How to set up an AI receptionist for a therapy practice

Knowing how to set up an AI receptionist for a therapy practice is less about flipping switches and more about a handful of deliberate decisions that decide whether the agent sounds like your front desk or like a generic bot renting your phone number. A personalized, self-learning agent can answer every call at once, in the caller's own language, and route the ones that need a human without hesitation, but only if it is configured around how your practice actually runs. This is a working guide to the eight things you set up, in the order they matter, with the specific choices behind each one. It is written for a behavioral health audience, so every step keeps the same hard boundary: the agent handles the front desk, and a clinician stays in the loop on anything clinical.

The reason to treat setup as a real project rather than an afternoon task is that the easy calls hide the hard ones. Any agent will book a clean Tuesday-at-three appointment and look finished. The setup work below is what determines how the agent handles the evening caller asking about a medication refill in Spanish, the distressed regular who needs a person immediately, and the prospective client comparing three practices in the same sitting. Each of the eight steps closes one of those gaps, and skipping any of them shows up later as a call that goes wrong in a way you never see.

Step 1: Name your agent and set its personality

The first decision is also the one practices overthink the most, so it helps to separate the two parts of it. Naming the agent is genuinely yours to make, and a real first name works better than a label like "the assistant," because callers relax faster when a conversation feels like it is with someone rather than something. Pick a name that fits your practice's tone, keep it consistent across your phone line and your website widget, and introduce it plainly at the top of every call so a caller always knows who they are speaking with.

The personality is where practices assume they have to do a lot of authoring, and they do not. Vybz agents ship with a predetermined personality that our clinicians and forward-deployed team built specifically for behavioral health, and it is designed to cater to everyone who calls a mental health practice rather than to a single imagined caller. That baseline is warm without being saccharine, calm under pressure, unhurried on a first call that clearly took courage to make, and precise on the administrative details where a wrong word costs a client or a claim. You are not writing a personality from a blank canvas, and that is deliberate, because a personality tuned for a first-time therapy caller is a different thing from a chirpy retail script, and getting it wrong is the fastest way to lose the person on the other end.

Two capabilities come built into that baseline personality and are worth understanding before you configure anything else, because everything downstream depends on them.

• The agent is genuinely multilingual and can talk to anyone. A caller who is more comfortable in Spanish, Mandarin, or Vietnamese is met in that language from the first sentence, without a menu, a callback, or a "press 2 for" detour. For a behavioral health practice this is not a nicety, because the moment a person reaches out for care is exactly the moment you do not want a language barrier deciding whether they book or hang up. A multilingual AI receptionist means the practice stops silently filtering out everyone who does not want to conduct a sensitive conversation in their second language.
• The personality is a foundation you adjust, not a fixed script. The baseline covers the tone every mental health caller needs, and your specifics layer on top of it: how formal your practice is, what the agent calls your clinicians, the small phrasing choices that make it sound like your office. You keep the parts that make a front desk feel like yours, and you inherit the parts that took a clinical team to get right.

Step 2: Let the agent personalize itself to each caller

A single fixed personality, however good, is still one voice for every caller, and real front desk staff do not talk to a frightened first-timer the way they talk to a brisk regular rescheduling between meetings. The second thing you set up is the agent's ability to adapt itself to the person actually on the line, in real time, within the same call.

The agent reads signals a good receptionist reads without thinking about it. It notices the caller's tone, whether they sound anxious, rushed, guarded, or relaxed, and it matches its own warmth and pacing to meet them. It picks up on patterns in how someone speaks and mirrors them: a caller who is brief and direct gets crisp, efficient answers, while a caller who is hesitant and searching for words gets more space and a slower, gentler cadence. It adapts to speed, slowing down for someone who needs time to think and keeping pace with someone who wants to get the booking done and move on. None of this changes what the agent is allowed to do or say, because the facts and the guardrails stay fixed. What changes is the delivery, so the same correct information lands the way that particular caller needs to hear it.

Consider a concrete evening at a practice. A caller named Marcus phones after hours, speaks quickly in Spanish, and asks whether he can get an early refill on a prescription before his next appointment. The agent switches to Spanish from his first sentence, matches his faster pace so he does not feel slowed down, and stays calm while it does the one thing that call requires: it recognizes that a medication question is clinical, does not attempt to answer it, and routes it to the right person under the escalation rules you set in a later step. The personalization did not change the boundary, and it never should. It changed whether Marcus felt understood in the thirty seconds before the handoff, which is the difference between a caller who trusts the practice and one who does not call back. This kind of caller-shaped adaptation is the same growing, human-like responsiveness we cover in our overview of personalized AI agents for behavioral health practices, where recent context stays vivid and each interaction is shaped by the person in front of it rather than an averaged script.

Step 3: Give the agent its guardrails

Guardrails are the most important thing you configure, and they are where a behavioral health setup diverges most sharply from a general receptionist. A guardrail is an explicit boundary on what the agent must not do, and unlike the personality, guardrails are not a baseline you inherit and leave alone. They are specific to your practice, and the whole point of them is that the agent respects them without exception, on every call, regardless of how a caller phrases the request.

Guardrails come in two shapes, and you will set both. The first shape is a hard prohibition, a thing the agent is never to do no matter what. The clearest examples in behavioral health are the clinical ones: the agent does not give medical or medication advice, does not offer any form of diagnosis or clinical opinion, does not counsel a caller in crisis, and does not attempt to talk someone through a mental health emergency. Those are not judgment calls the agent makes case by case, because they are fixed rules that stop it before it starts and route the call to a human instead. The second shape is a practice-specific instruction that reflects how your clinic chooses to operate: the agent does not quote a price for a service you prefer to discuss live, does not promise a specific clinician's availability without checking, does not commit the practice to accepting a plan before eligibility is confirmed, and does not share any detail about one client with anyone else who calls.

The reason to invest real time here is that a guardrail the agent respects perfectly is invisible, while a missing one is the failure you find out about from an upset caller. Depending on your practice, you can hand the agent as many specific guardrails and instructions as your operations require, and the more precisely you describe the lines it must not cross, the more reliably it handles the calls that live near those lines. When you cannot decide whether something needs a guardrail, the safe default is to add one and route the ambiguous case to a person, because in behavioral health an over-cautious handoff costs a minute while an over-confident answer can cost far more.

Step 4: Configure how and when the agent escalates

Guardrails tell the agent when to stop. Escalation tells it where the call goes next, and configuring the escalation paths well is what turns a stop into a smooth handoff instead of a dead end. This is the step that most determines whether callers who need a human actually reach one, so it deserves more than a single fallback number.

The core idea is that escalation should match the situation and the moment. A distressed caller during business hours needs a live person on the phone right now, so that path routes straight to whoever on your team owns urgent handoffs, with the context the agent already gathered so the caller does not start over. A routine question the agent cannot resolve during the day might warrant a quieter handoff. Something that arrives at two in the morning cannot ring a phone that nobody will answer, so it needs a path that reaches your team the way your team actually works after hours. A well-configured agent has several routes and picks the right one, rather than funneling everything into one channel that is wrong half the time.

The channels available to you cover the ways a front office already communicates, so you route each kind of escalation to wherever your team will see it fastest.

• A live call transfer during business hours, for the moments that need a person immediately, warm-handed off with the caller's context so nothing is repeated.
• A Slack message to the right channel or person, so a handoff shows up where your team already spends its day and gets picked up without anyone watching a separate inbox.
• An email to the owner of that kind of request, for things that need attention but not an interruption, with the full call summary attached.
• A custom workflow or automation, for the escalations that should kick off a defined sequence rather than a single notification, so an after-hours crisis path, a billing question, and a new-referral handoff can each trigger exactly the steps your practice has decided on.

The mix you choose depends entirely on your practice, and setting it up thoughtfully is what keeps the agent's escalation rate honest: the calls that genuinely need a human reach one quickly through the right channel, and the calls that do not are resolved by the agent without ever touching your team.

The diagram shows the shape of the routing; the chart below shows how the volume actually splits across those paths once a practice is running.

The channels above are the same ones our integrations connect into, so an escalation can post to your Slack, land in your email, or start an automation without anyone wiring it up by hand. You can see the full set of tools the agents plug into on our integrations page, and the day-to-day capabilities of the receptionist that runs these paths on the Front Office Agent page.

Step 5: Ground the agent in your practice memory

An agent with a good personality and clean guardrails still needs to know the facts of your practice to answer anything specific, and that knowledge lives in practice memory. This is what lets the agent draw on detailed, current information about your clinic to ground every answer rather than guessing, and it is the difference between an agent that sounds like it works at your practice and one that sounds like it read a brochure once.

Practice memory holds the operating reality a real receptionist carries in their head: which clinicians are on the roster and what each one treats, who is accepting new clients, the insurance panels each clinician is on and what to say when a plan is out of network, your hours and telehealth rules, your referral acceptance rules, and the exceptions that never fit a general script. When a caller asks whether a specific clinician takes their plan, the agent answers from your actual panel data instead of hedging. When two clinicians share a last name, the agent tells them apart from the caller's own record. The agent pulls exactly the relevant piece of knowledge for the question in front of it, which keeps its answers specific and correct on precisely the calls a generic bot fumbles.

The part that matters for setup is that seeding this memory is not a data-entry chore you inherit. Our forward-deployed team records your practice's specifics during deployment by watching how your front office actually runs, and from then on the memory keeps itself current through a self-learning loop rather than through you maintaining a knowledge base by hand. Confidential practice details live in a store that is isolated to your clinic and never shared with another, while transferable know-how lives in a shared base, so the agent can be grounded in your real exceptions without any risk of them reaching another practice. The full structure of that two-store memory, and the loop that folds each call's lesson back into it, is laid out in our deep dive on how Luna's memory and self-learning work, which is the deeper treatment of the system this step configures.

Step 6: Build follow-up plans that convert leads

The steps so far make the agent excellent at the calls that come in. The next two make it work on the calls and messages that go out, and follow-up is where a practice recovers the clients that a receptionist who only answers the phone quietly loses. A person searching for a therapist usually contacts several practices in the same sitting, and the one that stays in touch thoughtfully is often the one that gets the booking, so a follow-up plan is not an add-on but a core part of converting the demand you already paid to attract.

A good follow-up is a sequence, not a single nudge, and setting it up means deciding the cadence and the substance of each touch. Someone who called, asked about availability, and did not book yet is worth a warm, specific follow-up a day or two later that references what they actually asked about, rather than a generic "just checking in." Someone who started intake and stopped halfway needs a gentle reminder that makes finishing easy, not a guilt trip. Someone who asked about a plan you do not accept might get a follow-up pointing them to a clinician who does. The plan you build decides how many touches, how far apart, on which channel, and what each one says, and the agent runs the whole sequence for every qualifying lead so nobody falls through a gap because your front desk was busy that afternoon.

The reason to make these follow-ups detailed rather than perfunctory is that engagement is what actually moves a lead to a booked session. A single automated text that lands wrong reads as spam and does the opposite of what you want, while a short sequence that sounds like your practice genuinely remembered the person tends to recover a real share of the leads that would otherwise have gone silent. Because the agent runs the sequence grounded in the practice memory from the previous step, each follow-up can be specific to what that lead cares about, which is the whole reason a personalized agent converts where a mass blast does not.

Step 7: Set up routine check-ins for reminders that matter

The last piece of the outbound setup is routine check-ins, the recurring outreach that keeps existing clients engaged between sessions and reduces the drop-off that quietly erodes a behavioral health practice. Where follow-ups are about converting a new lead, check-ins are about the people already in your care, and they are the setup step most directly tied to the outcomes a clinician cares about.

A check-in is a scheduled, gentle touch on the things that matter to a person's care between visits, and you configure which ones make sense for your practice. Common examples in behavioral health include a reminder to take a medication on schedule, a short prompt about sleep for a client whose plan is tracking it, a nudge toward a between-session exercise a clinician assigned, or a simple "how has this week been" that gives a client a moment to say if something is off before the next appointment. These are administrative and supportive touches that a clinician has decided are worth sending, and they run on a cadence you set, so a client feels attended to without your team spending its day sending individual messages. The same channels that handle appointment work handle these behavioral health appointment reminders and check-ins, so a medication prompt or a sleep check goes out reliably without a human queuing it each time.

The boundary from the guardrails step holds here as firmly as anywhere. A check-in reminds, prompts, and gives a client an opening to flag a problem, and it never interprets the answer clinically. If a client responds to a "how has your week been" prompt with something that suggests they are struggling, that is an escalation, and it routes to a person under the paths you configured in step four rather than the agent attempting to help. Set up this way, routine check-ins extend a clinician's reach into the long stretch between appointments, which is exactly where engagement is either kept or lost, while keeping every clinical judgment with the human who is qualified to make it.

Putting the eight steps together

Set up in this order, the eight steps compound into an agent that behaves like a well-trained front desk rather than a phone tree. The name and personality make the first impression right, and the self-personalization makes it right for each specific caller. The guardrails and escalation paths make the agent trustworthy on the hard calls, stopping where it should and handing off cleanly through the channel your team actually watches. The practice memory makes every answer specific to your clinic, and the follow-ups and check-ins turn the agent from something that only answers into something that actively works to convert leads and keep existing clients engaged. Each step closes a gap that a generic setup leaves open, and the gaps are exactly where a behavioral health practice cannot afford to leak calls or clients.

The honest through-line across all eight is the boundary that never moves. Every step above makes the agent more capable at the front desk, and not one of them lets it drift into clinical work. It books, informs, reminds, follows up, and routes, and the moment anything looks clinical or looks like risk, it stops and a person takes over. That boundary is what makes the whole setup safe to run at a behavioral health practice, and it is why the receptionist, the appointment check-in agent, and the intake agent all operate inside the same guardrails you set once and can trust from then on.

If you would rather judge the setup by hearing it than by reading about it, the fastest way is to put an agent through a call yourself and notice how it handles the moments a generic bot fumbles, from a language switch to a medication question to a distressed handoff. You can do exactly that with our live agent demos, and the difference the eight steps make tends to be obvious within the first hard question.

Also read

• The AI co-worker who runs your practice operations and reports back nightly: the receptionist is one member of the team, and this covers Luna, the colleague who works your inbox, EHR, sheets, and KPIs behind it.
• How accurate is AI at complex front desk operations?: what actually determines whether the agent you just set up gets the hard calls right.
• Do clients trust AI for therapy intake and booking? What we found: the survey evidence that your clients are ready for the setup this playbook describes.
• Switching EHR systems: the migration challenges and how Luna moves your records: the other big setup project a practice faces, with the costs, the timeline, and a checklist for the weeks before go-live.