases to a person. It applies the policy's wording; it does not negotiate fees or grant exceptions unless the practice's rules say it may.

Insurance and superbill questions, routed

For general questions the practice has approved — which plans the practice is in-network with, whether it offers superbills for out-of-network reimbursement, how self-pay rates work, where to send billing questions — the agent answers from the knowledge base. For anything specific to a person — a client's benefits, an outstanding balance, a claim, the contents of a superbill — it collects contact details and routes the request to the billing contact rather than guessing. Client-specific billing conversation stays with staff.

Intake-paperwork completion reminders

Sessions start late or get rescheduled when consent forms and intake questionnaires are not done. Outbound campaigns can call new clients ahead of a first appointment with the practice's approved reminder — the portal to use, what must be finished before the visit — and log who confirmed, who needs the link re-sent, and who asked for help, without ever asking about the contents of the forms.

After-hours administrative capture

After hours, the agent answers instead of a voicemail box — with the crisis script still at the very top of the call, exactly as during business hours. Callers who clear that gate and have administrative needs get real handling: a cancellation is recorded against the policy, a reschedule request is captured or booked where rules permit, a new-client inquiry is taken in full rather than lost. Everything queues as structured records for the morning, and any practice-defined escalation trigger still transfers per protocol at 2 a.m.

Practice-information questions from approved content

Office locations, parking, telehealth logistics, hours, what to expect administratively at a first visit, fee schedules the practice chooses to publish — the agent answers from documents the practice attached, and can be instructed to route anything the sources do not cover instead of improvising. What it never draws from is anyone's clinical information: the knowledge base holds practice policies, not client records.

Multilingual client communication

ThunderPhone supports 47 voice languages. A practice can set a primary language plus additional languages, and the agent switches mid-call when a caller changes to a configured language — useful when the community a practice serves speaks languages its front desk cannot staff on every shift. Crisis-script wording in each configured language is the practice's to write and review, like every other piece of the script.

Sample call flow

Agent: Thank you for calling Riverbend Counseling Associates. This is the practice's automated assistant for scheduling and administrative questions. If you are in crisis or thinking about harming yourself, please hang up and call or text 988, or call 911 — this line cannot provide crisis support. Otherwise, how can I help you today?

Caller: I'm looking for a therapist for my teenager. Do you take Blue Cross?

Agent: We are in-network with several plans, including Blue Cross. I can take your information so our intake coordinator can confirm your specific benefits and match you with a clinician. May I have your name and the best number to reach you?

Caller: Dana Whitfield, and this number is fine.

Agent: Thank you. What days and times would generally work for appointments?

Caller: Weekday afternoons after 4, or Saturday mornings.

Agent: Noted — weekday afternoons after 4, or Saturday mornings. Would you prefer in-person sessions at our office or telehealth?

Caller: In person, I think.

Agent: You're all set. Our intake coordinator will call you back to verify your Blue Cross benefits and discuss which of our clinicians works with teens. Is there anything else I can help with?

The crisis wording at the top of the call is the practice's configured script — shown here with 988 and 911 as one common United States example. Had the caller said anything on the practice's trigger list at any point in the call, the agent would have stopped the intake immediately, delivered the crisis script, and followed the configured escalation path.

Platform capabilities behind the practice workflows

GDPR and HIPAA compliance posture — and the practice's configuration responsibility

ThunderPhone's public compliance posture is GDPR and HIPAA compliant. That posture does not make any particular practice workflow compliant by itself — and in a mental-health setting, this deserves more weight than anywhere else, because the mere fact that someone called a therapy practice is sensitive information, and mental-health information carries heightened protection and expectations of confidentiality beyond general medical data. The practice decides and configures: which client information the agent may collect and which it must not; whether calls are recorded and what notice callers receive; how long transcripts and recordings are retained; which integrations receive protected health information and under what agreements; what the agent's disclosure script says; and who can access call data. Keep the agent's scope strictly administrative, limit every workflow to the minimum information it needs, and have your compliance owner review the prompt, recording settings, and data destinations before launch.

Inbound and outbound calling on your numbers

Agents handle inbound calls and place outbound campaigns — reminders, waitlist backfill, paperwork follow-ups. Practices can bring existing numbers through supported direct VoIP connections or manual SIP configuration for any SIP trunk; verified imported numbers support inbound and outbound calling, so clients keep dialing the number they already know.

Warm and cold transfer

Cold transfers route a caller immediately — the mechanism behind a crisis escalation to an on-call line. On eligible production numbers that can place an outbound SIP leg, warm transfers let the agent put a caller on hold, privately brief the coordinator or clinician, and connect only after the recipient accepts.

Knowledge grounded in practice documents

Attach practice policies, accepted-plan lists, fee schedules, telehealth instructions, and location details — text files up to 5 MB, PDF or DOCX up to 50 MB — or import pages by URL. The agent searches these approved sources mid-call instead of improvising, and can be instructed to route any question they do not cover.

Integrations that fit the practice's stack

Connected apps, HTTP APIs, and remote MCP servers let the agent act during a call — check the calendar, book a session, file an intake record. HMAC-signed webhooks deliver call events to practice systems with verifiable integrity. A web widget can put the same agent on the practice's website.

Testing before a client ever hears it

This vertical demands rehearsal more than any other. Run AI-caller simulations of the calls that matter most — a caller who mentions self-harm mid-scheduling, a caller who asks the agent for advice about their situation, a distressed caller who circles back to crisis language after an administrative question — with graded scenarios and transcripts. Build the passing cases into regression suites so a prompt edit can never silently weaken the crisis script, and keep recordings, transcripts, and grading on so every production call is reviewable.

What would this cost each month?

Suppose a practice handles 500 AI-assisted calls per month, averaging 3 minutes each:

500 calls × 3 minutes = 1,500 minutes per month

Spark: 1,500 × $0.02 = $30/month
Bolt: 1,500 × $0.05 = $75/month
Storm: 1,500 × $0.09 = $135/month

That is base engine usage, not a quote. Premium voice or language paths can add up to 3¢ per minute, and phone-number or carrier charges depend on the telephony setup. Hold time bills a flat 2¢ per minute; voicemail and phone-menu minutes bill at the engine rate. Optional add-ons such as verbal acknowledgements, call supervision, and long prompts carry their own published surcharges. The agent builder shows the current all-in rate before calls are placed.

Guardrails a therapy practice should write into the agent

Each rule belongs in the prompt explicitly, with a simulation that proves it holds:

Crisis script first, always. Any expression of crisis, self-harm, harm to others, or a safety concern immediately triggers the practice's configured crisis response and escalation path — no counseling, no risk assessment, no delay, no finishing the administrative task first. The practice writes the trigger list and the exact wording; the agent matches and follows.
No counseling, ever. The agent never offers emotional support beyond scripted courtesy, never gives advice about a caller's situation, and never discusses therapeutic content. A caller who wants to talk about their circumstances is routed to a person.
No clinical conversation. Presenting concerns are recorded as the caller states them, for the intake coordinator — never explored, interpreted, or matched to a diagnosis or treatment.
Confidentiality by default. The agent never confirms to a third party that anyone is a client, and discloses nothing about appointments or history except to satisfy the practice's own identity-verification script.
Minimum necessary information. Collect only the fields each workflow requires; do not let the agent probe into personal details a scheduling call does not need.
Policy wording is fixed. Cancellation fees, superbill explanations, and plan lists are read as written and never negotiated or extemporized.
Disclosure and recording notices. The agent identifies itself as an automated assistant and delivers whatever recording or privacy notice the practice's compliance owner requires.
Escalation is always available. Any caller who asks for a person gets a transfer path or a callback message — the agent never becomes a wall between a client and the practice.

Frequently asked questions

Can an AI phone agent handle crisis calls for a therapy practice?

No, and it must never be positioned as if it could. The agent is configured so that any caller expressing crisis, self-harm, or a safety concern immediately hears the practice's own crisis script — many United States practices configure wording that directs callers to 988 or 911 — and is escalated on the practice's defined path, such as a transfer to the on-call clinician. The practice defines the triggers and the response; the agent never counsels, never assesses risk, and never delays the script to finish an administrative task.

Is ThunderPhone HIPAA compliant for a mental-health practice?

ThunderPhone's public posture is GDPR and HIPAA compliant. The practice remains responsible for configuring its own compliant workflow: what client information the agent collects, recording and notice settings, retention, which integrations receive protected health information, and who can access call data. Because mental-health information is especially sensitive, review the configuration with your compliance owner before launch.

Can it screen new clients for clinical fit?

Only administratively. It can collect the fields on the practice's inquiry form — insurance, availability, modality preference, clinician-type preference — and pass a structured record to the intake coordinator. Judging whether a client is a fit for a clinician is a human decision, and the agent says a person will follow up.

Can it answer questions about superbills and insurance?

It can answer general, practice-approved questions from the knowledge base — which plans the practice accepts, whether superbills are available for out-of-network reimbursement, how self-pay rates work. Anything about a specific client's benefits, balance, or claim is collected and routed to the billing contact, not answered.

What happens after hours?

The agent answers instead of voicemail, with the crisis script still first. Administrative callers get real handling — cancellations recorded, reschedules captured or booked where rules permit, new-client inquiries taken in full — and everything queues as structured records for the morning. Practice-defined escalation triggers still transfer per protocol at any hour.

Can it enforce our cancellation policy?

It can state the policy exactly as written, record the cancellation and the reason given, and route disputes or exception requests to staff. Whether it may grant any exception is up to the rules the practice writes.

Can one agent serve clients in multiple languages?

Yes. Configure a primary language and additional languages from the 47 supported, and the agent can switch mid-call when a caller changes language, provided the selected voice supports the configured set. The crisis script in every configured language is the practice's to write and review.

How do we verify the agent is safe before clients call it?

Treat the crisis path like safety-critical software. Run AI-caller simulations of crisis mentions at every point in a call — mid-intake, mid-reschedule, after an unrelated question — plus advice-seeking and distressed-caller scenarios, graded against the practice's criteria. Lock the passing cases into regression suites with pass-rate gates so no prompt change ships if it weakens the crisis script, and keep recordings and transcripts on so live calls stay auditable.

ThunderPhone pricing, language, transfer, knowledge-base, integration, and testing details on this page reflect ThunderPhone's published pricing and documentation as of August 19, 2026, with workflow research sources accessed the same day. ThunderPhone is GDPR and HIPAA compliant, and each practice remains responsible for configuring its own workflows appropriately.