By Mahan Kenway
Disclosure: This article was researched and written by an autonomous AI assistant for Mahan Kenway. The cover is an original AI-generated conceptual illustration, not a Microsoft product image.
An AI assistant that answers customer questions may need to make a dozen small decisions before it writes a single sentence. Is this an outage or a billing question? Should a human review it? Which model or tool should handle the next step?
On November 17, Microsoft announced that Microsoft-Decision-1.1 is generally available in Foundry, as well as through OpenRouter and Vercel. Instead of producing a conversational response, it evaluates a bounded question and returns a decision with numerical confidence information. Version 1.1 also adds image inputs and, Microsoft says, runs more than twice as fast as its predecessor at the same cost.
That is a useful architectural shift, not a guarantee that an application will suddenly make reliable decisions. The practical question is where a specialized model actually helps, how its API differs from chat, and what happens when its confidence is wrong.
A judge, not a chatbot
A generative model can return JSON containing a label, but it still has to generate that response. A decision model is designed around a different contract: the application supplies a situation and a defined question, and the service returns a structured decision rather than a written explanation.
Microsoft’s Foundry documentation describes three question types:
noul: a probability for a yes-or-no condition.
choice: one selected option and probabilities for the available options.
score: a value on an ordered scale, derived from the probabilities assigned to its levels.
You can ask multiple questions about the same state in one request. For example, a support system could ask both which team should receive a ticket and whether the issue is urgent. The answer can drive routing without asking the model to compose the customer reply; a generative model can still handle that later.
That distinction matters for debugging. A typed output is easier to integrate, but this model does not give a written rationale. If an auditor needs an explanation, a numerical answer alone is insufficient. A separate explanation generated later should not be treated as proof of the decision model’s actual reasoning.
What a request looks like
The Foundry endpoint is a decision API, not the usual chat-completions endpoint. Microsoft’s guide places it at the resource endpoint followed by /providers/microsoft/v1/systemone. The model field is the name of your deployment, which may differ from the underlying model name. Authentication can use Microsoft Entra ID or an API key.
Provider formats are not interchangeable. For example, Vercel documents its own Decisions API and an AI SDK interface; do not assume that a Foundry request body can be sent unchanged to a different gateway.
A compact request body for ticket routing could look like this:
{
"model": "my-decision-deployment",
"state": "Customers receive a 500 error after checkout.",
"questions": {
"destination": {
"type": "choice",
"instructions": "Which team should investigate first?",
"criteria": {
"engineering": "Application errors and outages",
"billing": "Invoices and charges",
"review": "Not enough information to choose safely"
}
},
"urgent": {
"type": "noul",
"instructions": "Is a widespread service outage described?"
}
}
}
The review option is deliberate. A fixed-choice model can still pick an option when the available information is poor. Your application must define what happens when the model selects review, when the top probability falls below your threshold, or when the API is unavailable. Those are product decisions, not settings the model can safely make on your behalf.
Version 1.1 can also assess images alongside text. Microsoft’s examples use an assets object containing base64-encoded images and reference those assets in the request. The guide specifies image dimensions of 280–1,568 pixels and up to eight rendered image occurrences per question. That makes screenshot classification or a damage-photo triage prototype plausible, but it also sends image data to the provider. Review data-handling requirements and avoid sending sensitive content that the workflow does not need.
A probability is most useful when it is calibrated: among many cases assigned roughly the same confidence, the proportion that are correct should be roughly consistent with that confidence. Even then, a well-calibrated model can be confidently wrong on individual cases, especially after a shift in input data.
Microsoft says it improved calibration and tested 1.1 on 102 internal benchmarks containing more than 385,000 questions held out from training. The company also reports a score of 68 on Artificial Analysis’s Decision Index v2.1. The independent index combines several evaluations, including document decisions, tool calling, knowledge, reasoning and consistency under rewording. Its score is not simply “percent correct” and should not be read as a production success rate.
For your own workload, build a labeled test set from representative requests. Measure per-category errors, not just overall accuracy. Try different wording and option order; Microsoft explicitly warns that both can affect scores. Decide separately how costly a false positive and a false negative are, and test thresholds against those costs. For borderline or unfamiliar cases, route to review or a suitable fallback.
This is particularly important for safety filters. Microsoft notes that the model can miss harmful content and flag benign content. It also cautions against using its output as the sole basis for consequential decisions about people. A decision score should not replace permissions, policy enforcement, human review where appropriate, or a hard-coded rule protecting a destructive action.
Pricing is low, but the whole workflow still has a cost
Microsoft lists Foundry pricing at $0.042 per million input tokens for Global Standard, $0.046 for the US Data Zone, and $0.050 for EU and APAC Data Zones. Output tokens are included at no additional charge under that pricing. Those are rates in Microsoft's launch announcement, not a promise about what another gateway will charge. The Azure pricing page does not yet separately list 1.1, so verify the live deployment quote and region before budgeting.
For a simple illustration, one million requests averaging 1,000 billed input tokens each would amount to one billion input tokens, or $42 in model input charges at the published Global Standard rate. That excludes the rest of the system: upstream and downstream models, storage, networking, retries, and human review. Image-related billing and actual token usage should be measured with your chosen deployment rather than guessed from image file size.
Microsoft offers Global Standard and selected-region Data Zone deployments. Its deployment guide says Global Standard processing may occur in supported Azure regions, while Data Zone processing stays within the selected zone. Neither option guarantees lower latency for your application. Measure end-to-end latency, including network overhead and retries, using your own traffic.
A sensible first pilot
Start with a high-volume decision that is easy to reverse, such as routing internal support tickets. Run the decision model in shadow mode beside the existing workflow, compare its choices to human labels, and inspect disagreements. Only then consider automatic routing for categories that meet a measured threshold, with a clear review path for the rest.
Keep the model outside authorization boundaries. Even if it recommends a tool, an independent permission check should decide whether that tool may actually run. Log the model version, decision, confidence and outcome in a privacy-conscious way, and re-evaluate after changing the model, the wording or the mix of incoming requests.
Decision 1.1 is interesting because it makes a common agent-building step explicit: choosing among actions is a different job from writing an answer. Its value will depend less on a headline benchmark than on whether developers can make that choice measurable, auditable and safe in their own systems.