Finding an agent is easy. Using it well is the skill. Learn to judge an Agent Store listing on governance and cost, and to route each task to Chat, Researcher, Analyst, or Excel Copilot.
What you'll learn
Find and add an agent through the Agent Store's Agents → All Agents path
Judge a listing on its governance record. A good description alone doesn't decide it.
Classify an agent's cost as covered, metered, or unresolved before a pilot
Route a task to Chat, Researcher, Analyst, or Excel Copilot by its dominant need
The Agent Store puts an Add button beside almost any listing, including one that promises to "find every team insight automatically." The button answers only whether you can start the access flow. It says nothing about approval, suitability, or the risk of the work you have in mind.
Finding an agent comes first
You discover agents in the Microsoft 365 Copilot app through Agents → All Agents. If direct addition is allowed, the path continues as All Agents → Add. Otherwise the app shows a request-approval flow. An agent may be visible, or even added to your personal view, without being approved or enabled for the team. Personal addition is your action. Tenant-wide use is an administrator's decision.
Treat a listing as the start of an evaluation. If direct addition isn't available, use the request flow shown in the app. When no such flow appears, follow your organization's usual service-request route. Record what happened on screen. A submitted request remains pending until an access decision arrives. Selecting an agent gives you somewhere to begin the conversation, while responsibility for using it and acting on its output stays with you.
Broad access is a stop sign
If a listing you don't recognize asks for access to all your team's sites, chats, and messages, stop there. Hold it for review and escalate.
Read the fields that are actually there
Listings don't all show the same information. Work from the evidence on screen and record any gap instead of completing it from assumption. Capture the exact name, the stated purpose in your own words, and the access action shown, whether that's Add, a request flow, or nothing usable. When publisher status matters, check whether Microsoft built the agent. Write Not shown for any field the listing leaves blank.
Agents built by Microsoft is a separate Microsoft Support directory, outside All Agents. All Agents tells you what's available in your tenant. The directory gives you separate publisher evidence. If the exact name appears there, you can confirm that Microsoft built it. If it doesn't, record it as "unresolved": a failed search doesn't prove an agent isn't Microsoft-built.
Use a short governance check before you trust the description. Does the stated purpose fit a low-risk task? Is the publisher known? Is the requested access proportionate? Can you identify who owns the agent and how it would be removed? The description is a claim about the product. Those governance fields tell you whether the proposed use has support behind it. For a newly added agent, begin with safe, checkable material before any real work data.
Test a new agent safely· copilot-chat
Bad example
Rewrite this announcement to sound more professional.
Good example
Rewrite this short announcement for clarity. Keep every date, amount, deadline, named team, and required action unchanged. If information is missing, point it out instead of inventing a detail. Announcement: "Starting September 1, employees must submit travel expenses within 10 business days of returning. Receipts are required for expenses above $25. Questions should be sent to Finance Operations."
Why this works: A cleaner announcement you can audit fact by fact (same dates, threshold, and team), which tells you whether a newly added agent preserves meaning before you trust it with anything real.
Resolve the cost before a pilot
Store presence doesn't tell you how an agent is billed. Before repeated or team-wide use, put the cost into one of three states. Covered means a billing owner has confirmed the interaction is included in your licensed treatment. Metered means a documented meter applies. The Copilot Studio pay-as-you-go meter, for example, is $0.01 per message. That figure belongs to that meter and can't be carried over to every agent in the store. Use Unresolved when neither treatment can yet be established.
Cost state
What it means
Before a pilot
Covered
Confirmed under your licensed treatment
Keep the license evidence and who confirmed it
Metered
A documented meter applies (e.g. Copilot Studio's $0.01/message)
Record the meter, the billing owner, and an approved usage boundary
Unresolved
Neither is established
Get a billing-owner decision first, do not run the pilot
"Unresolved" is a legitimate thing to write down. Piloting still needs a resolved cost before it starts. You can record it and still be blocked, because the gate for running an agent needs either a confirmed covered treatment or a named meter with an owner and a boundary. The listing often can't answer the cost question. When it can't, get the answer in writing from your license or billing owner.
Unresolved doesn't authorize a pilot
Writing down that an agent's cost is unresolved is the correct record when the evidence isn't there. It still fails the gate to run a pilot. Resolve it with a billing owner before you commit a team to the tool.
Route by input, reasoning, and destination
Ready-made agents overlap because each accepts natural-language prompts. Their useful difference is the work at the center of the task. Researcher, Chat, Analyst, and Excel Copilot suit different inputs and outputs. Before choosing, identify the primary input, the reasoning required, and where the result needs to remain.
If your dominant need is…
Start with…
A quick answer, draft, summary, or brainstorm
Copilot Chat
Multistep research across work and web, with citations
Researcher
Calculations, patterns, or charts from uploaded files
Analyst
Formulas, transformations, or charts inside a workbook
Excel Copilot
The practical test is fast. If a good paragraph finishes the job, start with Chat. If you need several sources compared into a cited report, that's Researcher. If uploaded data needs calculating or charting, that's Analyst. If the result has to live in a workbook (a new formula column, a chart that stays with the sheet), that's Excel Copilot, where current documentation calls the workflow Edit with Copilot (formerly "Agent Mode"). A real task can even chain them: analyze a CSV in Analyst, reproduce the verified calculation in Excel Copilot, draft the explanation in Chat, and open Researcher only if the explanation needs outside evidence.
Keep the result in the workbook· excel
Bad example
Add a Gross Margin column to this workbook and highlight low values.
Good example
In the Sales table, add a Gross Margin column using (Revenue - Cost) / Revenue. Format it as a percentage with one decimal place, highlight values below 20%, and don't change the Revenue or Cost columns. Show me the proposed formula before you apply it.
Why this works: A new margin column that stays in the workbook for further spreadsheet work. The result had to live in the sheet, which made Excel Copilot the right surface.
Match the method to the job
The last habit is the cheapest and most overlooked: reach for the smallest method that produces the evidence you need at acceptable risk. A manager asking for the effective date in one known, accessible policy document needs a manual lookup. Researcher and an Agent Store trial would only add work.
Know when to escalate, too. A job that starts as a one-document date lookup and grows into comparing four policy versions, preserving citations, and explaining conflicting definitions has changed shape. Change the route to Researcher, and record why. A decision trail like "started as a lookup, became a cross-source comparison, moved to Researcher" is what lets a reviewer follow your reasoning later. Pick the smallest tool that works, and re-route deliberately when the task outgrows it.
Draft with a hard source boundary· copilot-chat
Bad example
Write a short status update from these project notes.
Good example
Using only the project notes I've pasted below, draft a 150-word status update for a department director: progress completed, one current risk, two decisions needed, and the next milestone. Don't use web information or invent dates or results, and mark anything missing as needs-input. [paste your notes here]
Why this works: A concise, on-message draft that stays inside your source and flags gaps instead of filling them with plausible invention. Chat is the right surface when the job is mostly writing. Research and math call for a different route.
Try it yourself
Hold or trial this agent?
Governance is a decision you can practise on paper. Evaluate this fictional Agent Store listing your team is considering for summarizing low-risk pilot notes.
01
Read the listing. Name: Team Insight Pro. Description: "Find every team insight and recommend next actions automatically." Publisher: not stated. Requested access: all team sites, chats, and messages. Sharing, lifecycle owner, removal process, evidence, and policy status: not stated.
02
Score it on seven criteria: task fit, publisher, access, sharing, lifecycle, evidence, and policy. Write "Not shown" where the listing is silent rather than guessing a favorable default.
Hint: Broad access plus an unknown publisher is the combination that should stop you.
03
Choose one disposition (Reject, Hold for review, or Trial with conditions) and justify it in one sentence.
04
If you'd ever trial it later, list the conditions first: one narrow task, fictional data, minimum access, success criteria, an owner, a review point, and a stop condition.
A "Hold for review" decision backed by the unresolved governance fields and the broad-access request.
Key takeaways
Discovering or even adding an agent doesn't approve it. Tenant-wide approval is an administrator's decision.
Record only what a listing shows, and write "Not shown" for the rest. Judge it on governance, not its description.
Classify cost as covered, metered, or unresolved. An unresolved cost blocks a pilot even when it's accurately recorded.
Route by dominant need: Chat to draft, Researcher to synthesize, Analyst to calculate, Excel Copilot to keep results in a workbook.
Use the smallest method that produces the evidence, and re-route deliberately when a task outgrows it.
Check your understanding
1. Where do you confirm whether an agent is built by Microsoft?
2. A discoverable agent promises to "find every team insight," but its publisher, sharing, and lifecycle owner are unknown and it requests access to all team sites and chats. What's the best decision?
3. You read that a Copilot Studio meter is $0.01 per message. What does that tell you about the cost of an arbitrary Agent Store agent?
4. A manager wants the effective date from one known, accessible policy document, and the answer must be directly verifiable. What's the best starting point?
5. You need to add a margin formula column that stays in an existing workbook for further spreadsheet work. Which surface fits best?
Frequently asked questions
Terms used in this lesson
Agent Store
The Microsoft 365 Copilot surface for discovering agents, reached through Agents → All Agents.
agent
A task-focused AI assistant used through Microsoft 365 Copilot. Selecting one gives you a starting point. The responsibility for the result stays with you.
source boundary
A statement of which information an agent is allowed to use: the files or content in scope, and whether the web is included.
surface
The app or agent in which you run a prompt (Chat, Researcher, Analyst, or Excel Copilot), each with a different strength.
cost state
How an agent's billing is classified before use: covered, metered, or unresolved.
