Decision Models vs LLMs in Microsoft Foundry

Key insights

• The video defines decision models as AI that picks from options instead of writing text.
• It compares decision models with LLMs and covers when to use them together.
• Microsoft benchmarks discussed in the video include accuracy, speed, and calibration.
• Live demos include TypeSafe Jev, Microsoft-Decision-1, a Foundry agent, and Azure Logic Apps Automation.

Rafsan Huseynov recently published a video called Decision Models vs LLMs: When to Use Each. The video explains a newer AI model type that does not generate text, then shows how it differs from large language models and how both can work together in workflows. The main takeaway is practical: use a decision model when software needs a structured choice, and use an LLM when it needs generated language.

The video is organized around a simple problem many teams face in AI projects: not every task needs free-form text. Sometimes the system only needs a classification, a route, a score, or a yes/no decision. Huseynov uses Microsoft examples, benchmarks, and live demos to show how that split changes the design of an application or agent.

What's new

The core topic is decision models, which the video describes as AI models that do not write text. Instead, they choose an answer from a list that you provide and return a probability for each option. That makes the output structured and easier for software to use directly.

The video also covers Microsoft-Decision-1 in Microsoft Foundry, plus a Foundry agent that uses Decision-1 as a tool. Another demo shows an email triage workflow in Azure Logic Apps Automation, where Decision-1 makes the decision and the LLM writes the reply. Huseynov also walks through benchmarks for accuracy, speed, and calibration.

Why it matters

This matters most for builders who need reliable decisions inside apps, agents, or workflows. If the task is to pick a category, route a request, rank items, or decide whether another model should be called, a decision model can fit better than a general-purpose text generator. That can reduce unnecessary generation and make the application’s next step easier to automate.

For admins and solution builders, the point is not that decision models replace LLMs. The video frames them as different tools for different jobs. LLMs still fit summarization, explanation, and other open-ended language tasks. Decision models fit bounded decisions where the expected answer comes from a known set of options.

The video description also signals a workflow design pattern that many Microsoft 365, Power Platform, and Azure teams will recognize: let one model decide, then let another model write. In the example shown, the decision model handles triage and the LLM handles the reply. That separation can help teams keep the decision step structured while preserving the flexibility of generative text where it is actually needed.

How it works

According to the material, the workflow starts by defining the options or labels the decision model can choose from. The model then evaluates the input and returns a structured result, such as a selected option and a probability for each choice. The video presents this as a different output shape from an LLM, which normally generates text rather than selecting from a fixed set.

The demos show several ways to use that pattern:

1. Try TypeSafe Jev in the playground.
2. Use Microsoft-Decision-1 in Microsoft Foundry.
3. Insert Decision-1 as a tool inside a Foundry agent.
4. Build an Azure Logic Apps Automation flow where Decision-1 makes the decision and the LLM drafts the response.

The video also mentions benchmarks for accuracy, speed, and calibration. Calibration here means how well the model’s confidence matches real outcomes, which matters when software depends on probabilities or scores. The result is a model choice that is meant to be more predictable for structured tasks than a general text model.

Limitations and open questions

The material does not provide a full technical explanation of training, evaluation methods, or exact benchmark results. It mentions pricing as a chapter topic, but the supplied description does not include the actual price details. It also does not spell out licensing, tenant availability, or regional rollout status.

The video and supporting notes make clear that decision models are for bounded decisions, not open-ended conversation or writing. That means they are not a substitute for an LLM in tasks like summarizing, rewriting, or explaining text. The material also does not show every possible decision pattern in detail, only examples such as routing, ranking, classification, scoring, and evaluation.

Microsoft-Decision-1 is described in the research notes as available in public preview through Microsoft Foundry, but the video description itself does not repeat that status. Readers who want to adopt it should verify current documentation before planning production use, especially for workflow-critical scenarios.

Sources

• Rafsan Huseynov video on decision models vs LLMs
• Microsoft documentation for Decision-1 in Foundry

Frequently asked questions

What is a decision model in this video?

It is an AI model that does not write text. It picks an answer from a list you provide and returns a probability for each option.

When should I use a decision model instead of an LLM?

Use a decision model for bounded tasks such as classification, routing, ranking, scoring, or evaluation. Use an LLM when you need generated text, summaries, rewrites, or open-ended answers.

Does the video show decision models and LLMs working together?

Yes. One demo uses Decision-1 to make the decision while the LLM writes the reply in an Azure Logic Apps Automation workflow.

Is Microsoft-Decision-1 generally available?

The supporting research notes describe it as public preview in Microsoft Foundry, but the video description itself does not state availability details. Current documentation should be checked before use.