Blog Post

Azure Integration Services Blog
6 MIN READ

Announcing Decision APIs as actions in Azure Logic Apps

DivSwa's avatar
DivSwa
Icon for Microsoft rankMicrosoft
Nov 17, 2026

Today at Microsoft Ignite 2026, we're announcing the public preview of Decision actions powered by decision models in Logic Apps Automation. Three new built-in actions, Yes/No, Choice, and Score, let a workflow ask a focused question about any content and get back a typed answer it can act on right away. The actions and the Microsoft-Decision-1 model behind them are fully managed, so there's no model to deploy, no output format to prompt for, and no response to parse.

Many steps in a business process come down to a judgment call. Is this message a refund request? Which team should handle this ticket? How severe is this alert? Until now, you had two options: write rigid rules, or call a general-purpose large language model, prompt it for a format, and parse the text it returns. Decision actions give you a third option, designed for exactly these questions.

Background: Decision models

AI applications and agents make many small judgments in the course of their work. These judgments generally take one of three forms:

  • Yes-or-no checks, such as whether an email is spam, whether a customer wants to cancel a subscription, or whether a job application meets the minimum qualifications.
  • Selections from a known set, such as which language a message is written in, which expense category a receipt belongs to, or which product a customer is asking about.
  • Ratings on an ordered scale, such as how positive a product review is, how closely a student’s answer matches the answer key, or how formal the tone of a draft is.

These judgments have typically been handled by general-purpose generative models, which return free-form text that the application must then interpret, even though the set of valid answers is known in advance.

Decision models are designed for these tasks. Each request supplies the content to evaluate and one or more questions, and each question declares its answer type: the probability that a condition is true, a selection from a fixed set of options, or a score on an ordered scale. This approach offers two advantages over generative models.

Typed responses. Every response follows the schema the caller defines. A probability is always a number between 0 and 1, a selection is always one of the options supplied, and a score always falls within the defined scale. Applications can use the answer directly, without output-format instructions in the prompt and without parsing or validation code.

Token efficiency. Because the model returns a value rather than text, each answer is a single output token instead of a paragraph. Requests are also smaller, because they don't need formatting instructions or examples of the expected output, and a single request can answer several questions about the same input.

This category of model is new. Microsoft introduced Microsoft-Decision-1 in the Microsoft Foundry model catalog on October 9, 2026, in public preview, and several other providers have released comparable models in recent weeks. Microsoft-Decision-1 is intended for classification, routing, evaluation, and workflow control. It accepts text input and is priced at $0.042 per million input tokens, with no charge for output.

 

Why decision models matter for AI Automation

A common pattern in Logic Apps today uses a chat completion action to classify content, a Compose or Parse JSON action to normalize the response, and a Switch or condition action to route on the result. Decision actions replace the first two steps of that pattern. You define the answer options in the action itself, the response is constrained to those options, and every answer carries a probability that downstream actions can evaluate.

Three actions, one for each kind of question

Yes/No

Yes/No answers whether a condition is true, which makes it a good fit for gates and checks such as refund requests, spam, or a missing purchase order number. It returns the probability that the answer is yes, which you compare with a threshold in a Condition action. For example, ask "Is the customer asking for a refund?" and send high-probability messages to the refunds queue.

 

 

Choice

Choice picks the one option from your list that fits best. Use it for routing and classification when the categories have no natural order, such as support teams, document types, or review topics. It returns the selected option, a probability for each option, and an overall confidence value, so you can point a Switch action at the result. Include an Other option so that input that fits none of your categories isn't forced into the wrong one.

 

 

Score

Score places the input on an ordered scale you define, such as low to high urgency, severity, or negative to positive sentiment. It returns a probability-weighted score across your levels, along with a confidence value and per-level probabilities, so a result can fall between two levels. A mixed product review, for example, can land between Neutral and Positive, which gives you a finer reading than a single label.

 

 

How it works in the designer

Decision actions appear under AI and agents in the designer. Every Decision action takes the same three inputs:

  1. State: the content to evaluate, as text or JSON. This value usually comes from the trigger or an earlier step, such as a ticket body or the output of a previous action.
  2. Question or instructions: what you want the model to decide, such as "Is the customer asking for a refund?" You can also add rules here, such as how to break a tie.
  3. Options or levels: for Choice, a stable key and a short description for each option. For Score, the levels in order from lowest to highest. For Yes/No, you can optionally describe what yes and no mean.

The outputs appear as dynamic content, so you can branch on them with the actions you already use. Each answer, its probability, and its token usage are recorded in run history, so you have a clear record of how every item was handled and cost is visible from the first run.

 

What you can build

Any step that turns on a judgment call is a candidate. You can use one action on its own or combine all three on the same input, so a workflow can gate, route, and prioritize in a single pass.

Scenario

Yes/No

Choice

Score

Customer support triage

Is the customer asking for a refund?

Which team should handle this: Billing, Shipping, Technical support, or Account access?

How urgent is this ticket?

Review analysis

Does the customer ask for a reply or replacement?

Is the review mainly about product quality, delivery, price, or service?

How positive is this review overall?

Bug report triage

Does the report include steps to reproduce?

Which product area does it affect?

How severe is the impact on users?

Lead qualification

Is this a genuine inquiry from a potential customer?

Which product are they most interested in?

How ready are they to buy?

Content moderation

Does the post share someone's personal contact information?

What kind of issue does it have, if any?

How serious is the problem?

Designed for human oversight

Decision actions help you automate judgment calls responsibly. Every answer includes a probability or confidence value, so you can send low-confidence results to a person or a review queue instead of acting on a guess. Run history records each question, answer, and probability, which gives your team an auditable record of how the workflow reached each outcome.

 

 

 

Get started

Decision actions are available today in public preview in Logic Apps Automation.

The quickest way to start is from a template. We've published templates that show each action in a complete workflow. Find them in the Logic Apps template gallery - Templates | Logic Apps Automation

 

To build your own workflow instead:

  1. In the workflow designer, add an action, and then look under AI and agents for Yes/No, Choice, or Score.
  2. Specify the state, the question, and the options or levels.
  3. Add a Condition or Switch action that uses the result.

We'd love to hear what you build and what you'd like to see next. Share your feedback or request a feature here 

Updated Nov 17, 2026
Version 3.0