AI Receptionist

The call arrives on an ordinary phone number at a programmable-voice provider. The provider streams the caller’s audio to the application in near real time, the speech is transcribed, a language model decides what to say or do next, and synthesized speech is streamed back into the same call. Phone audio is narrowband: Twilio’s Media Streams carry μ-law audio at 8,000 Hz, and voice-agent platforms that connect to Twilio require their speech input and output to be configured for that format.

• • Telephony: receives the call and streams audio both ways over a WebSocket.
• • Speech-to-text: turns the caller’s speech into text the model can read.
• • Language model: interprets the request, chooses a reply and decides when to call a tool.
• • Text-to-speech: speaks the reply back into the call.

Typical jobs are answering routine questions (hours, service area, what the business does), qualifying the caller, capturing name, number and reason for calling, booking a time, and writing a summary afterwards. Actions such as booking use function calling: the model emits a structured request, for example to check availability on Tuesday morning, and the application validates it, runs it against the real calendar and returns the result to the model before it replies.

A language model answering from its training data alone can confabulate, which NIST defines as producing “confidently stated but erroneous or false content”, and a caller has no easy way to check what they were told. Receptionists should answer from material the business supplies (services, hours, policies), retrieved at call time, and should be explicitly allowed to say they do not know and offer a callback rather than guess. NIST also notes that a language model could falsely claim to be human, which is a reason to configure the agent to be honest about what it is.

Anything a caller says is input to the model, so a caller can attempt prompt injection (“ignore your instructions and…”). OWASP’s guidance for LLM applications applies directly: give the model only the tools it needs, with the minimum permissions, handle privileged operations in code, require human approval for high-impact actions, and enforce authorization in the downstream system rather than trusting the model to decide.

• • Allow: reading published business information, checking open slots, booking into open slots, taking a message.
• • Keep behind a person: refunds, price changes, cancelling other customers’ appointments, reading other customers’ records.

This is general information, not legal advice. In the United States, federal wiretap law permits recording when one party to the call consents, but some states require every party’s consent; California Penal Code § 632, for example, prohibits recording a confidential communication “without the consent of all parties”. Announcing at the start of the call that it may be recorded is common practice for that reason. Outbound calls carry separate rules: in February 2024 the FCC confirmed that AI-generated voices are an “artificial or prerecorded voice” under the TCPA, so calls that use them require the called party’s prior express consent unless an emergency purpose or exemption applies. That matters if an AI system places callbacks, not only when it answers.

DoubleXL Answer is an AI receptionist for home-services and appointment-based businesses. It answers the business’s line 24/7, qualifies the caller and answers routine questions. It does not write to the calendar itself: when a caller wants an appointment it either texts them the business’s booking link after the call or captures their details and preferred times for the team to confirm. Every call is saved as a transcript with a short summary, and the business sets what the assistant says, what it must never say, and how it handles appointment requests. Current plans are listed on the DoubleXL pricing page.