Microsoft Decision-1 is a decision model that answers typed questions about text. It selects categories, assigns scores against a rubric, and estimates yes/no probabilities. Applications can use those answers to classify information or suggest a workflow's next step, with several questions sharing the same input.
Microsoft Decision-1 became available on AI Gateway on October 9, 2026. You can call it with Gateway credentials without setting up an Azure account or deployment.
Copy link to headingHow does Microsoft Decision-1 work?
Microsoft built Decision-1 by post-training Qwen3.5-9B for single-pass decision scoring. Given defined answer options, the model scores the alternatives. Its structured output lets application code read a result without extracting a label from a generated paragraph.
You supply the evidence and define what each question means. For example, a documentation team might ask whether feedback describes outdated instructions or a confusing explanation. Those categories need descriptions that distinguish them, plus an option for feedback that doesn't identify either problem.
This is the broader decision-model approach. The model interprets the evidence within the supplied criteria; your application decides what consequence to attach to its answer. Selecting an issue category could suggest a review queue, but it doesn't establish that a reported product problem is real.
Copy link to headingWhat kinds of answers can it return?
Through the AI SDK decision API, Microsoft Decision-1 supports three question types:
Score levels use zero-based positions. With four levels, the result falls between 0 and 3 and can lie between two levels. That number represents a position on your rubric, so it needs the rubric alongside it to be interpretable.
Multiple questions can examine the same text in one request. For documentation feedback, the issue category and whether setup was blocked answer different questions. Keeping them separate avoids forcing every report into a combined category such as "outdated and blocking."
Copy link to headingWhat can you use Microsoft Decision-1 for?
Decision-1 fits tasks where the useful output comes from a defined set. Beyond documentation feedback, consider these applications:
Group product research comments by the friction they describe, preserving an "unclear" category for comments without enough detail.
Assess an AI-generated answer against a supplied policy or reference passage before showing it to a reviewer.
Rank candidate incident records against the current incident description so responders can inspect the most relevant records first.
Suggest which stage of a workflow should handle an input, while application code checks whether the proposed transition is allowed.
Microsoft describes internal uses in Xbox research feedback analysis, Copilot quality assessment, and incident knowledge retrieval. Those examples illustrate the kinds of questions a decision model can answer; they don't establish performance on a different application's data.
For complete examples across several models, the AI Gateway Decision API use cases include a release-note audience classifier and a notification filter. Each example turns a bounded prediction into an application suggestion.
Copy link to headingHow can you try Microsoft Decision-1 through AI Gateway?
Use the model ID microsoft/microsoft-decision-1. Gateway exposes it through the AI SDK decision interface, its OpenAI-compatible Decisions API, and its TypeSafe-compatible API. An OpenAI-compatible request selects Microsoft's model through that ID; the request format does not determine the model provider.
The following example reads a documentation feedback report. It asks which problem the reader describes and whether that problem prevented setup. The Boolean question explicitly treats missing evidence of a block as false, so its output means "a block was reported," not "the reader probably failed."
Use Node.js 22.18 or later and install the packages in your project directory:
npm install ai @ai-sdk/gatewayThe experimental decision API requires ai 7.0.128 or later. Configure AI_GATEWAY_API_KEY in your server environment, or use Vercel OIDC, following the decision quickstart.
import { experimental_decide as decide } from 'ai';import { gateway } from '@ai-sdk/gateway';
const state = { feedback: 'The setup guide still says to open Settings > Tokens, but that page ' + 'was removed. I could not create a token and stopped the installation.',};
try { const result = await decide({ model: gateway.decisionModel('microsoft/microsoft-decision-1'), state, questions: { issue: { type: 'choice', instructions: 'Identify the main documentation problem reported in the feedback. ' + 'Use unclear when the feedback does not identify one.', criteria: { discovery: 'The reader cannot find the relevant documentation.', explanation: 'The reader finds the page but cannot understand it.', outdated: 'The documented steps no longer match the product.', unclear: 'No specific documentation problem is identifiable.', }, }, blocked: { type: 'boolean', instructions: 'Does the feedback explicitly say a documentation problem ' + 'prevented the reader from completing setup?', criteria: { true: 'Setup was not completed because of the documentation problem.', false: 'Setup was completed, or the feedback does not establish a block.', }, }, }, maxRetries: 0, abortSignal: AbortSignal.timeout(30_000), });
console.log(JSON.stringify(result.answers, null, 2));} catch { console.error('Decision failed. Keep the feedback in the review queue.'); process.exitCode = 1;}Save the example as review-docs-feedback.mjs, then run it with node followed by the filename. To load credentials from .env.local, place --env-file=.env.local between node and the filename.
To use Vercel OIDC authentication, run vercel link to connect your working directory to a Vercel project. Then run vercel env pull .env.local to save the project's development environment variables, including its OIDC token, to .env.local. Load the file with Node's --env-file=.env.local option, leaving AI_GATEWAY_API_KEY unset so the SDK uses the token. Repeat the command when the token expires after 12 hours.
The supplied report should produce the outdated-instructions category and a high probability that setup was blocked. In a live smoke check, it returned outdated and a blocked probability of approximately 0.999. Feedback that described confusion but confirmed successful installation returned explanation with a blocked probability below 0.001. Vague feedback returned unclear. These synthetic checks verify the example's behavior on those inputs, not its accuracy across real feedback.
The code prints answers without changing a queue or notifying anyone. Your documentation dashboard could display the suggested category alongside the original report, letting an editor confirm the issue before assigning work. If the request fails, the example prints an error and exits with a nonzero status.
Copy link to headingWhat should your application validate before acting?
Probability values describe the model's predictions under your questions and criteria. High probability does not independently verify a reader's claim. In the example, the model can identify that someone reported a removed settings page, but the documentation team still needs to check the current product.
Before using the output to assign work automatically, test labeled feedback that reflects your users. Include reports where a reader later completed setup, complaints that mention several issues, and comments too vague to classify. Review wrong predictions with high probabilities as well as uncertain ones, since a threshold alone won't catch every error.
Keep any action rule separate from the model call. For example, a team might initially show every prediction to an editor, then automate a narrow category after measuring its errors. Store the original feedback with the answer so reviewers can correct the suggestion without reconstructing its context.
Gateway decision requests appear in logs and count toward configured budgets. Those controls track requests and spend; measuring decision quality still requires expected answers for your own cases.
Copy link to headingFrequently asked questions
Copy link to headingIs Microsoft Decision-1 a language model or a decision model?
Microsoft Decision-1 is a decision model built by post-training Qwen3.5-9B for decision scoring. Its decision interface returns bounded answers such as category selections and probabilities, rather than an open-ended written response.
Copy link to headingDo you need an Azure account to use Microsoft Decision-1?
No. Through AI Gateway, you can use Gateway credentials without creating an Azure account or deploying the model in Azure. The Gateway model ID is microsoft/microsoft-decision-1.
Copy link to headingCan Microsoft Decision-1 answer several questions at once?
Yes. You can submit multiple typed questions about shared input in one request. This lets an application classify a report and assess a separate yes/no condition without making separate calls for those questions.
Copy link to headingDoes a high probability mean the answer is verified?
No. The probability is a model prediction based on the supplied evidence and criteria. Test predictions against labeled examples before choosing how much automation to allow, and check external facts through the systems that own them.
Copy link to headingCan you call Microsoft Decision-1 with the OpenAI-compatible Decisions API?
Yes. AI Gateway's OpenAI-compatible Decisions endpoint accepts Microsoft's Gateway model ID. The compatible request format lets you use that API shape while Microsoft Decision-1 supplies the answers.