1.3 Chat Experiences vs. Agent Experiences
Key Takeaways
- A chat experience needs no setup and reasons across everything you can already reach; an agent experience is configured once and narrowed to the sources it was given.
- The Work IQ toggle at the top left switches Copilot Chat between work data through Microsoft Graph (on) and the public web in the Bing search index (off).
- An agent carries three things a bare chat does not: instructions, knowledge sources, and starter prompts.
- Declarative agents use Copilot own orchestrator and models and rely on user-initiated interactions; custom engine agents choose their own models and can trigger actions without direct user input.
- Agents respect existing Microsoft 365 permissions, so an agent never grants a colleague access to content they could not already open.
1.3 Chat Experiences vs. Agent Experiences
Quick Answer: A chat experience is the general-purpose Copilot conversation — one assistant, broad reach, whatever you happen to ask today. An agent experience is Copilot deliberately narrowed: a named assistant that carries its own instructions, its own knowledge sources, and its own starter prompts, built for one job that recurs. Both run on the same models and neither can see content you could not already open. The difference is scope and preparation, not power.
This is a blueprint bullet in its own right — "understand the difference between a chat experience and an agent experience" — and it is tested as a routing decision rather than as a definition. An item will describe a business situation and ask which one fits. Get the underlying distinction right and those items answer themselves.
The Chat Experience
Copilot Chat is the open-ended surface. You start a conversation, you supply whatever context the question needs, and the conversation ends when you close it. Nothing is configured in advance, and nothing persists as a reusable asset except the transcript.
What the chat is grounded in depends on one control. With a Microsoft Copilot license, the Work IQ toggle at the top left of the Copilot interface switches the grounding:
| Work IQ | What Copilot is grounded in |
|---|---|
| On | Work data available to you, reached through Microsoft Graph — OneDrive documents, email, and other organizational content you have permission to open |
| Off | Data from the public web in the Bing search index, combining generative AI with current public information |
That single toggle is the whole story of chat grounding, and it explains the most common real-world complaint: someone asks Copilot about an internal project, gets a generic web answer, and concludes Copilot "cannot see" their files. Work IQ was off.
Three characteristics define the chat experience:
- Zero setup. There is nothing to build before you use it.
- Context is supplied per turn. If you want Copilot to use a specific file, you reference it in that prompt. Next week you reference it again.
- Breadth is the point. Chat is the only surface that can reason across a meeting, a workbook and a contract in the same question.
The Agent Experience
Microsoft's own framing for business users is plain: "Agents extend the capabilities of Microsoft Copilot by letting you customize your Copilot experience to fit your specific needs."
An agent is Copilot with three things attached in advance that a bare chat does not have:
- Instructions — a standing brief describing how it should behave, what it should refuse, what format to answer in.
- Knowledge — the specific SharePoint sites, folders, files or connectors it is allowed to draw on.
- Starter prompts — the suggested questions shown to anyone who opens it, so a colleague understands what it is for without being trained.
Because those three are configured once, the agent produces the same framing on every use and for every user, which is precisely what an ad-hoc chat cannot guarantee.
Where Agents Come From
| Source | What it is |
|---|---|
| Agent Store | The in-product catalogue where you browse and add agents published by Microsoft, external publishers, or your own organization |
| Microsoft-installed agents | Specific agents such as Researcher and Analyst are preinstalled and pinned by default |
| Agent Builder | The lightweight authoring surface inside Microsoft 365 Copilot where a business user builds their own (Chapter 5) |
Two Kinds of Agent
The exam is a business-user exam, so the second row here is background rather than a build target — but the distinction explains why some agents need admin approval and others do not.
| Declarative agents | Custom engine agents | |
|---|---|---|
| What varies | Instructions, knowledge and prompts only | The AI models and the orchestration itself |
| Models used | Copilot's own orchestrator and models | Chosen by the maker |
| How they run | User-initiated interactions | Can trigger actions automatically, without direct user input |
| Built with | Agent Builder, SharePoint, or Copilot Studio | Copilot Studio and the Microsoft 365 Agents Toolkit |
| Availability | Designed for individuals and small teams | Must be published and approved by the organization first |
Chat or Agent? The Decision
| Dimension | Chat experience | Agent experience |
|---|---|---|
| Setup required | None | Configured once, then reused |
| Knowledge scope | Everything you can already reach | Narrowed to the sources it was given |
| Consistency | Varies with how you phrase the question | Same instructions apply every time |
| Who benefits | You, in this session | You and everyone you share it with |
| How you reuse it | Retype or save the prompt | Open the agent |
| Permissions | Your own | Still your own — an agent grants nothing new |
The practical rule fits in one line: a one-off question is a chat; a repeated question over a fixed body of source material is an agent.
| Situation | Right experience | Why |
|---|---|---|
| "What did we agree in yesterday's steering call?" | Chat | Asked once, about one artifact |
| "Answer benefits questions from the approved 2026 handbook" | Agent | Recurring, many askers, one authoritative source |
| "Compare this quote against the signed SOW and last month's invoice" | Chat | Spans three unrelated artifacts, asked once |
| "Give new joiners consistent onboarding answers" | Agent | The value is that everyone gets the same answer |
| "Summarise this 60-page report I just received" | Chat | Nothing recurs; there is nothing to configure |
On the Exam: Two distractor families appear constantly. The first claims an agent lets colleagues reach files they otherwise could not — agents respect existing Microsoft 365 permissions, and if a user has no access to a SharePoint site the agent surfaces nothing from it. The second claims an agent is "more accurate" or "more powerful" than chat. A declarative agent uses the same orchestrator and the same models; it is more consistent and more focused, not more capable.
A human resources department wants to deploy an AI assistant that answers employee questions exclusively using the verified 2026 Employee Benefits Handbook, without pulling unverified drafts or outdated policy files from other SharePoint sites. Which solution is most appropriate?
A user opens Copilot Chat, asks "what are the risks flagged in our Q3 delivery plan?", and receives a generic answer about project risk management with no reference to any company document. The user definitely has access to the plan. What is the most likely cause?
Which statement correctly distinguishes a declarative agent from a custom engine agent?