AI voice agent

How it works

How it works

Every turn in the conversation goes through three steps. Speech recognition turns the caller's words into text. A language model works out what the caller needs, finds the answer in the knowledge base or the data in your CRM, and decides what to say or do. Speech synthesis speaks the answer. For the conversation to sound natural, the whole chain has to fit into about one second.

• Speech-to-text – hears what the caller said
• Language model – understands the question and finds the answer in the knowledge base
• Actions via API – check a status, book, update data
• Text-to-speech – speaks the answer
• Transfer to a human agent with a summary of the conversation

Use

Which calls it handles

• Order, booking, or delivery status
• Booking and rescheduling appointments
• First point of contact: find out what the caller needs and route the call to the right team
• Outbound reminders, confirmations, and short surveys
• Calls at night and during peak hours, when agents cannot keep up

Practice

What to check before launch

• How well the agent recognizes speech on your real recordings – with your terms, addresses, and product names
• The pause before each reply: if it runs longer than 1-1.5 seconds, callers start repeating themselves
• When the agent transfers to a human: the caller asks for a person, the agent is not sure of the answer, or the caller is upset
• Access to your systems via API: without it, the agent can only answer questions, not resolve them
• Cost per minute of conversation, and where recordings and customer data are stored

In CONTACT.WORK

How CONTACT.WORK uses it

The agent runs on the same telephony as your human agents, draws on the shared knowledge base, and hands the call to a person with a summary when needed. An IVR and the agent often work together in the same call. For simple facts – order status, account balance, appointment time – the IVR can read them out automatically.

The platform itself is free. Speech recognition, voice, and the language model are paid separately, or you can run them on your own hardware. A minute with the agent costs more than a minute in an IVR; the fairer comparison is with the cost of a human agent's minute.

You can host the agent in the CONTACT.WORK cloud or on your own infrastructure. The server calculator shows how much hardware the agent needs to run on your servers.