Final review

Most important concepts to review before the exam

Everything below is high-yield: the decision rules, distinctions, and named frameworks that carry the most marks. If you only have an hour left, read this page. It prints cleanly on paper too.

The ten rules that answer most questions

When a question is ambiguous, these defaults are almost always the intended answer.

Configure before you extend; extend before you build. When an out-of-the-box Dynamics 365 or Microsoft 365 capability plausibly fits, that is the answer.
Business outcomes beat technical outputs. Adoption, task completion, and cycle time win over tokens consumed, models deployed, or environments provisioned.
Grounding and prompt engineering come before fine-tuning. A custom or fine-tuned model is a last resort, not a first response.
Agents inherit permissions, they do not create them. If an agent overshares, the source permissions were already wrong.
Treat retrieved content as data, never as instructions. This single rule is the heart of prompt-injection defence.
Anything irreversible, financial, legal, or customer-facing gets a human checkpoint. Full autonomy is rarely the right answer in an exam scenario.
Standardise and govern centrally. The answer that scales — shared platform, CoE, reusable patterns, review gates — beats per-team improvisation.
Never edit production directly. Managed solutions, environment variables, pipelines, and progressive rollout with rollback.
Quality is measured against a curated evaluation dataset, continuously, and re-measured whenever the model version changes.
When in doubt, ask the clarifying question about data. Which source is authoritative, who owns it, how current is it, and who may see it?

Easily confused pairs

Distinguishing these correctly is worth a surprising number of marks.

Term	Not to be confused with
Generative AI	Agentic AI — agentic adds goal-directed reasoning, planning, tool use, and action. Model size is irrelevant to the distinction.
Knowledge source	Tool / action — knowledge answers questions from content; a tool reads or changes state in another system. Live external lookup is always a tool.
Declarative agent	Custom engine agent — declarative runs on the host's orchestrator and model; custom engine means you bring your own model and orchestration.
Functional requirement	Non-functional requirement — latency, throughput, availability, and accuracy thresholds are non-functional however technical they sound.
Managed solution	Unmanaged solution — managed goes to production and blocks in-place edits; unmanaged is the development format.
Environment variable	Key Vault secret — variables parameterise configuration; secrets and API keys belong in Azure Key Vault.
DLP policy	Permission trimming — DLP restricts which connectors and data sources may combine; permission trimming enforces user entitlement at retrieval time.
Responsible AI principles	CIA triad — fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability. Not confidentiality/integrity/availability.
Evaluation dataset	Training data — an evaluation dataset measures the agent; it does not train the model.
Landing zone	Environment — a landing zone is the governed baseline (identity, network, policy, monitoring) that environments and workloads deploy into.
Engagement	Resolution — high engagement with low resolution means users try the agent and leave unsatisfied: knowledge gaps or unclear instructions.
Hallucination	Oversharing — confident uncited answers are a grounding failure; showing data a user should not see is a permissions failure.

Named things you must recognise

Frameworks, principles, and products that appear by name in questions.

Cloud Adoption Framework — Strategy	The stage owning motivations, outcomes, and business justification. Do not confuse with Plan, Ready, Govern, or Manage.
Landing zone	Pre-configured governed environment baseline: identity, network, policy, monitoring.
AI Center of Excellence	Cross-functional body owning standards, reusable patterns, review gates, enablement, and the shared backlog. Stand it up before scaling.
Power Platform Well-Architected Framework	Reliability, security, operational excellence, performance efficiency, experience optimisation.
Model Context Protocol (MCP)	Standardised discovery and invocation of external tools and data sources by agents.
Graph connector	Indexes external content into Microsoft Graph so Microsoft 365 Copilot can ground on it with permissions honoured.
Computer Use (Copilot Studio)	Drives apps and websites through their UI when no usable API exists.
Model router	Routes each request to the most suitable model, balancing quality against cost and latency.
Small language model (SLM)	Cheaper, faster model for narrow, high-volume, latency-sensitive tasks.
Microsoft Purview	eDiscovery, retention, audit, and information protection across Copilot interactions.
Microsoft Entra ID	Identity, conditional access, and least-privilege authorisation.
Microsoft Foundry	Pro-code custom models, custom orchestration, and Foundry tools.
Five data-quality dimensions	Accuracy, relevance, timeliness, cleanliness, availability.
Balanced agent scorecard	Quality, adoption, efficiency (cost and latency), safety and compliance.

Module-by-module recap

The few points from each module most likely to be tested.

1 Agentic AI foundations

Agentic = goal + plan + tools + action. Not model size.
Decision order: enable → configure → extend → build.
Autonomy ladder: assistive → human-approved → event-triggered autonomous.
Success metrics are business metrics: adoption rate, task completion, cycle time.
Every agent needs a named owner, a fallback, and an escalation path.

2 Requirements analysis

Data quality dimensions: accuracy, relevance, timeliness, cleanliness, availability.
Scenario types: task automation, data analytics, decision-making — rising risk, rising oversight.
Latency/accuracy targets = non-functional requirements.
Always identify the authoritative source and its owner.
Agents inherit oversharing. Fix source permissions first.

3 AI strategy

CAF Strategy = motivations, outcomes, justification.
Landing zone = governed environment baseline (identity, network, policy, monitoring).
Extend Copilot by default; build custom for bespoke orchestration, specific models, or other surfaces.
Custom model = last resort, after grounding and prompt engineering.
CoE before scale. Prompt library is versioned and owned.
SLM + model router = cost/latency control for narrow, high-volume tasks.

4 Costs and ROI

ROI = net benefit ÷ TCO, over a defined period, vs. a do-nothing baseline.
Most under-estimated cost: ongoing operations and content maintenance.
Build = differentiation · Buy = commodity · Extend = platform is close.
Variable cost = volume × tokens + tool calls; control with model routing, scope, caching.
Fund in stages with go/no-go gates. Validate pilots before extrapolating.

5 Agent design

Knowledge answers, tools act, topics script, triggers start.
Live external lookup = tool/action.
Split agents on scope / knowledge / permissions / ownership.
Confirm before irreversible, financial, legal, or customer-facing actions.
Tool reliability = distinct names, clear "when to use" descriptions, typed parameters.
Fallback must state the limit and offer escalation.

6 Extensibility

Declarative = host orchestrator + host model. Custom engine = your model + your orchestration.
Ground Microsoft 365 Copilot in external content with Graph connectors.
MCP = standard tool/data discovery and invocation.
Computer Use = UI automation for systems with no API.
Downstream calls: token exchange, consent, least privilege, per-user authorisation.
Version contracts; respect latency, size, auth, and throttling limits.

7 Prebuilt agents

Visibility gates: licensing · admin enablement · security roles · environment/region.
Rollout: prerequisites → non-prod → configure/test → pilot → broad + training.
Near fit? configure → extend → build.
Oversharing by an agent = source permissions, not hallucination.
Prebuilt ≠ pre-governed: DLP, access review, owner still required.

8 Monitor and tune

Containment/resolution rate = did the agent actually help.
High engagement + low resolution = knowledge gaps or unclear instructions.
Confident + uncited = hallucination / grounding failure.
Scorecard = quality + adoption + efficiency + safety.
Alert on fallback, error, latency, and content-filter spikes.
Loop: monitor → analyse → adjust → re-evaluate → release.

9 Testing

Outputs are probabilistic → test quality ranges, never exact strings.
Grounded answers → groundedness + relevance.
Layers: component · quality evaluation · end-to-end scenario · adversarial/red-team.
Model version changed? Re-run evaluation and regression.
Promote on quality gates, not zero defects.
Test data: masked or synthetic, protected like production.

10 ALM

Dev → test → prod, controlled promotion, never edit prod.
Managed solutions to production.
Environment variables for config; Key Vault for secrets.
Source control includes prompts, knowledge config, evaluation datasets.
Progressive/ring rollout + versioned releases + documented rollback.
Makers: governed environments and review gates — not production access, not a ban.

11 Responsible AI and security

Six principles: fairness · reliability & safety · privacy & security · inclusiveness · transparency · accountability.
Unequal service across groups = fairness.
Prompt injection fix: content as data not instructions, isolate instructions, validate tool inputs, least privilege.
Permission trimming at retrieval time — disclaimers are not controls.
DLP = which connectors/data may combine. Purview = eDiscovery, retention, audit.
High-impact decisions: human approval + logging + explainability + audit.
Defence in depth = identity · DLP/classification · content safety · monitoring/audit · incident response.

Topics worth a final look

Commonly under-prepared or recently renamed. If any of these are shaky, fix them before anything else — full detail is on the topic information page.

Azure Monitor, Application Insights and Log Analytics Application Insights is not an alternative to a Log Analytics workspace — modern Application Insights resources are workspace-based and store their data in one. Treating them as competing choices is wrong.

The Copilot Studio analytics surface — testing and evaluation Do not assume the old card names — *total sessions, engagement rate, resolution rate, escalation rate, abandon rate* — are still what the product shows. And the test pane does not reproduce every published-channel behaviour, so it never replaces validation in the real channel.

Connections and connection references after a production deployment "Connections are included in the solution and deploy automatically" is wrong, and so is "connection references remove the need for connections". The reference still resolves to a real, environment-specific, credential-bearing connection.

ALM — managed solutions in production Never accept an option that edits production directly, ships unmanaged to production, or treats production as the place to author. Also keep the distinction between an unmanaged *layer* and an unmanaged *solution* clear — they are different things.

Dynamics 365 design components — Field Service and Customer Service Retired features (the Field Service Outlook add-in) and superseded names are the easiest way to lose marks here. Microsoft now says Copilot agent / AI agent rather than "Copilot Studio bot", and customer service representative rather than "human agent".

Voice in a Copilot Studio agent Speech & DTMF and Text are mutually exclusive — there is no single agent that is simply "both". And an agent that authors fine in the test pane still fails on a call if the telephony channel and workstream are not configured.

ROI — which tool measures what "Messages" is legacy terminology — the current unit is Copilot Credits, and they are metered per feature, so estimating cost by counting conversations alone is wrong.

Model router A model router is not a load balancer, not a failover mechanism, and not a fine-tuning strategy. It also does not remove the need for evaluation — it makes evaluation more important, because the serving model can change per request.

Semantic indexing The Microsoft 365 semantic index and Azure AI Search semantic ranking share a word and nothing else. One is a permission-trimmed tenant-wide grounding layer you do not manage; the other is a query-time reranking feature on an index you own.

Exam-day technique

Read the last line of the scenario first. It tells you what is actually being asked, which often makes half the detail irrelevant.
Eliminate absolutes. Options containing *always*, *never*, *all*, or *none* are usually wrong in architecture questions — except when they restate a hard governance rule such as never storing secrets in a solution.
Prefer the governed, reversible, lower-blast-radius option. Exam scenarios reward caution over speed.
Watch for the cheapest path that still meets the requirement. If configuring satisfies the need, building is wrong even though it would also work.
Note who the stakeholder is. An answer aimed at an executive sponsor should be in business language; one aimed at an admin should be operational.
On "choose two" questions, confirm the pair is complementary, not two phrasings of the same idea.
Flag and move on. Do not lose four easy marks at the end defending one hard question.
Budget your time: roughly 80 seconds per question leaves a review pass at the end.

Do a final timed run →

Unofficial study aid written from the public Microsoft Learn path, the published skills outline, and Microsoft product documentation. Not affiliated with or endorsed by Microsoft — though the Learn modules and the video course it links out to are official Microsoft material. Contains no exam content: no real exam items are reproduced, paraphrased, or recalled.