AB-100 cheat sheet

Microsoft CAF AI adoption sequence: Strategy → Plan → Ready → Adopt → Govern → Manage; the Plan phase output is a prioritized use case portfolio, a roadmap, and a skills plan.

A use case qualifies for build only if it has a named process owner, a baseline KPI measurable before launch, permission-trimmed source content, and a scope one agent can complete end to end.

Agent spectrum by build effort: Microsoft 365 Copilot (no build) → declarative agent (instructions + knowledge + actions on the Microsoft 365 orchestrator) → custom engine agent (own orchestrator; Copilot Studio, Azure AI Foundry, or Microsoft 365 Agents SDK).

Choose Copilot Studio when the solution is low-code, grounded in Microsoft 365 and Dataverse content, and needs Power Platform connectors; choose Azure AI Foundry Agent Service when it needs multi-model choice, pro-code orchestration, or Azure-native tools such as code interpreter and file search.

Microsoft 365 Copilot is licensed per user; Copilot Studio agent consumption is metered separately through prepaid capacity or pay-as-you-go, so pilot sizing must include projected session or message volume.

Responsible AI Standard principles - fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability - each require an entry in the use case impact assessment.

Copilot answer scope equals the signed-in user's existing permissions; no agent design can surface content the user could not already open.

Autonomy rule: unattended triggers are acceptable when the resulting action is reversible or an approval step is inserted; irreversible or financially material actions require human-in-the-loop.

Primary cost drivers per use case: session or message volume, model choice, retrieval index size, and number of external tool calls.

Agent anatomy to specify during design: instructions, knowledge sources, tools/actions, topics, orchestration mode, channels, authentication, fallback and escalation paths.

Generative orchestration selects topics, tools, and knowledge at runtime from their name and description text - descriptions act as prompts and must state when to use and when not to use each item.

Classic orchestration routes on topic trigger phrases; each topic needs distinct, non-overlapping phrases, and unmatched utterances fall through to the fallback topic.

Knowledge source options include SharePoint, Dataverse, uploaded files, public websites, Azure AI Search indexes, and Microsoft Graph connectors; disable general-knowledge answers when responses must stay inside approved content.

RAG pipeline: chunk → embed with an embedding model → index in Azure AI Search (vector, hybrid, semantic ranker) → retrieve top-k chunks → ground the prompt; index refresh cadence determines answer freshness.

Tool design rules: one tool per intent, minimal OpenAPI schema, idempotent operations where possible, and explicit structured error responses so the orchestrator can retry or escalate.

Authentication decision: no authentication for anonymous public content, Microsoft Entra ID when the agent reads user-scoped data, manual OAuth for third-party systems without Entra integration.

Decompose broad domains into connected (child) agents: the parent routes, each child holds its own instructions, knowledge, and tools, keeping each instruction set within the model's effective budget.

Design the exit path first: fallback topic wording, escalation trigger conditions, and the conversation variables (reason code, transcript reference) that must survive handoff.

ALM: build in a development environment as an unmanaged solution, export a managed solution to test, then promote to production; Power Platform pipelines automate the same promotion path.

Environment strategy: separate development, test, and production environments; apply managed environment settings to test and production for sharing limits, solution checker enforcement, and usage insights.

Data loss prevention policies are environment-scoped and classify connectors as business, non-business, or blocked - an agent cannot combine connectors across those groups.

Channel rules: Teams and Microsoft 365 Copilot publishing require an app manifest and admin approval for org-wide availability; websites and custom apps use the web channel or Direct Line.

Authentication is configured per channel and must match the identity used to trim data; Teams and Microsoft 365 Copilot support Microsoft Entra ID single sign-on.

Observability: Copilot Studio analytics reports sessions, engagement, resolution, and escalation rates; connecting Application Insights adds transcripts, tool-call traces, latency, and errors.

Gate every publish on an evaluation run over a fixed test set with expected answers, scoring groundedness, completeness, and correctness against the prior version.

Governance: Microsoft Purview sensitivity labels, retention, and audit apply to agent-generated content, and audit logs record agent and administrator activity.

Operate a weekly loop: review unresolved and escalated sessions, adjust knowledge sources or instructions, then republish through the same pipeline so changes stay traceable.