Module 22

Voice Agents for Businesses

Build AI phone agents that answer and make calls for clinics, salons, shops and service companies, with the right platform, phone number, tools, costs and legal care.

≈ 7 h studyDraftedSources: Vendor docs and pricing pages for Vapi, Retell AI, ElevenLabs, Twilio; independent 2026 pricing comparisons; TRAI and FCC public guidance; Module 12; checked 23 Oct 2026

On this page

What a voice agent does
How a call works
Platforms
The law and consent (summary, not legal advice)
Designing the conversation
Build it: platform + n8n tools
Labs
Metrics
Exercises with model answers
Key takeaways
Check your understanding
Recommended resources

Costs and rules vary by country

Per-minute prices change often and depend on the voice, AI model and phone carrier. Telecom and consent rules differ by country. Check each vendor's pricing page and get legal advice before calling real customers. Examples are hypothetical; screens are illustrative wireframes.

22.1 What a voice agent does

A voice agent is an AI that talks on the phone (or in a browser or app). It listens, understands, answers, and can take actions such as booking an appointment, then hands the call to a person when needed.

Receptionist

Answer after-hours calls
Hours, prices, directions
Book or reschedule

Reminders

Appointment and delivery reminders
Confirm or reschedule in the call

Lead qualification

Call back web leads within minutes
Ask 3–5 questions, book a sales call

Order and status lines

Where is my order?
Return policy, store stock

Surveys and feedback

Short NPS or post-visit calls
Summaries to a sheet

Not a good first use

Medical or legal advice
Debt collection without legal review
Anything needing judgement or empathy at length

Good first use cases. Start with short, repetitive calls where the caller's goal is clear and mistakes are easy to fix.

22.2 How a call works

Inside a voice agent call. Voice platforms join these pieces, handle interruptions (barge-in) and pauses, and keep the delay to about a second. Some setups use a single speech-to-speech model instead of separate listen and speak steps.

22.3 Platforms

Platform	Style	Pricing model	Typical all-in cost per minute	Good for
Vapi	Developer platform, very flexible	Platform fee (about $0.05/min) plus your chosen speech, model, voice and phone providers	about $0.13 to $0.32	Custom builds, choosing each component
Retell AI	Developer-friendly with a strong dashboard	Usage per minute, phone line included	about $0.07 to $0.31	Fast setup, receptionists and outbound
ElevenLabs Agents	Best-known for natural voices	Plans with included minutes, extra minutes billed; phone carrier separate	about $0.08 and up	Voice quality, many languages
Bland AI	Enterprise phone agents	Per minute	varies	High-volume calling
Synthflow	No-code builder	Plans with minutes	varies	Non-technical teams
Twilio (ConversationRelay and others)	Build-your-own on a phone platform	Telephony plus your AI costs	varies	Engineering teams wanting full control

Voice agent platforms (checked 23 Oct 2026; verify pricing). All-in cost includes speech-to-text, the AI model, the voice and the phone line. Figures are ranges from several 2026 comparisons, not quotes.

Phone numbers in India. Providers such as Exotel, Plivo and Knowlarity offer Indian virtual numbers and connect to voice platforms (often via SIP). Indian numbers require KYC documents. International platforms may offer Indian numbers only through a partner, so check availability before choosing a platform.

Indian languages. Check each platform's supported languages and test Hindi and regional-language calls with real speakers. Accents, code-mixing (Hinglish) and background noise affect accuracy.

22.3.1 What it costs: a worked example

A clinic receives 1,000 calls a month, averaging 3 minutes. At $0.15 per minute all-in, that is 1,000 × 3 × $0.15 = $450 a month (about ₹38,000 at ₹84 to the dollar), plus a phone number fee, the platform's plan, and staff time for handoffs. Compare it with the cost of missed calls and of staff answering them.

Before any real call

Tell callers it's an AI at the start, and say if the call is recorded.
Outbound calls need prior consent. In India, commercial calls fall under TRAI's commercial communication rules (registration, DND preferences, consent). In the US, the FCC treats AI-generated voices in robocalls as "artificial" under the TCPA, so prior express consent is required.
Personal data: collect the minimum, store recordings securely and delete them on schedule (India's DPDP Act, GDPR in Europe).
Offer a human option ("press 0 or say 'agent'") and honour do-not-call requests.

22.5 Designing the conversation

A good receptionist call (salon, hypothetical). Read back every detail before acting. Spoken numbers and names are easy to mishear.

Voice prompt rules

Replies under two sentences; no lists, tables or URLs (they don't work in speech).
Say numbers the way people speak them ("four-thirty in the afternoon").
Handle silence ("Are you still there?") and interruptions gracefully.
Give a clear way out: "If you'd like to talk to someone, just say 'staff'."
Never guess. "I'm not sure, let me get a colleague to call you back."

22.6 Build it: platform + n8n tools

A voice platform's agent setup (illustrative wireframe, not a real product screenshot). Names and layout differ between Vapi, Retell AI and ElevenLabs, but every platform has these settings.

The tool behind 'check_availability' (n8n). Keep tool responses short and plain, because the agent reads them aloud. Respond within a couple of seconds or callers hear silence. Node types shown: Action · Logic (filter, IF, router) · Output · Trigger (starts the flow).

22.7 Labs

Lab 1 · Browser test call (free trial credits, 30 minutes)

Sign up for Vapi, Retell AI or ElevenLabs Agents and create an agent from a receptionist template.
Write the instructions and first message for Glow Salon (hypothetical), with hours, prices and three services.
Click talk/test and make five test calls, including one where you interrupt and one where you ask for a human.

Lab 2 · Availability and booking tools (60 minutes)

Build the n8n check_availability workflow above and a book_appointment workflow that creates a Google Calendar event.
Add both as tools in the voice platform, with clear descriptions and parameters.
Test a full booking. Check the calendar event appears.

Lab 3 · Post-call summary (20 minutes)

Set the platform's post-call webhook to an n8n workflow that writes caller, intent, outcome and summary to a Sheet and posts transfers to Slack.

Lab 4 · Inbound phone number (paid, 20 minutes)

Buy or import a phone number (platform or Twilio; in India via a provider such as Exotel or Plivo, which needs KYC). Attach it to the agent and call it from your phone. Add the AI disclosure and recording notice to the first message.

Lab 5 · Test plan (30 minutes)

Write 20 test scenarios (booking, rescheduling, unclear request, background noise, Hindi caller, angry caller, wrong number). Run them, score each (resolved, transferred correctly, mistakes), fix the instructions and re-test.

22.8 Metrics

Metric	What it tells you
Containment / resolution rate	Share of calls fully handled by the agent
Correct transfer rate	It handed off the calls it should have
Booking conversion	Calls that ended in a booking
Response latency	Time from caller stopping to agent speaking; aim for about a second
Transcription accuracy	Mis-heard names and numbers
Average handle time and cost per call	Efficiency and budget
CSAT or complaints	Caller experience

22.9 Exercises with model answers

Exercise · Price it

A car service centre gets 600 calls a month, 4 minutes each. Platform quotes $0.12 a minute all-in. Staff handle 20% of calls after transfer. Monthly AI cost?

Model answer

600 × 4 × $0.12 = $288 a month for agent minutes (assuming the transfer happens after most of the call; transferred minutes may continue to be billed until the agent leaves the call). Add the number rental and platform plan, and staff time for the 120 transferred calls.

Exercise · Fix the script

The agent says: "Our services include haircut ₹500, colour ₹2,500, spa ₹1,800, visit glowsalon.com/book or press 1 for…" What's wrong?

Model answer

Too long, a list and a URL in speech. Better: "We do haircuts, colour and spa treatments. Which one are you interested in?" Then give the one price asked for, and offer to book or send a WhatsApp link.

22.10 Key takeaways

A voice agent = phone + speech-to-text + AI model + text-to-speech + tools + handoff.
Choose Retell AI or ElevenLabs for speed, Vapi or Twilio for control, a no-code builder for non-technical teams.
Budget all-in per-minute cost, not the headline platform fee.
Disclose AI, get consent for outbound calls, and follow local telecom and privacy rules.
Read back details, keep replies short, and test with real speakers before going live.

22.11 Check your understanding

Check your understanding

Why can the all-in cost be several times the platform's headline per-minute fee?

Taxes only
Speech-to-text, the AI model, the voice and the phone line are billed too
Because of the logo
It can't

Why: Each component has its own per-minute or per-token price.

What should an outbound sales voice agent have before calling?

A funny greeting
Prior consent from the person and compliance with telecom rules
A longer script
A website

Why: Consent and telecom regulations apply to automated and AI calls.

Which reply style works best on a phone call?

A bulleted list with links
Short sentences, one question at a time
A long paragraph
A table

Why: Listeners can't scan; short, single questions are easiest to follow.

22.12 Recommended resources

Vapi docs · Retell AI docs · ElevenLabs Agents docs · Twilio Voice
TRAI for Indian commercial-communication rules · FCC on AI-generated voices in robocalls
Module 12 (agentic AI) for tool safety and evaluation

← PreviousModule 21: AI Chatbots for Business: WhatsApp, Slack and Web Next →Module 23: Learning Paths, Courses and Communities