AB-100 — Microsoft Certified: Agentic AI Business Solutions Architect Cheat Sheet

Compact AB-100 Cheat sheet for designing secure, governed, business-aligned agentic AI solutions with Microsoft technologies.

Use the tables for a quick pre-exam check. Expand a topic’s notes for explanations, examples, and additional distinctions.

Scope and study context
ItemReference
Vendor/providerMicrosoft
Certification titleMicrosoft Certified: Agentic AI Business Solutions Architect (AB-100)
Exam codeAB-100
Candidate mindsetTranslate business outcomes into secure, governed, measurable agentic AI solutions across Microsoft services.
Common exam angleChoose the right architecture, Microsoft service, control boundary, data grounding pattern, evaluation approach, and operational model.

AB-100 scenarios are likely to reward architecture decisions more than feature memorization. Think in this order:

  1. Business objective: What outcome, user group, process, and success metric?
  2. Agent fit: Should the solution be an agent, copilot extension, workflow automation, or standard app?
  3. Grounding: What data is authoritative, fresh, permission-trimmed, and auditable?
  4. Actions: What tools, APIs, connectors, and approvals are needed?
  5. Safety and governance: What can the agent do, for whom, under what policy?
  6. Evaluation: How will quality, risk, and business value be measured before and after release?
  7. Operations: How will the solution be monitored, improved, and controlled over time?

The exam mindset is architectural: you are not just choosing an AI feature. You are translating a business outcome into a secure, governed, measurable agentic solution that can interact with people, data, tools, and enterprise processes.

After this Cheat Sheet, move into IT Mastery practice with original practice questions. For each question:

  1. Identify the business goal before reading the answers.
  2. Mark the key constraint: security, cost, governance, autonomy, data freshness, or user workflow.
  3. Eliminate answers that rely only on prompting for security or compliance.
  4. Prefer the least complex architecture that satisfies all requirements.
  5. Review the detailed explanations for both correct and incorrect choices.
  6. Use topic drills for weak areas such as grounding, tool execution, governance, or evaluation.

Core Agentic AI Concepts

ConceptExam-ready meaningHigh-yield distinction
AgentAI system that can reason over goals, use tools, retrieve context, and take steps toward a task.More autonomous than a single prompt-response chatbot.
CopilotUser-facing AI assistant, often embedded in a Microsoft experience.Usually assists a human; may be extended with plugins, connectors, or agents.
Tool useAgent calls APIs, functions, connectors, workflows, or external systems.Requires strict authorization, input validation, and audit logging.
GroundingSupplying the model with trusted enterprise data.Reduces hallucination but does not eliminate need for evaluation.
RAGRetrieval-augmented generation: retrieve relevant content, then generate an answer grounded in that content.Good for knowledge-heavy scenarios; not enough for transactional actions.
OrchestrationLogic that plans steps, calls tools, manages state, and handles fallback.Can be low-code, pro-code, or hybrid.
Human in the loopHuman review, approval, or escalation before or after agent action.Essential for high-impact, irreversible, regulated, or ambiguous decisions.
GuardrailsPolicies, constraints, filters, prompt instructions, validation, and monitoring.Must be layered; prompt instructions alone are weak controls.
EvaluationTesting agent quality, safety, grounding, tool accuracy, and business outcomes.Should happen before release and continuously after deployment.
Responsible AIDesign discipline for fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.Architecture concern, not just a policy document.

Microsoft Service Selection Matrix

RequirementPreferWhyWatch for
Build low-code business agents for Teams, Microsoft 365, customer service, or internal workflowsMicrosoft Copilot StudioFast agent creation, topics, actions, connectors, governance through Power Platform ecosystem.Do not choose pro-code services when low-code connectors and governance meet the requirement.
Extend Microsoft 365 Copilot with enterprise data or actionsCopilot Studio, Microsoft Graph connectors, plugins/actionsKeeps experience close to Microsoft 365 users and permission-aware content.Respect existing Microsoft 365 and Graph permissions.
Build custom, pro-code agentic apps with fine-grained orchestrationAzure AI Foundry, Azure AI Agent Service, Azure OpenAI Service, Semantic KernelMore control over model selection, tools, retrieval, orchestration, app integration, and DevOps.Requires deeper engineering, security, and operations design.
Add enterprise search grounding over documents and structured indexesAzure AI SearchIndexing, hybrid/vector search, semantic ranking capabilities, security filtering patterns.Poor chunking, stale indexes, or missing ACL trimming cause wrong answers or data exposure.
Use Microsoft 365 content as knowledge sourceMicrosoft Graph connectors, Microsoft 365 data access patternsAligns with Microsoft 365 content and identity.Avoid bypassing user permissions by copying broad data into an unprotected index.
Store business records for Power Platform agentsMicrosoft DataverseBusiness data model, security roles, Power Platform integration.Model data and roles intentionally; do not treat it as a simple file store.
Orchestrate business process approvalsPower Automate or Azure Logic AppsWorkflow, connectors, approvals, integration triggers.Agent should not silently perform high-risk actions when approval is required.
Expose enterprise APIs safely to agentsAzure API ManagementCentral API gateway, policies, throttling, transformation, auth enforcement.Do not let agents call raw backend APIs without contracts and controls.
Run custom tool logicAzure Functions, Azure Container Apps, or app servicesEncapsulates action logic behind secure interfaces.Validate all model-provided inputs; assume tool arguments can be unsafe.
Protect secrets and keysAzure Key Vault, managed identitiesCentral secret management and identity-based access.Avoid hard-coded secrets in prompts, code, environment files, or connectors.
Monitor apps and agentsAzure Monitor, Application Insights, service-specific analyticsTelemetry, traces, failures, latency, tool-call behavior.Log enough for diagnosis without storing sensitive prompts unnecessarily.
Govern data, classification, and compliance controlsMicrosoft PurviewData governance, sensitivity labels, data loss prevention, audit and risk controls.Governance must cover data sources, retrieval stores, outputs, and user access.
Secure cloud posture and workloadsMicrosoft Defender for Cloud, Microsoft Sentinel where applicableThreat detection, posture management, security operations.Agentic systems expand attack surface through tools and data paths.

Solution Pattern Selection

PatternChoose whenAvoid whenKey controls
FAQ/knowledge copilotUsers need answers from trusted content.User needs complex action execution or multi-step planning.RAG, citations, content freshness, permission trimming.
Task assistantUser remains in control and agent drafts, summarizes, or recommends.Process requires unattended execution.Clear UX, confidence display, review before submission.
Transactional agentAgent updates systems, creates tickets, sends messages, or executes workflows.Actions are irreversible and no approval path exists.Tool authorization, input validation, audit trail, human approval.
Autonomous workflow agentAgent decomposes work and executes multi-step processes.Risk, ambiguity, or compliance requires human decisioning.Bounded scope, state machine, rollback, exception handling, continuous monitoring.
Multi-agent systemSpecialized agents handle planning, retrieval, domain tasks, or review.A single agent or deterministic workflow is sufficient.Clear responsibility boundaries, message contracts, loop prevention.
Copilot extensionUsers already work in Microsoft 365 or Teams and need contextual actions.Standalone app UX or custom orchestration is required.Graph permissions, connector governance, tenant policies.
Deterministic workflowRules are stable and explainability is more important than flexible reasoning.Inputs are unstructured or require language understanding.Use Power Automate, Logic Apps, or standard app logic before adding an LLM.

Reference Architecture

    flowchart LR
	    U[User or business process] --> UX[Copilot, Teams, app, or workflow]
	    UX --> ORCH[Agent orchestration]
	    ORCH --> LLM[Model endpoint]
	    ORCH --> RET[Retrieval layer]
	    ORCH --> TOOLS[Tools and actions]
	    RET --> DATA[Enterprise data sources]
	    TOOLS --> API[APIs, connectors, workflows]
	    API --> SYS[Systems of record]
	    ORCH --> EVAL[Evaluation and guardrails]
	    ORCH --> OBS[Telemetry and audit]
	    SEC[Identity, access, policy, governance] --> UX
	    SEC --> ORCH
	    SEC --> RET
	    SEC --> TOOLS
Notes and examples
LayerArchitect decisionsCommon traps
User experienceWhere users work, what they can see, how approvals appear, fallback route.Building a separate app when Teams, Microsoft 365, or existing business app integration is better.
OrchestrationLow-code vs pro-code, planning strategy, tool routing, memory/state, retries.Letting the model control everything without deterministic boundaries.
Model accessModel family, deployment region, prompt design, token budget, latency target.Assuming the largest model is always best.
RetrievalIndex design, chunking, metadata, ranking, security filters, freshness.Loading all content into prompts or using one global index without ACLs.
Tools/actionsAPI contracts, schemas, validation, idempotency, approval gates.Exposing broad admin APIs directly to an agent.
Identity/securityUser-delegated access vs app-only access, managed identity, RBAC, secrets.Using a privileged service account for all users.
GovernanceData classification, DLP, audit, retention, responsible AI review.Treating prompts and generated outputs as outside governance.
OperationsMonitoring, evaluation, feedback loops, incident response, cost tracking.No way to reproduce or diagnose bad agent behavior.

Quick review: architecture checklist

Before answering an AB-100 scenario, mentally check:

  • Outcome: Is the business problem clearly defined?
  • Users: Who interacts with the agent and through which channel?
  • Data: What sources ground the response?
  • Permissions: Does the user have rights to the data and actions?
  • Actions: What systems can the agent modify?
  • Autonomy: Is approval required?
  • Safety: What can go wrong, and what prevents it?
  • Evaluation: How is correctness and safety measured?
  • Operations: How is the agent monitored and updated?
  • Governance: Who owns the agent, data, tools, and lifecycle?

Requirements-to-Architecture Decision Table

If the scenario says…Architectural implication
“Answers must reflect the user’s permissions”Use permission-trimmed retrieval; avoid broad shared indexes unless ACLs are enforced.
“Users need citations or source links”Use RAG with document references and answer-generation constraints.
“The agent must update records”Expose constrained tools/actions; add validation, authorization, audit, and rollback or compensation.
“The business process requires manager approval”Insert human approval through Power Automate, Logic Apps, or app workflow before final action.
“Data changes frequently”Prefer live API retrieval or frequent index refresh; document freshness strategy.
“Highly sensitive data is involved”Apply least privilege, data classification, encryption, DLP, private networking where appropriate, and restricted logging.
“Nontechnical business teams will maintain the agent”Favor Copilot Studio and Power Platform governance if requirements fit.
“Custom orchestration and complex integration are required”Favor Azure AI Foundry, Azure AI Agent Service, Semantic Kernel, and pro-code services.
“The agent must scale across departments”Design reusable tools, standardized evaluation, environment strategy, governance, and lifecycle management.
“The output is used for consequential decisions”Require human review, explainability, documented limitations, bias/safety testing, and auditability.
“The agent needs to call many enterprise systems”Use API Management, connector strategy, identity design, throttling, error handling, and integration monitoring.
“The solution must be supportable by IT operations”Include telemetry, alerts, runbooks, versioning, incident process, and ownership model.

Agent Scope and Autonomy Controls

Autonomy levelDescriptionSuitable examplesRequired controls
InformAgent answers questions only.Policy Q&A, knowledge search, onboarding help.Grounding, citations, feedback, content safety.
DraftAgent prepares content but user submits.Email draft, case summary, proposal outline.User review, clear generated-content labeling.
RecommendAgent proposes a decision or next best action.Support routing, risk triage, sales recommendation.Explanation, confidence cues, human accountability.
Act with approvalAgent prepares an action; human approves.Refund, record update, external message.Approval workflow, audit, action preview.
Act within boundsAgent executes low-risk actions under predefined constraints.Create ticket, schedule internal meeting, update noncritical metadata.Policy constraints, validation, monitoring, rollback.
AutonomousAgent plans and executes multi-step workflows with minimal human input.Narrow, repeatable operational processes.Strict scope, kill switch, incident response, continuous evaluation.
Notes and examples

High-yield exam trap: Do not equate “agentic” with “fully autonomous.” Many business solutions should use limited autonomy with human approval.

Data Grounding and RAG Reference

Design areaGood choicePoor choice
Source selectionAuthoritative systems of record, curated knowledge bases, approved Microsoft 365 content.Random exports, stale files, duplicated uncontrolled content.
IndexingChunk by meaning, preserve metadata, include source URI, owner, sensitivity, timestamps.Huge chunks, missing metadata, no source traceability.
RetrievalCombine keyword, vector, semantic, filters, and reranking where useful.Rely on embeddings alone for all queries.
SecurityEnforce user or group permissions at retrieval time.Retrieve everything, then ask the model not to reveal restricted content.
FreshnessRefresh indexes based on business need or query live APIs for volatile data.Treat static snapshots as current.
AnsweringRequire grounded answers, citations, and “I don’t know” behavior when evidence is missing.Force an answer even when retrieval confidence is low.
EvaluationTest groundedness, relevance, citation accuracy, and refusal behavior.Test only with happy-path questions.
Notes and examples

RAG Failure Modes

SymptomLikely causeArchitect response
Correct source exists but answer is wrongPoor chunking, ranking, or prompt instructions.Tune chunking, metadata, ranking, and answer constraints.
Agent exposes restricted contentMissing ACL trimming or overprivileged index access.Enforce identity-aware retrieval and least privilege.
Agent says “I don’t know” too oftenRetrieval recall too low or content not indexed.Improve index coverage, synonyms, hybrid retrieval, query rewriting.
Agent hallucinates citationsPrompt allows unsupported citations or source mapping is weak.Generate citations only from retrieved source IDs.
Slow responsesLarge context, many tools, inefficient retrieval.Reduce context, cache safe data, optimize tool calls, use appropriate model.
Stale answersIndex refresh not aligned to data change rate.Refresh more often or retrieve live from source system.

Identity, Access, and Security

TopicExam-ready rule
Least privilegeGive the agent, tools, and users only the permissions required for the task.
User-delegated accessUse when results/actions must reflect the signed-in user’s permissions.
App-only accessUse only when the app has its own controlled service permissions and business justification.
Managed identityPrefer for Azure-hosted workloads accessing Azure resources.
SecretsStore in Azure Key Vault or managed secret stores; do not place secrets in prompts or code.
API securityPut controlled contracts in front of systems; validate input and output.
Network controlsUse private access patterns where required by the organization’s security model.
Data protectionClassify data, apply DLP, encryption, retention, and access reviews according to policy.
Prompt injectionTreat retrieved content and user input as untrusted; isolate instructions from data.
AuditabilityLog user, tool, action, approval, result, and correlation IDs where appropriate.
Notes and examples

Delegated vs App-Only Access

Use delegated access when…Use app-only access when…
The answer must be trimmed to the user’s rights.A background process performs a controlled organizational task.
The user is responsible for the action.The application identity is responsible for the action.
Different users should see different results.All authorized users receive the same approved service behavior.
Microsoft 365 or Graph data must follow user permissions.A curated service index or backend API has its own authorization layer.

Common trap: Using one highly privileged app identity to retrieve all enterprise data can violate least privilege and produce data leakage.

Data security and identity controls

For AB-100, security is not an afterthought. Agentic systems can amplify access because they combine natural language, data retrieval, and action execution.

Control areaWhat to design for
IdentityUse enterprise identity such as Microsoft Entra ID where appropriate
AuthorizationApply least privilege to users, agents, connectors, and service identities
Permission trimmingEnsure retrieved content respects the user’s access rights
SecretsStore credentials securely; do not place secrets in prompts or code
Data loss preventionApply policies to prevent sensitive data movement to unauthorized systems
Audit loggingRecord important prompts, tool calls, approvals, and outputs where appropriate
Environment separationSeparate development, test, and production environments
Conditional accessConsider user, device, network, and risk context where relevant
Data lifecycleAddress retention, deletion, classification, and sensitivity labels when required

Security traps

  • Using a high-privilege service account that lets the agent retrieve or modify more data than the user can access.
  • Treating prompts as a security boundary.
  • Allowing unrestricted tool calls from untrusted user input.
  • Returning retrieved content without checking sensitivity or access.
  • Logging sensitive prompts or responses without retention and access controls.
  • Forgetting that connectors and plugins may expand the agent’s effective attack surface.

Tool and Action Design

Design elementRecommendation
Tool contractUse explicit schemas, required fields, allowed values, and clear descriptions.
AuthorizationEnforce outside the model in the API/workflow layer.
ValidationValidate model-provided parameters as untrusted input.
IdempotencyDesign repeated calls not to duplicate harmful actions.
ConfirmationPreview high-impact actions before execution.
Error handlingReturn structured errors the orchestrator can handle safely.
LoggingRecord action intent, parameters where safe, actor, approval, result, and correlation ID.
RollbackProvide compensation or reversal for actions where possible.
Rate controlProtect backend systems from loops or excessive calls.
Notes and examples

Example tool contract style:

{
  "name": "create_support_ticket",
  "description": "Create a support ticket after user confirmation.",
  "parameters": {
    "type": "object",
    "properties": {
      "customerId": { "type": "string" },
      "severity": { "type": "string", "enum": ["low", "medium", "high"] },
      "summary": { "type": "string" }
    },
    "required": ["customerId", "severity", "summary"]
  }
}

Architectural point: the schema helps the model call the tool correctly, but the backend still must enforce permissions, validation, and business rules.

Prompt, Instruction, and Memory Controls

AreaGood practiceAvoid
System/developer instructionsState role, allowed tasks, forbidden tasks, escalation rules, citation rules.Hiding security-critical logic only in natural-language instructions.
User inputTreat as untrusted.Letting user override policy, identity, or tool restrictions.
Retrieved contentTreat as data, not instructions.Allowing documents to redefine the agent’s behavior.
Conversation historyKeep only relevant context.Sending unnecessary sensitive history into every request.
MemoryStore explicit, consented, useful facts where appropriate.Persisting sensitive inferred data without governance.
Output formatUse structured output for tool calls and downstream systems.Free-form output when a deterministic schema is required.
Notes and examples

Prompt Injection Defense Layers

LayerControl
Instruction hierarchyKeep trusted system instructions separate from user and retrieved content.
Retrieval filteringRetrieve only authorized, relevant content.
Tool allowlistOnly expose tools needed for the current task.
Parameter validationReject unsafe or unauthorized tool arguments.
Action approvalRequire human approval for sensitive actions.
Output filteringDetect unsafe, unsupported, or policy-violating output.
MonitoringReview suspicious prompts, failed guardrails, and abnormal tool usage.

Responsible AI and Risk Controls

Microsoft Responsible AI principleArchitecture implications
FairnessTest for uneven quality or harmful outcomes across user groups and scenarios.
Reliability and safetyEvaluate failure modes, fallback behavior, and robustness before production use.
Privacy and securityMinimize data exposure, secure identities, protect prompts and outputs.
InclusivenessDesign accessible experiences and support diverse user needs.
TransparencyTell users they are interacting with AI, show sources, explain limitations.
AccountabilityDefine owners, approval paths, audit trails, incident handling, and review process.
Notes and examples

Responsible AI and risk controls

Microsoft exam scenarios commonly reward designs that balance usefulness with responsible AI practices.

RiskExampleMitigation
HallucinationAgent invents a policy or priceGrounding, citations, evaluation, refusal behavior
Prompt injectionDocument tells the agent to ignore prior instructionsContent filtering, instruction hierarchy, tool restrictions, retrieval sanitization
Data leakageAgent exposes confidential recordsPermission trimming, DLP, identity-aware retrieval
Unsafe actionAgent cancels an order or changes payroll incorrectlyApproval workflow, validation, rollback, scope limits
Bias or unfairnessAgent treats customers or employees inconsistentlyEvaluation sets, human review, policy constraints
OverrelianceUsers accept AI output without judgmentUX cues, confidence, citations, training
Compliance breachSensitive data leaves approved boundaryGovernance, classification, approved connectors, monitoring
Poor explainabilityUser cannot understand why a recommendation was madeShow sources, reasoning summary, decision factors where appropriate

Evaluation and Acceptance Criteria

Evaluation typeWhat to testExample acceptance question
GroundednessOutput supported by retrieved sources.“Can every factual claim be traced to approved evidence?”
RelevanceAnswer addresses the user’s actual task.“Did the agent solve the requested problem without irrelevant content?”
Tool accuracyCorrect tool selected with valid parameters.“Did the agent call the right API with authorized inputs?”
SafetyHarmful, disallowed, or sensitive outputs blocked.“Does the agent refuse or escalate unsafe requests?”
Permission behaviorUsers only see and act on permitted data.“Do two users with different rights get appropriately different results?”
RobustnessHandles ambiguous, adversarial, and incomplete requests.“Does the agent ask clarifying questions when needed?”
LatencyResponse time fits the user workflow.“Is the solution usable during the business process?”
Cost efficiencyModel, retrieval, and tool usage align with value.“Can a smaller model or cached retrieval meet the need?”
Business outcomeProcess metric improves.“Did the agent reduce handling time, errors, or backlog?”
Notes and examples

Test Set Design

IncludeWhy
Happy-path examplesConfirms intended behavior.
Edge casesReveals brittle assumptions.
Ambiguous promptsTests clarification and refusal behavior.
Restricted-content promptsTests permission trimming.
Prompt injection attemptsTests instruction isolation and tool safety.
Stale or missing data casesTests uncertainty handling.
High-impact action requestsTests approval and audit controls.
Realistic user languageTests production readiness beyond lab prompts.

Governance and Lifecycle

PhaseArchitect checklist
DiscoverBusiness goal, users, data sources, risk level, success metrics, ownership.
DesignPattern selection, Microsoft services, identity, retrieval, tools, controls.
BuildEnvironments, source control, CI/CD, prompt/version management, secure connectors.
EvaluateOffline test sets, red-team prompts, responsible AI review, user acceptance.
DeployRelease gates, tenant/environment policies, monitoring, rollback plan.
OperateTelemetry, feedback review, model/index/tool updates, incident response.
ImproveAnalyze failures, update prompts, refine retrieval, add tests, retire unused features.
Notes and examples

Environment and ALM Considerations

ConcernPractical reference
SeparationUse separate development, test, and production environments.
VersioningVersion prompts, agent definitions, tools, workflows, and evaluation sets.
Change controlTreat prompt and tool changes as production-impacting changes.
RollbackKeep a known-good agent or configuration available.
OwnershipAssign business owner, technical owner, data owner, and support owner.
DocumentationRecord data sources, permissions, limitations, known risks, and escalation paths.

Governance and lifecycle

AB-100 scenarios may describe multiple departments building agents. The architect should prevent uncontrolled sprawl.

Governance needDesign response
Many teams creating agentsDefine standards, templates, review gates, and ownership
Sensitive data accessRequire classification, DLP, permission trimming, and approved connectors
Production deploymentUse environment strategy, testing, approvals, and monitoring
Business-user maintainabilityProvide low-code governance, managed environments, and change control
ReuseCreate shared connectors, prompts, policies, evaluation sets, and components
Compliance evidenceMaintain logs, approvals, evaluation results, and documentation

Observability and Troubleshooting

SignalWhat it tells you
User feedbackPerceived usefulness, missing features, bad answers.
Prompt/response tracesReason for poor answer or unsafe behavior; handle sensitive logging carefully.
Retrieval diagnosticsWhich documents/chunks were used and why.
Tool-call logsAction selection, parameters, failures, latency, repeated calls.
Approval logsHuman review patterns and bottlenecks.
Safety eventsRefusals, content filter events, prompt injection attempts.
Cost and usage metricsModel, retrieval, and tool consumption trends.
Business KPIsWhether the agent improves the intended process.
Notes and examples

Troubleshooting Decision Table

ProblemFirst checksLikely fix
Agent gives generic answersNo grounding, weak retrieval, vague prompt.Add authoritative data, tune retrieval, require citations.
Agent uses wrong toolTool descriptions overlap or orchestration lacks constraints.Clarify tool schemas, restrict tools by intent, add tests.
Agent fails for some usersPermission issue, connector configuration, identity mismatch.Verify delegated/app permissions and data ACLs.
Agent takes unsafe actionMissing approval, weak validation, overbroad tool access.Add approval gate, reduce permissions, enforce backend policy.
Agent loops or calls tools repeatedlyNo step limit, poor error handling, ambiguous state.Add maximum steps, deterministic workflow states, retry policy.
Good in test, poor in productionTest set not representative, data drift, user behavior differs.Expand evaluations with real-world cases and monitor drift.
High latencyToo much context, too many tool calls, slow backend.Reduce context, optimize retrieval, parallelize safe calls, cache.

Common AB-100 Traps

TrapBetter answer
“Use an LLM for everything.”Use deterministic workflow or standard automation when rules are clear.
“Grounding guarantees correctness.”Grounding helps; evaluation, citations, and fallback are still required.
“Prompt instructions are security controls.”Enforce security in identity, retrieval, API, and workflow layers.
“Agent autonomy is always desirable.”Match autonomy to risk; use approval for high-impact actions.
“One index for all enterprise content is simpler.”Simpler can be unsafe; design permission-aware retrieval.
“The biggest model is the best architecture.”Select model based on quality, latency, cost, context, and task complexity.
“Business users can own the whole solution alone.”Low-code still needs governance, data ownership, security, and lifecycle management.
“Logs should capture everything.”Capture diagnostic value while minimizing sensitive data exposure.
“Evaluation is a one-time launch task.”Agent quality changes as prompts, tools, data, and user behavior change.
“APIs exposed to agents can trust agent inputs.”Validate and authorize every tool call as untrusted input.
Notes and examples

Common exam traps

TrapBetter reasoning
Selecting the most advanced model by defaultStart with business need, risk, cost, latency, and grounding
Replacing all workflows with agentsUse deterministic automation for deterministic processes
Treating a proof of concept as production-readyAdd security, testing, monitoring, ALM, and governance
Solving data security with prompt instructionsEnforce identity, authorization, DLP, and permissions in architecture
Assuming retrieved content is correctUse authoritative sources, metadata, freshness, and evaluation
Allowing broad tool accessUse least privilege, schema validation, approvals, and audit logs
Ignoring user experiencePut the agent in the user’s actual workflow
Measuring only technical accuracyAlso measure business impact and adoption
Forgetting operational ownershipAssign owners for data, agent behavior, connectors, and support
Overbuilding custom solutionsUse Microsoft platform capabilities when they satisfy requirements

Compact Design Checklist

Before choosing an answer in an AB-100 scenario, verify:

  • Business fit: The solution solves a measurable business problem.
  • Pattern fit: Agent, copilot, RAG, workflow, or traditional app is justified.
  • Microsoft service fit: Low-code vs pro-code choice matches ownership and complexity.
  • Grounding fit: Data is authoritative, current, permission-aware, and traceable.
  • Tool fit: Actions are limited, validated, authorized, observable, and recoverable.
  • Identity fit: Delegated vs app-only access is intentional.
  • Risk fit: Autonomy level matches impact and uncertainty.
  • Responsible AI fit: Safety, transparency, privacy, fairness, and accountability are designed in.
  • Evaluation fit: Test cases include realistic, adversarial, permission, and edge scenarios.
  • Operations fit: Monitoring, ownership, support, incident response, and improvement loop exist.

The core AB-100 mindset

For most scenarios, think in this order:

  1. Business outcome — What measurable process, decision, or user experience needs improvement?
  2. User and workflow context — Who uses the agent, where, and at what point in the process?
  3. Agent capability — Should the solution answer, recommend, summarize, plan, or act?
  4. Grounding and data — What trusted enterprise data is required, and how are permissions enforced?
  5. Tools and actions — Which systems can the agent call, and what approvals are required?
  6. Risk controls — How will the design reduce hallucination, prompt injection, data leakage, unsafe actions, and compliance risk?
  7. Evaluation — How will quality, safety, cost, latency, and business impact be measured?
  8. Operations — How will the agent be monitored, updated, versioned, and governed after deployment?
    flowchart LR
	    A[Business outcome] --> B[User workflow]
	    B --> C[Agent role and autonomy]
	    C --> D[Grounding data]
	    D --> E[Tools and actions]
	    E --> F[Security and governance]
	    F --> G[Evaluation and testing]
	    G --> H[Deployment and operations]
	    H --> I[Continuous improvement]

High-yield Microsoft solution positioning

AB-100 questions often test whether you can select an appropriate Microsoft-aligned architecture pattern, not just name a product.

Need in the scenarioLikely architectural directionWatch for
Employees need AI assistance inside Microsoft 365 work patternsExtend or configure Microsoft 365 Copilot experiences where appropriateDo not build a separate custom app if the requirement is mainly productivity-context assistance
Business users need a low-code conversational agent for internal or customer workflowsConsider Microsoft Copilot Studio with connectors, topics, actions, and governanceConfirm data sources, authentication, escalation, and environment strategy
Solution needs advanced custom orchestration, model selection, evaluation, or AI application lifecycle controlConsider Azure AI Foundry and related Azure servicesHigher flexibility usually means more design responsibility
Agent must retrieve enterprise knowledge from documents or indexed dataUse a grounded retrieval pattern, often with search/indexing and permission controlsDo not rely on the model’s general knowledge for enterprise facts
Agent must update records, submit requests, or trigger workflowsUse tools, APIs, connectors, Power Automate, Logic Apps, or custom servicesRequire approvals, auditability, idempotency, and error handling
User only needs deterministic workflow automationUse workflow automation instead of an agent when natural-language reasoning is unnecessaryDo not overuse agentic AI for fixed rule-based processes
Sensitive or regulated process with material business impactAdd human-in-the-loop, policy checks, logging, and constrained tool executionPrompt instructions alone are not sufficient control

Key terms to know cold

TermPractical meaningExam trap
AgentAI system that can reason over context, use tools, and pursue a task through multiple stepsTreating every chatbot as a true agent
CopilotAI assistant experience embedded in a user workflowAssuming a copilot always has permission to all enterprise data
GroundingSupplying trusted context to improve relevance and reduce unsupported answersConfusing grounding with model training
RAGRetrieval-augmented generation: retrieve relevant content, then generate an answer grounded in itAssuming RAG automatically solves permissions, freshness, or quality
Tool / actionCallable capability such as an API, connector, workflow, or functionForgetting authorization, validation, and rollback behavior
OrchestrationControl logic that coordinates prompts, tools, memory, routing, and multi-step tasksLetting an agent freely call tools without constraints
Human-in-the-loopHuman review, approval, or escalation before or during an actionAdding approval too late for high-risk actions
Prompt injectionMalicious or accidental instructions that try to override system behavior or exfiltrate dataBelieving “ignore malicious instructions” is a complete defense
EvaluationMeasuring correctness, groundedness, task success, safety, cost, and latencyTesting only with happy-path demos
ObservabilityLogs, traces, metrics, and feedback used to operate the solutionIgnoring production monitoring after launch

Agent autonomy decision rules

The safest architecture usually gives the agent the minimum autonomy needed to create value.

Autonomy levelWhat the agent can doBest forRequired controls
InformAnswer questions or summarize grounded contentKnowledge lookup, policy Q&A, document summarizationSource citation, permissions, freshness checks
RecommendSuggest next steps but not executeSales coaching, support triage, process guidanceExplanation, confidence indicators, user validation
DraftPrepare messages, tickets, plans, or records for user reviewEmail drafts, proposal content, case notesHuman review, content safety, audit trail
Act with approvalExecute an action after explicit approvalSubmit request, update CRM, create ticketApproval workflow, input validation, logging
Act independentlyExecute bounded low-risk tasks automaticallyRoutine routing, notifications, data enrichmentStrict scope, monitoring, rollback, exception handling

Fast rule

If the action is irreversible, external-facing, financial, legal, safety-related, privacy-sensitive, or reputationally risky, design for human approval or strong deterministic controls.

Requirements analysis: what the architect must clarify

Scenario questions often include extra detail. Focus on the requirement that changes the architecture.

Requirement areaQuestions to askDesign implication
Business valueWhat KPI improves: cycle time, accuracy, cost, satisfaction, throughput?Defines success metrics and prioritization
UsersInternal employee, customer, partner, admin, frontline worker?Determines identity, channel, UX, and access model
DataWhich systems are authoritative? How fresh must data be?Determines grounding, indexing, connectors, and sync strategy
ActionsDoes the agent only answer, or can it modify systems?Determines tool design, approvals, and audit needs
RiskWhat can go wrong if the agent is wrong?Determines autonomy, testing depth, and controls
ComplianceAre there privacy, retention, residency, or industry constraints?Determines governance and data handling
OperationsWho owns, monitors, and updates the solution?Determines ALM, telemetry, support, and change management

Choosing between knowledge, workflow, and agent patterns

Scenario cluePreferWhy
“Users need answers from internal documents”Grounded knowledge copilot / RAG patternRetrieval and source grounding are central
“Users need a request submitted after a conversation”Conversational agent plus workflow/actionNeeds both natural language and system execution
“Every request follows fixed rules”Deterministic workflow automationAgent reasoning adds unnecessary risk
“Agent must plan across multiple systems”Tool-using agent with orchestrationRequires state, tool selection, and error handling
“Business users must maintain conversation flows”Low-code agent designMaintainability by non-developers matters
“Developers need full control over models, evaluation, and deployment”Custom Azure AI architectureFlexibility, testing, and lifecycle control matter

Grounding and retrieval review

Grounding is central to enterprise agentic AI. The agent should use trusted sources instead of unsupported model knowledge when answering business-specific questions.

Retrieval design checklist

Design pointGood practiceCommon mistake
Source selectionUse authoritative repositories and systems of recordIndexing outdated, duplicate, or unofficial files
ChunkingSplit content into useful, semantically coherent unitsChunks too large to retrieve precisely or too small to preserve context
MetadataStore labels such as owner, date, department, sensitivity, product, regionFailing to filter by region, role, or effective date
PermissionsEnforce user access at retrieval and response timeReturning content the user cannot access
FreshnessDefine sync/index update requirementsAssuming indexed content is always current
CitationsProvide references where usefulProducing confident answers with no traceability
FallbackSay when information is unavailable or uncertainForcing an answer when grounding fails
Notes and examples

RAG versus fine-tuning

NeedBetter fitReason
Answer from changing enterprise documentsRAG / groundingEasier to update content without retraining
Adapt style, format, or task behaviorPrompting or fine-tuning, depending on needBehavior may not require retrieval
Add private facts to responsesRAGFacts should remain in controlled data sources
Improve specialized output consistencyPrompt templates, examples, evaluation, or fine-tuningDepends on volume, stability, and governance
Reduce hallucination about company policyGrounded retrieval plus refusal/fallback behaviorTraining alone does not guarantee current policy accuracy

Designing tool-using agents

Agents become business-critical when they can act. That increases value and risk.

Tool design principles

PrincipleWhy it matters
Narrow tool scopeReduces damage from bad planning or malicious input
Explicit schemasImproves reliability and validation
IdempotencyPrevents duplicate actions during retries
Confirmation stepsProtects high-impact operations
Error handlingLets the agent recover or escalate gracefully
AuditabilitySupports compliance, troubleshooting, and user trust
Rate limits and quotasProtects downstream systems
TimeoutsPrevents long-running or stalled agent tasks
Rollback or compensationHandles partial failure in multi-step workflows
Notes and examples

Tool-call decision path

    flowchart TD
	    A[Agent wants to call a tool] --> B{Is the tool needed?}
	    B -- No --> C[Answer without action]
	    B -- Yes --> D{Is user authorized?}
	    D -- No --> E[Refuse or escalate]
	    D -- Yes --> F{Is action high risk?}
	    F -- Yes --> G[Request approval]
	    F -- No --> H[Validate inputs]
	    G --> H
	    H --> I{Validation passes?}
	    I -- No --> J[Ask clarification or stop]
	    I -- Yes --> K[Execute tool]
	    K --> L[Log result and report status]

Orchestration and multi-agent thinking

Not every problem needs multiple agents. The exam may describe complex workflows to test whether you can decompose responsibilities safely.

PatternUse whenCaution
Single agentOne primary user task, limited tools, clear contextAvoid overloading with unrelated responsibilities
Router patternDifferent tasks require different specialized flows or agentsRouting logic must be tested
Planner-executorAgent plans steps and uses tools to complete themNeeds constraints, monitoring, and recovery
Human escalationThe agent cannot safely or confidently complete the taskEscalation path must be part of the design
Deterministic workflow plus AI stepAI is needed for summarization, classification, or drafting inside a stable processDo not make the entire workflow probabilistic
Multi-agent collaborationDistinct specialized roles improve quality or maintainabilityMore complexity, latency, cost, and testing burden

Prompt and instruction hierarchy review

Prompts are important but should not carry the entire architecture.

Prompt layerPurpose
System/developer instructionsDefine role, boundaries, policies, and response style
Grounding contextSupplies retrieved facts or task-specific data
User requestDescribes the current user’s goal
Tool outputsProvide results from external systems
Conversation history or memoryPreserves relevant context across turns

Prompt engineering traps

  • Hiding business rules only in a long prompt when they should be enforced by code, workflow, or policy.
  • Asking the model to “always be accurate” without grounding or evaluation.
  • Including sensitive credentials or secrets in prompt text.
  • Allowing retrieved documents to override higher-priority instructions.
  • Using one giant prompt instead of modular instructions, tools, and evaluation.

Evaluation: what to measure

A strong architect defines how the agent will be evaluated before full deployment.

Metric areaWhat it measuresExample evaluation method
Task successWhether the agent completes the intended workflowTest cases with expected outcomes
GroundednessWhether answers are supported by retrieved sourcesCompare response claims to source documents
Retrieval qualityWhether the right content is retrievedPrecision/recall review against known queries
SafetyWhether harmful, disallowed, or sensitive responses are blockedRed-team prompts and policy tests
Tool accuracyWhether tool calls use correct parameters and sequenceSimulated API tests and logs
LatencyHow quickly the solution respondsPerformance tests under realistic load
CostToken, compute, search, and workflow execution costUsage telemetry and budget thresholds
User satisfactionWhether users trust and adopt the solutionFeedback, surveys, support tickets
Business impactWhether the KPI improvedBefore/after process metrics
Notes and examples

Evaluation traps

  • Testing only a few demo prompts.
  • Ignoring negative tests and adversarial prompts.
  • Measuring answer fluency instead of correctness.
  • Not testing permissions with different user roles.
  • Not retesting after prompt, model, data, or tool changes.
  • Failing to include real business edge cases.

Monitoring and operations

Production agentic AI requires continuous operation, not one-time deployment.

Operational concernWhat to plan
TelemetryPrompts, responses, retrievals, tool calls, latency, errors, user feedback
VersioningTrack changes to prompts, models, tools, indexes, and workflows
RollbackAbility to revert a bad prompt, connector, index, or model change
Incident responseProcess for unsafe output, data leakage, or failed actions
Cost managementBudgets, quotas, alerts, usage reporting
Model updatesRegression testing when model behavior changes
Content updatesIndex refresh strategy and source ownership
Support modelWho handles failed sessions, escalations, and user issues
GovernanceReview process for new agents, tools, connectors, and data sources

Scenario-based decision table

Use this table when a question asks for the “best” design.

If the question emphasizes…Choose the answer that prioritizes…
Least custom developmentExisting Microsoft copilot or low-code capabilities
Maximum control and custom AI lifecycleAzure-based custom architecture and evaluation
Business-user ownershipCopilot Studio / low-code maintainability with governance
Enterprise data Q&AGrounded retrieval with permissions and citations
Process executionTool/action integration with validation and approvals
High-risk actionsHuman-in-the-loop and constrained automation
Compliance and privacyData governance, identity, audit, DLP, retention
Accuracy problemsBetter grounding, evaluation, retrieval tuning, fallback
Production readinessMonitoring, versioning, rollback, support
AdoptionUser workflow fit, training, feedback, measurable value

Final readiness check

You are closer to exam-ready when you can explain why an architecture is appropriate, not just name the Microsoft service involved. Focus your final review on scenario tradeoffs: when to use an agent, how to ground it, how to constrain its actions, how to secure it, and how to prove it works.

Next step: use the AB-100 question bank for targeted topic drills, then complete mixed mock exams with detailed explanations to test whether you can apply these decision rules under exam-style pressure.

Put the review into practice