Not every step in an AI application needs a large language model to generate a response. Many high-volume application steps are decisions: classify an input, select an option, assign a score, verify a condition, prioritize an item, or control a workflow. To suit these tasks, a new class of models have been developed to address these fast, structured, requests.

We are excited to be adding Microsoft-Decision-1 to the Foundry Models Catalog. Our newest, decision model for applications that need to choose among predefined options rather than generate open-ended text.

What are decision models?

As AI applications mature, the industry is moving beyond the idea that every model interaction needs to result in generated content. Developers are increasingly building systems composed of different models, tools, and application logic, each suited to a different part of the workload. Within those systems, many of the most frequent operations are not generation problems at all. They are decisions about what should happen next.

Decision models were designed for this emerging layer of the AI stack. Rather than producing an open-ended response, they evaluate a defined set of possibilities and return structured signals that an application can use to take action. That makes them particularly relevant as developers move toward more dynamic AI applications, where models need to work together and applications must continuously determine where requests should go, which actions should be taken, and when additional processing is needed.

For developers, this creates an opportunity to be more intentional about where and how different models are used. Instead of relying on a general-purpose generative model for every step, developers can match the model to the task, using generative and reasoning models where their capabilities are needed and decision models for focused, repeatable choices. The result is a more modular approach to AI application design, where generation, reasoning, and decision-making become complementary building blocks for creating intelligent systems.

Now available in public preview, Microsoft-Decision-1 is a decision model built on Qwen3.5-9B and designed for applications that need to make fast, structured choices among predefined options. It brings a purpose-built decision capability to the Foundry Models catalog, giving developers another building block for designing AI applications where different steps may benefit from different types of models.

• Built for predefined decisions. Microsoft-Decision-1 supports use cases such as agent controls, model routing, intent analysis, data labeling, AI judging, incident response routing, data validation, and content classification.
• Structured for application logic. The model produces predefined answers and confidence signals that applications can use to select the next action or route uncertain results for further review.
• Designed for lower-latency, lower-cost workflows. Microsoft-Decision-1 is intended for simple, repeatable, and time-sensitive decisions that do not require the generative capabilities or extended reasoning of a larger language model.

You can explore these capabilities in the Foundry Playground. This example below shows Microsoft-Decision-1 checking urgency, selecting a support team, and scoring priority for a customer request. 

Microsoft-Decision-1 in the Microsoft Foundry playground.

Microsoft-Decision-1 represents an area of continued model innovation at Microsoft. The team will rebase future iterations of the model on OpenAI models and Microsoft’s MAI models, bringing together advances from Microsoft’s own model portfolio with a model architecture purpose-built for decision-making. Over time, this work could create opportunities to further explore how Microsoft models can be specialized for the high-volume decisions that sit inside applications and agentic systems.

For developers, Microsoft-Decision-1 expands the choices available when building applications and agents with Foundry. Rather than using the same model for every step, developers can combine generative and reasoning models with a decision model and choose the capability that best fits each part of the workflow. That model diversity is increasingly important as AI applications evolve from single-model experiences into systems that coordinate specialized models, tools, and actions.

Enterprise applications make decisions throughout a workflow: identifying a customer’s intent, selecting a tool, routing an incident or checking whether a response meets specified criteria. For example, a support application could classify a request as a billing question, a technical issue or an account-access problem. An agent could evaluate a proposed response against a quality rubric and determine whether to accept it, revise it or request review.

 

These decisions can sit alongside generative and reasoning steps in the same application. A language model might draft a response, while Microsoft-Decision-1 evaluates it against defined criteria. With Microsoft-Decision-1 in the Foundry catalog, developers can explore decision workloads through the platform they use to discover and deploy models across various scenarios:

• Classification: Assign inputs to defined categories, such as customer intents or feedback themes.
• Routing: Select a model, tool, destination or workflow from a set of options.
• Evaluation and verification: Assess responses or inputs against specified criteria.
• Workflow control: Decide whether to continue, retry, stop or send a case for review.

A different model for a different kind of workload

For developers, decision models introduce a different way to think about optimization. An AI application may make thousands of small choices around a comparatively small number of complex generative or reasoning tasks. Running every one of those steps through the same general-purpose model can overlook an opportunity to optimize the system around the work each step actually needs to perform.

That pattern is showing up across Microsoft. Xbox Research used Microsoft-Decision-1 to categorize more than 10,000 pieces of open-ended feedback into researcher-defined themes, reporting quality competitive with GPT-5 while running 80–100 times faster. Other Microsoft teams have also evaluated the model for assessing Copilot response quality and supporting adaptive replanning in Microsoft Discovery.

These scenarios matter because they are not simply classification benchmarks. They are examples of decisions embedded inside larger applications. As AI architectures become more modular, developers can consider generation, reasoning, and decision-making as distinct workloads, then select models based on what each step actually requires.

For benchmark methodology, measured results and additional examples, read the Command Line blog here.

For tasks with predefined outcomes, evaluate Microsoft-Decision-1 using representative inputs and the quality, latency and cost requirements of your application. Compare decision accuracy, including ambiguous cases, latency under expected traffic and cost per completed decision. Where confidence estimates inform automation or review thresholds, validate their calibration on your own data.

Start with a decision in your application: The following Python example shows how to compare support-routing decisions with expected labels and measure request latency. Replace the illustrative messages and category definitions with examples from your own workload.

Before running: Deploy Microsoft-Decision-1 in your Microsoft Foundry subscription and configure FOUNDRY_BASE_URL and FOUNDRY_API_KEY. Set FOUNDRY_MODEL to the model or deployment identifier specified in the quickstart. Keep credentials outside the source code. Model requests incur applicable usage charges.

Integration note: Confirm the route and authentication header before running it against your deployment.

#!/usr/bin/env python3 """Classify support requests with Microsoft-Decision-1.""" import json import os import urllib.error import urllib.request from collections.abc import Mapping from statistics import median from time import perf_counter from typing import Any from azure.identity import DefaultAzureCredential OPTIONS = { "billing": "Charges, invoices, refunds, or subscription payments", "technical": "Software errors, bugs, or integration failures", "account": "Sign-in, password, or account-access problems", } EXAMPLES = [ ("I was charged twice.", "billing"), ("The integration crashes during checkout.", "technical"), ("I cannot sign in after resetting my password.", "account"), ] AZURE_ENDPOINT = os.environ["AZURE_ENDPOINT"].rstrip("/") DEPLOYMENT_NAME = os.environ["DEPLOYMENT_NAME"] TIMEOUT_SECONDS = 60 TOKEN_SCOPE = "https://cognitiveservices.azure.com/.default" CREDENTIAL = DefaultAzureCredential() class DecisionAPIError(RuntimeError): """Raised when the API can't return a valid classification.""" def read_choice( payload: Any, options: Mapping[str, str], ) -> str: if not isinstance(payload, dict): raise DecisionAPIError("The API returned an invalid response.") answers = payload.get("answers") if not isinstance(answers, dict): raise DecisionAPIError("The API returned no answers object.") answer = answers.get("team") if not isinstance(answer, dict) or answer.get("type") != "choice": raise DecisionAPIError("The API returned an invalid answer.") choice = answer.get("choice") if not isinstance(choice, str) or choice not in options: raise DecisionAPIError( f"The API selected an unsupported team: {choice!r}." ) return choice def predict( text: str, options: Mapping[str, str], ) -> str: if not text.strip(): raise ValueError("text must not be empty.") if len(options) < 2: raise ValueError("options must contain at least two choices.") body = json.dumps( { "model": DEPLOYMENT_NAME, "state": text, "questions": { "team": { "type": "choice", "instructions": ( "Which team should handle this customer support " "request? Select exactly one team based on the " "primary problem." ), "criteria": dict(options), } }, } ).encode("utf-8") access_token = CREDENTIAL.get_token(TOKEN_SCOPE).token request = urllib.request.Request( f"{AZURE_ENDPOINT}/providers/microsoft/v1/systemone", data=body, headers={ "Authorization": f"Bearer {access_token}", "Content-Type": "application/json", "Accept": "application/json", }, method="POST", ) try: with urllib.request.urlopen( request, timeout=TIMEOUT_SECONDS, ) as response: response_body = response.read() except urllib.error.HTTPError as exc: raise DecisionAPIError( f"The API returned HTTP {exc.code}." ) from exc except urllib.error.URLError as exc: raise DecisionAPIError( f"Couldn't reach the API: {exc.reason}." ) from exc try: payload = json.loads(response_body) except (json.JSONDecodeError, UnicodeDecodeError) as exc: raise DecisionAPIError( "The API returned invalid JSON." ) from exc return read_choice(payload, options) def main() -> None: correct = 0 latencies_ms = [] for text, expected in EXAMPLES: start = perf_counter() predicted = predict(text, OPTIONS) elapsed_ms = (perf_counter() - start) * 1000 correct += int(predicted == expected) latencies_ms.append(elapsed_ms) print( f"Expected={expected}, predicted={predicted}, " f"latency={elapsed_ms:.1f} ms" ) accuracy = correct / len(EXAMPLES) print(f"Accuracy: {accuracy:.1%}") print( "Median request latency: " f"{median(latencies_ms):.1f} ms" ) if __name__ == "__main__": main()

This small dataset illustrates the evaluation process. For meaningful results, use a larger, representative labeled dataset and inspect errors by category. The measured latency includes the client call and network time; sequential requests do not measure latency under production load. Assess cost and confidence calibration separately.

Your application defines the available options and resulting actions. Use evaluation results to determine where automation is appropriate and where additional validation or human review is needed.

Pricing

Get started

Explore Microsoft-Decision-1 in the Foundry catalog to make your first request. We look forward to seeing how developers use Microsoft-Decision-1 to bring structured decisions into their applications and agents through Microsoft Foundry.

Copilot Notebooks is an AI-powered workspace for a project or topic to help individuals and teams move work forward. It brings your Microsoft 365 work context into one workspace where you can collect references, ask questions grounded in your project context, uncover key insights across your work, and turn those insights into shareable outputs. Coming to Frontier in the coming weeks, this update helps you create notebooks, work with Copilot, and create or edit artifacts without leaving Notebooks. Try Copilot Notebooks today at https://aka.ms/copilotnotebooks.

Create Notebooks with ease

Copilot can help get you started by creating a new Copilot Notebook at your request. With new expanded references support and a UX refresh that keeps your notebooks organized, you can set up a notebook for your workflow.

Create a new notebook by just describing your goal to Copilot, rather than setting it up manually. Simply describe what kind of notebook you'd like to make — “I’m working on pitching to a new customer in [industry] and need to get more familiar with the industry and its biggest players” — and notebook creation with Copilot can put together relevant references across your work data and the web to build out that notebook for you.

The notebook then isn’t just an empty container, it’s a workspace pre-filled with relevant references. More reference types means Markdown, TXT and RTF files can be added alongside the existing supported source types. So documentation, notes, logs, and transcripts are also included along with your documents, emails, and meetings.

As your project progresses, the new UX update brings a refreshed experience to the Microsoft Copilot app and OneNote to help keep your notebook organized and easy to use by grouping and reordering content and artifacts from the navigation pane.

The new UX update is currently available in Frontier, with notebook creation with Copilot coming in the next few weeks. Expanded reference types are generally available.

Get started:

• “I’m working on pitching to a new customer in [industry] and need to get more familiar with the industry and its biggest players. Create a notebook on the industry in 2026, including a competitive analysis and overview.”

Work with Copilot in Notebooks

Once your notebook is set up with the right context, Copilot can help move work forward by managing project information, carrying out simple tasks, and automating recurring work. You can also choose the model Copilot uses to tailor its response to your needs.

Copilot can now help carry out tasks. Copilot does simple work for you — Use Copilot to manage references and keep your notebooks current, synthesize information across project context, and create and edit single files like Word, Excel, or PowerPoint documents from chat. And now scope your prompt to just your Notebook only or change to "Work and web" for broader context beyond your references.

 

For frequent tasks, you don’t have to repeatedly prompt Copilot again and again. Schedule reoccurring Copilot tasks, like creating a daily morning brief or weekly status deck, into simple automations so you just have to prompt Copilot once. Create new custom automations, use scheduled task templates, adjust settings like frequency and task instructions, and monitor active and inactive automations.

When you’re looking for a certain response quality, model choice lets you specify the model Copilot uses in your Notebooks chat for your prompt, whether you want it to “think deeper” or choose a specific GPT model. More models coming soon.

Delegating simple work and model choice are currently available in Frontier, with simple task automations coming in the next few weeks.

Get started:

• "Find and add relevant references from this past week based on [Project X] to my notebook."
• "Every Friday at 4:00 PM, Summarize activity across [Topics or Workstreams] from [Time Period], including notable updates, decisions, actions, risks, and unresolved issues. Create a crisp PowerPoint deck that gives a clear weekly narrative, calls out what changed, and closes with follow-ups for the coming week."

Create and edit artifacts from Notebooks

With more ways to create, edit, and customize artifacts, you can complete your workflow and get more work done without leaving Notebooks.

Notebooks doesn’t just start and end with adding references and working with Copilot. Now, with the full power of Office, create, edit, and collaborate on documents, spreadsheets, and presentations without breaking your flow. View documents inline, and edit Word, Excel, PowerPoint files directly on the canvas or by prompting Copilot to add new content such as an intro paragraph, bullet points, or a graph, or adjust formatting such as fonts, colors, and layouts. Teams can also invite others to collaborate and continue refining the same file with Copilot. Read more about Office in Copilot here.

Deliverables aren’t always simple Word, Excel, and PowerPoint files. Go beyond traditional Office deliverables with new support for interactive reports. Create, live-edit, and collaborate with your team and Copilot on visual reports for data trends, monthly updates, and other dynamic content. You can also create PDFs from your notebook content through chat for a polished file like a one-pager you can save or send.

In addition to the other visual and interactive artifact formats, the updated audio overview offers another way to engage with your notebook content. They now draw from more reference types in your Notebook like Outlook emails, Teams meetings and chats, OneNote and Loop pages, and plain text (.txt) files, for a more comprehensive overview. New audio formats are now available such as “Critique and Debate.” You can also navigate longer overviews with topic-based episodes and save audio overview files to OneDrive with inherited sensitivity labels.

 

Copilot can also suggest artifacts for your project, proactively recommending relevant Word, Excel, or PowerPoint files based on WorkIQ and your notebook content, then lets you quickly generate them. More artifact types to come.

The power of Office, interactive reports, and PDF files are currently available in Frontier, with new audio overview updates coming in the next few weeks. Suggested artifacts are generally available.

Get started:

• "Update this project plan into an executive-ready document by adding a summary, action items, and next steps."
• "Create a rich interactive report from this notebook, complete with executive summaries, visual insights, filters, drill-down sections, and source-grounded references, so stakeholders can quickly understand the key findings and take action."
• "Create a one-page PDF executive summary of this notebook, highlighting the key findings, decisions, and recommendations to share with leadership."

Review our previous blog about Copilot Notebooks

What’s New in Notebooks | June 2026 | Microsoft Community Hub

 

Microsoft Office has defined productivity for the PC era, driving the shift from desktop software to cloud computing and collaboration-based work. Now, we’re defining productivity in the AI era with Microsoft Copilot. The Copilot app is where you can move from intent to outcome. Start with an idea, and work with AI to bring it to the finish line.

Today, we’re bringing the full power of Office into the Copilot app, combining the rich creation and collaboration capabilities of Word, Excel, and PowerPoint with an AI-first way of working. You can start in either Chat or Cowork to explore an idea or create a first draft, then continue editing and collaborating directly within the same experience. Office in Copilot allows you to move from an idea to a finished document, spreadsheet, or presentation without losing context - within Office in Copilot or if you switch to the standalone apps. 

That continuity matters because work is often complex: documents have multiple contributors, presentations are built from branded templates, and spreadsheets power important business decisions. By bringing Office into Copilot, people can use the editing, fidelity, and collaboration capabilities they already rely on without leaving their AI-powered workflow.

Copilot also gives users flexibility within that workflow: they can select the model that best fits the task at hand or let Auto choose the right model automatically.

 

Built for teamwork

Great work rarely happens in a single prompt. A launch brief becomes a working document, a budget model changes as assumptions shift, and a presentation takes shape through rounds of feedback. The Copilot app is designed for that full journey—not just the first draft.

Give Copilot a prompt, like “Turn these research notes into a customer-ready launch brief with an executive summary, key messages, risks, and next steps.” Copilot creates a Word document, which opens side by side with your Copilot conversation.

Your conversation history is retained, and you can continue to iterate on the document with follow-up prompts like “Make the opening more direct,” or “Add a table comparing the three rollout options.” You also have full editing capabilities, so you make direct edits using familiar controls available in the canvas itself.

When it’s time to bring others in, you can work with teammates in the same view with built-in sharing, comments, coauthoring, and presence. Any changes that you or others make are visible in real time. Copilot retains everything in the same file with no separate drafts to reconcile. The result is a connected experience where creating, editing, and collaborating happen in one place, in real time.

A home for your work

Sometimes the next step starts by building on work you’ve already created. Library in the Copilot app brings your recent files across Word, Excel, and PowerPoint, as well as pages and AI-generated images, whether they were created in Copilot or in the Office apps. You can quickly find and reopen an existing file, continue editing it alongside Copilot, or even start a new document, spreadsheet, or presentation directly from Library.

Because every file is stored in OneDrive, the work remains connected wherever you pick it up next. Files retain their formatting, comments, sharing permissions, and version history, while inheriting the security labels, compliance, and governance controls organizations already trust. From a first prompt to the latest team edit, your work stays easy to find, ready to continue, and grounded in the same familiar Office file.

New experiences across Microsoft 365

We're also continuing to improve Copilot across Microsoft 365,  ensuring the latest is coming to Office in Copilot.

In PowerPoint, Copilot helps keep presentations on-brand automatically, makes it easier to refine slides and images, and quickly apply custom templates. In Excel, users can review edits made by Copilot, understand what changed, and create visualizations with chart recommendations. Across Microsoft 365, Skills and connectors bring specialized expertise directly into the flow of work, from financial analysis in Excel to legal workflows in Word, while enabling organizations to scale best practices through reusable, role-specific experiences. The result is a work experience that connects specialized AI assistance with the tools people use every day.

Office, wherever work happens

For decades, Office has helped people turn ideas into documents, data into decisions, and presentations into action. As Copilot becomes a new way to work, we're bringing familiar Office creation and collaboration experiences directly into that environment.

This is another step toward a more unified experience where AI and productivity tools work together seamlessly.  From the first idea in the Copilot app to the final deliverable in the Office apps, people can move from intent to outcome with less friction and more momentum.

Office in Copilot is rolling out today for Frontier customers in Chat and Cowork.

---

BONUS: AI Changes Management, Not Leadership

In this BONUS episode, we explore what AI changes for software leaders, and what it leaves firmly in human hands. Mateo Bervejillo, CEO of Xseed Solutions and author of Back to Basics, brings the view of a manager of managers: AI can help us process information faster, but leadership still happens through clarity, judgment, and relationships.

From Attorney to Software CEO: Clarity as a Leadership Skill

"What are we building? Why are we building it? What's the benefit for the client? What's the business value?"

 

Mateo started his career as an attorney, then moved through operations, vendor management, business development, and finally into software company leadership. That path gave him an unusual advantage in technology: an obsession with clear definitions. In law, unclear scope creates risk. In software, unclear scope creates waste, frustration, and products that miss the point. For Mateo, leadership starts before execution, with crisp answers to simple questions: what problem are we solving, why does it matter, and what would success look like?

AI Removes Busywork, But It Does Not Replace Leadership

"There's a part about management that is moving information around, up and down, sideways. That is changing drastically because of AI."

 

Mateo sees AI transforming the administrative side of management: gathering data, processing updates, preparing reports, and helping managers stay on top of follow-ups. But he draws a firm line between management mechanics and leadership. AI can make managers faster at moving information, but it cannot choose the right people, create trust, provide meaningful feedback, or understand what motivates each person. The danger is that organizations may conclude they need fewer leaders because the reporting work became easier, while the human work actually becomes more important.

The One-on-One Cannot Become a Slack Check-In

"If you have one manager leading 50 people, that one-on-one that was so important now becomes a Slack message."

 

One of the risks Mateo calls out is the pressure to flatten organizations and stretch managers across too many people. On paper, AI makes that look possible. In practice, it can turn leadership into status collection. Mateo described his own weekly rhythm: face-to-face conversations, walking meetings, and one-hour one-on-ones where the work is to listen, ask questions, understand what is happening in the person's life and work, and keep the company aligned around its North Star. That is not overhead. It is the work that keeps teams healthy before problems become escalations.

Use AI as a Sparring Partner, Not as the Thinker

"The problem of the blank page goes away with AI. You can have a conversation."

 

Mateo uses AI as an assistant for podcast preparation, idea generation, research, outreach, and even book writing. When drafting Back to Basics, he used AI to challenge the structure and ask what was missing, which led to several chapters he had not planned. The point is not to outsource judgment. It is to remove the fear of starting, create better first drafts, and pressure-test thinking. For Scrum Masters and agile coaches, that means AI can help design workshops, prepare questions, and explore options, but the leader still has to understand the context and make the call.

Outcomes Keep Leaders Away From Micromanagement

"Too much focus on the inputs can be misleading and you forget about the outcome, which is where the client sits."

 

AI gives managers more visibility into activity, and that can quickly become control. Mateo's antidote is to keep the organization focused on outputs and outcomes instead of inputs. Story points, process steps, and activity metrics can be useful, but they can also pull leaders into managing the wrong thing. Clients care about outcomes. Teams need clarity about the result, then room to decide how to get there. When leaders become obsessed with the how of every input, they move from creating clarity into micromanagement.

Managing Up With AI, Managing People With Care

"Today we are going through a time where we need to stay close to people more than ever and make it personal."

 

In Mateo's book, leadership is framed across four dimensions: managing up, managing yourself, managing teams, and managing sideways. Through the AI lens, he sees the biggest immediate gain in managing up: clearer communication, faster delivery, better follow-up, and fewer things falling through the cracks. But when it comes to managing teams, the basic human work becomes more important. People are anxious about AI and jobs, even when they are not saying it openly. Leaders need to stay close, make the relationship personal, and create the conditions where people can thrive.

About Mateo Bervejillo

Mateo Bervejillo is CEO of Xseed Solutions, a Montevideo-based software development and staff augmentation company helping US firms build high-performing LATAM engineering teams. An attorney by background, he brings a results-first view to tech leadership. He wrote Back to Basics, a practical leadership guide for new managers in technology. LinkedIn.

 

You can link with Mateo Bervejillo on LinkedIn, learn more about Xseed Solutions, and listen to Mateo's podcast, The Future Of The Future.

 

Download audio: https://traffic.libsyn.com/secure/scrummastertoolbox/20261010_Mateo_Bervejillo_BONUS.mp3?dest-id=246429

Share Episode         
         
We welcome David Burns, Head of Open Source at BrowserStack and previously Mozilla, this week to help explain how can browser and mobile app testing be so bad still in 2026. "Even if we had to do it all over again, we probably still wouldn't have gotten it right", might just be the biggest learning. We review some of the core challenges that cause even the most experienced web engineers pain and how some companies are attempting to avoid them.

         

We also review shares lessons from Mozilla’s 'War on Orange,' demonstrating why chasing a 100% green CI suite at scale is a fool's errand, and why engineering teams must instead manage delta tolerances and adhere to a strict testing pyramid. Additionally, we investigate the physical hardware nightmares of running 30,000 real phones in data centers; from battery swelling and custom Faraday cages to complying with Apple EULA requirements using racks of Mac Minis. And finally, we discuss WebMCP, Project Fugu, and why LLM browser agents still struggle against edge-case browser behaviors like the Back-Forward Cache (BF Cache).