Microsoft AI Apps and Services

This is the single most testable distinction in the domain, and the one people get wrong.

Microsoft 365 Copilot uses your organizational data and the web. It requires a paid add-on license on top of a Microsoft 365 subscription.

Microsoft 365 Copilot Chat uses the web, and users can supply organizational data themselves. It does not require the add-on license.

The difference in one line: the licensed product reaches into your tenant's content automatically; the unlicensed one does not, though a user can paste content in. That distinction drives cost, rollout scope, and the whole "who gets a license first" conversation.

Microsoft Graph is what makes the licensed version useful. It holds information about users, their activity, and the organizational content they can access, and it brings that personalised context into the prompt: a user's mail, chats, documents, and meetings.

Two things to hold onto:

• Copilot only surfaces data the user already has permission to access. The permission model is inherited, not bypassed.
• Semantic indexing over Graph data is what makes retrieval find conceptually relevant material rather than only keyword matches.

Copilot Search is the universal search layer across Microsoft 365 and connected non-Microsoft sources, integrated so a user can search and then move into chat for follow-up.

Capability differs sharply by application, and the exam tests that rather than treating "Copilot" as one uniform thing.

• Word: drafts new content, summarizes, answers questions about the document, light editing commands. Word files can also serve as grounding material.
• PowerPoint: builds a deck from a prompt or from an existing Word document using enterprise templates; summarizes; adds slides and applies deck-wide formatting.
• Excel: suggests formulas and chart types, surfaces insights about data in the sheet. Narrower than people assume.
• Outlook: summarizes threads, drafts replies pulling from other content the user can access, and offers coaching on clarity and tone.
• Teams: summarizes chat threads, and in meetings answers questions from the live transcript. Note the scoping: in chat it uses only that single thread, and in meetings only the transcript.
• Loop, Whiteboard, OneNote, Forms: drafting and idea organization in each.

The Teams scoping detail is a fair exam target because it is counterintuitive: meeting Copilot knows the transcript, not your email.

Researcher and Analyst are the two specialized agents to distinguish. Reach for research when the task is gathering and synthesising across many sources; reach for analysis when the task is reasoning over data to produce findings. Matching the agent to the shape of the task is the tested judgment.

Copilot Studio is a graphical, low-code tool for building agents and agent flows. Its stated appeal is that you do not need a data scientist or a developer. You describe the agent in plain language, give it instructions, knowledge sources, tools, and triggers, then test and publish.

An agent in this sense is a scoped version of Copilot: it coordinates a language model with instructions, context, knowledge sources, topics, tools, and triggers to do a defined job. A helpdesk ticket agent, an HR agent that answers from your policy library.

Flows automate repetitive work and can be triggered manually, on a schedule, by an event, or by an agent as a tool. They can include human review steps: worth remembering, because that is where your approval gate lives.

The build-buy-extend decision is the manager-level judgment here:

• Use what ships when the need is close to what Copilot already does in the apps. Cheapest, no maintenance.
• Extend Copilot with an agent when you want Copilot's existing behavior plus your own data, instructions, or a specific domain. You keep the existing orchestration.
• Build a custom agent when you need your own branding, your own channels, your own end-to-end flow, or integration with external APIs and events.
• Buy when a mature product already does it and the integration cost is lower than the build.

The bias should be toward the top of that list. Every step down adds build cost and, more importantly, ongoing maintenance that someone has to own after you have moved on.

The integrated-stack argument worth being able to make: staying inside the Microsoft surface means the permission model, compliance tooling, and audit trail are inherited rather than reconstructed. That is a risk-mitigation case, not just convenience, and it is the framing the exam favors.

Copilot serves people doing knowledge work inside Microsoft 365. Foundry is the platform side: where you build AI into your own applications and processes rather than into someone's Word document.

The toolset spans model access and selection, enterprise search over your own content, and vision services for images and documents. The exam-relevant framing is not the API detail but which surface a given business process belongs on.

The decision usually resolves cleanly:

• The work happens inside Microsoft 365 apps, by people, on documents and mail → Copilot.
• The work is a process or product: a customer-facing application, a document pipeline, something embedded in a line-of-business system → Foundry.
• The work is Microsoft 365-shaped but needs your own data or behavior → extend Copilot with an agent before reaching for Foundry.

Matching a model to a need is the same discipline as everywhere else: weigh capability against latency and cost for the specific task rather than defaulting to the largest model. A classification step and a long-form drafting step have very different requirements, and using one model for both usually overpays for one of them.

The platform argument for Foundry is scalability and security you would otherwise have to build: managed scaling, an established security and compliance posture, and integration with the identity and governance tooling already in place. Assembling that yourself is the hidden cost of the do-it-yourself alternative.