Microsoft Releases Decision 1.1 for Typed AI Routing | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Microsoft Releases Decision 1.1 for Typed AI Routing
Now generally available in Foundry, Microsoft’s decision model returns probabilities, fixed choices and ordered scores for applications that need a bounded answer.
Microsoft has launched Microsoft-Decision-1.1, a specialized model now available in Microsoft Foundry designed for classification, routing, and ranking tasks. Unlike generative LLMs that produce prose, this model returns structured, typed outputs including probabilities, fixed choices, and weighted scores. By providing predictable data shapes like noul gates or multi-level severity scores, developers can integrate AI directly into conditional logic and automated workflows. This architectural separation allows applications to maintain strict control over policy, thresholds, and escalation logic, ensuring that automated decisions remain testable, auditable, and aligned with specific business requirements.
Microsoft announced on November 17, 2026, that Microsoft-Decision-1.1 is generally available in Microsoft Foundry. The model handles classification, routing, ranking, grading and binary decisions, and Microsoft says it can also serve as a judge in Foundry Evaluations. Its defining feature is the output: a typed decision instead of a generated explanation. (Microsoft announcement)
A support system could use it to select a team, an agent could use it to choose a tool, and an evaluation pipeline could use it to grade an answer. In each case, the application supplies a bounded question and receives values it can compare, rank or use in a conditional branch.
The useful distinction for developers is architectural. Decision 1.1 gives the application a predictable response shape, but the application still owns the policy: which labels exist, what severity means, when a probability is sufficient and when to escalate.
How Decision 1.1 Differs From a Generative LLM
Microsoft’s documentation says Decision 1.1 “doesn't generate a free-form response or written rationale.” It accepts text, JSON and image inputs, then returns numerical decisions for named questions. Image inputs are referenced through assets associated with the request. (Microsoft Learn)
A generative LLM can write a classification label or produce structured output. Decision 1.1 exposes a narrower interface built around three decision types:
Question type
Returned result
Typical application
noul
A probability from 0 through 1 for a yes-or-no condition
Gates, filters and checks
choice
Exactly one selected option, plus a probability for every option
Classification, routing and model or tool selection
score
A value on an ordered scale, plus probabilities for its levels
Grading, prioritization and ranking
This removes the need to extract a decision from explanatory prose. It also means there is no written rationale to show a reviewer. If a workflow needs both a score and an explanation, Decision 1.1 alone does not provide both.
I’d view that as a useful separation of responsibilities. A routing component should have a small, testable output contract. A component responsible for explaining a decision has a different job. Keeping those roles separate makes it easier to inspect where a pipeline fails.
A typed result still needs accuracy testing. A cleanly formatted choice can be wrong, and a numerical probability needs validation before it controls an important action.
The Three Question Types and Their Practical Uses
Noul: Turn a Judgment Into a Gate
A noul question asks whether a condition is true. Microsoft documents examples such as detecting whether a tool call touches production data, whether a document contains prompt injection, or whether a support reply promises a refund.
The result is a probability between 0 and 1. Your application decides what to do with it.
Optional true and false criteria let developers define the boundary more precisely. For example, a refund check can distinguish a commitment to return money from a general discussion of billing.
Microsoft recommends selecting thresholds against labeled examples before automating a gate. That matters especially when the consequences are asymmetric: allowing a risky action and unnecessarily blocking a harmless action may have very different costs. (Microsoft Learn)
Choice: Select One Route, With an Escalation Path
A choice question selects exactly one option from a fixed set. Documented uses include assigning support tickets, classifying intent, selecting a model or tool, and sorting alerts by likely cause.
Each option has a description. Those descriptions are part of the decision specification, so labels such as billing and engineering benefit from explicit boundaries.
The response includes probabilities for all options. Microsoft recommends treating a low top probability as a reason to escalate to a person or another model.
I’d keep that escalation path explicit. Because choice must select an option, receiving a label is not sufficient evidence that the request fits the available categories. A poorly covered request can still produce a valid-looking route.
Score: Put Items on an Ordered Scale
A score question defines levels from lowest to highest, such as:
Cosmetic
Minor
Major
Critical
Microsoft documents scoring for incident severity, answer quality, helpfulness and frustration. The important implementation detail is that the returned value is a probability-weighted average of the level indexes.
For four levels, the indexes run from 0 through 3, and the result can fall between them. It is not necessarily an integer category or a percentage.
A Request Can Ask Several Questions
Decision 1.1 accepts a state and one or more named questions. That allows an application to ask for a team assignment, an outage check and a severity score using the same input.
The following illustrative JSON follows the documented request structure. It shows a payload, not an executed request:
{
"state": "Checkout returns HTTP 500 for every customer.",
"questions": {
"production_outage": {
"type": "noul",
"instructions": "Does this ticket describe a production outage?",
"criteria": {
"true": "A live service is unavailable",
"false": "No live service outage is described"
}
},
"team": {
"type": "choice",
"instructions": "Which team should handle this ticket?",
"criteria": {
"billing": "Charges, invoices, and refunds",
"engineering": "Bugs, errors, and outages",
"support": "How-to and account questions"
}
},
"severity": {
"type": "score",
"instructions": "How severe is this incident?",
"criteria": [
"Cosmetic",
"Minor",
"Major",
"Critical"
]
}
}
}
The application would read the corresponding entries in the response’s answers object. Microsoft says those entries use the question names supplied in the request; the response also includes the deployed model name and token usage. The documented quickstart provides the deployment and API-call instructions.
Stable question names make it easier to connect decisions to downstream logic. In this example, team could drive routing while severity controls queue order. Those are application policies, not actions performed automatically by the model.
Why an Average Score Can Hide an Important Difference
The scoring formula creates a consequential limit: different probability distributions can produce the same score.
Consider two hypothetical outputs on the four-level severity scale:
Probability distribution
Weighted score
100% probability of Major, index 2
2.0
50% probability of Minor, index 1, and 50% of Critical, index 3
2.0
The arithmetic is straightforward: the second case gives 0.5 × 1 + 0.5 × 3 = 2.0. These are explanatory examples, not measured model results.
A queue sorted only by the average would treat the cases equally. An incident policy concerned with possible critical outages might need to handle them differently.
That is why retaining the per-level probabilities can be useful even when a dashboard displays only one number. The average supports ordering, while the distribution preserves information relevant to an escalation rule.
Microsoft advises validating scores on labeled examples and using them primarily for relative ordering and thresholds, rather than treating them as absolute, calibrated ratings. A score of 2.4 is therefore best understood in the context of the scale and workload that produced it. (Microsoft Learn)
Getting Started in Microsoft Foundry
Microsoft’s quickstart walks through deploying Decision 1.1, calling the decision API and building a support-ticket classifier.
The documented prerequisites are:
An Azure subscription with a valid payment method.
Access to Microsoft Foundry and a Foundry project.
Permission to create and manage model deployments. Microsoft lists Cognitive Services Contributor as a suitable role.
Microsoft Entra ID authentication, which Microsoft recommends, or an API key.
Python 3.10 or later for the quickstart.
The guide also specifies the Azure Identity library:
This command installs the authentication library used by the quickstart. It does not deploy the model or submit a decision request.
Deployment starts in the Foundry model catalog. Select Microsoft-Decision-1.1, create the deployment, and follow the guide’s endpoint and authentication instructions. The documentation lists the model name as Microsoft-Decision-1.1, its model version as 1, and its API type as Decision. These identifiers matter when selecting the deployment and connecting application code. (Microsoft Learn)
For an initial implementation, I’d start with an observable workflow such as ticket routing. Record the proposed route and compare it with labeled examples before letting it assign tickets automatically. That creates a way to inspect category boundaries and escalation thresholds without making a high-consequence gate the first experiment.
Foundry Evaluations Still Needs a Defined Rubric
Microsoft says Decision 1.1 can act as a judge in Foundry Evaluations. That gives its scoring interface a second role: evaluating outputs from other AI models. (Microsoft announcement)
A useful evaluation still needs a defined question. “How good is this answer?” leaves substantial room for interpretation. A scale tied to concrete qualities gives the evaluator a clearer target.
The same principle applies to changes over time. If a team changes a level description, threshold or label definition, it should recheck the result against its labeled examples. The numerical output makes comparisons convenient, but the meaning of the number depends on the rubric.
Deployment Location Matters More Than a Headline Ratio
Microsoft documents two supported deployment types:
Deployment type
Processing scope
Practical consideration
DataZoneStandard
Keeps inference processing within the selected data zone; available in selected regions
Useful for locality-sensitive workloads and potentially shorter routing
GlobalStandard
Can process inference in any supported Azure region
Provides global capacity, with potentially higher or more variable latency
Microsoft recommends creating a Foundry resource close to the application and users, then calling that resource’s endpoint. It also explicitly advises benchmarking the workload because deployment type does not guarantee lower latency. (Microsoft Learn)
Microsoft’s broader launch roundup describes Decision 1.1 as 50 times faster and 79 times cheaper than GPT-6.1 Sol. Those are vendor-reported comparisons; the headline ratios alone do not provide the workload and serving details needed to forecast a particular application’s latency or bill. (Microsoft AI)
For implementation decisions, the more useful starting point is the documented contract: bounded outputs, named questions, returned probabilities and token usage. Measure those against the requests, regions and error costs your application actually has.
Frequently Asked Questions
3 questions
1
Does Decision 1.1 Generate Explanations?
No. Microsoft documents typed numerical decisions without free-form responses or written rationales. Applications needing explanations require another component.
2
Can Decision 1.1 Choose a Model or Tool?
Yes. Model, tool and agent selection are documented uses for choice questions. The application supplies the options and handles the selected route.
3
Is a Score of 2 the Same as “Major”?
Not necessarily. On a four-level scale, 2 is the index of Major, but a weighted combination of other levels can also produce 2. Check the returned level probabilities when that distinction matters.