Voice Agent Script Builder
Turn "we want an AI to answer the phone" into a complete, approved call script — greeting to goodbye, including every escape hatch.
Outcome: After completing this template, you can hand a developer or vendor a full voice-agent script: greeting and consent line, intent menu, approved answers, booking and message flows, escalation rules, a forbidden-topics list, fallbacks, and a closing — all reviewed and signed off.
Page 1 — About This Template
What this is
Context: A voice agent is only as good as its script. Vendors can supply the voice and the plumbing; only you can supply what it's allowed to say, what it must never touch, and when it hands off to a person. This template is the ten-block skeleton of a complete call script, a fully worked example, and a blank version to fill in.
Explanation — the ten blocks of every good script:
| # | Block | What it does |
|---|---|---|
| 1 | Greeting + consent line | Answers, names the business, discloses it's an AI assistant, states recording notice |
| 2 | Identification | Collects caller name and callback number early (so a dropped call is never a lost call) |
| 3 | Intent menu | Finds out why they're calling — conversationally, not "press 1" |
| 4 | FAQ answers block | Word-for-word approved answers to your most common questions |
| 5 | Booking flow | Checks availability, books within your rules, confirms details back |
| 6 | Message-taking flow | Captures a complete message when booking/answers aren't possible |
| 7 | Escalation / handoff lines | Routes emergencies and human-requests immediately, with graceful wording |
| 8 | Forbidden-topics list | What the agent must never discuss, and the exact deflection line for each |
| 9 | Fallback lines | What it says when it doesn't understand — and the two-strikes rule |
| 10 | Closing | Confirms outcomes, sets expectations, ends warmly |
Writing rules that make voice scripts work:
- Write for the ear: short sentences, contractions, no bullet-speak. Read every line aloud before approving it.
- One question at a time. Callers can't hold three prompts in memory.
- Confirm by repeating back anything that matters (numbers, dates, spellings).
- Plan for interruption. Callers talk over agents. Every block should survive being cut off — the agent answers the interruption, then resumes or reroutes.
- The word "agent" (or "person," "representative", "human") is sacred. It transfers, immediately, from anywhere in the call.
Who this is for
Office managers, practice managers, and owners preparing a voice agent for intake, FAQs, and scheduling — whether building with a vendor, with us, or exploring what's involved.
When to use it
After you've listed your call types (see the article in Related), before any vendor configuration session. Bring the completed script to that meeting; it converts a vague project into a buildable one.
Warnings
- Every FAQ answer becomes gospel. The agent will repeat what you write thousands of times — get prices, policies, and hours verified before launch, and put one person in charge of updates.
- Recording and consent notices are legally required in many places, and rules differ (including two-party-consent regions). Have your consent line checked against current local requirements before launch — this template's wording is a starting point, not legal advice.
- Health, legal, and financial contexts carry extra duties. A medical or dental practice script should collect the minimum and keep clinical matters entirely human (verify current requirements with your compliance advisor).
- This script does not make the agent "done." Weekly transcript review and a wrong-answer log are part of operating it, forever.
Page 2 — COMPLETED EXAMPLE (Fictional)
Bright Smile Dental — a fictional family practice: Dr. Lee and Dr. Moreno, five staff, office manager Tanya. Hours Mon–Thu 8–5, Fri 8–1. The agent handles after-hours and overflow calls. All details invented for illustration.
Block 1 — Greeting + consent line
"Thanks for calling Bright Smile Dental. I'm the practice's automated assistant — I can book appointments, answer common questions, or take a message for the team. This call may be recorded for quality. If this is a dental emergency, say 'emergency' at any time. How can I help?"
Design notes: discloses AI in sentence two; recording notice before any personal info is collected; emergency path offered up front. Consent wording verified for our state — Tanya re-checks annually.
Block 2 — Identification
"Happy to help with that. Can I get your first and last name?" …
"Thanks, {name}. And the best number to reach you, in case we get cut off?" …
"Got it — {number}. Let me read that back: {number-repeated}. Is that right?"
Rule: name + callback captured before any flow. If caller declines: "No problem — I can still answer general questions, but I'll need a callback number to book or take a message."
Block 3 — Intent menu (conversational)
"What can I do for you today — are you looking to book or change an appointment, ask a question about the office, or leave a message for the team?"
| Caller intent (heard as) | Route to |
|---|---|
| "book," "appointment," "cleaning," "reschedule," "cancel" | Block 5 — Booking |
| "hours," "insurance," "cost," "where," "parking," "new patient" | Block 4 — FAQs |
| "pain," "swelling," "bleeding," "broke," "knocked out," "accident," "emergency" | Block 7 — Emergency, instantly |
| "bill," "charge," "refund," "wrong amount" | Block 6 — Message (billing tag), per forbidden list |
| "agent," "person," "human," "representative," anger signals | Block 7 — Handoff |
| Anything else / unclear | Block 9 — Fallback |
Block 4 — FAQ answers block (approved word-for-word)
- Hours: "We're open Monday through Thursday, 8 to 5, and Friday 8 to 1. Closed weekends."
- Location/parking: "We're at 412 Cedar Avenue, next to the public library. Free parking behind the building."
- New patients: "Yes, we're welcoming new patients! I can book your first visit right now — it's about an hour, and please bring a photo ID and your insurance card if you have one."
- Insurance: "We work with most major dental plans. I can't confirm what your specific plan covers — but if you leave your details, our front desk will verify your coverage and call you back before your visit."
- Cost questions: "Exact costs depend on the treatment and your insurance, so I don't want to guess. Our front desk can give you an estimate — shall I take a message, or book you in for an exam?"
- Cleaning frequency: "For most patients it's every six months, but Dr. Lee or Dr. Moreno will recommend what's right for you at your visit."
Note the pattern: anything plan-specific or clinical ends in a handoff, not an answer.
Block 5 — Booking flow
"Sure — is this for a cleaning and check-up, or something else?" → (if "something else" involves symptoms → Block 7 triage; otherwise continue)
"Are you a current patient with us, or will this be your first visit?" →
"Do mornings or afternoons work better?" →
"I have {option 1} or {option 2}. Which would you like?" →
"You're set, {name}: {service} with {provider} on {day, date} at {time}. You'll get a text confirmation shortly — reply to it if you need to change anything. Anything else I can help with?"
Rules: books only cleanings, check-ups, and consultations from the live calendar; max one appointment per call for new patients; no double-booking overrides; anything symptom-driven is offered the next available exam AND flagged for a morning callback.
Block 6 — Message-taking flow
"I'll make sure the team gets this first thing. You're {name} at {number} — and what's the message?" …
"Let me confirm: {summary}. Did I get that right?" …
"Perfect. The team is in {next business day/time} and billing questions are usually returned by early afternoon. Anything else?"
Every message logs: name, number, timestamp, category tag (billing / clinical-callback / insurance / other), and the agent's one-line summary plus full transcript link.
Block 7 — Escalation / handoff lines
- Emergency (any trigger word, any time): "That sounds urgent, and I want to get you to a person right away. Please hold while I connect you to our on-call line." → If no answer in 25 seconds: "I couldn't reach the on-call line. If this is severe — trouble breathing, uncontrolled bleeding, or a serious injury — please hang up and call 911. Otherwise, I've marked your message URGENT, and someone will call you back as soon as possible. Stay by your phone, {name}."
- Human requested: "Of course — one moment while I transfer you." → After hours: "Our team has gone home for the day — I'll take a priority message so a real person calls you first thing tomorrow. What should I tell them?"
- Upset caller (raised voice, repeated complaint, 'this is ridiculous'): "I'm sorry — I want to make sure a person handles this properly. Let me take down exactly what happened, and I'll flag it priority for the office manager." (Never defends, never argues.)
Block 8 — Forbidden-topics list (with deflection lines)
| Never discuss | Deflection line |
|---|---|
| Clinical advice, symptoms, diagnoses, "should I…" medical questions | "That's really a question for the doctor — I'll flag your message for a clinical callback, and if it's urgent, say 'emergency' and I'll connect you now." |
| Medication and prescription questions | Same clinical-callback deflection |
| Insurance coverage promises ("will my plan cover…") | "Our front desk verifies coverage directly with your insurer so you get an accurate answer — can I take your details?" |
| Billing disputes, refunds, charge amounts | "I want to get this right for you, so I'll pass it straight to our billing specialist — they usually call back by early afternoon." |
| Discounts or price negotiation | "I'm not able to adjust pricing — but I'll note your question for the office manager." |
| Opinions of other practices; staff personal information | "I can't help with that one, but is there anything about our office I can answer?" |
| Card numbers or payment over the phone | "For your security, I don't take payment details — our front desk handles payments directly. Shall I have them call you?" |
Block 9 — Fallback lines (two-strikes rule)
- First miss: "Sorry — I didn't quite catch that. Could you say it another way?"
- Second miss: "I'm having trouble understanding, and I don't want to waste your time. Let me take a message so the right person calls you back — or say 'agent' if you'd like me to try transferring you now."
(Never a third attempt. Never blames the caller.)
Block 10 — Closing
"You're all set, {name} — {outcome recap}. {Expectation line: 'You'll get a text shortly' / 'Expect a call tomorrow morning.'} Thanks for calling Bright Smile Dental — have a great {morning/afternoon/evening}!"
Page 3 — BLANK VERSION
Business name: ______________ Line(s) covered: ☐ After-hours ☐ Overflow ☐ All calls Prepared by: ______________
- Greeting + consent line (name the business, disclose AI, recording notice, emergency escape word):
_________________________________________________________________
- Identification (name + callback + read-back line; what if caller declines?):
_________________________________________________________________
- Intent menu (your conversational prompt + a route table: trigger words → block):
_________________________________________________________________
- FAQ answers (top 6–10 questions, word-for-word approved answers; mark any that must end in handoff):
_________________________________________________________________
- Booking flow (services it may book, calendar rules, confirm-back line, what it must never book):
_________________________________________________________________
- Message-taking flow (fields captured, confirmation line, callback expectation, category tags):
_________________________________________________________________
- Escalation / handoff (emergency trigger words → destination + no-answer fallback; human-request line; upset-caller line):
_________________________________________________________________
- Forbidden topics (list each + its exact deflection line — minimum: advice in your regulated area, disputes, payments, negotiation):
_________________________________________________________________
- Fallback lines (first miss, second miss, then what):
_________________________________________________________________
- Closing (recap + expectation + sign-off):
_________________________________________________________________
Quality Checklist — run before launch, then monthly
Accuracy
- [ ] Every FAQ answer verified by the person who owns that fact (hours, prices, policies)
- [ ] Booking rules match the real calendar and real services
- [ ] One named owner for script updates, with a change log
Tone
- [ ] Every line read aloud; sounds like your business, not a kiosk
- [ ] Agent discloses it's an assistant in the greeting — no pretending to be human
- [ ] Apology and upset-caller lines reviewed by whoever handles complaints today
Interruptions
- [ ] Tested: caller interrupts mid-greeting with their need — agent adapts
- [ ] Tested: caller changes topic mid-booking — agent follows, then returns or reroutes
- [ ] "Agent"/"person" tested from every block, including mid-booking
Routing
- [ ] Every emergency trigger word tested → correct destination, correct no-answer fallback
- [ ] Escalation accuracy target set (emergencies: 100%) and someone reviews every escalation weekly
- [ ] Message tags land where the right person actually looks
Fallback
- [ ] Two-strikes rule works; third miss never happens
- [ ] Nonsense input, silence, and background noise all tested
- [ ] Forbidden topics probed with sneaky phrasings ("hypothetically, would my plan cover…")
| Owner | Approver | Date completed | Version |
|---|---|---|---|
| 1.0 |
Safety note: Scripts touching health, legal, or financial contexts should collect the minimum personal information needed, keep advice entirely human, and get compliance review before launch — consent/recording rules and industry data duties vary by location (verify current requirements). And keep the human review loop running after launch: ten sampled transcripts a week plus every escalation is the difference between an asset and a liability.
Practice (15 minutes): Role-play your draft with a colleague: you play the agent and may only say what's written; they play three callers — a routine booking, an upset billing call, and an emergency. Every place you're forced to improvise is a gap in the script. Fix those first.
Next step: Hear these blocks working in a live agent at /demos/voice-agent — then bring your completed script to a call at /contact, or explore deployment options at /solutions/voice-agents. A finished script turns Implementation from weeks of discovery into configuration.