Meet AI and the Copilot UniverseBeginner13 min

Which Copilot Do You Need?

A one-minute method for routing any task to the right Copilot, knowing when a job needs an agent, and verifying an answer before you trust it, the habit that matters most.

What you'll learn

  • Route a task to the right Copilot in under a minute
  • Tell a one-off chat apart from a job that needs an agent
  • Name the three ways Copilot output goes wrong
  • Decide whether to use, edit, hold, or discard an answer

Before you start

On this page

You know the family. The question you'll face in practice is simpler: which Copilot fits this task? In most cases, the location of the evidence gives you the answer in under a minute.

A one-minute way to pick a Copilot

Leave the product names aside for a moment. Describe who's doing the work, the result they need, and where the trustworthy information lives. The evidence location usually decides the surface. Start with the Copilot closest to both the source and the job.

Your job Start with Why
Personal questions or personal content Consumer Copilot The account and content are personal
Work question needing only public or pasted info Copilot Chat No work data required
A draft or summary that needs your permitted work files Microsoft 365 Copilot The answer depends on work-grounded evidence
Writing, explaining, or debugging code GitHub Copilot The work lives in a repository
Investigating a security alert Security Copilot The work centers on security operations
A reusable assistant other people will use Copilot Studio (an agent route) It's a governed solution designed for reuse

The evidence location points to the surface. Company files and email lead to Microsoft 365 Copilot. A code repository leads to GitHub Copilot, while security telemetry leads to Security Copilot. When two options seem to fit, ask which system actually holds the source of truth.

Take a real one: an operations manager needs a weekly launch-readiness brief built from the team's project files, meeting notes, and status emails. Personal? No. Answerable from public info? No, it lives in permitted company files. So the answer is Microsoft 365 Copilot, with one access check attached: does she have permission to open those files? A license doesn't repair missing permissions, and the surface is only right if the evidence is reachable.

Two questions settle most of it

Where does the trustworthy information live? That picks the surface. Is this a one-time answer or something you'll reuse? That decides whether you need a chat or an agent. Answer those two and you've routed the task.

Let Copilot help you route the task· copilot-chat
Bad example

Which Copilot should I use for my task?

Good example

I need to [describe your task in one sentence]. Ask me up to three questions to identify the audience, the result I need, and where the authoritative information lives. Then recommend which Microsoft Copilot to start with and name one thing that would change your recommendation.

Why this works: A short back-and-forth ending in a recommended Copilot plus the condition that would switch it: a decision you can defend, not a guess.

When do you need an agent instead of a chat?

Most tasks belong in a conversation. You open Copilot, describe the result, and work with the answer. An agent becomes useful when a saved setup will handle the same job for many people, such as recurring policy questions, an intake form, or a question asked fifty times a month.

An agent stores a role, instructions, the knowledge it may use, and sometimes tools it can call. Other people can reuse that setup without explaining the task again. Complexity isn't the deciding factor. A single chat can take a transcript and a rulebook and produce a detailed analysis. It remains a one-time conversation unless the setup is saved for reuse.

Begin in chat. Once repeat use is clear, Microsoft offers Agent Builder inside Microsoft 365 Copilot Chat for simple cases and Copilot Studio for more control. You'll meet both later in the course. For now, use chat for a one-off request and consider an agent when the team will repeat the task.

An agent is a saved setup, not a fancy prompt

Ask one thing to tell them apart: could I hand this to a colleague to reuse tomorrow without re-explaining it? If the role, instructions, and knowledge are saved for reuse, it's an agent. If it lives and dies in one conversation, it's a chat, however clever.

Three ways Copilot can go wrong

A confident tone says nothing about correctness. Learn these three failure modes and you'll know what to look for:

  • Fabrication: a plausible claim that isn't supported by your source: a date that never appeared in the notes, a number copied wrong, a "next step" the source only implied.
  • Bias: uneven wording, emphasis, or omission: reporting four criticisms while dropping three compliments, or turning "some comments" into "everyone thought."
  • Staleness: information that was true once and isn't now: an old deadline, a renamed product, a superseded policy.

Reviewing asks whether the answer addresses the task, uses the right tone, and follows the format. Verifying asks whether each important claim appears in an allowed source. A polished draft can pass review and still fail verification because of one number. Before you send, decide from, or act on an answer, locate the support for each important claim. If sources conflict, use the newest authoritative one. When it leaves a point unresolved, mark that point "needs confirmation" instead of carrying an older rule forward.

Spend your checking time according to consequence. Skim a low-risk detail. Confirm a medium-risk quantity or owner. Don't use a high-risk claim that could create a false deadline, commitment, or decision until you've checked it. Evidence and risk are separate questions. A number can be copied correctly and still deserve close attention when a decision depends on it.

Force an honest blank instead of a guess· copilot-chat
Bad example

Write a status update from these notes: [paste notes].

Good example

Draft a status update for my manager using only the notes below. Use the headings Progress, Risks, and Next steps. If a date, owner, or number isn't in the notes, write "Not stated in the source" and do not infer or fill it in. Notes: [paste notes].

Why this works: An update where every fact traces to the notes and every gap is labeled, so you can see at a glance what's known versus what still needs chasing.

Make Copilot audit its own draft· copilot-chat
Bad example

Check the draft above and fix anything wrong.

Good example

Audit the draft above against only the notes I gave you. For each factual claim, quote the supporting line from the notes or write "Not found," rate the risk if it's wrong as low, medium, or high, and say whether to keep, fix, or remove it. Don't add any new facts.

Why this works: A claim-by-claim table that surfaces unsupported statements and ranks them by consequence: your checklist for what to verify yourself.

Use, edit, hold, or discard

Once you've checked an answer, put it into one of four buckets and move on:

  • Use: the important claims are verified, current, and fairly framed. Ship it.
  • Edit: the facts hold up, but the wording or format needs your hand.
  • Hold: one important claim can't be confirmed yet. Don't act until it is.
  • Discard: it can't be reliably repaired from what you have. Start over.

A classic trap hides in that decision: watch what each date modifies. "Send the update before Monday" is a reporting deadline, not proof the event itself happens Monday. Treating one as the other is the kind of fabrication that slips past a careless read and into a calendar.

Match the effort to the stakes. A throwaway brainstorm barely needs a glance. Anything touching money, deadlines, commitments, people, or compliance earns a real check. Then place the answer in Use, Edit, Hold, or Discard based on what the source supports.

Raise the bar with the stakes

When an answer affects money, a deadline, a commitment, a person's standing, or a compliance rule, treat every date, number, and name as unverified until you've matched it to a source. A single wrong figure in a polished paragraph is the failure mode that costs you.

Try it yourself

Route three real jobs

Turn the decision method into muscle memory on your own work. About eight minutes, no license required.

  1. 01

    Write down three real tasks from your world, for example, "summarize our permitted project files for a status email," "debug a failing test," and "investigate a suspicious sign-in."

  2. 02

    For each, answer the three questions: who's doing it, what result they need, and where the authoritative information lives.

    Hint: The evidence location usually decides it. Company files, a code repo, or security telemetry each point somewhere different.

  3. 03

    Pick the starting Copilot from the decision table, and name one change that would move you to a different one.

  4. 04

    For the one you'd do first, write the one claim in its answer you'd verify before sending, and where you'd check it.

Three tasks routed with a defensible reason each, plus a verification plan for the one you'll run first: the full module skillset in one page.

Key takeaways

  • Route by where the authoritative evidence lives. The nearest fit usually wins.
  • A one-off answer is a chat. A task many people reuse is a candidate for an agent.
  • Complexity doesn't make something an agent. A saved, reusable setup does.
  • Copilot goes wrong three ways, fabrication, bias, and staleness, so verify instead of only reviewing.
  • After checking an answer, decide to use, edit, hold, or discard it.

Check your understanding

  1. 1. A security analyst must investigate a suspicious mailbox rule using alerts and telemetry. Which Copilot is the stronger first choice, and why?

  2. 2. An employee pastes a transcript and governance rules into one chat and asks for a detailed action register. It's complex but won't be reused. Is this an agent?

  3. 3. A draft says a make-up session happens "on Monday," but the notes only say the status update is due before Monday. What kind of error is this?

  4. 4. An older policy allowed exceptions. The newer version is silent about them. How should a current summary treat the exception process?

  5. 5. You've verified an answer, but one important claim still can't be confirmed. Which decision fits?

Frequently asked questions

Terms used in this lesson

agent
A saved, reusable Copilot configuration (a role, instructions, and allowed knowledge) that people use repeatedly instead of re-explaining a task each time.
authoritative evidence
The trustworthy source a result must be checked against. Its location usually decides which Copilot to use.
staleness
When information was accurate once but no longer applies: an old deadline, a renamed product, a superseded policy.
verification
The evidence check: confirming each important claim appears in an allowed source, as opposed to a broad review of tone and format.

Further reading