#AI Horizons 25-03 – AI Foundry Innovations

Microsoft Azure continues to expand its AI capabilities leveraging AI Foundry with several significant innovations designed to enhance developer productivity and business automation. These new features represent a substantial evolution in how organizations can implement and scale AI solutions across their technology stack.

Responses API and Computer-Using Agent

The newly introduced Responses API from Azure OpenAI unifies the best capabilities from chat completions and assistants APIs into a single, cohesive experience. This stateful API streamlines complex AI interactions and supports the groundbreaking computer-use-preview model.

Responses API: Simplifying Complex AI Interactions

The Responses API delivers several key benefits for developers:

• Seamless Integration: Enables AI applications to retrieve information, process data, and execute actions within a unified framework
• Tool Calling: Integrates multiple AI tools in a single API call, significantly enhancing operational efficiency
• Enterprise-grade Privacy: Maintains data protection according to Azure’s established security standards

Computer-Using Agent: AI That Navigates User Interfaces

The Computer-Using Agent (CUA) represents a significant advancement in AI automation capabilities:

• Autonomous UI Navigation: Interacts directly with graphical user interfaces to automate complex tasks
• Dynamic Adaptation: Adjusts actions based on UI changes, reducing dependence on brittle automation scripts
• Cross-Application Execution: Functions across both web and desktop applications using natural language commands
• No API Dependencies: Integrates disparate systems without requiring pre-built API connections

The CUA can perform a wide range of operations including opening applications, clicking buttons, completing forms, and navigating multi-page workflows. This capability enables organizations to automate processes across previously disconnected systems.

Access Note: The computer-use-preview model requires registration, with access granted based on Microsoft’s eligibility criteria. Even customers with access to other limited-access models must specifically request access for this model through the computer-use-preview limited access model application.

Service Enhancements

Provisioned Spillover (Preview)

Azure OpenAI Service has introduced Provisioned Throughput Unit (PTU) spillover, now in public preview. This feature:

• Manages traffic fluctuations on provisioned deployments by routing excess traffic to a designated standard deployment
• Ensures consistent performance during high usage periods
• Maintains uninterrupted critical processes
• Provides flexibility to handle unexpected traffic bursts or seasonal demand peaks

For technical implementation details, see Azure’s documentation on managing traffic with spillover for provisioned deployments.

Content Filtering Configurations

Azure now provides more granular control over content filtering:

• In addition to deployment-level content filtering, developers can now specify custom configurations at the individual request level
• This is implemented via a request header for each API call
• For implementation details, refer to the documentation on using content filters

Databricks Foundation Model Update

Databricks Foundation Model APIs have expanded to include Anthropic models, providing developers with:

• Access to state-of-the-art open models from Databricks-hosted serving endpoints
• Multiple access options:
  • Pay-per-token: Suitable for initial experimentation but not optimized for high-throughput applications
  • Provisioned throughput: Recommended for production workloads requiring high performance and security
  • Batch inference: Designed for batch processing workloads using AI Functions

Why It Matters

These innovations transform how developers and IT professionals can implement AI across their organizations:

1. Reduced Integration Complexity: The unified Responses API eliminates the need to manage multiple API interactions, accelerating development cycles
2. Automation Beyond APIs: The Computer-Using Agent enables automation of legacy systems and applications without available APIs or integration points
3. Resource Optimization: Provisioned spillover allows organizations to maintain performance during peak periods without over-provisioning resources
4. Flexible Control: Request-level content filtering provides more precise governance over AI outputs
5. Expanded Model Access: The addition of Anthropic models to Databricks broadens the AI toolkit available to developers

For development teams, these advancements mean faster implementation, broader automation capabilities, and more flexible resource management—ultimately enabling more sophisticated AI applications with less development overhead.

Organizations implementing these technologies can expect to see improvements in operational efficiency, reduced manual processes, and enhanced ability to automate complex workflows across previously siloed applications.