🤖

Azure AI Foundry Overview and Demo

Jun 11, 2025

Summary

  • This session, led by Scott Hanselman and Yina Arenas, focused on demonstrating Azure AI Foundry’s capabilities using the Hansel Minutes podcast production as an example.
  • Key topics included an overview of AI models, the new agentic platform, and observability tools for monitoring and managing AI applications.
  • Several live demos highlighted model selection, multi-agent workflows, efficient cost management, and end-to-end security and monitoring, including new features announced at Microsoft Build.
  • Attendees saw practical automation of podcast workflows, with an emphasis on human oversight, efficiency, and cost reduction.

Action Items

  • None explicitly assigned during this session.

Introduction and Overview of Azure AI Foundry

  • Azure AI Foundry is an open, flexible, and secure platform enabling developers to infuse AI into any application, whether new or existing.
  • Three core focus areas: AI models, an agentic platform, and observability tools.
  • The session used the Hansel Minutes podcast workflow as a practical example to showcase Foundry’s capabilities in automating complex, repetitive tasks.

Models: Catalog, Selection, Fine-Tuning, and Local Execution

  • Azure AI Foundry now offers access to over 10,000 models, including from OpenAI, DeepSea, Mistral, Meta, and others, with easy switching and unified access.
  • The new “Model Router” feature helps route queries to the most appropriate model, optimizing for cost and efficiency (up to 60% cost reduction vs. using single large models).
  • Fine-tuning and model distillation capabilities demonstrated—for example, reducing cost of transcribing and summarizing 1,000 podcast episodes from about $100 to $1.50 using a dedicated, smaller model.
  • "Foundry Local" allows running models on local machines, not just in the cloud, demonstrated by running a summarization task on a laptop GPU.
  • Demonstrated leaderboards for model benchmarking and the use of own datasets to evaluate model results.

Agents: Workflow Automation and Orchestration

  • Clarified definitions: Agents orchestrate tasks, using tools and sometimes multiple models, supporting memory, multi-modal input/output, and human-in-the-loop workflows.
  • Recommended building composable, single-purpose agents that can be orchestrated together for complex workflows.
  • Two workflows supported: connected agents (API-like tool invocation) and multi-agent workflows (for processes with human steps).
  • General Availability announced for Azure AI Foundry Agent Service, enabling declarative agent creation, enterprise integration, tool use, and interoperability with other agent platforms.
  • Showcased multi-stage podcast workflow automation: guest intake, bio generation, transcript creation, show notes drafting, link verification, and agentic search across 20 years of podcast content.
  • Emphasized grounding and restricting agent responses to the relevant corpus and domain, highlighting efficiency, security, and cost benefits.

Observability, Monitoring, and Security

  • Foundry provides tools for experimental tracking, production monitoring, debugging, and reliability—including integration with OpenTelemetry, Azure Monitor, Application Insights, and third-party tools like Grafana.
  • Agents gain unique enterprise identities (integration with Entra), supporting scoped permissions and detailed governance.
  • Security features include data labeling, data protection, and full integration with Microsoft Purview and Defender.
  • Evaluation metrics for agent performance (e.g., intent, relevance, task adherence) can be incorporated into CI/CD pipelines, with automated evaluation runs upon deployment.
  • Emphasized the importance of observability in distributed AI systems and maintaining human oversight in automated workflows.

Practical Demonstrations: End-to-End Podcast Automation

  • Live demos covered the full podcast production flow: transcribing audio, generating show notes, verifying links, assembling summaries, and producing YouTube-ready transcript files.
  • Highlighted efficient use of fine-tuned and distilled models tailored for specific tasks (e.g., show notes generation, not general-purpose chat).
  • Showed error cases and transparent evaluation of model outputs, reinforcing the importance of continuous monitoring and improvement.
  • Announced upcoming support for markdown in Notepad as a productivity booster.

Use Cases, Customer Impact, and Future Opportunities

  • Over 70,000 customers are using Azure AI Foundry, with over 10,000 leveraging the agent service to create millions of agents.
  • Practical applications include guest sourcing automation, localization/translation, and social content generation.
  • Encouraged attendees to consider how to apply AI Foundry’s automation features to reduce “dull, dirty, or dangerous” work in their own domains.

Decisions

  • General Availability of the Azure AI Foundry Agent Service — Announced to enable enterprise-ready, declarative agent workflows.
  • Integration of Observability and Security — Committed to supporting industry standards (OpenTelemetry, Entra, Purview, Defender) for monitoring and governance.

Open Questions / Follow-Ups

  • None explicitly noted; session concluded with an invitation for attendees to explore further and a call to action to apply these tools.