Microsoft Decision-1 Is Now on Vercel AI Gateway, and It Changes How You Handle Logic in Your App
A new AI model just landed on Vercel's AI Gateway that doesn't write text or generate images. It makes decisions. Here's why that's a bigger deal than it sounds for designers building their first apps.
Microsoft's Decision-1 model is now available on Vercel's AI Gateway, and it does something refreshingly specific: instead of generating prose, it answers structured questions about text with probabilities, choices, and scores. For designers building vibe-coded apps, this is the piece that's been quietly missing.
Why "just make a decision" is harder than it sounds
Most AI models are generalists. You give them a prompt, they give you a blob of text. That's great for drafting copy or generating component ideas, but it's a poor fit when your app needs to act on something. Should this support message go to the billing team or the technical team? Does this user-submitted review meet your quality bar? Is this input a yes or a no?
Until now, the common workaround was to prompt a general-purpose model and then parse its reply, hoping it said "billing" and not "the billing department" or "definitely billing-related." Fragile at best. Decision-1 is built specifically for this job. It returns structured answers with probability scores for each option, so your app knows not just what the answer is but how confident the model is. That confidence score is what lets you build a sensible fallback: if the model isn't sure, flag it for human review instead of acting blindly.
What the model actually does
Decision-1 handles three question types:
Boolean questions are simple yes/no calls. Is this message a complaint? Does this text contain a phone number?
Choice questions ask the model to pick from named categories you define. You supply the options and a short description of each, and the model picks the best fit, with a probability score attached to every option.
Score questions ask the model to rate an input against an ordered set of criteria you write. Think of it as a rubric evaluator.
The clever part is that you can bundle multiple questions about the same input into a single request. One call, several answers. That keeps things tidy and cheap.
Access is through Vercel's AI Gateway using the model name microsoft/microsoft-decision-1. You don't need a separate Azure account or any Microsoft-side setup. One AI Gateway API key is all it takes. Calls show up in the Gateway's logs and count toward any spending budgets you've set, sitting alongside your other model usage in one dashboard.
How a designer could actually use this
Say you're building a simple feedback tool for a client. Users submit responses, and you want to triage them automatically before they land in a spreadsheet.
With Decision-1, you could set up a Choice question that routes each response to one of three buckets you define: "feature request" "bug report" or "general praise." You write a one-line description for each bucket, pass in the user's text, and get back a category plus a confidence score. If confidence is above a threshold you pick, the app routes it automatically. Below that threshold, it drops into an "uncertain" pile for a human to look at.
That workflow would have previously required either a lot of brittle prompt engineering or a custom classifier someone had to train. Now it's a single API call with no machine-learning background required.
Other realistic starting points for designers:
- Content moderation. Score user-generated content against a rubric of your community guidelines before it goes live.
- Onboarding routing. Ask a Boolean question to detect whether a new user's goal matches a particular use case, then show a tailored next step.
- AI output review. Use a Score question to evaluate whether a generated response meets a quality bar before surfacing it to the user, a pattern sometimes called "LLM-as-judge" (meaning the model grades another model's output).
The AI SDK decision API (experimental_decide) is the most straightforward entry point, and the documentation covers Choice, Score, and Boolean question formats with examples.
What to keep in mind
A few honest caveats. The model is described as being designed for classification and routing tasks, not open-ended reasoning. The quality of your results will depend heavily on how clearly you write the criteria you give it. Vague category descriptions will produce vague, low-confidence answers. Treat writing those descriptions as a design problem: the clearer the brief, the better the output.
The experimental_ prefix on the SDK method is also worth noticing. That label signals the API surface could still change, so build with that in mind if you're shipping something to real users soon.
That said, having a model purpose-built for structured decisions, with confidence scores, no extra account setup, and usage visible in the same dashboard as everything else, is a genuinely useful addition to the vibe-coder's toolkit. The logic layer of your app just got a lot less painful to build.