Microsoft Releases AI Call Center Stack with Voice, SMS, and Memory (github.com)

Microsoft’s new Azure-based AI Call Center Stack packages speech, SMS, RAG and OpenAI LLMs into a ready-to-deploy proof-of-concept for automated customer voice interactions. The stack lets developers initiate outbound calls via a single API or expose a public phone number, uses Cognitive Services for STT/TTS/translation, streams conversations in real time (resume after disconnects), and persists transcripts, structured “claim” schemas, reminders and to‑dos in Cosmos DB with Redis caching. It integrates gpt-4.1 and gpt-4.1-nano for nuanced understanding, generates embeddings (ADA) for search/RAG, includes monitoring (Application Insights), feature flags, human-agent fallback, call recording, and supports brand-specific custom voices. A French demo and a GHCR container image + Bicep/Make deployment templates accelerate getting started. For AI/ML teams this is significant because it demonstrates an end-to-end pattern for production-ish voice agents: low-latency voice I/O, retrieval-augmented generation over internal documents, caching and queueing for scale, and guardrails like content filtering and jailbreak detection. Key trade-offs and technical implications include LLM cost/performance choices (gpt-4.1 variants with higher price for better performance), the need to provision LLM endpoints to cut latency, and careful RAG/security controls for sensitive customer data. Note: the project is a POC intended for experimentation rather than turnkey production use.