AI voice agent development for phone lines that answer, qualify and book

What is AI voice agent development?

AI voice agent development is the work of building a phone agent that listens to a caller, understands what they want, replies in a natural voice and takes an action, such as booking, transferring or logging a message. The agent combines three rented AI services with your own business rules and systems.

Vapi's documentation describes the same three building blocks every voice assistant uses: speech-to-text to turn the caller's words into text, a large language model to decide what to say and do, and text-to-speech to speak the reply. Telephony connects all of this to a real phone number. The development work is in the parts around those services: the script, the tools the agent can call, the transfer rules, the consent greeting, the integrations, the logging and the cost controls.

A good voice agent is narrow and predictable. It does not try to hold open-ended conversations about anything. It knows its job, for example “book HVAC service visits and take emergency messages”, and politely hands everything else to a person or a voicemail. That focus is what makes it reliable enough to put on your main business line.

• Telephony: the phone number, call routing, transfers and recordings.
• Speech-to-text: real-time transcription of the caller.
• Language model: understanding, script logic and tool calls.
• Text-to-speech: the agent's voice.
• Your systems: calendar, CRM, help desk, order database.

What can an AI phone agent do for a US small business?

It can answer every call, including after hours and during rushes, qualify callers with the same questions every time, book appointments into live calendars, send confirmation texts, transfer urgent or high-value callers to a person, and write a clean summary of every conversation.

Home services companies are the classic fit. An HVAC or plumbing business in Texas or Arizona gets its heaviest call volume on the hottest or coldest days, exactly when technicians are too busy to pick up. A voice agent can triage “no cooling” emergencies to the on-call technician and book routine tune-ups into open slots.

Medical and dental offices use voice agents for overflow and after-hours scheduling, with care around health information. Law firms use them for structured intake before a consultation. Property managers use them for maintenance requests. Restaurants and salons use them for bookings and common questions. In each case, the value comes from calls that would otherwise go to voicemail and never call back.

Businesses in these fields often need a matching website too; see our pages on HVAC website design and plumber website design.

Twilio, Vapi or Retell: which stack should AI voice agent development use?

Use a voice platform such as Vapi or Retell when you want to launch quickly with standard call flows; use Twilio with custom orchestration when you need unusual logic, strict control over data paths, or the lowest possible per-minute cost at high volume. We build on all three and choose after seeing your call flows.

Vapi describes itself as a developer platform for building voice AI agents that can make and receive calls and integrate with existing systems, with a choice of providers and models for each component, including OpenAI, Anthropic, Google, Deepgram and ElevenLabs. It supports single assistants and multi-assistant “squads”.

Retell's documentation describes a platform to build, test, deploy and monitor voice and chat agents with telephony, prompts, tools and analytics built in. It lets you buy US or Canadian numbers directly or bring your own telephony through SIP trunking, and lists ready-made integrations with CRMs such as Salesforce, HubSpot, Dynamics 365, GoHighLevel and Zoho, plus post-call analysis.

Twilio provides the phone network layer, numbers, call control and recordings, and leaves the AI orchestration to your own code. That is more work, but every step is yours to shape and log. For many projects we start on a platform for the pilot and keep the option to move orchestration into your own code later if volume or requirements justify it.

How a voice agent call works: latency, interruptions and hand-offs

A caller speaks, speech-to-text streams the words, the model decides a reply or tool call, text-to-speech speaks it, and the loop repeats. The whole loop has to feel like a normal conversation, which means responses must start quickly and the agent must stop talking when the caller interrupts.

Latency is the biggest quality issue in AI voice agent development. Long pauses make callers talk over the agent or hang up. We reduce it by choosing fast models for the conversational turns, keeping prompts short, streaming speech out as soon as the first words are ready, and running slow tasks, such as CRM lookups, with a short filler phrase instead of silence.

Call-recording consent: one-party and all-party consent states

If your voice agent records or transcribes calls, assume you need to tell callers at the start of every call. Federal law allows recording when one party consents, but several states require the consent of all parties, and callers may be in any state.

Under 18 U.S.C. § 2511(2)(d), it is not unlawful for a person to intercept a communication when they are a party to it or one party has given prior consent, unless the purpose is criminal or tortious. California Penal Code § 632 applies to anyone who, without the consent of all parties to a confidential communication, records it. Florida's statute § 934.03(2)(d) makes interception lawful when all parties have given prior consent. Washington's RCW 9.73.030 requires the consent of all participants and says consent is obtained when one party announces to the others, in a reasonably effective manner, that the conversation is about to be recorded.

Twilio's recording documentation also reminds users that recording calls requires compliance with consent laws, naming California's Invasion of Privacy Act, and recommends consulting legal counsel. That is our advice too: we build the announcement, logging and retention settings; your lawyer confirms the wording and policy for the states you serve.

• A clear greeting line, such as “This call is answered by an automated assistant and may be recorded and transcribed.”
• Consent event logged with the call record.
• An option for callers who object: transfer to staff or take a message without recording, per your policy.
• Retention settings for recordings and transcripts that match your own policy.

Read the source texts directly: California Penal Code 632 and Washington RCW 9.73.030.

Can an AI voice agent make outbound calls legally?

Outbound calling with an AI voice is far more regulated than answering inbound calls, and we only build it after your counsel has signed off on consent and call lists. Most of our AI voice agent development work is inbound for that reason.

The Telephone Consumer Protection Act, at 47 U.S.C. § 227(b)(1)(B), makes it unlawful to initiate a call to a residential telephone line using an artificial or prerecorded voice to deliver a message without the prior express consent of the called party, subject to listed exceptions. Section 227(b)(1)(A) separately restricts calls using an automatic dialing system or an artificial or prerecorded voice to certain numbers, including cellular numbers. An AI-generated voice raises exactly these questions.

Where outbound calls are appropriate, for example returning a call the customer requested or reminding a patient who opted in, we build consent checks into the dialer: calls only go to contacts with a recorded consent source, quiet hours are respected, and opt-out requests are stored and honored. Whether a given campaign is lawful is a legal question for your attorney, not something a developer should decide.

Connecting an AI voice agent to your calendar and booking rules

The agent books through the same calendar your staff uses, with your rules in code: which services take how long, which technician or provider handles which job, buffer time between visits, and which ZIP codes you serve. It never invents a free slot; it asks the calendar.

For Google Workspace or Microsoft 365 calendars we use their official APIs. For scheduling tools used in dental, medical, salon and field service businesses, we connect through the tool's API where one exists. If your scheduling system has no API, the agent can take a structured booking request and send it to staff for confirmation, which is slower but honest.

Callers hear two or three concrete options rather than a long list, pick one, and get a text confirmation with the time and address. If they need to reschedule later, the same agent can find the booking by phone number and move it. Every booking is written back to your CRM with the call summary, so staff see the full picture.

Field teams that also need a technician app can combine this with field service app development.

AI voice agent development with CRM integration and call summaries

After each call, the agent writes a structured record to your CRM: caller name and number, reason for calling, answers to your qualifying questions, outcome, and a short summary with a link to the recording or transcript if you keep them. That turns every phone call into usable pipeline data.

We map the agent's questions to real CRM fields, so a lead from the phone looks just like a lead from your website form. Existing contacts are matched by phone number, so a returning customer's call is attached to their history rather than creating a duplicate. Hot leads can trigger a text to the right rep instantly.

HubSpot, Salesforce, Zoho and similar CRMs all have APIs we can write to. If your CRM is a spreadsheet, we can write there too, though a proper pipeline will serve you better. The same routing rules that assign web leads by territory can assign phone leads, which is one reason voice agents and CRM projects often go together.

AI voice agent cost per minute: what you actually pay for

Running costs are usage-based and come from several providers at once: the phone number and call minutes from the telephony provider, speech-to-text minutes, text-to-speech characters or minutes, language model tokens, and, on a platform such as Vapi or Retell, a platform fee per minute. Each provider bills you directly.

We do not publish provider rates because they change often and depend on your plan and volume. Instead, during the pilot we measure real calls and give you a cost-per-call and cost-per-minute breakdown by component, plus a forecast for your expected call volume. That is the only number that matters when you compare against an answering service or a receptionist's hours.

Costs are easy to control once you can see them. Short, focused scripts mean fewer minutes. Fast, smaller models handle routine turns; larger ones are used only where needed. Spam and robocall filtering prevents paying for junk calls. Hard limits on call length stop an agent from looping. Spending alerts in each provider account catch surprises early.

How much does AI voice agent development cost to build?

With us, a pilot starts from US$600 over 2–4 weeks: one phone number, one call flow, calendar booking or message taking, summaries to email or your CRM, and a test period. A fuller build with several call flows, locations, a dashboard for staff to review calls and edit answers, and deeper integrations starts from US$900 over 6–12 weeks.

US providers price AI voice agent development in very different ways, from monthly platform subscriptions with setup fees to large custom projects, so quotes vary widely. What drives our estimate is the number of call flows, the number and quality of integrations, whether you need a custom dashboard, whether recordings and transcripts need special retention handling, and how much testing your call types demand.

• Call flows: each distinct script, such as booking, emergency and billing, adds design and testing.
• Integrations: calendar, CRM, help desk, practice management or order systems.
• Dashboard: call review, answer editing and analytics for your staff.
• Languages: English first; other languages need your approved scripts and extra testing.
• Volume: high volume can justify custom orchestration to lower per-minute costs.

Designing the conversation: scripts, guardrails and fallbacks

Write the script like you would train a new receptionist: what to say first, which questions to ask in which order, what the agent must never say, and when to hand over. The model handles natural phrasing; the script keeps it on task.

We write guardrails in plain language and in code. The agent does not quote prices unless you give it a price list to read from. It does not give medical, legal or financial advice. It does not promise arrival times the calendar cannot support. When it is unsure, it says so and offers a person or a callback. When a caller is angry or distressed, it transfers.

Testing an AI phone agent before it answers real customers

Test with scripted calls, messy calls and real staff before any customer hears the agent. We build a test set of 40 to 80 scenarios from your real call types, including accents, background noise, callers who change their mind, people who give information out of order, and callers who just want a human.

Your staff then call the agent for a few days and flag anything wrong. We review every transcript, fix prompts and rules, and repeat. Only then do we route a share of real calls, often after-hours calls first, because the risk is lower and the benefit is immediate.

After launch, a weekly review of a sample of calls continues during the maintenance period. Common adjustments include clearer questions, better handling of addresses and names, faster transfers and tighter answers. Platforms like Retell include testing and analytics tools; on a Twilio-based build, we set up our own call review screen.

Health, legal and personal data on AI voice agent calls

If callers will share health, legal or financial details, design for that before the first build. For healthcare, that means choosing telephony, speech and model providers that will sign the agreements your obligations require, limiting what the agent asks, and controlling where recordings and transcripts are stored and for how long.

We build access controls, encryption, audit logs and data minimization into the agent and its dashboard. We do not describe our work as HIPAA-compliant or certified, and we do not sign business associate agreements on your behalf: compliance is your responsibility, confirmed by your own counsel, and the provider agreements are between you and each provider.

Many medical and dental offices choose a narrower design: the agent handles scheduling and general questions, asks only for the minimum needed to book, and transfers anything clinical to staff. Our HIPAA-aware app development page explains how we approach regulated data in general.

Working with an India-based AI voice agent development team from the US

Our evenings in India overlap with US Eastern mornings, and early Pacific calls are possible. That works well for voice projects: we review the previous day's call transcripts while you sleep and send you a summary with proposed changes before your business day starts.

Every account is yours: the phone numbers, the telephony provider, the voice platform if used, the AI model provider and the cloud hosting. We work through access you grant. Quotes are in USD, paid by wire, Wise or PayPal against milestones in your written quote, and nothing is billed before you approve the estimate.

We do not visit your office or install phone hardware; if your desk phones or phone system need changes, your phone vendor handles that part. See how the wider arrangement works on our offshore development team page.

Worked example: AI voice agent development for a hypothetical Phoenix plumbing company

Say a plumbing company in Phoenix with six trucks misses a large share of evening and weekend calls. Customers who reach voicemail call the next company on the list. The owner wants emergencies routed to the on-call plumber and everything else booked for the next day.

An AI voice agent development plan for them would include: a greeting that states the call is answered by an automated assistant and may be recorded; three questions to separate emergencies (active leak, no water, sewage backup) from routine work; ZIP code check against the service area; for emergencies, a warm transfer to the on-call phone with a spoken summary; for routine jobs, two or three next-day slots from the dispatch calendar and a text confirmation; and a record in the CRM for every call.

A pilot like this would start from US$600, running on a voice platform with the company's own number forwarded after hours. Per-minute running costs would be measured during the first two weeks and reported per call. This is a hypothetical scenario to illustrate scoping, not a client story or a promised result.

AI voice agent development checklist before launch

Use this list to judge any voice agent proposal, from us or anyone else. If the answer to any item is “we'll figure it out later”, the launch is not ready.

• The agent's single main job, written in one sentence.
• Script with question order, approved answers and forbidden topics.
• Recording and AI disclosure greeting, reviewed by your counsel for the states you serve.
• Transfer rules: who gets which calls, and what happens if nobody answers.
• Calendar rules: services, durations, buffers, service areas.
• CRM field mapping and duplicate matching by phone number.
• Retention settings for recordings and transcripts.
• Per-minute cost measured on test calls, with spending alerts set.
• Test set of real call types passed, including angry and confused callers.
• A staff member who reviews a sample of calls every week.

Send your answers, even partial ones, through the contact page and you will receive an itemized AI voice agent estimate in USD.