Voice AI development: AI voice agents built on your phone lines

Start by writing down the ten call types that make up most of your volume and what a good outcome looks like for each. Then connect a phone number through a carrier such as Twilio or a SIP trunk, pick speech to text, a language model and a voice, and give the agent tools to read and write your calendar or CRM. Test on recorded calls, then launch on a share of traffic.

There is no single best model. For most phone agents a fast, smaller model such as GPT 5 mini, Claude Haiku or Gemini Flash handles the conversation, because response speed matters more than depth on a call. Speech to speech options such as the OpenAI Realtime API cut latency further. We choose by testing candidates on your own recorded calls.

A custom AI receptionist for one line typically costs $10,000 to $30,000 to build, an integrated voice agent $30,000 to $75,000, and a contact centre voice platform from $75,000. Running costs are separate: our typical estimate is 8 to 25 US cents per minute for speech, model, voice and telephony, plus support from $1,500 per month. You get one fixed build price in writing after a free scoping call.

An AI receptionist is one kind of AI voice agent. It answers inbound calls, handles common questions, takes messages, books appointments and transfers to staff. An AI voice agent is the broader category: it can also run customer service flows, make outbound reminder or qualification calls, take payments through a compliant flow, or sit inside a contact centre alongside human agents. The engineering underneath is largely the same.

It is worth it when you miss calls that turn into lost bookings, or when staff spend hours a day on the same five questions. With fewer than ten mostly complex calls a day, an answering service or AI receptionist subscription may be cheaper. A custom build pays off when the agent must book into your own systems and follow your rules exactly.

People expect a reply roughly 200 milliseconds after they stop speaking in normal conversation, so any delay beyond about one second feels awkward on a phone call. We engineer for a voice to voice response under about one second, measured from the moment the caller stops talking to the first sound of the reply. That budget covers end of speech detection, transcription, the language model, speech synthesis and the telephone network.

Yes, and it should. We build warm transfers where the agent tells the caller what is happening, passes a short summary to the staff member or contact centre queue, and hands over the call through SIP or your phone system. Transfers trigger on request, on low confidence and on sensitive topics. After hours, the agent takes a detailed message instead.

Often yes, with conditions. In the US the FCC ruled in 2024 that AI generated voices count as artificial voices under the TCPA, so marketing calls need prior express written consent. The UK treats automated marketing calls under PECR as needing specific consent, and the UAE restricts telemarketing hours, requires consent and bans calls to numbers on its Do Not Call Registry. Inbound answering is far less restricted.

Yes. We build bilingual agents that detect the caller’s language in the first few seconds and continue in it, including switching mid call when a caller mixes Arabic and English, which is common in the Gulf. Accuracy varies by dialect, so we test Gulf, Levantine and Egyptian Arabic samples from your callers before choosing providers.

We can, with the person’s documented, informed consent and a written agreement on how and where the voice is used and how the consent can be withdrawn. Providers such as ElevenLabs require verification that the speaker agreed. Several US states, including Tennessee under its ELVIS Act, protect voice likeness. Many clients choose a licensed stock voice instead, which avoids the question entirely.

Use a platform when your calls are simple, volume is moderate and the built in integrations cover your systems. Platforms such as Vapi and Retell are quick to start and we often build on them. Choose custom engineering when you need deep integration with practice management or legacy phone systems, strict data residency, Arabic dialect tuning, or lower cost per minute at high volume.

The best AI voice agent for customer service is the one that resolves your most common call types end to end without making callers repeat themselves. That depends less on the brand and more on integration with your order, booking or ticketing system, a clean handoff to people, and measured accuracy on your own calls. Compare vendors on resolution rate, latency, cost per minute and data terms, not on demo voices.