AI call agent

The call arrives through a telephony interface. Speech is converted to text continuously, a language model decides on the reply, and a synthetic voice speaks it. For the conversation to feel natural, this chain has to run in well under a second. So we keep replies short, allow interruptions and load customer data while the call is still connecting.

The real value comes from actions: the agent checks free slots in the calendar, creates a record in the CRM or reads out the status of an order. It only works with the systems we have explicitly connected.

Calls with a clear goal and few variants work best: booking, moving or cancelling an appointment, opening hours and standard questions, status checks, taking callback requests, pre-qualifying enquiries. An agent is also reachable around the clock and absorbs peaks when the line would otherwise be busy.

Complaints, negotiations and anything where the caller expects a human are not a fit. The agent recognises these calls and transfers them — with a short summary, so nobody has to explain their issue twice.

Speech technology is now the smaller part. Most of the work is in the integrations and in testing real conversation paths.

• The number of requests the agent should close on its own
• Systems it accesses: calendar, CRM, ticketing, knowledge base
• Languages and accents of your callers
• Running costs: telephony and model usage are billed per minute and depend on call volume

Callers are told at the very start that they are speaking to an AI. The EU AI Act requires it, and it lowers the hang-up rate because nobody feels deceived. Whether calls are recorded, how long transcripts are kept and who may read them is agreed before launch.

After go-live we review conversations: where did callers ask again, where did they hang up, what was handed to people? Those cases shape the next version. How we fit such agents into existing processes is described on the AI automation page.