[KongchangAI]
· 2 min read· 1,332 words

Microsoft Decision-1.1 Launches on AI Gateway: A Structured Decision Model Explained

Microsoft Decision-1.1 Launches on AI Gateway: A Structured Decision Model Explained

Microsoft's Decision-1.1 decision model launches on AI Gateway, returning structured answers instead of free-form text.

Microsoft AI's Decision-1.1 model is now on AI Gateway, focusing on structured answers for routing tickets, choosing workflow steps, and scoring AI responses. It supports Boolean, Choice, and Score question types with probabilities or numeric ratings, lets multiple questions share one input, and is accessible via AI SDK, the OpenAI-compatible Decisions API, or AI CLI. Requests are logged, budgeted, and support fallbacks, though the SDK interface remains experimental.

Microsoft AI's Decision-1.1 model is now available on AI Gateway. Unlike generative models that output free-form text, Decision-1.1 focuses on returning structured answers for decision-oriented tasks such as routing support requests, selecting the next step in a workflow, and scoring AI responses.

Decision-1.1 now available on AI Gateway

Why We Need a "Decision Model"

Most conversational models are used to write copy or answer questions, producing natural-language paragraphs. But in engineering systems, many steps require a clear choice or numeric value, not a passage of text that can be interpreted in multiple ways. For example: should a customer support ticket go to the billing team or the technical team? Which node should the approval workflow move to next? Is the model's answer high or low quality?

Decision-1.1's approach is to turn these judgments into programmable return values. Developers define the question and the options, and the model returns a structured result that downstream code can use directly, eliminating the need to parse intent out of natural language. This reduces integration complexity and makes the decision process easier to log and audit.

"Structured Output" is the key background for understanding this type of model. Traditional large language models generate free-form text by default; even if the prompt says "answer only yes or no," the output may still include explanatory sentences that require regex or a second parsing pass to extract the result. Structured output uses constrained decoding or function calling to force the model to return according to a predefined schema—such as a JSON object, an enum value, or a numeric value. Decision-1.1 productizes this capability further: developers don't need to design a JSON Schema themselves; they only declare the question type (Boolean/Choice/Score) and candidate criteria, and the model returns an answer object that can be deserialized directly. This approach moves the reliability of output format from prompt engineering to the API contract layer, making decision logic easier to test, version, and audit, and avoiding the ambiguity and fragility of natural-language parsing.

Three Return Types and Batch Questions

According to the official documentation, Decision-1.1 supports three question types. Boolean returns the probability that a condition is true; Choice returns the selected category along with the probability distribution across each option; and Score gives a numeric rating on a developer-defined scale. The presence of probabilities and scores lets callers set thresholds or apply weighting rather than being forced to accept a black-and-white conclusion.

Another noteworthy detail is that multiple questions can share the same input in a single request. This means the system can simultaneously determine department ownership, urgency, and a handling recommendation for the same ticket text, reducing redundant model calls.

The probability distribution returned by the Choice type deserves elaboration. The model gives a probability value for each candidate category, typically normalized via softmax so that the values sum to 1. Callers can use this to set confidence thresholds: for example, route to human review when the highest probability is below 0.6, or distribute tickets weighted by probability. The Boolean type returns the probability that the condition is true rather than a hard boolean, which likewise supports threshold-based judgment. The Score type outputs a numeric value on a developer-defined continuous scale (such as 1-5 or 0-100), suitable for ranking or scoring AI response quality, risk level, and so on. This "probabilistic output" design distinguishes the decision model from a classifier—it preserves uncertainty information so that downstream systems can handle cases by confidence level rather than being forced to accept a potentially wrong definitive conclusion.

Access Methods and Migration Cost

Decision-1.1 follows Decision-1's request format; to switch, you only need to change the model ID to microsoft/decision-1.1. Entry points include the AI SDK, the OpenAI-compatible Decisions API, and the AI CLI. The official examples note that TypeScript requires ai 7.0.128 or higher, and Python requires OpenAI Python SDK 3.26.0 or higher.

In the code example, developers issue a decision request through the experimental_decide function, passing in a state that describes the current situation, declaring each question's type, instructions, and candidate criteria in questions, and finally reading the choice field from the returned answers. Note that this interface is still marked as experimental in the SDK and may change later.

The experimental_decide interface mentioned in the article is marked as experimental, meaning its function signature, parameter structure, and return format may undergo incompatible changes in later SDK versions. For production systems, this requires developers to isolate the integration: wrap decision calls in a dedicated adapter layer so that changes to the experimental interface don't spread into business code; also lock the SDK version and read the changelog before upgrading. In addition, the three entry points—AI SDK, the OpenAI-compatible Decisions API, and the AI CLI—correspond to different use cases: the SDK suits in-application integration, the OpenAI-compatible API makes it easy to reuse existing client code, and the CLI suits scripted testing and debugging. Developers should choose the appropriate entry point for their deployment environment and note the differences in minimum SDK version requirements between TypeScript and Python.

Operations and Observability

Decision requests appear in request logs, count toward budget, and support decision fallbacks. For production systems, these capabilities determine whether the model can be safely placed in critical paths: logs help troubleshoot misjudgments, budget controls costs, and the fallback mechanism provides a safety net when a decision fails.

Developers can try Decision-1.1 in the model playground; details on request format and question types are available in the decision model documentation.

"Decision fallbacks" refers to the mechanism that automatically switches to a backup model or preset rules when the primary decision model call fails, times out, or returns a result that doesn't satisfy constraints. In critical paths, fallback is a safeguard for availability: for example, falling back to a rules engine when the primary model is unavailable, or switching to another model version and retrying. The accompanying request logs record each decision's input, question type, returned answer, and probability, making it easy to investigate misjudgments after the fact and analyze distribution drift; budget accounting incorporates decision calls into unified cost tracking, preventing high-frequency batch decisions from quietly driving up expenses. Together, these three capabilities answer the core questions of a production environment—can errors be detected, can losses be contained, and are costs controllable—making them prerequisites for a decision model to move from experimentation to production.

Boundaries of Applicable Scenarios

A structured decision model is not meant to replace generative models, but to fill in the category of scenarios where a definite option must be given. When business rules are clear and the candidate set is limited, having the model return categories and probabilities directly is more reliable than having it write an explanation that a program then parses. Conversely, if the task itself requires open-ended creation or long-text reasoning, this type of model is not suitable.

Decision-1.1's value lies in embedding model capabilities into the process control layer, turning AI output from "a suggestion you can reference" into "a result you can execute directly."

One-line Summary

Microsoft's Decision-1.1 decision model is now on AI Gateway, returning structured answers instead of free-form text.

Paragraph Summary

Microsoft AI's Decision-1.1 model is now available on AI Gateway. Unlike generative models that output free-form text, it focuses on returning structured answers for decision tasks such as routing tickets, selecting the next workflow step, and scoring AI responses. The model supports three question types—Boolean, Choice, and Score—returning probabilities or numeric ratings, and callers can set thresholds or apply weighting; multiple questions can share the same input in a single request, reducing redundant calls. It follows the Decision-1 format, requiring only a model ID change, and can be called via the AI SDK, the OpenAI-compatible Decisions API, or the AI CLI. Decision requests are logged, counted toward budget, and support fallback, but the interface is still marked as experimental in the SDK. The article emphasizes that this type of model fills the scenario where a definite option must be given, rather than replacing generative models.

#Microsoft Decision-1.1#AI Gateway#structured output#decision model#AI SDK
Share:

Related articles