A voice agent that onboards patients and books appointments

Client Industry

Multi-location healthcare · clinic network

Key Challenge

Manual, phone-bound patient onboarding & scheduling

Core Tech Stack

Dialogflow CX · Vertex AI · Twilio · Cloud Functions

Deployment Model

Event-driven serverless on Google Cloud

SoftmaxAI has developed a production-quality conversational voice agent which acts as a front end for a clinic network's inbound telephone line, handling patient intake and the booking of appointments end to end over the public switched telephone network (PSTN). The agent identifies callers using their incoming number, distinguishes between new and returning patients, and manages the whole conversation on Google Cloud's serverless platform — only transferring the call to human staff when it is genuinely necessary to do so.

• Caller status was ambiguous. Before any intake could begin, the staff had to manually decide whether each caller was a new or returning patient — a repetitive triage procedure on every single call.
• Onboarding was slow and error-prone. The practice of obtaining demographic and intake information by voice left room for errors during transcription and for the creation of inconsistent records.
• Scheduling demanded live availability. Booking required a real-time check against the actual calendar availability — not against a fixed rota — while also preventing double-booking when calls came in at the same time.
• Conversations wandered off-script. Callers tend to interrupt and ask unpredictable questions; a strict IVR decision tree is unable to maintain a coherent dialogue when the conversation moves away from the happy path.

The agent operates as a modular, event-driven pipeline on Google Cloud, following a data flow from live telephony inputs, through conversational processing and serverless integration, and then to voice serving.

• Spec before code. Before writing a single line of production code, we prepared a specification which had been approved by the client and that included mappings for every call flow, patient state, intent taxonomy, escalation rule, and data-handling boundary.
• Hybrid orchestration. Dialogflow CX manages the conversation by means of an auditable state machine to ensure reliable dialogue flows, and Vertex AI's generative fallback maintains the coherence of the dialogue when callers go off-script.
• Serverless, event-driven integration. Cloud Functions provide scheduling and record-handling logic in the form of stateless, independently scalable endpoints which are invoked on demand from the conversation and do not require always-on middleware to be maintained.
• Human-owned decisions. The escalation thresholds, the conversational tone, and the way PHI was handled remained the responsibility of humans, and were validated against realistic call scenarios until the behavior had become dependable before go-live.

• Eliminated manual caller triage. The agent is able to identify callers using their inbound number at the edge and automatically routes new and returning patients, thus eliminating a repetitive task from each call.
• End-to-end self-service onboarding. New patients are guided through a structured intake conversation by the agent, resulting in records that are consistent and without the need for staff to transcribe.
• Real-time scheduling orchestration. Appointments are made in accordance with the current calendar availability via the scheduling API, eliminating the need for back-and-forth coordination and replacing it with a single automated process.
• Graceful human handoff. Routine calls are handled automatically; those that are unclear or sensitive are properly passed on to a human, enabling staff to focus their attention on the ones that actually require it.
• Horizontally scalable by design. Because the architecture is stateless and serverless, it is able to handle a high volume of calls at the same time without having to provision resources or re-architect the system — the amount of capacity increases in accordance with demand.