Microsoft Releases Decision-1 Model for AI Agent Workflow Control
The new scoring model provides calibrated probabilities for routing and classification tasks in enterprise applications.
Key points
- Microsoft-Decision-1 delivers median latency roughly 35 times faster than GPT-6 Sol.
- The model was post-trained from Qwen3.5-9B and evaluated on 150,000 blind questions.
- Internal tests showed competitive accuracy against larger models at significantly lower cost.
Microsoft released Microsoft-Decision-1, a specialized model designed to help artificial intelligence agents determine their next actions. According to Help Net Security, the tool provides calibrated probability scores for fixed answer options in specific situations. It aims to integrate decision intelligence into existing workflows securely.
The model handles routing, classification, prioritization, and verification tasks. Achint Srivastava, vice president of software engineering at Microsoft, stated it supports these functions in trusted environments. It is not built for open-ended text generation, translation, or summarization.
Technical Performance and Benchmarks
Microsoft evaluated the system across 36 benchmarks containing nearly 150,000 questions. These benchmarks were kept blind from the training process to ensure valid testing. The company reported the model operates 2.5 times faster than H2O-Lightning-4B v1.1.
In terms of latency, Microsoft-Decision-1 processed requests about 35 times faster than GPT-6 Sol. The probability scores indicate confidence levels for each potential answer. Developers can use these scores to decide whether to act, defer decisions, or request human review.
Internal Use Cases and Costs
Microsoft tested the model on tasks such as sorting customer feedback and evaluating AI answers. Xbox researchers used it to categorize over 10,000 pieces of feedback from surveys and social platforms. The results matched the quality of GPT-6 Sol while running 14 times faster.
The Copilot team assessed the model’s ability to judge chat and agent responses. Microsoft reported performance competitive with GPT5.6 Luna, with processing speeds 100 times faster. The company noted the processing cost was one two-hundredth of the alternative.
What should you do?
Developers should review the model’s limitations before integration. Microsoft advises validating confidence thresholds on examples that match intended use cases. Teams should monitor how the model handles edge cases in their specific workflows.
What to do and how to stay safe: Microsoft
- Evaluate the model’s accuracy on datasets representative of your specific application domain before deployment.
- Set strict confidence thresholds to determine when the system should defer decisions to human reviewers.
- Monitor latency and throughput metrics to ensure the model meets performance requirements for real-time workflows.
- Verify that the model is not used for tasks outside its scope, such as open-ended generation or translation.
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Is Microsoft-Decision-1 capable of generating open-ended text or summaries?
No, the model is designed strictly for decision-scoring tasks like routing and classification, not for text generation or summarization.
How does the speed of Microsoft-Decision-1 compare to GPT-6 Sol?
According to Microsoft, the new model is approximately 35 times faster than GPT-6 Sol in median latency benchmarks.
What base model was used to create Microsoft-Decision-1?
The system was post-trained from the Qwen3.5-9B model for single-pass scoring capabilities.
