AB-100 cram sheet
Every flashcard fact for AB-100, condensed into one scannable page and grouped by exam domain. Built for the last hour before the exam — and for printing: use your browser's print or save-as-PDF and the navigation, sidebar, and controls drop away.
D1 · Plan 25–30%
| What is the role of an AI Center of Excellence? | Centralize AI governance decisions, responsible AI accountability, and regulatory monitoring across all AI projects. |
|---|---|
| What separates CAF AI maturity Level 2 from Level 3 in the skills and data readiness framework? | Level 2 = model selection, deployment and endpoint management, data cleaning; feasible use cases include custom analytical AI on Foundry Tools, a generative chat app without RAG, AutoML, and fine-tuning. Level 3 adds proficiency in prompt engineering, data chunking, query processing, and grounding data for indexing; it unlocks RAG apps in Foundry and training/deploying ML models. Level 4 adds infrastructure-level training on VMs, AKS, or Container Apps. |
| What four regional decisions does CAF AI Ready call out for production AI workloads? | Use at least two regions where availability and criticality justify the cost (load balance endpoints with API Management); verify AI service availability per region before deployment; evaluate regional quota limits and contact Azure support proactively for capacity above standard quotas; co-locate RAG data with the models to cut latency; and replicate fine-tuned models, RAG datasets, and training data to a secondary region for continuity. |
| In CAF AI Ready, how do you separate internet-facing from internal AI workloads, and where do AI resources live? | Create separate management groups for internet-facing ("online") and internal-only ("corporate") AI workloads, apply AI-specific Azure Policy to each (starting from Azure landing zone baseline policies), and deploy AI resources into workload-specific application landing zone subscriptions — not platform subscriptions — so they inherit workload governance and workload teams retain autonomy. |
| What four areas does the CAF AI Strategy phase define? | AI use cases, AI technology strategy, data strategy, and responsible AI strategy. |
| What does turning OFF "Allow ungrounded responses" do? | It blocks any response generated in a turn where the agent didn't call a knowledge source or tool, and answers are returned only when they include an in-text citation. Requires generative orchestration. It doesn't guarantee the model never blends in general knowledge, and it also disables follow-up questions. |
| Name the six phases of the Cloud Adoption Framework AI adoption process. | AI Strategy, AI Plan, AI Ready, Govern AI, Secure AI, and Manage AI. |
| How does Copilot Studio capacity enforcement work, and how do you avoid being blocked? | Capacity is pooled tenant-wide and enforced monthly; unused credits do not carry over. At 125% of prepaid capacity, custom agents are disabled (ongoing conversations finish) and admins are notified. Agent flow enforcement is separate — it blocks new agent flow runs while the agent keeps serving non-flow interactions. Remedies: reallocate environment capacity, buy more credits, or link an Azure subscription for pay-as-you-go, which exempts that environment from enforcement. |
| What are the three levels of agent architecture complexity, and which is the recommended default? | Direct model call (a single prompt, no agent logic) → single agent with tools → multiagent orchestration. Use the lowest level that reliably meets requirements. A single agent with tools is often the right default for enterprise use cases; set iteration limits to guard against infinite tool-call loops. Multiagent is justified by cross-domain scope, per-agent security boundaries, or parallel specialization. |
| What are connected agents in Foundry Agent Service, and what depth limit applies? | A main agent delegates to purpose-built subagents using natural-language routing, with no custom orchestrator or hardcoded logic. Maximum depth is 2 — a parent can have many sibling subagents, but subagents can't have their own subagents (exceeding it raises an Assistant Tool Call Depth Error). Citations from subagents aren't guaranteed. |
| Copilot connectors vs Power Platform connectors as knowledge — one-line rule? | Copilot connectors = index-then-answer: copy content into the Microsoft Graph semantic index for broad, citable, reusable knowledge. Power Platform connectors = live API bridge: real-time facts and transactions with no data movement. Most enterprises use both. |
| What are the headline Copilot Credit billing rates in Copilot Studio, and when are they zero-rated? | Classic answer 1, generative answer 2, agent action 5, tenant graph grounding for messages 10, agent flow actions 13 per 100 actions; text/generative AI tools bill at basic 1, standard 15, premium 100 per 10 responses. Employee-facing use in Copilot Chat, Teams, or SharePoint by a Microsoft 365 Copilot licensed user under that user's identity is zero-rated, within fair use limits. |
| What is Copilot Studio? | A low-code/no-code SaaS tool for building conversational agents and automation flows that connect to many data sources and publish across channels (web, Teams, mobile). |
| Which Dynamics 365 apps store data natively in Dataverse, and which need extra configuration? | Customer engagement apps (Sales, Customer Service, Human Resources, and so on) store data directly in Dataverse. Finance and operations apps currently require the Data Integrator to be configured to make their business data available in Dataverse — plan for it in any cross-app agent design. |
| When should you build a custom engine agent instead of a declarative agent? | When you need custom orchestration/complex workflows, your own or domain-specific models, proactive triggering, agent-to-agent collaboration, or availability outside Microsoft 365. |
| Which Microsoft Purview capability does CAF recommend to protect generative AI applications and data? | Microsoft Purview Data Security Posture Management (DSPM) for AI. |
| When should you build a declarative agent for Microsoft 365 Copilot? | When workflows stay inside Microsoft 365, you want fast/low-code setup, and you want to reuse Copilot's orchestrator and models, inheriting M365 security, compliance, and responsible AI. |
| What security constraint applies to Copilot Studio event triggers on autonomous agents? | Event triggers can only use the agent maker's credentials, and every authenticated action must work with those credentials for the agent to run unattended — so published agents may let users reach data through the maker's authorization. Triggers require generative orchestration, count toward Copilot Credit consumption, and can be blocked by admin data policies. |
| What is a Fabric data agent and what are its hard limits? | The conversational analytics component over governed OneLake data — up to 5 data sources (lakehouse, warehouse, semantic model, KQL database, mirrored DB, ontology, Microsoft Graph) using NL2SQL / NL2DAX / NL2KQL. Read-only; no unstructured files (.pdf/.docx/.txt); English only; responses capped at 25 rows and 25 columns; up to 100 example queries per source; can't change the LLM. |
| What are the three layers of Fabric IQ and the two core items that deliver them? | Unified data with OneLake (shortcuts, mirroring, catalog — a governed, multicloud source of truth and the distribution layer to Fabric, Foundry, and Copilot Studio), business intelligence with Power BI semantic models (measures, hierarchies, dimensions), and operational intelligence with ontology (entities, relationships, properties, rules, actions). The two core items are ontology (preview) and semantic model. |
| SFT vs RFT vs DPO vs distillation — when does each apply? | SFT: labeled prompt/completion pairs when there are finite correct ways to solve a task. RFT: complex or dynamic problems with many valid solution paths, scored incrementally by a grader. DPO: alignment to subjective tone/style preferences from binary preferred/non-preferred pairs, no reward model needed. Distillation: fine-tune a smaller model on a larger model's outputs to cut cost and latency. SFT then DPO can be stacked. |
| What is a Foundry IQ knowledge base and what does agentic retrieval add? | A configurable multi-source knowledge layer — knowledge sources (Blob Storage, SharePoint, OneLake, public web) plus retrieval parameters — that multiple agents can share. Agentic retrieval decomposes a question into subqueries, runs them in parallel, semantically reranks, and returns extractive answers with citations, while synchronizing ACLs, honoring Purview sensitivity labels, and running queries under the caller's Entra identity. |
| How do generative AI and analytical AI/ML differ in typical use cases? | Generative AI creates content (text, code, images) and boosts knowledge work; analytical AI/ML automates data-heavy tasks and produces predictive insights (e.g., forecasting). |
| What dimensions should you assess when reviewing grounding data? | Accuracy, relevance, timeliness, cleanliness, and availability. |
| What is the Fabric data agent governance/intent precedence order? | Organizational intent (tenant policies) > role-based intent (workspace governance and permissions) > developer intent (instructions, example queries, source config) > user intent (prompts). Higher layers always override lower ones, so policy beats developer instructions. |
| In Copilot Studio, how many SharePoint URLs and public websites can be used as knowledge in generative vs classic orchestration? | Generative: 25 SharePoint URLs and 25 websites (Dataverse and connector sources unlimited). Classic: 4 SharePoint URLs and 4 website URLs, 2 Dataverse sources with up to 15 tables each. Uploaded files don't count toward the 25-source generative search limit. |
| What is a maker-checker loop and what two controls must it have? | A group chat variant (also called evaluator-optimizer, generator-verifier, critic or reflection loop) where a maker agent produces output and a checker agent evaluates it against criteria, pushing it back with feedback until approval. It requires clear acceptance criteria so the checker makes consistent pass/fail decisions, and an iteration cap with defined fallback behavior — escalate to a human reviewer or return the best result with a quality warning. |
| Which interoperability standard does CAF recommend for cross-system AI data ingestion? | Model Context Protocol (MCP). |
| Which dimensions do Foundry model leaderboards compare, and how should an architect read them? | Quality index (averaged accuracy across reasoning, knowledge, QA, math, coding), safety (attack success rate — lower is better), performance (latency TTFT and throughput GTPS), and cost. Read them together: a high-quality but high-latency model may not suit real-time apps. Start from the scenario leaderboard when your use case maps to one, then evaluate the shortlist on your own data. |
| What is Microsoft Foundry model router? | A trained model that routes each prompt in real time to the most suitable underlying LLM, optimizing cost and latency while maintaining comparable quality through a single deployment. |
| What is the trade-off of marking a Copilot Studio knowledge source as an "Official source"? | Official sources are treated as trusted and used directly, with a distinctive indicator on responses — but the feature isn't compatible with generative orchestration, so the agent must run classic orchestration to use it. |
| Key limits for Microsoft 365 Copilot organizational prompts. | Up to 1,000 published prompts per tenant; bulk import capped at 5 MB and 100 prompts per file (UTF-8 CSV); up to 4 pinned prompts; ~3 hours to appear in the prompt lab; managed by an AI Administrator or Search Editor in the Microsoft 365 admin center. Field limits: Title 35, Display prompt 132, Prompt 8,000. |
| Sequential vs concurrent vs group chat vs handoff vs magentic — one line each on coordination and what to watch for. | Sequential: linear pipeline, deterministic order; early-stage failures propagate. Concurrent: parallel on the same input; needs conflict resolution and is resource-intensive. Group chat: shared thread with a chat manager; loops, keep to three or fewer agents. Handoff: dynamic delegation, one active agent at a time; infinite handoff loops and unpredictable routing. Magentic: manager agent builds and adapts a task ledger; slow to converge and stalls on ambiguous goals. |
| Give an example of a prebuilt agent and what it does. | The Sales Agent for Microsoft 365 Copilot automates lead management across Dynamics 365 or Salesforce, schedules meetings, and reaches out to customers. |
| What are the five Microsoft "best practices" bullets for prompt construction? | Be specific (restrict the operational space); be descriptive (use analogies); double down (repeat key instructions before and after primary content); order matters (recency bias applies even to few-shot example order); give the model an out (e.g. "respond with 'not found'"). |
| What is the Copilot Studio / AI hub prompt library? | A collection of predesigned prompt templates that encode prompt-engineering best practices, searchable by job type (Finance, HR, Legal, Sales, IT…) and task (Analyze, Classify, Create, Extract, Summarize). Makers copy and adapt them, then choose model, temperature, inputs (text/image/document), Text or JSON output, and optionally one Dataverse table. |
| Name five Microsoft-documented prompt engineering techniques for improving grounding and accuracy. | Provide grounding context; ask for inline citations near the statements they support; break the task down into steps; specify output structure (e.g. entity/relationship tuples); repeat instructions at the end to counteract recency bias. Also: prime the output with a cue, add clear syntax/separators, and give the model an "out". |
| Which external framework do Microsoft's six responsible AI principles align with? | The NIST AI Risk Management Framework. |
| What belongs in TCO for an AI agent versus the benefit side of an ROI analysis? | TCO: token consumption, storage, grounding connections, hosting, licensing. Benefits: reduced handle time, higher CSAT/retention, faster onboarding. |
| What is the key context-window limitation of model router? | The effective context window is limited by the smallest underlying model; use the model subset feature to select models that support larger contexts. |
| How do you ensure model router automatic failover works in a custom deployment? | Select a model subset with at least two models, since the subset also serves as the fallback set. |
| What are the three Microsoft Foundry model router routing modes? | Balanced (default; cost+quality), Quality (max accuracy, ignores cost), and Cost (max savings within a wider quality band). |
| When should you turn on "Tenant graph grounding with semantic search" for a Copilot Studio agent? | When not all end users have Microsoft 365 Copilot licenses but the tenant has at least one, and the agent grounds on SharePoint or Copilot connectors. Requires user authentication set to "Authenticate with Microsoft", carries extra Copilot Credit cost, and may slightly increase latency. |
| What are the three Microsoft AI consumption/service patterns? | SaaS (Copilots), PaaS (Foundry development platform), and infrastructure (VMs/AKS for bring-your-own-model). |
| When is a small language model (SLM) the right choice? | For narrow, well-defined tasks needing low latency, constrained compute, on-device or cost-sensitive deployment. |
| Guidance for temperature and top_p? | Temperature 0–2: low (~0.2) for focused, concrete output such as legal or contractual text; higher (~0.7) for creative output. Top_p controls randomness differently. General recommendation: alter one of the two at a time, not both. |
| What three categories help assess where AI agents fit? | Task automation, data analytics, and decision-making. |
D2 · Design 25–30%
| When adding an A2A connection in Copilot Studio, what URL do you enter and where does the name/description come from? | Enter the agent's communication endpoint URL — NOT the agent card URL. If a valid agent card exists at the standard /.well-known path, Copilot Studio auto-populates name and description from it; otherwise enter them manually (the description tells the main agent when to use the other agent). Authentication options: None, API key, or OAuth 2.0. |
|---|---|
| What metadata travels in a Copilot Studio A2A message payload? | A unique contextId, message IDs, locale, the full chat history (not just the latest utterance), and message content parts (text, tool calls, other metadata). The contextId is what allows context management across agents, with multiple tasks living inside a single context. |
| What is the Agent2Agent (A2A) protocol and what does it let a Copilot Studio agent do? | An open standard (Linux Foundation) defining a contract for agent-to-agent communication. It lets an orchestrator send whole tasks to external agents, pass rich structured metadata, and receive responses in a predictable format — so Copilot Studio delegates tasks rather than just calling APIs. Unlike a plain HTTP connector it supports multi-turn interactions and is interoperable across frameworks. |
| MCP vs A2A: who orchestrates, and how visible are the tools? | MCP: the host orchestrates which tools to invoke and synthesizes the final output — strong central control. A2A: the invoked agent uses its own chain-of-thought and orchestration, and its tools/APIs are opaque to the requesting agent. Use MCP for tool and data access; use A2A for cross-platform agent messaging where agents are or must be opaque to each other (e.g. two different engineering organizations). |
| What is the agent feed in a model-driven app, and what are its two sections? | A task-based collaboration surface at the top of the sitemap where users supervise agent work. 'Needs Attention' holds tasks logged by request_assistance or invoke_data_entry that await user input; 'Completed' holds everything else, including log_for_review items the agent finished autonomously. An Insights panel aggregates tasks completed by agents vs by users over 7/14/30 days. |
| What changes for the agent feed on 1 May 2026? | The agent feed will support only agents that use the Power Apps MCP server to create tasks. Agents not onboarded lose the feed entirely — it won't appear in the model-driven app. Onboarding = add the Power Apps MCP Server tool in Copilot Studio, set Maker-provided credentials for trigger-driven autonomous runs, update instructions to call the tools, then save and publish. |
| How is running an agent flow metered against Copilot Studio capacity? | From a topic: one Classic answer plus the agent flow actions executed. Via generative orchestration: one Autonomous action plus the actions. Runs from the flow designer or the agent's test chat consume nothing. When prepaid capacity is exhausted new runs are blocked (in-flight runs finish) unless pay-as-you-go is enabled; Microsoft 365 Copilot licensed users are unaffected. |
| When should an architect choose an agent flow over a generative agent? | When the work needs deterministic, author-defined, fixed-sequence automation (validate, look up, write back) rather than model-decided branching. The agent calls the flow when needed. |
| How are agents in SharePoint secured and governed? | Agents are .agent files, so file permissions decide who can open or edit them, and responses are always trimmed to each user's own permissions on the underlying sites and libraries. Access requires a Microsoft 365 Copilot licence or a pay-as-you-go billing policy scoped to a security group. Site owners can nominate one main agent per site. Admins use restricted access control, restricted content discovery (which removes agent features from a site without changing site access), Purview DLP, and the Copilot Control System agent inventory to block agents in Copilot Chat. |
| In the Power Apps AI hub, when do you choose a custom AI Builder model over a prebuilt one, and which model types are custom? | Use custom models when the data is unique to your business; use prebuilt models for scenarios common across businesses. Custom: document processing, prediction, object detection, plus custom category classification and entity extraction. Prebuilt: invoice, receipt, contract, ID and business card processing, text recognition, key phrase extraction, language detection, sentiment analysis, text translation, text generation. Reach them via Power Apps or Power Automate > … More > AI hub > AI models. |
| How do you let users ask questions of structured finance and operations data? | Add the data as a knowledge source: finance and operations virtual entities, or native Dataverse tables populated via a sync utility such as dual-write. Those sources can be added to Copilot for finance and operations apps (in-app sidecar), to custom Copilot Studio agents, or — via a published declarative agent — to Microsoft 365 Copilot. Answers respect the data the user has access to. |
| In classic orchestration, how are topics selected and when is knowledge used? | Topics are selected by matching the user's message to trigger phrases (via NLU). Knowledge is used only as a fallback when no topic matches, or when called explicitly from a topic. |
| What does the code-first path for generative pages give you that the maker portal does not? | Local development with GitHub Copilot CLI, Claude Code, or another codegen tool via the Power Platform Skills model-apps plugin (/genpage). Available worldwide in public clouds (the maker experience is limited to US, GB, AU, SG), gives access to newer frontier models, generates package.json and genpage.d.ts for IntelliSense, deploys with generative page PAC CLI commands, and handles localization automatically. Requires Node.js 18+ and an authenticated PAC CLI session. |
| What is the computer use tool in Copilot Studio for, and what is its key prerequisite? | It lets an agent operate websites and desktop apps via a virtual mouse/keyboard (vision + reasoning) to automate tasks where no API exists. It requires generative orchestration to be turned on. |
| What does enabling access control on a computer use tool actually enforce? | It blocks interactions with websites/apps not on the allow list. It does NOT prevent the model from opening them; actions on non-allowed sites simply fail. |
| How is computer use billed per step? | 5 Copilot Credits per step on a standard model, or 15 Copilot Credits per step on a premium model. |
| Where can computer use store the sign-in credentials it needs at runtime? | In Power Platform internal encrypted storage (no setup) or in a customer-provided Azure Key Vault referenced by the tool. |
| Name three computer use machine-hardening best practices. | Use a dedicated/isolated machine; apply least-privilege to the run account; and restrict web/app access with allow lists (trusted websites and specific desktop apps). |
| Copilot Studio offers two custom entity types. What are they and when do you use each? | Closed list entities — small, manageable value lists with simple labels, optional synonyms and smart matching. Regular expression (regex) entities — pattern matching for things like tracking IDs, licence numbers, credit cards or IP addresses; NLU and CLU use .NET regex syntax, NLU+ uses JavaScript syntax; use (?i) for case-insensitive matching and | to combine patterns. |
| List four capabilities of Copilot in Dynamics 365 Sales. | Record summarization, recent-changes catch-up, meeting preparation, and email assistance (also news updates and SharePoint information assistance). It only accesses data the signed-in user can see. |
| Name three DTMF design controls in a Copilot Studio voice agent. | 1) Termination key — none, * or #; optional and not counted in the input length. 2) Digit length — a fixed length or a range; wrong lengths cause a reprompt. 3) DTMF caching — key presses are buffered for the next prompt so experienced callers can key ahead; cached keys are dropped if the next turn disallows barge-in. Unmapped global keys trigger the 'Unknown dial pad press' system topic. |
| What are the three tool categories of the Dynamics 365 ERP MCP server, and how is the agent scoped? | Data tools (OData/SQL CRUD over data entities — most efficient for standard operations), Form tools (open forms, set control values, click controls, save — for non-CRUD business logic and run-time calculated values, via server APIs not a browser), and Action tools (api_find_actions / api_invoke_action against developer-written ICustomAPI classes). Scoping is dynamic: every call returns only the menu items, entities, actions, and view-model fields the authenticated identity's security role can access; calls outside the role are rejected. Allowed MCP clients (Entra client IDs) controls which platforms may connect. |
| How do you add company knowledge to the finance and operations in-app help sidecar? | In Copilot Studio on the paired Dataverse environment, open the 'Copilot for finance and operations apps' agent → Knowledge tab → Add knowledge → upload files (PDF, RTF, Word) → wait for Status = Ready → Test → Publish. For SharePoint or other source types you must author your own topic in that agent. Dataverse virtual entities published from finance and operations apps are NOT supported as knowledge sources here. AI general knowledge (model + Bing) is off by default and is only consulted after custom sources are exhausted. |
| What does generative help and guidance in finance and operations apps ground on by default, and how does it answer? | By default it grounds only on published Microsoft product documentation for finance and operations apps, honouring each user's language preference. The pipeline is: query rewrite → document search using the Bing index restricted to learn.microsoft.com → response generation with citations. It is delivered as the sidecar, the primary Copilot interface in these apps. |
| Which system topic provides the knowledge fallback (generative answers) in a classic-orchestration agent, and what happens to it under generative orchestration? | The Conversational boosting system topic. Under generative orchestration the agent does not use it for knowledge search, so customizations to it are ignored. |
| What transfers when you move a generative page between environments in a solution? | Generative pages are solution-aware (they appear as UX Agent Project rows) and are pulled in via the app's sitemap dependency. Only the first prompt and the last published code move; the full agent conversation and revision history stay in the source environment. Pages created during preview must be loaded in the designer once to trigger migration to the solution-aware data model. |
| What is a generative page in a model-driven app, and what are its hard limits? | An AI-generated page: you describe it in natural language (optionally attaching an image), and the app agent produces React/TypeScript code plus the business logic. Limits: Dataverse tables only, max 6 tables per page, prompt max 50,000 characters, no collaborative editing (one maker at a time), a fixed list of supported data types, and model-driven apps only — not canvas. |
| Glossary vs synonyms when designing business terms for Copilot in Dynamics 365 Sales. | Both are configured on the Copilot in Dynamics 365 Sales agent under Knowledge > SalesSpecificQnA, then published. Synonyms = alternative names for specific columns in your tables (e.g. 'opportunity revenue' → the custom revenue column). Glossary = business terms plus descriptions, so Copilot can interpret organisation-specific concepts and guidance for words like 'top', 'best', or 'we'. Prerequisite: the Copilot Studio Author role. |
| Declarative vs custom engine agent: what is the core difference? | Declarative agents use Copilot's built-in orchestrator and model (inheriting M365 compliance/RAI, low-code, no extra hosting). Custom engine agents bring your own orchestrator and models, support proactive workflows and external channels, and you own compliance/hosting. |
| Which agent type supports proactive messaging and agent-to-agent collaboration? | Custom engine agents. Declarative agents rely on user-initiated interactions and are designed for individual use. |
| What is the recommended way to connect an existing MCP server to a Copilot Studio agent, and which transport is supported? | Use the MCP onboarding wizard on the agent's Tools page. Copilot Studio supports the Streamable transport; SSE is deprecated and unsupported after August 2025. |
| How is access to MCP server tools governed in Power Platform? | MCP connectivity relies on Power Platform connectors, so any DLP/data policy regulating connectors also governs the MCP server and its tools. |
| Choose between conversational language understanding (trigger phrases) and generative orchestration: which handles unanticipated phrasing without enumerating utterances? | Generative orchestration — it selects components by description and natural language, so authors don't have to predict every way a user phrases a request. |
| What are the limits and the recognition rule for a Question node set to 'One of multiple entities'? | Up to five entities, no external entities, and only one entity of each type. It recognises a single entity per turn, and if several are present it takes the first entity in the node's configured list. The result variable is a record with one element per configured entity — branch with 'is not Blank'. |
| What are open list (dynamic inline) entities used for, and what are their constraints? | Values are populated at runtime from an external source (Dataverse, Excel, a database or Power Fx) so the recognisable set can vary per user or session — account names, payees, prescriptions. Most common with classic voice agents. Back them with a table variable of no more than 100 entries; use records with DisplayName plus a Synonyms array to support synonyms. Voice agents can auto-assign DTMF keys to the options. |
| What is the default orchestration mode for newly created Copilot Studio agents, and what can it do that classic cannot? | Generative orchestration is the default. Unlike classic, it can select and chain a combination of topics, tools, other agents, and knowledge, and auto-generate questions to fill missing inputs. |
| Name the three Power Apps MCP server tools and when to use each. | log_for_review — passive oversight of high-confidence/low-risk work the agent already completed (lands in Completed). request_assistance — the agent is blocked or the stakes are high; creates a Needs attention task and waits asynchronously for the human, then resumes via callback. invoke_data_entry — extract structured records from unstructured input (email, PDF, docx, images); the record is NEVER created without user review and acceptance. |
| Name the five pillars of the Power Platform Well-Architected Framework. | Reliability, Security, Operational Excellence, Performance Efficiency, and Experience Optimization. |
| Which Power Platform Well-Architected pillar is unique compared to Azure WAF, and what does it cover? | Experience Optimization — creating meaningful, useful experiences that ensure successful business outcomes. |
| What is proactive slot filling, and how do you force a Question node to ask anyway? | The agent listens across turns and fills multiple entity slots from a single utterance ("hiking boots under $100" fills both category and budget), skipping the corresponding Question nodes. To always prompt, open the node's Question behavior panel and set Skip question to Ask every time. |
| What is a prompt action (AI prompt) in Copilot Studio used for? | A reusable generative operation with defined inputs/outputs (e.g., summarize or extract) that topics and agents can invoke. |
| What is the Sales agent in Microsoft 365 Copilot relative to Copilot in D365 Sales? | It is the evolution of the seller Copilot experience and will become the preferred way to access Copilot assistance across Microsoft 365 and Dynamics 365 Sales. |
| What does turning on Smart matching for a closed list entity do? | It applies the language model's fuzzy logic to the entity's values: the agent autocorrects misspellings and expands matching semantically (the docs' example matches "softball" to "baseball"). It is separate from synonyms, which you define explicitly per value. |
| In generative orchestration, what is the single most important factor for selecting a topic, tool, agent, or knowledge source? | Its description. Clear, specific, non-overlapping descriptions let the orchestrator match user intent accurately; overlapping descriptions can cause multiple topics to fire. |
| How many trigger phrases does Copilot Studio recommend per topic, and which orchestration mode actually uses them? | 5 to 10 short trigger phrases (short phrases beat long sentences). Classic orchestration matches user input against trigger phrases; generative orchestration selects topics from their descriptions instead. Phrases can be uploaded or downloaded as a text file (max 3 MB, one per line, append or replace). |
| Contrast basic voice mode and streaming mode for Copilot Studio voice agents. | Basic voice mode: Speech → Text → NLU/NLU+ → classic orchestration → Speech. Pick it for custom or neural voices, fine-grained recognition control, DTMF-heavy flows and minimum cost — but it supports classic orchestration only. Streaming mode: one real-time model handles audio in and audio out, giving ultra-low latency, natural turn-taking and strong mixed-language handling, at the cost of fewer voices, fewer customisation points and shallower reasoning depth. |
| What are the five key strategies in Power Platform Well-Architected recommendation XO:10 (conversational user experience)? | Explain the AI's capabilities up front; understand the user's intent; optimise how the AI interprets input (handle both structured commands and open-ended language, break complex requests into steps); guide the user through goal-oriented conversation flows that handle interruptions; and design fallback mechanisms — clarifying questions, contextual follow-ups, alternative suggestions, and escalation after repeated misunderstandings. |
D3 · Deploy 40–45%
| In a Dataverse environment, what governs access to model-driven apps and components? | Dataverse security roles take over and apply to all model-driven apps and components. |
|---|---|
| Which evaluators does Microsoft recommend combining for agent applications? | Tool Call Accuracy, Task Adherence, Intent Resolution, Rubric, and content safety — with risk and safety evaluators (hate and unfairness, sexual, violence, self-harm) added for every application. |
| When a Foundry Hosted agent is deployed, what identity does it receive? | The service provisions compute, assigns a dedicated Microsoft Entra ID (agent identity), and exposes a dedicated endpoint. Permissions must be granted to that identity — they are not inherited from the project identity. |
| What is the ALM mechanism for moving agents/flows/tables across environments, and what must each environment include? | Solutions (export/import). Every participating environment must include a Dataverse database. |
| Managed vs unmanaged solutions: where is each used? | Unmanaged in development (editable); managed in test/production (locked, no direct edits). Source control is the source of truth. |
| Which CI/CD platform does Microsoft cite for automating Power Platform build/test/deploy? | Azure DevOps, which can also be combined with in-product pipelines. |
| For generative-orchestration agents (no classic topics), which analytics replace topic analytics? | Conversation outcomes (Resolved/Escalated/Abandoned) and themes on the Analytics page. |
| How long are Copilot Studio session details and transcripts retained vs aggregate analytics? | Session detail and transcripts: last 28 days. Aggregate analytics data: up to 360 days. Times are in UTC. |
| Where do Copilot Studio maker audit logs surface, and what adds real-time alerting? | Microsoft Purview for maker audit logs; Microsoft Sentinel for monitoring and alerts on agent activities. |
| Where is comprehensive Power Apps activity logging integrated for compliance/SIEM? | The Office Security and Compliance Center, with an API consumed by SIEM vendors, covering Dataverse and Microsoft 365. |
| What Microsoft Entra objects underpin agent identities, and who is accountable for one? | Agent identity blueprint, blueprint principal, agent identity, and agent user. Copilot Studio creates a Microsoft Copilot Studio agent identity blueprint in the tenant and an agent identity per agent (deleted with the agent). A human *sponsor* is accountable for lifecycle and access; sponsorship transfers to the sponsor's manager if they leave. |
| Which service connection types do Power Platform Build Tools support for Azure DevOps? | Service principal via workload identity federation (recommended, and required in practice for MFA-enforced tenants), service principal with client secret, and username/password (no MFA support). Build Tools 2.0 is Power Platform CLI based; task versions can't be mixed in one pipeline. |
| Continuous evaluation versus scheduled evaluation in Foundry — what is the difference? | Continuous evaluation scores sampled live production responses at a set sampling rate. Scheduled evaluation runs a chosen evaluation template against a fixed test dataset on a schedule to detect system drift against a stable baseline. |
| What are the Copilot Studio enforcement thresholds for prepaid capacity? | General overage enforcement triggers at 125% of prepaid tenant capacity: custom agents are disabled (ongoing conversations aren't cut off) until capacity increases or resets. Agent flow enforcement is separate — when prepaid capacity is fully consumed, new agent flow runs are blocked while the agent keeps serving non-flow interactions. |
| List the Copilot Studio agent evaluation test methods. | General quality, Compare meaning, Tool use, Keyword match, Text similarity, Exact match, and Custom. All except General quality require expected answers, keywords, or expected tools; a case missing them returns Invalid. |
| Which four operational signals does the Foundry Agent Monitoring Dashboard surface, and what thresholds does the guidance flag? | Token usage, latency (above ~10 seconds suggests throttling, complex tool calls, or network issues), run success rate (below 95% warrants investigation), and evaluation metrics from sampled outputs — plus red teaming results when scheduled scans are enabled. |
| Which control keeps Copilot Studio generative AI data inside the US? | Disable data movement across geographic locations for generative AI features (admin setting). |
| Who owns solution objects deployed by a Power Platform pipeline? | The deploying identity. For standard deployments that's the requesting maker; for delegated deployments it's the service principal or pipeline stage owner. Deploy-time sharing (security roles, canvas apps, cloud flows) applies only on an object's first deployment — manage access afterwards through security groups. |
| What can you NOT do with unsaved (draft) changes to a Foundry agent? | You can test them in the agents playground, but you cannot view conversation history, monitor performance, or run full evaluations — and unsaved changes are lost when you leave the portal. |
| How does an architect give a customer control over encryption keys for a Copilot Studio environment? | Enable customer-managed encryption keys (CMK) for the environment. |
| Why do architects restrict the default environment, and how do they route makers elsewhere? | All tenant users get Environment Maker in the default environment; use environment routing to send makers to a safe build environment. |
| What are the two evaluation levels for Foundry evaluators, and what is the constraint? | turn (individual responses, the default) and conversation (whole multi-turn interactions). All evaluators in a single run must support the level that is set — incompatible levels cannot be mixed. |
| Name the three stages of AI application lifecycle evaluation in Microsoft Foundry. | Base model selection (benchmarks), pre-production evaluation (datasets, edge cases, AI red teaming agent), and post-production monitoring (operational metrics, continuous evaluation, scheduled evaluation, scheduled red teaming, Azure Monitor alerts). |
| What structured feedback mechanism does Microsoft suggest on AI-generated content? | Buttons letting users flag content as inaccurate, harmful, or incomplete, providing a signal for analysis and improvement. |
| Name the three gated extensions in pipelines for Power Platform and their triggers. | Pre-export step required (OnDeploymentRequested → UpdatePreExportStepStatus); Is delegated deployment (OnApprovalStarted / OnDeploymentCompleted → UpdateApprovalStatus, called by the delegate's own connection); Predeployment step required (OnPreDeploymentStarted / OnPreDeploymentCompleted → UpdatePreDeploymentStepStatus). |
| Where is Dataverse Git integration intended to be used? | In developer environments, syncing unmanaged solutions to an Azure DevOps Git repository — not in test or production, where you deploy built managed artifacts via pipelines. Benefits include source control as the source of truth, SDLC auditing, short-lived rehydratable dev environments, and fusion team collaboration. |
| What does Power Platform DLP do, and what classifications exist? | Classifies connectors as Business / Non-Business (can't be mixed in one app) or Blocked (unusable); prevents data exfiltration. |
| Which Microsoft asset accelerates tenant-wide Power Platform governance and adoption monitoring? | The Center of Excellence (CoE) Starter Kit, plus admin connectors and PowerShell cmdlets. |
| In pipelines for Power Platform, what type of environment should the pipelines host be, and must it be a managed environment? | The host should be a production environment, but it does not itself have to be a managed environment. Developer environments used for dev/test also don't. All other environments used in pipelines must be managed environments. An environment can belong to only one host, and the host can't double as a development environment for the same solution. |
| What happens to an AI Builder model when its solution is imported into another environment? | Only the published version of the model is installed. You cannot train a new version of an imported document processing, object detection, or entity extraction model because the training dataset is not part of the solution — create a new model in the environment that holds the data instead. |
| Which test types does the Power CAT Copilot Studio Kit support? | Response match, attachments (such as Adaptive Cards), topic match (needs Dataverse), generative answers (needs AI Builder and Application Insights), multi-turn (ordered cases in one conversation context), and plan validation (expected tools in a generative orchestration plan). |
| How do unmanaged and managed layers interact in Dataverse? | There is one unmanaged layer holding all imported unmanaged solutions and ad-hoc customisations, and it sits above all managed layers. Managed layers stack in install order (last installed on top), with the system solution at the base. Uninstalling a managed solution reveals the one below it. |
| What does Customer Lockbox provide in Copilot Studio? | Just-in-time, explicit customer approval before Microsoft engineers can access customer data. |
| What measurement sequence does Microsoft recommend for harms? | Start manual (priority issues, spot-checks), then automate for large-scale coverage and ongoing regression monitoring. |
| Which Dataverse component types are merged, and what happens to the rest? | Only model-driven app, form, and site map components are merged. Every other component type uses 'top wins', where the top layer — base, patch, or a staged (_Upgrade) pending upgrade — defines runtime behaviour. Applying an upgrade flattens the layers into a new base. |
| List the four Responsible AI mitigation layers and an example of each. | Model (fine-tuning), Safety system (Guardrails/content filters), Application (metaprompt + UX), Positioning (transparency/docs). |
| How does Azure Machine Learning provide an audit trail for custom models? | The model registry versions models by name (incrementing on re-registration) with metadata tags; job history snapshots the code, data, and compute used to train; lineage logs who published a model, why it changed, and when it was deployed; and Event Grid publishes lifecycle events such as model registration, deployment, and data drift for automation and alerting. |
| What prerequisite enables DAU/MAU active-user metrics in Copilot Studio analytics? | The agent must require authentication (Settings > Security > Authentication); anonymous users can't be counted as unique. |
| Which Copilot Studio items are NOT solution-aware and need post-deployment steps downstream? | Azure Application Insights settings, manual authentication settings, Direct Line / web channel security settings, deployed channels, and sharing (with makers or end users). Put these on the release checklist — a clean solution import will not warn you. |
| Which three plans does Microsoft require for the Operate stage of a generative AI deployment? | Phased delivery plan, incident response plan, and rollback plan (plus telemetry and feedback channels). |
| Why can't a maker slip a change past QA in a Power Platform pipeline? | The solution is exported once, when the deployment request is submitted. Both managed and unmanaged artifacts are stored in the pipelines host, the system prevents tampering, and the same managed artifact per version must traverse the stages in sequential order — no re-export happens for later stages. |
| What stops a customisation from bypassing the QA stage of a Power Platform pipeline? | The solution is exported once when the deployment is requested, and that same artifact is reused for every stage in sequential order. The system does not re-export and prevents tampering with the exported artifact. |
| What four elements make an effective Copilot prompt? | Goal (what you need), context (why you need it or how you will use it), source (which data, files, or plugins to use), and expectations (the format or audience for the response). Prefer positive instructions and iterate on follow-up prompts. |
| Which application-level techniques reduce open-ended or manipulated responses? | Structure inputs/outputs, limit input/output length, use prepared/predetermined responses, and metaprompt tuning. |
| Which criteria does the Copilot Studio General quality method score, and what happens if one is missed? | Relevance, groundedness, completeness, and abstention. A response must meet all of them to count as high quality; if any one criterion is unmet the response is flagged for improvement and scores lower. |
| What is the purpose of red teaming and stress-testing in the Identify stage? | Intentionally probe the system to find its limitations, risk surface, and vulnerabilities, producing a prioritized list of harms. |
| Name the four stages of the Microsoft Responsible AI lifecycle. | Identify, Measure, Mitigate, Operate (aligned with the NIST AI Risk Management Framework). |
| What happens to inference calls after a Foundry model version is retired, and what notice do you get? | All inference returns 410 Gone. GA models get at least 60 days' retirement notice (email to subscription owners with active deployments, plus Azure Service Health advisories); preview models get at least 30 days and are force-upgraded or terminated. Retirement dates aren't extendable. |
| What is the Regression suite automation tool (RSAT) and what does it need? | It converts finance and operations Task recorder recordings into automated tests with no code; parameters live in Excel decoupled from the steps. It needs a Windows machine with Excel, an Azure DevOps project, and an Azure DevOps Test Manager or Test Plans licence. Version 2.3+ supports parallel execution. |
| What does the SharePoint sensitivity-label feature show in a Copilot Studio agent? | The highest sensitivity label applied to sources used in the response, plus per-reference labels in chat. |
| What two roles let a reviewer see Analytics plus transcript content without editing the agent? | Analytics Viewer sharing role + Bot Transcript Viewer security role. Analytics Viewer can only be shared with individuals, not groups. |
| What do the status values 10, 20, and 30 mean in pipeline extension actions? | 10 = pending (set by the system), 20 = complete the step, 30 = reject — the deployment doesn't proceed and its status is set to failed. Maker-facing comments explain a rejection; admin-only detail goes in the *Properties parameters. |
| What are the four Application Insights logging toggles for a Copilot Studio agent? | Enable logging (incoming/outgoing messages and events); Log conversation details (user ID, user name, message text — and, with OpenTelemetry tracing on, tool input arguments and outputs in spans); Log sensitive Activity properties; Node execution events (an event per topic node executed). |
| What are the size and retention limits for a Copilot Studio test set? | Up to 100 test cases per test set, each imported question up to 1,000 characters, import files in CSV or TXT, generation source files up to 5 MB — and test results are retained in Copilot Studio for 89 days (export to CSV to keep them longer). |
| Name the testing types in the Copilot Studio agent testing strategy. | Development time (unit), core scenario (happy path), knowledge, regression, adversarial, performance and load, security and compliance, and accessibility and UX — governed by shift-left, iterate and refine, test before production, and automate where possible. |
| What are the three versionUpgradeOption values for a Foundry Standard deployment? | OnceNewDefaultVersionAvailable (upgrade when a new default is set), OnceCurrentVersionExpired (upgrade only at retirement), and NoAutoUpgrade (never — the deployment stops working at retirement). Provisioned deployments are never auto-upgraded and must be migrated manually. |
| Why must a Foundry prompt-based agent change be saved as a new version rather than edited in place? | Each saved version is immutable — modifying an existing version requires saving a new one. Immutability preserves version integrity, prevents accidental overwrites, and enables directing requests at a specific version for controlled rollout and rollback. |
| How do you get a built-in monitoring dashboard for a Copilot Studio agent in Application Insights? | Open the Application Insights resource → Monitoring → Workbooks → the Copilot Studio Dashboard from the gallery. It surfaces total conversations, latency, exceptions, tool usage, and topic analytics, and it's editable and shareable (viewers need at least Reader on the resource). |
Use it as a check, not a first pass
If a line here is the first time you're meeting a fact, this sheet is doing the wrong job for you — go back to the domain notes and the quizzes. It works best as a rapid confirmation that recall is already automatic.