Microsoft Decision-1 Lands on AI Gateway: Revolutionizing AI Workflows with Probabilistic Decisions

Microsoft's Decision-1 model has officially landed on AI Gateway. This AI model, purpose-built for "decision-type" tasks, is offering developers a whole new way to work—letting AI not only generate content but also help you make choices.

What Is Decision-1?

Decision-1 is an AI model from Microsoft focused on structured decision-making. Unlike traditional text-generation models, it answers "typed questions"—the user poses a clear question, and the model returns probabilities, options, and scores rather than a block of free-form text.

This design makes Decision-1 particularly well-suited to three scenarios: classifying support requests, choosing the next step in a workflow, and evaluating AI-generated answers against scoring criteria. Developers can bind multiple questions to the same input, completing several judgments in a single request and dramatically boosting processing efficiency.

Three Question Types Covering Mainstream Decision Needs

Decision-1 supports three core question formats, each with probabilistic output, letting applications decide based on confidence whether to act automatically or escalate to a human.

Boolean questions: Answer "yes or no" questions, ideal for binary judgments such as whether a piece of content violates policy.

Choice questions: Pick the most appropriate answer from a set of named categories. For example, when an invoice payment failure request comes in, you can ask "which team should handle this," and the model will return a probability distribution across options like "billing" and "technical."

Score questions: Rate the input against an ordered set of criteria, well-suited to quality assessment or risk grading.

The official code example is very intuitive—just call the decide function, passing in the model name microsoft/microsoft-decision-1, the input state, and a set of questions, and you get back a structured answer object.

Probabilistic output and confidence thresholds are key concepts to understand when using Decision-1. The probability the model assigns to each option reflects how confident it is in that judgment given the current input—not a guarantee of absolute correctness. Applications typically set a threshold: only when the highest probability exceeds the threshold does the system execute an action automatically; otherwise it routes to human review or a fallback. The value of this mechanism is that it turns "is the model reliable" from a binary right-or-wrong question into a tunable risk-control parameter. Teams can set the threshold high or low based on their business's tolerance for errors, striking a balance between automation rate and accuracy.

Integration: One API Key Unlocks Everything

The barrier to entry for Decision-1 is quite low. Developers can call it in three ways: the AI SDK's decision API, the OpenAI-compatible Decisions API, and the TypeSafe-compatible API.

The most important point: you only need one AI Gateway API key—no separate Azure account, and no need to deploy the model ahead of time. For teams already using AI Gateway to manage other models, Decision-1 is almost a "zero-cost" integration.

AI Gateway is an intermediary layer that sits between applications and underlying model services, exposing models from different vendors and protocols through a unified set of standard interfaces. Developers only need an API key issued by the gateway to call any model mounted behind it, without having to separately apply for accounts on each cloud platform or configure authentication and quotas on their own. Beyond unified access, a gateway typically handles cross-cutting concerns like request routing, failure retries, rate limiting, usage metering, and cost aggregation, so teams don't need to change business code when swapping or adding models. Decision-1 being exposed through the gateway means it reuses this existing infrastructure—which is why the article emphasizes "zero-cost integration."

Coordination with AI Gateway: Trackable Costs

The unified model access, usage tracking, and cost management capabilities that AI Gateway itself provides apply equally to Decision-1. All decision calls show up in the logs and count toward budget statistics, on equal footing with other model requests. This means teams can see decision-model call frequency and cost share on a single dashboard, making cost optimization easier.

The Practical Value of Probabilistic Decisions

The truly interesting part of Decision-1 is its probabilistic output. Traditional classification models often give just one label, and applications can only accept the result wholesale. Decision-1 returns a probability for each option, giving the application layer more room for control—when a judgment's confidence falls below the threshold, it can automatically hand off to a human instead of forcing execution.

This "decisions with uncertainty" model is highly practical for scenarios like customer-service ticket routing, approval-workflow automation, and quality control on AI output. Developers can start by browsing the list of available models or reading the decision-model guide to quickly integrate Decision-1 into existing workflows.

Who Should Use Decision-1?

If your application needs to choose among multiple options, classify inputs, or score against criteria—while still keeping the flexibility of a human fallback—Decision-1 is worth a try. Especially for teams already using AI Gateway, the integration cost is almost negligible.

One-line Summary

Microsoft's Decision-1 decision model lands on AI Gateway, supporting classification, choice, and scoring with probabilistic structured output.

Paragraph Summary

Microsoft's Decision-1 model is now live on AI Gateway. It isn't a text-generation model but one focused on structured decisions: users ask typed questions, and the model returns probabilities, options, and scores. It supports boolean, choice, and score formats, and can bind multiple questions to the same input to make several judgments at once—useful for ticket classification, workflow branching, and AI output quality assessment. Integration requires just one AI Gateway API key, with no Azure account or pre-deployment needed, and can be called via the AI SDK, OpenAI-compatible, and TypeSafe-compatible interfaces. Since all calls are logged and counted against budgets, teams can track decision-model usage and cost on a unified dashboard. Probabilistic output also lets applications auto-execute or escalate to a human based on confidence thresholds, balancing automation efficiency with risk control.