"Hi, I found you on Psychology Today. Do you work with anxiety? How much is a session? Do you take Blue Cross? Do you have anything after 6?" If you are a psychologist or therapist in private practice, you have received this message dozens of times, usually in the middle of a session, and answered it two hours later. By then half of those people had already booked with someone else. The other half turned into eight rounds of email or phone tag to find a time.
This is where AI for psychologists makes the most sense as automation: first contact, scheduling and reminders. Not the content of therapy. This post shows how to build that flow with the ethical limits designed in from the start, not patched on later.
The line the flow never crosses
Every automation in a therapy practice should pass three questions. If any answer is "yes," the automation stops there and you step in.
- Does this require clinical listening? Any account of distress, personal history or request for emotional help.
- Could this be a crisis? Suicidal thoughts, violence, substance use, despair. Even mentioned in passing.
- Does this produce a clinical document or a professional decision? A referral, a letter, a report, a recommendation about which approach fits a specific person.
Everything left over is logistics: fee, insurance, telehealth or in person, hours, office address, the video link, reminders, receipts and superbills, the cancellation policy. That is what an AI agent can handle, and it is a lot.
The full flow: from the first "hi" to the session reminder
| Step | Who does it | What happens |
|---|---|---|
| 1. A message arrives (web chat, text or email) | Automatic filter (no AI) | Scans for risk words. If it finds one, it sends a fixed message with 988 and 911 and notifies you immediately |
| 2. First contact | AI agent | Answers fee, sliding scale, insurance networks, telehealth or in person, approach, ages served, only from your knowledge base |
| 3. Booking the initial consultation | Agent + calendar | Offers up to 3 open slots and books the free 15-minute phone consult or the intake session |
| 4. Before the session | Automation (your EHR) | Reminder 24 hours before with confirm or reschedule; for telehealth, the link to the client portal |
| 5. After the session | Automation (your EHR) | Receipt or superbill and, if relevant, payment |
| 6. Any personal content | You | The agent replies "I will leave that for [name] to talk about with you" and stops |
Notice the AI never asks "what brings you to therapy?" It sounds caring, but it opens a door the bot cannot hold. People answer with their life story, and that story ends up in a system that was never designed to store it. Ask only what the calendar needs. Your intake paperwork, inside the EHR, is where clinical history belongs.
A real example: the agent handling a first contact
I tested the system prompt below with typical messages. This was the response, with a fictional practice's details:
Person: hi, do you treat anxiety? how much are sessions? i really need help, i havent slept properly in weeks and yesterday i had a panic attack at work
Agent: Hi, thanks for reaching out. This is the virtual assistant for Dr. Sarah Klein's practice. I'm sorry you're going through this. Dr. Klein sees adults, in person in Denver and by telehealth anywhere in Colorado, and anxiety is one of the areas she works with. Sessions are US$ 175; she is out of network and can provide a superbill. For a free 15-minute phone consultation, I have Tuesday 11/17 at 12:30 p.m., Wednesday 11/18 at 5:00 p.m. or Thursday 11/19 at 8:00 a.m. Would one of those work? If at any point you don't feel safe, call or text 988, or call 911.
Person: wednesday 5 works
Agent: Done: Wednesday 11/18 at 5:00 p.m., phone consultation. To hold it, please send your full name and the best phone number to reach you. Dr. Klein will talk with you then.
Three things the prompt guaranteed: it did not ask about the panic attack, it gave no sleep tips and it kept 988 visible because the message showed intense distress, without making it dramatic. The phrase "panic attack at work" also triggered the alert to the psychologist, who can decide to call the person before Wednesday.
The system prompt (copy and adjust)
You are the virtual assistant for the practice of [NAME], [LICENSE TYPE], licensed in [STATE(S)]. You are NOT a therapist and you do NOT provide support, listening or emotional guidance.
You only: give fee, sliding scale, insurance, telehealth or in-person options, approach, ages served and cancellation policy using the KNOWLEDGE BASE below; offer up to 3 open slots; collect name and phone number for booking; reschedule and confirm appointments.
Non-negotiable rules:
1. Never ask why the person is seeking therapy and never ask for details about their situation.
2. If the person shares something personal, reply with one short sentence of respect ("I'm sorry you're going through this") and return to scheduling. Do not comment, advise or offer techniques.
3. If there is any sign of risk (death, suicide, self-harm, violence, "I can't do this anymore", hopelessness): reply "If you are not safe right now, call or text 988, or call 911. I have let [NAME] know so they can reach you as soon as possible." and trigger ALERT.
4. Never promise outcomes, improvement timelines or a number of sessions.
5. If asked whether you are a bot, say you are the virtual assistant.
6. If a question is not in the KNOWLEDGE BASE, say [NAME] will answer and trigger HANDOFF.
7. Do not confirm whether any specific person is or was a client.
8. Short messages, plain language.
KNOWLEDGE BASE:
[FEE, SLIDING SCALE, INSURANCE PANELS, TELEHEALTH STATES, APPROACH, AGES SERVED, OFFICE HOURS, CANCELLATION POLICY, LINK TO PRIVACY NOTICE]
Do not rely on the prompt alone for risk. Put a simple keyword filter in front of the AI (in n8n it is an "if contains" node) that fires the alert even if the model misses. Two layers, because this is the one point in the flow where an error is not acceptable.
The pieces and the monthly cost
| Piece | What I would pick | Approximate cost |
|---|---|---|
| Calendar, reminders, portal | The EHR you already use (SimplePractice, TherapyNotes and others), which signs a BAA | Already in your EHR plan, roughly US$ 50 to 70/month |
| Channel | Web chat on your site, or SMS through a provider that offers a BAA for HIPAA-eligible services (Twilio is one) | Per-message SMS fees plus US carrier registration for business texting; check current rates |
| Automation | n8n, self-hosted on a server you control, so data does not pass through one more company | A small server from about US$ 6 to 12/month, or n8n Cloud from about €24/month (check) |
| AI model | A small model via API from a provider that will sign a BAA for your use, if any PHI can pass through | A few dollars a month at a solo practice's volume |
If you see 20 to 30 regular clients and get only a few new inquiries a week, I would not build the AI part. I would set up online booking and automatic reminders in the EHR, which cost almost nothing extra and solve most of the pain. The first-contact agent pays off when inquiries pile up (a growing directory profile, a group practice, ads running) and you lose people to slow replies.
What changes in your ethical duties when you automate
- Confidentiality. The agent's conversations are stored somewhere. Know where, who can access them and for how long. Keep only what scheduling needs. A prospective client's name plus "panic attack" is already protected health information once it sits in your systems.
- Business associates. Every vendor that touches that data on your behalf needs a BAA. No BAA, no PHI: that is the simplest rule to follow.
- Transparency. People need to know they are talking to a virtual assistant. Saying so in the first message avoids the feeling of being tricked, which breaks trust before the first session.
- Advertising. The agent's answers are also statements about your services. Promised outcomes, improvement timelines and comparisons with colleagues stay out, exactly as on your profile.
- Licensure. The agent should not book a telehealth session with someone located in a state where you are not licensed. Put your states in the knowledge base and have it ask where the person will be during sessions.
- State law. A few states now regulate AI in mental health directly. Illinois, for example, passed a 2025 law that bars AI from providing therapy and limits how clinicians use it. Administrative scheduling is usually the safe zone, but read your state's rule before you switch anything on.
An honest note: the APA, the ACA and state boards have been publishing guidance on technology and AI often. Before you go live, check your board's website and your association's latest guidance. You do not need permission to automate scheduling, but it is good to know what they are watching.
Notes and marketing: the two other sensitive points
Progress notes
Your records follow your state's rules and HIPAA. If you use AI to format a note, either paste your own de-identified summary or use an AI feature inside your EHR that is covered by its BAA, and review before you sign. I would not connect any automation that writes to the chart on its own. The chart is where your clinical judgment is recorded; it has to be yours.
Marketing
For posts and your directory profile, use AI freely with the rules inside the prompt. There is a ready-made model, with tested output, in the guide to AI for therapists, and more variations in ChatGPT for therapists.
Common mistakes when automating a practice
- A "caring" bot. Messages like "I'm here to listen" make people pour their hearts out to a machine. The agent should be warm and brief, not therapeutic.
- A pre-consult form that asks for the presenting problem in detail. If you want an initial history, do it in session or in intake forms inside your EHR, not in a Google Form connected to the bot.
- Risk detection only in the prompt. Models miss things. Keyword filter before the AI, plus an alert on your phone.
- Using your personal cell number with an unofficial texting setup. No BAA, no audit trail, and if it breaks you lose every client conversation at once.
- Nobody sees the alert. Decide when you check alerts and say it in the auto-reply ("I return messages within one business day").
When it is not worth building
- Full caseload, few new inquiries: a simple booking page is enough.
- Mostly child and family work, where a parent asks long, specific questions: a well-written standard reply and you answering personally work better.
- If you will not have time to watch the alerts: the risk filter only protects people if someone reads it.
If you want the flow running without setting up n8n, webhooks and prompts, the ready-made agents do exactly this part: first contact, scheduling and reminders, with the risk safeguards already configured.

Want to skip this part?
Ready-made AI agents for therapists: nothing to build, no prompts to write, they handle scheduling and first contact for you.
- Answers fee, insurance and availability questions and books the first consultation
- Sends reminders, confirms and reschedules sessions on its own
- Never offers support: at any sign of risk it sends 988 and 911 and alerts you
Step by step
- Write the knowledge base with fee, insurance, telehealth states, approach, ages served, hours and cancellation policy.
- Use the booking, reminders and portal already in your EHR, under its BAA.
- Pick the channel (web chat or SMS through a provider that signs a BAA) and connect it to n8n.
- Build the risk keyword filter in front of the AI, with an alert on your phone.
- Configure the agent with the system prompt limited to logistics.
- Turn on the 24-hour reminder with confirm and reschedule options.
- Review the first week of conversations and adjust the knowledge base.
Frequently asked questions
Is it ethical to use a chatbot for a therapy practice?
It can be, if the bot handles only logistics (fees, scheduling, reminders), says it is a virtual assistant, does not offer support and passes any personal or risk content to the clinician. Check your state's rules on AI in mental health first.
Should the agent ask why the person wants therapy?
No. People tend to share their whole story, and that sensitive information ends up in a system that is not your clinical record. Ask only what scheduling needs.
How does the agent handle a crisis message?
With two layers: a keyword filter before the AI and a rule in the prompt. Both send 988 and 911 and trigger an immediate alert to the clinician.
Do I need a BAA to automate scheduling?
If the vendor stores or processes protected health information for you, yes. Inquiry messages can contain PHI, so use vendors that sign a Business Associate Agreement or keep the data out of their systems.
How much does it cost to automate scheduling for a therapist?
Online booking and reminders usually come with your EHR plan. Adding an AI first-contact agent adds a small server or n8n plan, SMS fees and a few dollars of API use per month.
