Build a Custom Voice AI for Your Freight Forwarding Team

Most freight operations start with a standard IVR phone tree. This system can route calls but cannot answer questions. A customer calling for a tracking update must still wait for a human agent, which defeats the purpose of automation and ties up staff.

Off-the-shelf voice AI platforms from major telephony providers seem like the next step. However, they are trained on generic retail or support conversations. They fail when a customer provides a PRO number instead of a house bill of lading number. This ambiguity leads to incorrect lookups and customer frustration. These platforms also charge per minute, and a high volume of 30-second status calls quickly becomes more expensive than a full-time employee.

A 12-person forwarder we worked with tried a well-known AI contact center solution. Its transcription model had a 30% error rate on logistics-specific acronyms. Customers were constantly escalated to human agents, who then had to re-ask for the tracking information. The manual workload barely changed, but the company had a new $1,200 monthly software bill.

Syntora would begin an engagement by auditing your existing Transportation Management System (TMS) and other relevant data sources. We would then design and implement API connections to these systems.

The proposed system would use the Claude API's function-calling feature to define specific, permissible actions, such as get_shipment_status(tracking_id) or lookup_container_eta(container_id). This pattern ensures a reliable translation between natural language requests and your internal systems. We've built document processing pipelines using Claude API for financial documents, and a similar architectural pattern applies to freight forwarding documents and queries.

The core of the proposed system would be a Python application built with FastAPI. It would receive calls through a Twilio phone number, which would pass the audio stream to the service. A real-time transcription service would convert speech-to-text. The transcribed text would then be fed to a Claude 3 Sonnet model. We would engineer a system prompt specifically for freight forwarding terminology to ensure the model understands the context and nuances of your business. The model's response would be converted back to audio using a low-latency text-to-speech engine.

The application would be designed for deployment on AWS Lambda, providing automatic scaling capabilities. Call transcripts and interaction logs would be stored in a Supabase database. As part of the deliverable, a simple dashboard would be provided to track call volume and identify common query patterns.