Voice AI agent development
Easital Technologies Ltd. builds AI voice agents: software that answers and places real phone calls, understands the caller, speaks a reply and completes a task such as booking an appointment or qualifying a lead. We build the whole call path, from the phone number and the carrier to speech recognition, the language model, the voice and the record of each call.
An AI voice agent is a program that holds a spoken conversation over a phone line. It turns the caller’s speech into text, uses a language model to decide what to say and do, and turns the reply back into speech, in a loop that repeats for the whole call.
Four parts make up the loop: telephony that carries the audio, speech-to-text, a language model with access to tools, and text-to-speech. Through its tools it can check a calendar, create a booking, query a knowledge base, send an SMS or transfer the caller to a person. “Voice AI agent” and “AI voice agent” are two names for the same thing.
The system it replaces or sits in front of is the interactive voice response menu, the IVR.
| Traditional IVR | AI voice agent | |
|---|---|---|
| How the caller responds | Presses a key or says one of a few set words | Talks in ordinary sentences |
| How the call flows | A menu tree fixed in advance | The model picks the next question, within set goals and rules |
| Requests outside the script | Repeats the menu or queues the caller | Answers from a knowledge base, asks a follow-up or transfers with context |
| What it can complete | Routing and simple lookups | Bookings, lead qualification, messages and record updates |
| Record of the call | Menu choices, sometimes a recording | Recording, transcript and summary |
| Predictability | The same path every time | Wording varies, so it needs testing, limits and monitoring |
| Fits best | Short fixed choices and keypad entry | Callers who explain a need in their own words |
We build six things, separately or together, around your phone numbers and your systems.
An agent that answers a business line: it answers questions from your knowledge base, books appointments, takes messages and transfers the call when a person is needed.
Agents that place reminder, follow-up and lead qualification calls from a contact list, with calling hours, opt-out handling and voicemail detection.
Agents that work beside a human team: first-line triage, overflow, after-hours cover and transfers that arrive with a summary.
Numbers from a carrier, SIP trunks or your existing PBX, connected to the agent with transfer destinations, keypad input, SMS and a fallback route.
The actions an agent can take during a call: calendar bookings, CRM lookups and updates, knowledge-base queries and webhooks into your own software.
A recording, a transcript, a summary and an outcome for every call, in a dashboard your team can search and use to review the agent.
A build runs in seven steps and starts with your real calls. The agent goes live only after it passes test calls you have approved.
We list the reasons people phone, from real calls or call logs, and decide which call types the agent handles and which go straight to a person.
For each call type: the goal, the details to collect, the rules and what the agent must never say. The test calls include interruptions, background noise and off-topic questions.
The number, the carrier or SIP trunk, the transfer destinations and the route a call takes if the agent cannot answer.
Speech-to-text, model and voice are chosen per language. Each stage streams its output to the next, and turn detection is tuned on real audio.
Calendar, CRM, knowledge base and SMS are connected one at a time. A name, a date or a phone number is read back to the caller before it is saved.
Disclosure wording, whether calls are recorded, how long records are kept and how opt-outs work. You and your legal counsel approve these before launch.
The agent starts on a share of calls or limited hours. We read the transcripts with you, fix what fails and extend coverage.
Calldone is the voice system to judge us by. It is live, and its site lets a visitor talk to an agent in the browser.
Client project
Built by Easital for a client in the United States
A voice-AI platform whose agents place and receive real phone calls, book appointments and qualify leads. Its site states that every conversation comes back as a recording, a transcript and a summary, that numbers connect through telephony carriers or a customer’s own SIP or PBX, and that agents use tools for transfers, SMS, knowledge-base queries and webhooks.
Four problems decide whether callers accept a voice agent: the delay before it replies, interruptions, the handover to a person, and the telephone network itself. A text chatbot has none of them.
Every turn passes through speech-to-text, the model, text-to-speech and the network, and the caller hears the sum as silence. We stream each stage so the voice starts on the first words of the reply, keep prompts short and host the parts close together. We measure the delay per stage on your own calls and quote no figure in advance.
People talk over an agent as they talk over each other. The agent has to stop speaking at once and listen, without treating a cough, a “yes” or background noise as an interruption. This is tuned on recordings of real calls.
A transfer needs three decisions: when to hand over, where the call goes and what goes with it. We pass a summary of the call, so the caller does not start again, and define what happens when nobody picks up.
Phone audio carries less detail than app audio, which makes names and numbers harder to recognize. Carriers, SIP trunks and PBX systems differ on transfers, keypad tones and caller ID, and outbound calls must tell a person from a voicemail greeting.
A fifth problem is shared with every language-model product: the agent can say something untrue. We limit answers to your knowledge base, fetch facts such as prices through tools, and test for invented answers before launch.
The rules on automated calls, consent and call recording differ by country and, in the United States, by state. Reviewing them for your use case is your responsibility and your legal counsel’s. Easital builds the controls that the review calls for and does not give legal advice.
In the United States, the Telephone Consumer Protection Act (TCPA) and state telemarketing and recording laws are the usual starting points for that review.
Speech engines, models and carriers are chosen per project and per language, by testing them on your calls. The parts are listed by what they do.
Chosen for accuracy on phone audio in the languages your callers speak, and for how quickly a result comes back.
Chosen for how the voice sounds in each language, the delay before the first word and the price per minute.
We connect to the main cloud telephony carriers and to existing phone systems over SIP.
Any major commercial provider or an open-weight model. On a call, reply speed counts as much as answer quality, so the model is picked by test and can be replaced.
There are three ways to engage Easital for voice work.
One agent for your business, built through the seven steps above on your numbers and connected to your calendar, CRM and knowledge base.
Best for: a company whose phone line is a main source of bookings or leads.
A multi-tenant product of your own: agents, numbers, campaigns and call records, with customer accounts and subscription billing if you sell it to others. Calldone is a platform of this kind.
Best for: companies launching a voice-AI product or a managed AI receptionist service.
We review an agent that replies slowly, talks over callers or mishandles transfers: its call path, prompts, tools and recordings, followed by a written list of fixes in priority order.
Best for: an agent that works in a demo and loses real callers.
An AI voice agent agency sets up and maintains phone agents for businesses: it designs the conversation, connects phone numbers, configures speech and model providers, links the agent to calendars and CRMs, and reviews calls after launch. Some agencies only configure a hosted platform. Easital also engineers the system underneath.
A traditional IVR plays a fixed menu and reacts to key presses or a few set words. An AI voice agent understands ordinary sentences, asks follow-up questions and completes tasks such as bookings. An IVR suits short fixed choices. An agent handles varied requests and needs more testing and monitoring.
Vapi is a hosted platform for building voice agents. The alternatives fall into three groups: other hosted voice-agent platforms, managed services that set up and run the agent for you, and a custom build on your own speech, model and telephony accounts. A hosted platform is enough when you need a standard inbound or outbound agent soon, call volume is moderate and its integrations cover your systems. A custom build is warranted when voice is the product you sell, when per-minute platform fees weigh on your margin, when you need your own carrier contracts, SIP or PBX, or when call audio and transcripts must stay in infrastructure you control.
Usually, yes. Existing numbers can be forwarded or ported to a carrier the agent uses, and a PBX can reach the agent over SIP. Calldone, built by Easital for a client in the United States, supports numbers from cloud telephony carriers as well as a customer’s own SIP or PBX.
That depends on the country, the state and on whether the agent answers calls or places them. You need to review consent, disclosure and recording rules with legal counsel before launch. We build the controls that review requires: disclosure wording, recording settings, opt-outs, calling hours and retention periods.
Four meters run during every call: telephony minutes, speech-to-text, the language model’s tokens and text-to-speech. Longer calls, longer prompts and premium voices raise the cost per call. We measure cost per call from the first prototype. See LLM cost optimization.
Describe who calls, why they call and where a call should go when the agent cannot help. We answer by email with our questions and a suggested first step.
Easital is an AI and SaaS engineering company that takes AI software to production, and runs AI products of its own. Founded in 2019.