Realtime Voice AI AGENTS Will Explode in 2025 | SHOWCASE

Real-time voice AI agents are moving from demos to practical business workflows—using function calling to check availability, confirm bookings, and write results into a database during a live phone conversation. The showcase centers on an “AI Dental” receptionist that answers calls, gathers appointment details, queries a scheduling database for open slots, rejects unavailable times, and then books the customer automatically.

The system is built around a real-time API plus a tightly controlled system message that defines the agent’s role and the exact tools it can use. When a caller asks for a specific time—like “9 a.m. tomorrow”—the agent triggers a function call that checks the database for that exact date and slot. If the requested slot is already taken, the agent responds with an apology and offers alternatives by calling another function to list available slots. Once the caller selects an open time (for example, “11 a.m.”), the agent confirms the appointment, collects contact information and any special requirements, and records the booking.

A key operational detail is the pipeline that turns a live call into structured data. The conversation is recorded and converted into an MP3 file, then transcribed using Whisper. From the transcript, structured outputs extract the fields needed to fill the appointment schema—first and last name, appointment date, chosen slot, contact email/phone, and special requests (such as requesting “calm background music” or even a “beer” to be offered during the visit). Those extracted values are then saved to the database and reflected on a dashboard, demonstrating end-to-end automation from voice interaction to persistent scheduling records.

The agent’s behavior also depends heavily on instruction-following. The system message instructs the agent to greet and identify the caller’s needs, collect required information, use the “list available slots” tool when asked for options, use the “check availability” tool when a specific time is requested, and always read back the confirmed date and time. The creator notes that some fields can be buggy (special requirements and certain extracted details aren’t always perfect yet), but the overall booking loop works reliably enough to be compelling.

Beyond scheduling logic, the showcase highlights voice and personality flexibility. Using the real-time API playground, different voices (e.g., “Sage” and “Ash”) can be selected, and the agent’s spoken style can be adjusted with voice configuration settings such as accent and pacing. Quick call snippets show the receptionist persona shifting while still performing the same appointment workflow.

The takeaway for 2025 is less about a single dental use case and more about a repeatable pattern: real-time voice input, tool-based function calling for database-backed decisions, structured extraction for reliable record-keeping, and configurable voice/personality for user experience. With more testing before production, the approach points to a broader wave of voice agents that can handle customer interactions—answering questions, checking availability, and completing transactions—without human intervention.