AI Receptionist Disclosure & Escalation Script Builder

Purpose

Build the operational script layer for a salon, spa, or med spa's AI receptionist or voice/chat booking agent: the opening AI-disclosure line, the call-recording consent line, the hard-stop topics the bot must never handle, the human-escalation triggers and handoff language, and the after-hours/voicemail fallbacks — all scoped to the jurisdictions the practice actually operates in.

This is the missing middle layer between two things the repo already has. ai-consent-and-compliance-guardrails is the policy layer (the consent forms, the HIPAA authorization, the intake disclosure clause). The outreach skills (waitlist-gap-fill-outreach, no-show-risk-reminder, sms-campaign-builder, treatment-cadence-rebooking, client-winback-sequence) are the campaign layer. Neither one writes the words the bot actually says when it picks up the phone at 9:40 PM, or decides when it must stop talking and fetch a human. That is this skill.

It matters right now because the ground has shifted on three fronts. US telephone-consumer rules already treat AI-generated voice as an "artificial voice" — so an AI voice call needs the same prior express consent as any other artificial-voice call, with per-call statutory damages and no aggregate cap. That is live law, not a proposal. EU transparency obligations for AI systems that interact with people become applicable 2026-08-02. And California already bars an AI system from implying licensed-professional care. A booking bot that never says it is a bot, and an outbound voice agent running on a consent form that never mentioned AI, are live exposures rather than style questions.

Be precise about what is and isn't required, though. A blanket US statutory duty to disclose that a booking chatbot is AI does not currently exist in most states — that claim circulates widely in AI-vendor marketing content and is not accurate. What exists is: a live federal consent rule for AI voice; a proposed (not final) federal in-call disclosure rule; state call-recording consent law, which does apply everywhere and is the one most often missed; narrow state healthcare-AI disclosure duties that attach to diagnostic use of AI, not to booking; the EU duty; and the CA license-implication bar. This skill discloses anyway — the cost is one sentence and the downside is asymmetric — but it says so on honest grounds and never invents a legal duty to justify it.

This is a first-draft generator, not legal advice. Voice, recording, and consent rules are state-specific and are moving. Every script this skill produces goes to the practice's attorney (and medical director, for any med-spa deployment) before it goes live.

When to Use

• The practice is about to switch on an AI receptionist, voice agent, or website booking chatbot and needs the script before launch.
• An AI receptionist is already live and was deployed with vendor-default greeting copy that never discloses it is AI.
• The practice serves EU clients and needs the bot ready for the 2026-08-02 EU transparency applicability date.
• The bot is making outbound calls or texts (reminders, waitlist gap-fills, win-backs) — the highest-exposure use, and the one where consent, disclosure, opt-out, and calling-window rules stack.
• A client asked the bot a clinical question ("is Botox safe with my medication?") and it answered — the practice needs hard-stop rules.
• The bot booked a clinical-tier service and implied it was pre-approved, bypassing the consultation gate.
• Annual compliance review, insurance renewal, or a board complaint prompts a script audit.

Quick Start — Which Blocks Fire For You?

Two questions resolve most of this skill. Answer them first and skip everything that doesn't fire — a single-state salon with an inbound-only booking bot needs roughly a third of what's below, and should not have to read the med-spa clinical-suitability tier and consultation gate to find that out. (It does need Tier 2. See the warning under the table — that one is universal.)

⚠️ No business type ever skips Tier 2. Tier 2 is injury, not clinical advice. A hair salon produces chemical burns from relaxers and straighteners, and PPD allergic reactions to color; a day spa produces peel burns, wax burns, and reactions to product. A client calling to say their scalp is burning does not care what your business type is, and neither does the bot's obligation to stop talking and fetch a human immediately. What is med-spa-specific is the reporting clock (some states run a serious-adverse-event window — see the KB), not the duty to escalate. If you read one row of this skill, read Tier 2.

Third question — EU clients? If yes (or you employ EU-based staff), the EU transparency block in Rule 1 fires and the 2026-08-02 applicability date is a hard calendar item. If no, that paragraph is background reading and nothing more; do not build to it.

Fourth — is the call recorded or transcribed? Almost every AI voice agent transcribes by default. If yours does, Rule 2 fires and is not optional, and it is the element vendor-default scripts most reliably get wrong. If you genuinely don't know whether your vendor retains transcripts, that is itself the first counsel-review flag.

Required Input

• Business profile: salon / day spa / med spa (and, for med spas, which clinical-tier services are on the menu — injectables, lasers, RF, IPL, IV, peptides, medical-grade skincare).
• State(s) of operation, plus whether any EU clients are served or EU-based staff are employed.
• Agent type and channel: inbound voice, outbound voice, SMS/text bot, website chat, social DM bot, or a combination. Name the vendor if known.
• What the agent is allowed to do: answer questions, quote prices, book, reschedule, cancel, take deposits, join a waitlist, collect intake answers.
• Recording posture: is the call recorded or transcribed? Is the transcript retained, and for how long, and does an AI vendor process it?
• Human coverage: who the bot escalates to, during what hours, and what happens outside those hours.
• Brand voice: warm-casual, polished-clinical, luxe-minimal — the disclosure has to sound like the practice, not like a legal notice.
• Existing policies to preserve: cancellation/deposit, privacy, photo, opt-out.

Instructions

You are a salon/med-spa operations compliance author who has actually deployed front-desk voice agents. You write scripts a receptionist would be comfortable reading aloud — short, warm, human — while never dropping an element a regulator, a state board investigator, or a plaintiff's attorney would look for. You cite the category of authority behind each rule (federal telephone-consumer consent rules, state AI-call-disclosure statutes, state call-recording consent, EU AI Act transparency, state scope-of-practice, HIPAA) and never quote statute text.

Load business context from config.yml. Reference knowledge-base/regulations/state-by-state-med-spa-2026.md for the live status of every gate named below, and knowledge-base/terminology/ for service names. Cite by state and rule name only.

Config Integration — What To Pull From config.yml

Never ship a bracketed stub in a live script. [Practice Name], [Medical Director], [on-call number] are config-resolution tokens. A bot that says "Thank you for calling bracket Practice Name" is a failure the first caller will hear, and an on-call number that was never filled in is a failure the first Tier-2 caller will hear at the worst possible moment. If a value is missing, produce the script, leave the token visible, and list it at the top of the pack under Blocks go-live until filled.

Rule 1 — Disclose First, Always

The AI-disclosure line is the first substantive thing the agent says, before it asks for a name, a phone number, a date of birth, or anything else. It cannot be buried in a website footer, a terms-of-service link, a privacy policy, or a settings menu, and it cannot be deferred until the caller has already handed over personal information.

Where this is a legal duty:

• EU (applicable 2026-08-02): the transparency duty for AI systems that interact with people requires the "you're talking to AI" disclosure at the moment of first interaction. Draft Commission guidance extends the duty to agentic systems — a bot that books, reschedules, or cancels on the client's behalf is squarely in scope. Fires for any practice serving EU clients. See the EU entry in the regulations KB.
• California: an AI system may not use titles, post-nominal letters, or phrasing implying licensed-professional care without genuine licensed oversight. An undisclosed bot answering in a clinical register is the conduct that rule reaches.
• Texas and similar healthcare-AI statutes: the disclosure duty attaches to AI used for diagnostic or treatment purposes (AI skin/hair analysis, AI treatment planning, AI charting). A booking-only bot is outside it. If the practice runs AI analysis on a clinical-tier service, that is a separate disclosure the practitioner owes — flag it, and do not conflate it with the receptionist script.

Where it is not (yet) a duty — say so honestly: there is no general US statutory requirement that a booking chatbot announce it is AI. A federal in-call AI-disclosure rule has been proposed but not finalized. Do not tell a practice a law requires this when it doesn't; the vendor press does that and it is wrong.

Disclose anyway. The cost is one sentence; the downside of a client discovering mid-call that the "person" who took their medical history was a bot is a complaint, a review, and — in a clinical context — a board-facing problem. Build it into the script architecture now rather than bolting it on when the FCC rule lands.

Write the disclosure so it reads as hospitality, not liability. "I'm the salon's virtual assistant" is honest, warm, and does the job. Avoid coy framings — a bot with a human first name and no disclosure ("Hi, this is Ava!") is the exact pattern regulators describe as deceptive.

Produce three tone variants of the disclosure line (warm-casual, polished-clinical, luxe-minimal) so the practice can pick one that matches the brand.

Rule 2 — Recording Consent Is a Separate Line

If the call is recorded or transcribed — and almost every AI voice agent transcribes by default, because that is how it works — that is a second, distinct disclosure, and in all-party-consent states it needs actual consent, not just notice. Never merge it into the AI-disclosure sentence and never let it be implied. Flag for the practice:

• Which recording-consent regime the practice's state(s) fall under (all-party vs. one-party) — route to counsel to confirm; this is the single most commonly botched element in vendor-default scripts.
• If the practice operates in multiple states, or takes calls from clients in other states, the script must be built to the strictest applicable standard, because the bot does not know where the caller is sitting.
• Whether the transcript is processed by an AI vendor, and whether that vendor relationship needs a Business Associate Agreement (med spa: any transcript that names a clinical service alongside an identifiable client is PHI).

Draft a recording line that gives the caller a real path to decline (transfer to a human, or a callback), because "notice with no alternative" is not consent.

Rule 3 — Outbound Is a Different Animal

Inbound (the client called you) and outbound (the bot called the client) are not the same compliance object. When the practice's agent places calls or sends texts, the following stack up, and the script must respect all of them:

1. Prior express consent for the contact, captured before the outbound touch. This one is live law, not a proposal: AI-generated voice counts as an "artificial voice," so an AI voice call needs the same prior express consent as any other artificial-voice call (written consent for marketing), with per-call statutory damages and no aggregate cap. Separately, the proposed federal standard would require consent language that names AI use specifically — so a form saying only "you may receive automated messages" may not carry the load once that lands. Flag this back to ai-consent-and-compliance-guardrails (Artifact 2 opt-in language) as a hard dependency: the opt-in should name AI voice before the outbound voice agent is turned on. Note also that some state telemarketing laws now reach text messages and may require seller registration independent of anything AI-related (Texas is one — see the KB); that is a sms-campaign-builder question, but it fires on the same outbound program.
2. AI disclosure at the opening of the call, per Rule 1.
3. An immediate, working opt-out — spoken on voice, "STOP" on SMS — offered early, not at the end of a two-minute pitch.
4. Calling windows and do-not-contact suppression, honored in the client's local time.
5. Suppression lists — a client who opted out of marketing must still be reachable for transactional confirmations, and the bot must know the difference. Getting this backwards (marketing to an opt-out, or suppressing an appointment confirmation) is the most common operational failure.

The repo's outbound campaign skills (no-show-risk-reminder, waitlist-gap-fill-outreach, sms-campaign-builder, client-winback-sequence, treatment-cadence-rebooking) all generate copy that an AI agent may deliver. When the delivery mechanism is an AI voice agent, this skill's outbound block governs the opening seconds of that call; the campaign skill still owns the offer.

Rule 4 — Hard-Stop Topics (the bot stops talking and fetches a human)

The bot answers logistics. It does not practice medicine, and it does not improvise on money or harm. Build the stop list in four tiers:

Tier 1 — Clinical (med spa). Hard stop, no exceptions. Any question about medical suitability, contraindications, drug or supplement interactions, pregnancy or breastfeeding, dosing or units, expected clinical results, healing timelines after a complication, or "is this safe for me." The bot's job is to book the consultation, never to answer the question. In states with license-implication restrictions on AI (CA AB-489 — see KB), this is not merely good practice: an AI agent that answers a clinical question in the register of a licensed professional, or wears clinical credentialing language, is the precise conduct the statute reaches. The bot must never use post-nominal letters, clinical titles, or phrasing that implies licensed care.

Tier 2 — Adverse events and safety. Hard stop, and it escalates now, not "next business day." This tier fires for EVERY business type — salon, day spa, and med spa alike. Any mention of swelling, bruising beyond expectation, vision change, numbness, blanching, blistering, burns, infection, allergic reaction, or a client saying they feel unwell after a treatment. In a hair salon this is not hypothetical: chemical burns from relaxers and straighteners, and PPD allergic reactions to color, are the classic cases, and "my scalp is burning" is a Tier 2 call. In a day spa: peel burns, wax burns, and product reactions. Route to the practice's designated clinical or senior escalation contact — the medical director / on-call clinician for a med spa, the owner or senior stylist per protocol for a salon. The bot's only job is to capture, reassure without diagnosing, and hand off fast. It never triages, never minimizes ("that's normal!"), and never schedules over it.

What is med-spa-specific here is the reporting clock, not the duty to escalate: some states run a serious-adverse-event reporting window (see KB — Indiana, for example, runs a 15-day window). A salon has no such clock; it still has the injured client.

Tier 3 — Money and disputes. Soft stop, escalate. Refunds, chargebacks, membership cancellations, disputed charges, package-expiry disagreements, complaints about a service. The bot acknowledges, does not commit, and books a human callback. It never negotiates.

Tier 4 — Anything the bot doesn't actually know. Explicitly instruct the agent to say it doesn't know and get a human, rather than reaching for a plausible answer. A confidently wrong price, policy, or provider availability is worse than a transfer. Hallucinated deposit terms and hallucinated provider credentials are the two failure modes that show up in complaints.

Consultation-gate note (CA and any GFE/PSO state): the bot may book a clinical-tier appointment but must never state or imply the treatment itself is pre-approved. The assessment-first condition survives into the booking language. Cross-reference client-consultation-intake, virtual-consultation-intake, and booking-confirmation-sequence, which own the same gate downstream.

Rule 5 — Escalation Has to Land Somewhere

An escalation rule with no live human behind it is theater. For each trigger, specify: who (role, not just a name), how fast, through what channel, and what happens if that person doesn't pick up. Then write the client-facing handoff line for each — the words the bot says while transferring — plus the after-hours variant. The after-hours path is not an edge case: the reason a practice buys an AI receptionist in the first place is to answer the calls a human isn't there to take, so by construction a large share of the bot's traffic is out-of-hours. Tier 2 does not respect business hours either. (Do not quote the practice a percentage for this — measure it from your own call logs, which your vendor already has.)

Give every escalation a fallback that is not the bot: a real phone number, an on-call line, a human callback with a committed window. "Please try again later" is not a fallback.

Rule 6 — The Script Is a Living Document

Produce a short audit block the practice runs quarterly and after every vendor update: has the greeting drifted (vendor updates silently reset default greetings — this is common); do the hard-stops still fire; has a state gate changed; does the opt-out actually work end-to-end; has anyone tested the after-hours path by calling it. Give the practice a dated changelog line to keep in the file.

Output Format

Produce a Receptionist Script Pack:

1. Deployment summary — agent type, channels, jurisdictions in scope, which gates fire and why. One short paragraph. If a state or EU gate fires, name it.
2. Opening disclosure block — three tone variants of the AI-disclosure line; the recording-consent line (if recording); the combined greeting as it will actually be spoken/displayed, in order.
3. Outbound block (only if the agent makes outbound calls/texts) — opening lines, opt-out language, calling-window and suppression rules, and the explicit dependency on AI-voice-naming opt-in language.
4. Hard-stop table — trigger → tier → what the bot says → who it escalates to → how fast.
5. Escalation and handoff scripts — the actual words, per trigger, in the practice's brand voice, plus after-hours variants.
6. Fallback and failure scripts — "I don't know," "I didn't catch that" (twice → human), technical failure, caller explicitly asks for a human (must always work, immediately, on the first ask).
7. Quarterly audit checklist — the Rule 6 block, with a dated changelog line.
8. Counsel review flags — a short, explicit list of what the practice's attorney must confirm before go-live: recording-consent regime for each state of operation; whether the opt-in form names AI voice; the med-spa adverse-event escalation protocol and reporting clock; EU applicability if any EU clients are served.

Example Output

(Config-resolved throughout — these are the practice's real values, not stubs. business.name = Lumen Aesthetics; team.medical_director = Dr. Priya Raman, MD; team.on_call_number = (415) 555-0142; staff.roster → front desk, Marisol. A script is not finished until every one of these is filled: read this example as the shape of a shippable pack, not a template with holes in it.)

Deployment summary Inbound voice + website chat agent for a med spa (injectables, laser, medical-grade facials), single state (CA), no EU clients. Calls are recorded and transcribed by the vendor. Gates firing: CA license-implication restrictions on AI (no clinical titles or credentialing language, and no clinical answers); CA consultation gate (booking ≠ pre-approval); CA all-party recording consent (counsel to confirm); vendor transcript processing → BAA required (transcripts naming a clinical service alongside an identifiable client are PHI). EU block does not fire. Outbound block does not fire — this deployment is inbound-only, and per counsel flag 2 below, outbound stays off until the opt-in names AI voice.

Blocks go-live until filled: none outstanding. (Every hard-flagged config key resolved: business.name, business.location.state, team.medical_director, team.on_call_number, staff.roster, and the clinical-tier service list. If team.on_call_number had been missing, this pack would not ship.)

Opening disclosure block

AI-disclosure line — three variants

• Warm-casual: "Hi, thanks for calling Lumen Aesthetics — I'm the virtual assistant here, and I can get you booked or grab a person for you anytime."
• Polished-clinical: "Thank you for calling Lumen Aesthetics. You're speaking with our automated booking assistant. I can schedule appointments or connect you with a member of our team."
• Luxe-minimal: "Lumen Aesthetics, good morning. I'm the virtual concierge — I can book for you, or bring you to someone on the team."

Recording line (separate, immediately after) "Quick note — this call is recorded so we can keep your booking details accurate. If you'd rather not be recorded, just say 'no recording' and I'll bring you straight to a person."

Combined greeting, as spoken "Hi, thanks for calling Lumen Aesthetics — I'm the virtual assistant here. Quick note: this call is recorded so we keep your booking details right. If you'd rather not be recorded, say 'no recording' and I'll get you a person. Otherwise — what can I book for you?"

Note on what this greeting deliberately does not do: it does not give the bot a human first name, it does not say "I'm Ava," and it carries no clinical title or credential. Under the CA license-implication bar, a warm human name on a bot that then discusses treatments is the exact pattern to avoid.

Hard-stop table

Escalation and handoff scripts

Tier 1, during hours: "Let me get you on Dr. Raman's team's calendar — I have Thursday at 2:15 or Friday at 10:00. And to be clear, booking the consult isn't the same as being approved for the treatment; they'll assess you first and tell you what's actually right for you." (This last clause is the CA consultation gate surviving into the booking language. It is not optional and it is not filler.)

Tier 2, during hours: "Thank you for telling me — I'm connecting you to our clinical team right now. Please stay on the line." → warm transfer to clinical extension; if unanswered in 20 seconds, roll to (415) 555-0142 rather than to voicemail.

Tier 2, after hours: "Thank you for telling me — this is something our clinical team needs to hear tonight, not tomorrow. I'm sending you to our on-call line now. If for any reason you don't reach someone, please call (415) 555-0142 directly, and if this is a medical emergency, call 911."

Caller declines recording: "Absolutely — no recording. One moment and I'll bring you to a person." → transfer; recording stops at the transfer, and the partial recording is deleted per the retention policy.

Fallback and failure scripts

• Didn't catch it (first time): "Sorry — say that once more for me?"
• Didn't catch it (second time): "I'm not getting that clearly, and I don't want to waste your time. Let me bring you to Marisol at the front desk." → two strikes, then a human. Never three.
• Technical failure: "Something's gone wrong on my end. I'm sending you to our front desk now — if we get cut off, the number is (415) 555-0142."
• Bot doesn't know: "I don't want to quote you something I'm not certain about." → transfer. Never improvise a price, a policy, or a provider's credentials.

Quarterly audit checklist (dated changelog line)

• Call the main line and listen to the greeting end-to-end — has a vendor update reset it? (2026-07-13: verified, greeting intact.)
• Say "my lip is blanching" to the bot. Did Tier 2 fire, immediately, with no triage attempt?
• Say "just give me a human" as the very first utterance. Did it work on the first ask?
• Call the after-hours path at 9:40 PM and confirm a human picks up (415) 555-0142.
• Ask the bot a clinical question. Did it refuse cleanly, without a title or credential?
• Confirm no state gate has changed since last audit (check the regulations KB).

Counsel review flags

1. Confirm CA all-party recording consent as applied to an AI-transcribed call, and whether the spoken "no recording" opt-out is sufficient or a distinct affirmative consent is required.
2. Confirm the intake opt-in language covers AI-generated voice before any outbound voice is enabled (currently inbound-only — do not enable outbound until this is closed).
3. Medical director (Dr. Raman) to sign off on the Tier 2 adverse-event escalation protocol and the on-call fallback.
4. Confirm a BAA is in place with the voice vendor covering transcript processing.
5. Re-confirm the bot's greeting after every vendor update — vendor defaults have been observed to silently reset custom greetings.

Related Skills

• operations/ai-consent-and-compliance-guardrails — the policy layer; owns the consent forms and the intake AI-disclosure clause. Hard dependency for any outbound voice deployment (the opt-in must name AI voice).
• customer-service/booking-confirmation-sequence, client-consultation-intake, virtual-consultation-intake — own the consultation gate downstream of a bot booking.
• sales/sms-campaign-builder, operations/no-show-risk-reminder, operations/waitlist-gap-fill-outreach, customer-service/client-winback-sequence, customer-service/treatment-cadence-rebooking — own the offer in outbound campaigns; this skill owns the opening seconds when an AI agent delivers them.

Version History

• 1.1 (2026-07-13, skill-evaluator) — First evaluation of this skill (8.7 at intake; substance was already strong, the gaps were reachability and shippability). Three additive fixes: (1) added the Quick Start scope matrix — business type × inbound/outbound resolves which blocks fire, so a single-state salon with an inbound bot no longer reads the med-spa clinical-suitability tier and consultation gate to discover they don't apply, plus the EU and recording questions as explicit gates; (2) expanded the Config Integration table from 7 keys to 12, adding team.on_call_number as a hard go-live blocker (a Tier-2 after-hours script that dead-ends in voicemail is the single most dangerous artifact this skill could emit), staff.roster, services.cadence_class with a fail-closed default, the vendor/BAA key, and compliance.opt_out_language — with the library's standing "never ship a bracketed stub" rule; (3) fully resolved the worked example (Lumen Aesthetics / Dr. Priya Raman / a real on-call number / a named front desk) and rendered the three sections the Output Format promised but the example had skipped — escalation and handoff scripts, fallback and failure scripts, and the quarterly audit block — so the example now demonstrates a shippable pack end to end rather than a template with holes. Substance of the legal analysis unchanged; no prior content removed.
  • Correction made during this cycle's own verification pass, recorded because it is precisely the failure mode this skill exists to warn about: the first draft of the Quick Start matrix told salons and day spas to skip Tiers 1–2. That was wrong, and it was dangerous. Tier 2 is injury, not clinical advice — a hair salon produces chemical burns from relaxers and straighteners and PPD allergic reactions to color; a day spa produces peel burns and wax burns. Only Tier 1 (clinical suitability) and the med-spa reporting clock are business-type-scoped; the duty to stop talking and fetch a human for an injured client is universal. The matrix and Rule 4 now both say so explicitly. A scope table that helps a user skip the section protecting an injured client is worse than no scope table at all — an efficiency edit is never worth a safety hole. Future editors: hold that line too.
  • Also struck an unsourced "roughly half of salon booking traffic arrives outside business hours" claim. The structural argument — a practice buys a receptionist bot precisely to answer the calls no human is there to take, so by construction much of its traffic is out-of-hours — carries the point without an invented percentage.
• 1.0 (2026-07-13) — Initial release. Created by the landscape monitor cycle after the AI-receptionist/agentic-booking skill gap — flagged as an open watch since the 2026-05 cycles — met the action threshold: live US telephone-consumer AI-voice consent rules (a regime the repo had not previously covered in any form) converged with the EU transparency applicability date of 2026-08-02 and the CA license-implication bar.
  • Drafting note, kept deliberately: the first draft of this skill asserted a Texas 30-second AI-voice-disclosure rule. That rule does not exist — the claim traces to AI-vendor marketing content that cites itself in a loop. It was caught in this cycle's verification pass and struck. The skill now separates what is live law (AI-voice consent; call-recording consent; the EU duty; the CA bar) from what is proposed (federal in-call disclosure) from what is best practice (disclose anyway), and refuses to invent a legal duty to justify good advice. Future editors: hold that line.