Eigen RadarAI
Archive
Analysis

Microsoft Decision-1 returns scores for choices applications supply

Microsoft has introduced Decision-1 in Foundry’s public preview for classification and workflow decisions. Developers supply the question and permitted choices; the model returns structured scores and confidence signals. Applications can use those signals to route a request or seek human review. Microsoft describes internal uses in customer feedback, response assessment and scientific workflows.

Artificial Intelligence··Midday
A game controller on a wooden desk beside a display grouping terrain textures, geometric forms and game objects. A different controller rests on the lower shelf.

Developers supply the choices and receive structured scores

Microsoft has introduced Decision-1 in the public preview of Foundry, its platform for building AI applications. Developers provide a text, a question and predefined choices. The model returns scores and confidence signals in structured form, including JSON, a format software can read directly. Its listed tasks include classification, intent analysis and routing requests to the appropriate team.[1], [2]

The model is built on Qwen3.5-9B, with further training by Microsoft. It evaluates supplied options without generating a written rationale. The application sets the choices, score thresholds and conditions that send a case to human review. Tasks requiring information absent from the input fall outside the scope described in the model catalog.[2]

A response can be generated before a separate quality check

Microsoft describes a workflow in which a generative model drafts a response and Decision-1 evaluates it against specified criteria. Other decisions include continuing a workflow, retrying a step, stopping or requesting human review. A Foundry Playground demonstration assesses a customer request’s urgency, selects a support team and assigns a priority score, placing the classifier inside a larger service workflow.[1]

Internal trials include feedback sorting and response assessment

Microsoft says Xbox Research, its game-research team, used Decision-1 to sort more than 10,000 pieces of open-ended feedback into themes researchers supplied. It also describes tests assessing Copilot responses and replanning in Microsoft Discovery, its scientific research platform. The model catalog warns against using it as the sole basis for consequential lending, employment or medical decisions.[1], [2]

References

  1. News sourceMicrosoftMicrosoft Decision-1 scores predefined choices instead of generating text↩1↩2↩3
  2. News sourcegihyo.jpMicrosoft Decision-1 assigns probability scores to supplied choices↩1↩2↩3
AI safety and regulationChip wars and the supply chain

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →