AB-731 Study Guide

Generative AI creates new content, such as text, images, code, and audio, from patterns it learned during training. Traditional AI usually classifies, predicts, or recommends based on existing data. For a leader, the practical difference is that generative AI produces a draft or an answer, where earlier AI mostly sorted things into categories or forecast a number.

What you need to know

• Generative AI produces original content in response to a natural language prompt; classic machine learning predicts a label or a value from structured data
• Large language models (LLMs) are the engines behind most generative AI, trained to predict the next token in a sequence
• Traditional AI examples: spam filtering, demand forecasting, product recommendations, fraud detection
• Generative AI examples: drafting an email, summarizing a meeting, writing code, generating an image
• Generative AI is a subset of AI, a broad field that also includes machine learning, computer vision, and natural language processing

Microsoft Learn resource: Introduction to generative AI and agents (opens in a new tab)

Not every problem needs generative AI, and part of a leader's job is knowing when it fits. Generative AI is strong at open-ended language and content tasks; a rules engine, a report, or a classic prediction model is often cheaper and more reliable for structured problems.

What you need to know

• Generative AI fits tasks like drafting, summarizing, rewriting, brainstorming, and answering questions over your own content
• It is a poor fit where answers must be exact and auditable every time, such as tax calculations or safety-critical control
• Match the tool to the task: Microsoft 365 Copilot for everyday work, Copilot Studio for custom agents, Microsoft Foundry when you need to build on models directly
• Start with a clear business outcome and success measure, then choose the solution, not the other way around
• Consider data readiness, cost, and risk before committing, not just the demo

Microsoft Learn resource: Create business value with AI (opens in a new tab)

A model is the trained engine that generates output. Most organizations start with a pretrained model that Microsoft or a partner already built, and only some go further and fine-tune one on their own examples. Knowing the difference helps you judge cost, effort, and when each approach is worth it.

What you need to know

• Fine-tuning changes the model's behavior; grounding leaves the model alone and feeds it your data when you ask
• Grounding is usually the first thing to try, because it is cheaper and keeps answers current
• Fine-tuning needs quality training examples, time, and ongoing maintenance

Microsoft Learn resource: Introduction to generative AI and agents (opens in a new tab)

Generative AI is usually billed by usage, and the unit of usage is the token. A leader needs to understand what drives the bill and how to weigh it against the value the solution delivers.

What you need to know

• A token is a chunk of text, roughly a few characters or part of a word; both the prompt you send and the response you get back count
• Longer prompts, bigger documents, and longer answers all raise token usage and therefore cost
• License-based products like Microsoft 365 Copilot bill per user per month; consumption-based services like Microsoft Foundry bill by tokens or usage
• ROI weighs the cost against measurable gains: time saved, faster turnaround, fewer errors, higher output
• Pilot with a defined group and measure before scaling, so ROI is evidence rather than a guess

Exam tip: If a question mentions "tokens," it is testing that you understand usage-based cost. If it mentions "per user per month," it is testing licensing.

Microsoft Learn resource: Create business value with AI (opens in a new tab)

Generative AI is powerful, but it has real limitations that a leader has to plan around rather than ignore. The exam expects you to name these honestly and know that Microsoft's response is grounding, human oversight, and responsible AI practices.

What you need to know

• Fabrication (sometimes called hallucination) is when a model produces confident but incorrect content; grounding on trusted data reduces it
• Reliability varies: the same prompt can produce different answers, so output needs review for high-stakes decisions
• Bias in training data can show up in output; representative data and testing help, but a human still checks
• Models have a knowledge cutoff and do not know recent events unless you ground them on current data
• Keep a human in the loop for anything consequential; treat output as a well-informed draft, not a final answer

Microsoft Learn resource: Introduction to generative AI and agents (opens in a new tab)

Generative AI creates value when it removes repetitive language work, speeds up a process, or lets a team do more without adding headcount. The exam frames this around scalability and automation.

What you need to know

• Automation: generative AI handles first drafts, summaries, and routine responses so people spend time on higher-value work
• Scalability: an AI solution can serve many users or handle spikes in demand without the linear cost of hiring
• Good early wins are high-volume, low-risk tasks: drafting, summarizing, translating, answering common questions
• Value shows up as time saved, faster response, higher consistency, and better employee experience
• The strongest cases pair a clear pain point with a way to measure the improvement

Microsoft Learn resource: Create business value with AI (opens in a new tab)

The quality of what you get out of generative AI depends heavily on what you put in. Prompt engineering is the practice of writing clear, specific instructions, and it is the single biggest lever most users have over output quality.

What you need to know

• A good prompt gives the model a goal, relevant context, a sense of the audience, and the format you want back
• Vague prompts produce vague answers; specific prompts produce useful ones
• Prompting is a skill your whole organization can learn, and it raises the value of every AI tool you already pay for
• Microsoft 365 Copilot offers a Prompt Gallery and prompt suggestions to help people get started
• Iterating on a prompt, refining it after seeing the first answer, is normal and expected

Microsoft Learn resource: Write effective prompts to achieve optimal results (opens in a new tab)

Beyond writing clearly, there are repeatable techniques that improve results, and a leader should recognize them well enough to encourage good practice across teams.

What you need to know

• Include context: reference the document, meeting, or data the answer should draw on
• Be specific about the format and length you want, such as a table, a summary, or five bullet points
• Give the model a role or perspective when it helps, for example "act as a financial analyst"
• Break complex requests into steps rather than asking for everything at once
• Review and refine: treat the first answer as a starting point and correct it with a follow-up prompt

Microsoft Learn resource: Write effective prompts to achieve optimal results (opens in a new tab)

Grounding is giving a model access to your organization's own data so its answers reflect your reality, not just its general training. It is how generic AI becomes useful for your business.

What you need to know

• Grounding connects a model to trusted sources, such as your documents, so answers are relevant and current
• It reduces fabrication because the model has real content to draw from instead of guessing
• Business requirements to weigh: which data to expose, who can see it, and how permissions are respected
• Microsoft 365 Copilot is grounded on your Microsoft 365 data through Microsoft Graph, honoring existing access rights
• Grounding keeps your data yours: it is used to answer your prompt, not to train the shared model

Microsoft Learn resource: Develop generative AI apps on Microsoft Foundry (opens in a new tab)

Retrieval-augmented generation, or RAG, is the most common pattern for grounding. Instead of retraining a model, you retrieve relevant content at the moment of the question and hand it to the model to answer from.

What you need to know

• RAG works in two steps: retrieve the relevant content, then generate an answer using it
• It keeps answers current without retraining, because it reads live data each time
• It lets a general model answer accurately about private or recent information
• Azure AI Search is the retrieval engine behind many RAG solutions on Microsoft Foundry
• RAG is usually cheaper and simpler than fine-tuning for keeping answers accurate and up to date

Exam tip: If a scenario needs answers grounded in current, private data without retraining a model, the answer is RAG.

Microsoft Learn resource: Develop generative AI apps on Microsoft Foundry (opens in a new tab)

AI is only as good as the data behind it. For grounding, fine-tuning, and machine learning alike, the quality and representativeness of your data shapes the quality of the results.

What you need to know

• Data quality matters more than volume: clean, accurate, well-organized data produces better answers
• Representative data reduces bias; data that overrepresents one group skews results toward it
• Data type matters: structured data suits classic machine learning, while unstructured text and documents suit generative AI and grounding
• Getting your content organized and labeled is often the real work before an AI project succeeds
• Poor or messy data is a common reason AI pilots disappoint, so data readiness is worth checking first

Microsoft Learn resource: Introduction to machine learning concepts (opens in a new tab)

For a business leader, adopting AI responsibly means protecting the data it touches. Microsoft's approach keeps your data protected while people work, so security enables adoption rather than blocking it.

What you need to know

• Enterprise AI like Microsoft 365 Copilot honors your existing identity, permissions, and data protection: it only surfaces content a user could already reach
• Your organization's prompts and data are not used to train the foundation models
• Data protection tools like sensitivity labels and data loss prevention continue to apply to AI-assisted work
• Security lets you say yes to AI with confidence, because the guardrails travel with the data
• Governance decisions, such as which data AI can reach, are set by your organization and respected by the service

Microsoft Learn resource: Implement a responsible generative AI solution in Microsoft Foundry (opens in a new tab)

Generative AI gets the headlines, but classic machine learning is still the better tool for many business problems, and the exam expects you to tell them apart.

What you need to know

• Machine learning predicts or classifies from structured, historical data: forecasting demand, scoring leads, detecting fraud
• It shines when you have lots of labeled examples and need a number or a category, not new content
• Generative AI shines when the task is producing language or content, not predicting a value
• Many real solutions combine the two: machine learning for the prediction, generative AI for the explanation
• Choosing machine learning where it fits is often cheaper and more accurate than forcing generative AI onto the problem

Microsoft Learn resource: Introduction to machine learning concepts (opens in a new tab)

A machine learning solution is not a one-time build; it is a cycle. A leader should recognize the stages, because each one needs people, data, and ongoing investment.

What you need to know

• Define the problem and the business outcome you want to predict or classify
• Prepare data: collect, clean, and label the examples the model learns from
• Train and evaluate the model, then test it against data it has not seen
• Deploy the model so applications and people can use its predictions
• Monitor and retrain over time, because real-world data drifts and accuracy degrades if you leave a model alone

Microsoft Learn resource: Introduction to machine learning concepts (opens in a new tab)

When AI is built into applications, it inherits the same security needs as any enterprise system, plus a few of its own. A leader should know the categories well enough to ask the right questions.

What you need to know

• Application security: protect the app and its connections the same way you would any business system
• Data security: control what data the AI can reach, and keep sensitive content protected with labels and policies
• Authentication: users sign in with their own identity, so the AI acts within their permissions, never beyond them
• Least privilege applies: an AI solution should reach only the data it needs for the task
• Auditing and monitoring let you see how AI is used and catch misuse early

Microsoft Learn resource: Implement a responsible generative AI solution in Microsoft Foundry (opens in a new tab)

The most valuable skill here is spotting where Copilot fits a real workflow. Copilot earns its keep on the language-heavy, repetitive parts of everyday work, so the exercise is matching common business processes to concrete Copilot use cases.

What you need to know

• Communication: draft and summarize emails in Outlook, catch up on long Teams threads
• Meetings: summarize a Teams meeting, list action items and decisions, recap what you missed
• Documents: draft a first version in Word, rewrite for tone, summarize a long report
• Data and slides: surface trends in Excel, build a first-draft deck in PowerPoint from a document
• The best early use cases are high-volume and low-risk, where a fast first draft saves real time

Microsoft Learn resource: Draft, analyze, and present with Microsoft 365 Copilot (opens in a new tab)

Microsoft uses the name Copilot across several products, and the exam expects you to tell them apart, especially the free web-grounded chat versus the licensed, work-grounded Microsoft 365 Copilot.

What you need to know

• The paid Microsoft 365 Copilot is the one grounded on your organization's content and embedded in Word, Excel, Outlook, Teams, and PowerPoint
• Microsoft 365 Copilot Chat gives web-grounded chat with commercial data protection to a broad set of users
• Enterprise data protection means prompts and responses are not used to train the models

Microsoft Learn resource: What is Microsoft Copilot? (opens in a new tab)

Copilot Chat is the conversational front door to Copilot, available in the browser and on mobile. A leader should know what it can do and how its data protection works.

What you need to know

• It is a chat experience for asking questions, drafting, and summarizing, reachable on the web and in the mobile app
• With a Microsoft 365 Copilot license, chat can be grounded on your work data through Microsoft Graph
• Without that license, Copilot Chat still offers web-grounded answers with enterprise data protection
• It supports uploading a file to ask questions about it, and can create images
• Mobile access means people can use it on the go, with the same data protection as the web

Microsoft Learn resource: Get started with Microsoft Copilot Chat (opens in a new tab)

Microsoft 365 Copilot is not one feature; it shows up inside each app tuned for the work that happens there. The exam expects you to match the app to what Copilot does in it.

What you need to know

• Word: draft, rewrite, and summarize documents, and turn a document into a starting point
• Excel: explore data, spot trends, and suggest formulas in natural language
• PowerPoint: build a first-draft presentation from a prompt or a document
• Outlook: draft replies, summarize long threads, and coach your tone before you send
• Teams: summarize meetings and chats, and surface decisions and action items

Microsoft Learn resource: Draft, analyze, and present with Microsoft 365 Copilot (opens in a new tab)

Copilot Studio is the low-code tool for building custom agents and extending Copilot. A leader should know it exists, what it is for, and that it does not require professional developers.

What you need to know

• Copilot Studio builds custom agents that answer questions and take actions using your own data and systems
• It is low-code, so business users and makers can build, not just developers
• Agents can be published to Microsoft 365 Copilot, Teams, websites, and other channels
• It connects to your data and to hundreds of systems through connectors
• It is how organizations extend Copilot beyond what comes in the box, for a specific process or audience

Microsoft Learn resource: Get started with Microsoft Copilot Studio (opens in a new tab)

Microsoft Graph is the connective layer that makes Microsoft 365 Copilot aware of your organization's content. Understanding it explains why Copilot's answers are relevant and how permissions are respected.

What you need to know

• Microsoft Graph is the gateway to your Microsoft 365 data: emails, files, chats, meetings, calendars, and people
• Copilot uses Graph to ground answers in your actual work content, not just the model's training
• The semantic index over Graph helps Copilot find the most relevant content for a prompt
• Graph honors existing permissions, so Copilot only surfaces content the user could already open
• It is the reason Microsoft 365 Copilot answers are specific to your organization rather than generic

Microsoft Learn resource: Microsoft Graph overview (opens in a new tab)

Microsoft's pitch is that an integrated AI stack, from Microsoft 365 Copilot to Copilot Studio to Foundry, is safer and easier to govern than a patchwork of disconnected tools. A leader should be able to explain that benefit.

What you need to know

• One identity, one permission model, and one data protection layer span the whole stack
• Enterprise data protection and responsible AI controls are built in, not bolted on
• Integration means data does not have to leave your trust boundary to be useful
• Governance and auditing are consistent, so risk is easier to manage than with scattered tools
• Buying into one ecosystem reduces integration cost and the safety gaps that appear between separate products

Microsoft Learn resource: What is Microsoft Copilot? (opens in a new tab)

Beyond Copilot, Microsoft offers a range of AI tools, and part of a leader's job is routing a business need to the right one. This skill is about choosing among the products, not mastering any single one.

What you need to know

• Everyday productivity for licensed users: Microsoft 365 Copilot inside the apps
• A custom agent for a specific process or audience: Copilot Studio
• Building on models or adding AI to your own applications: Microsoft Foundry and Foundry Tools
• Broad, web-grounded assistance for everyone: Microsoft Copilot or Copilot Chat
• Start from the business outcome and the users, then pick the product that fits, rather than fitting the process to a favorite tool

Microsoft Learn resource: Leverage AI tools and resources for your business (opens in a new tab)

Researcher and Analyst are specialized reasoning agents in Microsoft 365 Copilot for deeper, multi-step work than a normal chat turn. Knowing which to reach for is the point of this skill.

Researcher: Use it for complex, multi-step research that pulls together information from your work data and the web into a thorough, reasoned answer. It is the one to reach for when a question needs investigation, not a quick reply.

Analyst: Use it for data-heavy questions, where it reasons over your data step by step to produce analysis and insight, working through the problem the way a data analyst would.

Exam tip: Match the agent to the task: Researcher for deep research across sources, Analyst for reasoning over data. Both go beyond a standard Copilot chat answer.

Microsoft Learn resource: Microsoft Copilot - Service Descriptions (opens in a new tab)

Not every AI need calls for building something new. A leader should weigh buying a ready-made capability, extending Copilot, or building a custom solution, and know that extending is often the middle path.

What you need to know

• Buy: use Microsoft 365 Copilot as it comes when the built-in capabilities meet the need
• Extend: add your own data, actions, and agents to Copilot through the extensibility framework, using agents, connectors, and plugins
• Build: create a custom solution on Microsoft Foundry when you need full control over models and behavior
• Extending reuses Copilot's identity, security, and interface, so it is usually faster and safer than building from scratch
• The right choice weighs cost, time, control, and how unique the requirement really is

Microsoft Learn resource: Choose the best Microsoft 365 Copilot extensibility path for your scenario (opens in a new tab)

Foundry Tools are the building-block AI services you compose into custom solutions, for needs that Copilot does not cover out of the box. A leader should recognize the kinds of problems they solve.

What you need to know

• Foundry Tools are individual AI capabilities you add to your own applications, such as vision, language, and search
• They fit custom scenarios: processing documents, reading images, adding search over your content, building a customer-facing assistant
• They are consumption-based, so you pay for what you use rather than per user
• They target developers and solution teams, unlike Copilot, which targets everyday users
• Reach for them when the need is specific to your product or process and no packaged Copilot fits

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

The exam names three parts of the Foundry family, and a leader should be able to say what each one does in a sentence.

Azure Vision in Foundry Tools: Reads and understands images and video, including extracting text, detecting objects, and describing what is in a picture. Use it to automate anything that today relies on a person eyeballing an image or a scanned document.

Azure AI Search: Indexes your content so applications, and grounded AI, can retrieve the most relevant pieces on demand. It is the retrieval engine behind many RAG solutions.

Microsoft Foundry: The platform for building, deploying, and managing AI solutions, with a model catalog to choose from and tools to ground, evaluate, and monitor them responsibly.

Microsoft Learn resource: Microsoft Foundry documentation (opens in a new tab)

Microsoft Foundry offers a catalog of many models, and part of a leader's judgment is knowing that the biggest or newest model is not always the right one. The fit depends on the task, the cost, and the constraints.

What you need to know

• Foundry's model catalog includes models from Microsoft, OpenAI, and the open-source community
• Bigger models are more capable but cost more and can be slower; smaller models are cheaper and faster for simple tasks
• Match the model to the task: a small model may handle classification well, while a complex reasoning task needs a larger one
• Consider language support, cost, speed, and any data-residency or compliance constraints
• You can start with a capable model and optimize toward a cheaper one once the solution works

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

For custom AI, Microsoft Foundry's pitch to a leader is enterprise-grade scale and security without building the platform yourself. Know the benefits well enough to defend the choice.

What you need to know

• Scalability: the platform handles growth in users and demand without you managing the underlying infrastructure
• Security: it runs on Azure's identity, network, and compliance controls, so enterprise requirements are met
• Responsible AI is built in, with tools to evaluate models and filter harmful content
• One platform covers the lifecycle: choose a model, ground it, evaluate it, deploy it, and monitor it
• Consumption pricing means you scale cost with usage rather than paying up front

Microsoft Learn resource: Microsoft Foundry documentation (opens in a new tab)

Responsible AI is the practice of building and using AI in a way that is fair, safe, and trustworthy. For a leader it is not a compliance checkbox; it is what protects the organization's reputation and its people as AI spreads through the business.

What you need to know

• Responsible AI reduces the risk of biased, unsafe, or privacy-violating outcomes that damage trust
• It is an ongoing organizational commitment, not a one-time review at launch
• Microsoft publishes its own Responsible AI Standard and builds controls into its products
• Leaders set the tone: responsible AI succeeds when it is owned at the top, not left to individual teams
• Getting it right early is far cheaper than repairing trust after a public failure

Microsoft Learn resource: Embrace responsible AI principles and practices (opens in a new tab)

Governance is how an organization decides what AI it will use, for what, and within what limits. A leader should know the pieces that make up a workable AI governance approach.

What you need to know

• Set clear policies for acceptable use: what AI may be used for, and what is off-limits
• Decide what data AI tools may access, and keep that aligned with existing data protection
• Define who is accountable for AI decisions and who reviews high-impact use cases
• Provide guidance and training so employees know the rules and why they exist
• Review and update governance as tools, regulations, and risks change

Microsoft Learn resource: Embrace responsible AI principles and practices (opens in a new tab)

An AI council is a cross-functional group that steers AI strategy and keeps decisions aligned across the business. The exam expects you to know why it matters and who belongs on it.

What you need to know

• An AI council brings together leaders from IT, security, legal, HR, and the business, not just technologists
• Its job is oversight and alignment: setting priorities, approving high-impact use cases, and resolving cross-team conflicts
• It keeps AI strategy tied to business goals rather than scattered pilots
• It owns responsible AI policy and makes sure governance is actually followed
• Cross-functional membership means risks like privacy, compliance, and culture are considered together, not in silos

Microsoft Learn resource: Scale AI in your organization (opens in a new tab)

Microsoft frames responsible AI around six principles, and the exam expects you to know them by name. They are the yardstick you hold any AI solution against.

The six Microsoft responsible AI principles:

1. Fairness: AI systems should treat everyone equitably and avoid reinforcing bias
2. Reliability and safety: AI should perform consistently and safely, even in unexpected conditions
3. Privacy and security: AI should protect data and respect privacy throughout its lifecycle
4. Inclusiveness: AI should work for people of all abilities and backgrounds
5. Transparency: people should understand how an AI system works and its limitations
6. Accountability: people, not the AI, remain responsible for its outcomes

Exam tip: Expect a question that gives you a scenario and asks which principle applies. Bias in results points to fairness; an unexplained decision points to transparency; a human staying responsible points to accountability.

Microsoft Learn resource: Embrace responsible AI principles and practices (opens in a new tab)

Buying AI licenses does not change how people work; a dedicated adoption team is what turns a purchase into daily use. Know what the team is for and who is on it.

What you need to know

• An adoption team owns the rollout: planning, communication, training, and measuring progress
• It blends roles: an executive sponsor, IT, communications, training, and business unit representatives
• The executive sponsor matters most, because visible leadership support drives uptake
• The team runs the rollout in phases, learns from each one, and adjusts
• Its success measure is real usage and business impact, not licenses assigned

Microsoft Learn resource: Microsoft Copilot adoption and onboarding guide for IT admins (opens in a new tab)

Most AI rollouts stall for predictable reasons, and a leader who can name them can plan around them. The exam tests whether you recognize the usual barriers and their fixes.

What you need to know

• Lack of awareness: people do not know the tool exists or what it can do for them
• Lack of skills: people are not confident writing prompts or fitting AI into their work
• Trust and fear: worries about accuracy, privacy, or being replaced hold people back
• No clear use cases: without role-specific examples, people do not know where to start
• Weak leadership support: adoption stalls when leaders do not model the behavior themselves

Microsoft Learn resource: Scale AI in your organization (opens in a new tab)

Champions are enthusiastic employees who help their peers adopt AI, and a champions program scales adoption far beyond what a central team can reach alone.

What you need to know

• Champions are volunteers from across the business who learn the tools early and help colleagues
• They spread role-specific tips and real use cases that resonate more than top-down training
• They give the adoption team feedback from the front line about what is working and what is not
• Recognizing and supporting champions keeps the program energized
• Peer influence lowers the trust and skills barriers better than a mandate does

Microsoft Learn resource: Microsoft Copilot adoption and onboarding guide for IT admins (opens in a new tab)

Rolling out AI has consequences a leader has to anticipate, not discover later. This skill is about foreseeing the impact across data, security, privacy, and budget.

What you need to know

• Data: AI makes existing content far more discoverable, so oversharing and poor permissions surface fast; clean up access before rolling out
• Security: more AI touchpoints mean identity and access controls matter even more
• Privacy: be clear about what data AI uses and make sure it aligns with policy and regulation
• Cost: per-user licenses and consumption billing both add up, so plan and monitor spend
• The fix is preparation: tighten data governance and set a budget before scaling, not after

Microsoft Learn resource: Embrace responsible AI principles and practices (opens in a new tab)

Copilot is licensed in a few different ways, and a leader planning a rollout needs to know which model applies to which product so the budget is right.

What you need to know

• Included with a subscription: Microsoft 365 Copilot Chat comes with many Microsoft 365 plans, with enterprise data protection, at no added cost
• Per-user subscription: Microsoft 365 Copilot is a paid add-on licensed per user, billed monthly or annually
• Pay-as-you-go: agents built in Copilot Studio can be billed by consumption, so you pay for what is used
• The included Chat is a low-risk way to start broad, before committing to per-user licenses
• Match the model to the audience: per-user licenses for heavy users, consumption for occasional agent use

Microsoft Learn resource: Microsoft Copilot - Service Descriptions (opens in a new tab)

Custom AI built on Foundry is billed differently from Copilot, and a leader should know the two main models and when each makes sense.

What you need to know

• Pay-as-you-go: you pay for what you use, by tokens or transactions, with no upfront commitment
• Commitment tiers: you commit to a level of usage in advance for a lower unit price
• Pay-as-you-go suits pilots and unpredictable workloads; commitment tiers suit steady, high-volume production
• Consumption billing means cost scales with adoption, so forecasting usage matters
• Start pay-as-you-go to learn real usage, then move to a commitment tier once the pattern is clear

Microsoft Learn resource: Microsoft Foundry documentation (opens in a new tab)