Build a Voice AI System for Your Maintenance Team

Most property managers first try virtual receptionist services like Ruby. These services are staffed by humans who take messages and send emails, but they cannot create a structured work order in your PMS. A coordinator still has to manually re-enter the data. A single misheard address, like '13th Street' instead of '30th Street', wastes two hours of a technician's time and fuel.

Others attempt to build a simple IVR with a tool like Twilio Studio. The 'Press 1 for plumbing, Press 2 for electrical' flow is too rigid for real-world issues. It cannot understand a frantic tenant reporting that 'water is coming through my ceiling from the apartment above'. This urgent request gets routed to a generic voicemail, delaying the response by hours and leading to thousands in water damage.

The fundamental problem is that these tools cannot ask intelligent, clarifying questions. When a tenant reports a 'broken appliance', the system needs to ask 'Which appliance?' and 'What is it doing?'. Without this conversational ability, work orders are incomplete, forcing your team to make a follow-up call anyway and delaying repairs by a full day.

Syntora's approach to developing a voice AI system for property maintenance starts with configuring a dedicated phone number through a service like Twilio. When a tenant calls, an AWS Lambda function would receive the incoming call, integrating with a real-time transcription service to capture the conversation.

Initial transcription and issue classification, such as identifying plumbing or HVAC problems, would aim to occur rapidly, typically within a few seconds. We would implement Python with the FastAPI framework to manage the webhook from Twilio, providing a high-performance and scalable API for the system.

The live transcript would then be sent to the Claude API, using a carefully constructed prompt. This prompt would instruct the AI to function as a maintenance dispatcher: to extract essential information like the unit number, tenant name, contact details, issue category, and a comprehensive description. The AI would be designed to ask clarifying questions if critical details are missing, for example, 'Can you confirm the address and unit number for me?'. We anticipate that typical AI-driven conversations would take an average of 90 seconds.

Once all necessary data is collected, a separate Python function would format this information into a structured JSON payload. This payload would then be posted directly to your existing Property Management System (PMS) API, whether for platforms like AppFolio, Buildium, or a custom-built solution.

For logging and audit trails, Syntora would establish a Supabase database to record every call, transcript, and generated work order. The objective would be for successful work order creation within the PMS to happen very quickly after the call concludes.

Finally, the system would be configured to send an immediate confirmation SMS to the tenant and a Slack notification to the maintenance coordinator, including a direct link to the new work order. We would integrate `structlog` for structured JSON logging, which facilitates setting up CloudWatch alerts. These alerts could monitor for anomalies such as an API error rate exceeding 1% or Claude API latency going above 4 seconds. For an estimated volume of 1,500 calls per month, the typical AWS infrastructure costs are generally under $75.