A Voice AI Receptionist That Understands Accounting Clients

Most accounting firms first try their phone provider's built-in auto-attendant. This leads to the classic "Press 1 for..." menu, which frustrates clients and cannot handle any query that falls outside its rigid structure. A high-value prospect calling about partnership tax strategy will hang up before they navigate four menu levels to leave a voicemail.

Some firms then try virtual receptionist services like Ruby or Smith.ai. These are an improvement, but they are expensive, often costing over $500/month during busy seasons. The receptionists are shared resources who follow a generic script. They cannot answer firm-specific questions like "Is John the right person for trust and estate work?" and their primary function is to take a message, delaying a response by hours.

Off-the-shelf voice AI tools included with VoIP plans like RingCentral have limited intelligence. They can transcribe a voicemail, but they cannot check a partner's real-time availability in Calendly and book an appointment directly. They lack the deep API integrations needed to provide actual answers, so they just become a slightly more advanced answering machine.

Syntora's approach begins with a discovery phase to understand your specific operational needs and existing systems. We would then design and build the necessary integrations. This would involve using Python with httpx to connect to your scheduling tool, such as Calendly or Microsoft Bookings, and your practice management software, like Karbon or TaxDome. These integrations would allow the voice AI to check partner availability and query client status in real time.

The core of the system would be a FastAPI application, which we would deploy on a cloud platform like AWS Lambda. This application would receive call data from a programmable voice provider, such as Twilio. To interpret caller intent and drive the conversation, we would integrate with Anthropic's Claude 3 Sonnet API. This would enable the AI to differentiate between, for instance, a new lead and an existing client requesting a document, guiding each caller through appropriate conversational pathways.

To provide firm-specific answers, we would construct a knowledge base using a Supabase Postgres database with the pgvector extension. This database would be populated with your service descriptions, partner biographies, and answers to frequently asked client questions. When a caller asks about specific services, the AI would query this specialized knowledge store to deliver accurate and relevant information.

The delivered system would include observability features. We would configure structured logging using tools like structlog to help monitor system performance. If the AI system detects it cannot understand a caller after multiple attempts, or if an external API call to your calendar fails, the system would be designed to route the call to a human operator. It would also send an alert to your team's communication channel, such as Slack, detailing the issue. During our engagement, we would provide estimates for anticipated hosting costs, which typically remain low for systems processing a few thousand calls per month.