A traditional receptionist answers during staffed hours, follows a script, and hands off to colleagues. An AI receptionist can extend routine intake to nights and weekends, subject to phone-network availability, system capacity, and a tested overflow path. It should not be assumed to complete every call.
The core components are a speech recognition layer (STT), a language model that interprets the caller's intent and runs the qualifying conversation, a text-to-speech voice that speaks back, and integrations that drop the appointment into a calendar and write the lead record to a CRM.
A conversational agent can ask clarifying questions instead of forcing a phone tree. It still needs scope limits, tests for unsafe or regulated advice, human transfer, and a clear failure path; a prompt alone does not guarantee those controls.
Locally run speech and language components can operate on infrastructure the operator controls. Public phone numbers still require a carrier or SIP trunk, so the call path, records, consent, and provider terms need to be mapped before claiming data control or lower cost.