Most organizations that pay for Microsoft 365 Copilot are using maybe sixty percent of it. The missing forty percent sits one click away, under Agents in Copilot Chat: Researcher and Analyst, both GA, both included with the license you already have, both routinely ignored because nobody explained when to use them instead of the chat box they’re already typing in.
That’s the question this guide answers first, because it’s the one that matters.
The decision table
Standard chat is a sprinter: one retrieval pass, one answer, seconds. Researcher and Analyst are deliberately slower, multi-step, and they show their work. Use the wrong one and you’ll either wait minutes for what chat does in seconds, or get a shallow chat answer to a question that needed depth.
| Your task | Use |
|---|---|
| ”What did the email say?” / quick summary / draft a reply | Standard chat |
| Synthesize web sources + internal files + meetings into one structured, cited report | Researcher |
| ”Should we enter this market?” — needs reasoning across many sources, not one lookup | Researcher |
| Find patterns, drivers, or anomalies in a dataset | Analyst |
| One formula or chart in a spreadsheet you have open | Excel Copilot |
| Anything where you’ll forward the output to someone who’ll ask “says who?” | Researcher (built-in citations) |
The tell for Researcher: your question has the word and in it at the strategy level — competitors and our internal docs and what was said in meetings. The tell for Analyst: the answer is a number, a ranking, or a pattern, and you’d be suspicious of anyone who produced it without showing math.
Briefing Researcher: the clarifying questions are the job
Researcher runs multi-step research across the web and your work content — files, emails, meetings, chats — and produces a structured, cited report with visuals. It is slow by design; that’s the depth you’re buying. Which means a badly briefed run doesn’t just fail, it fails expensively, in minutes of your time.
A good brief covers three things:
- Scope — what’s in, and explicitly what’s out. “European market only, last 24 months, exclude anything pre-acquisition.”
- Sources — where to look and what to weight. “Prioritize our internal win/loss notes and the analyst reports in /Market-Research; use the web for competitor pricing only.”
- Format — the skeleton of the deliverable. “Five sections: landscape, top 3 competitors, our position, gaps, recommendation. Each claim cited.”
But here’s the part that separates power users from everyone else: Researcher asks clarifying questions before it runs, and answering them well is where quality is decided. Most people treat the questions as a speed bump — “yes, whatever you think” — and then wonder why the report is generic. The questions are Researcher handing you the steering wheel for a multi-minute, multi-source run. Answer with substance: real constraints, real audience, real decisions the report needs to support, names of documents you know exist. Two thoughtful sentences per question routinely turn a mediocre report into the one you actually forward.
If you remember one thing from this guide: the clarifying-questions step is the prompt. Everything before it was just the topic.
Analyst: read the Python, don’t trust the prose
Analyst does data analysis with chain-of-thought Python — it writes and runs actual code against your data, and shows it to you. That visible code is not a nerd garnish. It’s the entire trust model.
Here’s why it matters. The prose summary (“revenue declined primarily due to churn in the SMB segment”) sounds equally confident whether the underlying analysis was sound or garbage. The code can’t hide. You don’t need to be a programmer to do a useful review — you need to check three things:
- What did it actually load? Did it use all the rows, or silently drop ones with blanks? Dropped rows are the #1 source of confidently wrong conclusions.
- What did it group and filter by? If you asked about regions and the code grouped by country, your “regional” insight is an artifact.
- What did it do with missing or weird values? Filled with zeros? Excluded? Each choice changes the answer, and Analyst makes those choices for you unless you tell it otherwise.
Then make it grade itself: ask for a stated confidence level on each finding and what additional data would change the conclusion. Analyst answering “I’m moderately confident, but this dataset lacks tenure data which likely confounds the result” is worth more than any chart it draws.
And the standing caveat: Analyst inherits your data’s problems at full strength. Mixed types in a column, duplicate rows, a header row in the middle of the data — none of these stop it from producing a polished answer. Garbage in, confident garbage out. Clean the file (or at least ask Analyst to profile data-quality issues first) before believing anything downstream.
The workflow angle: agents as building blocks
Here’s the underrated part: both Researcher and Analyst are callable as steps in Copilot Studio workflows via the M365 Copilot node. That reframes them entirely. They stop being things a person remembers to run and become research and analysis as automation steps:
- A weekly workflow that runs a Researcher brief on a named competitor and posts the report to a Teams channel.
- An intake workflow where a submitted dataset gets an Analyst pass — anomalies and data-quality flags — before a human ever opens it.
- A deal-desk flow that assembles Researcher’s account research as the first step of every opportunity over a threshold.
The same briefing discipline applies, only more so: a workflow can’t answer clarifying questions interactively, so the brief baked into the node has to carry the scope, sources, and format on its own. Write it like a skill, not a chat message — see the SKILL.md guide; the craft is identical.
Honest limits
- Researcher is slow by design. Minutes, not seconds. That’s the trade for multi-step depth — but it means it’s the wrong tool for anything you need while someone waits on a call. Queue Researcher runs the way you’d queue a request to a human analyst: brief it, do something else, come back.
- Researcher is only as good as its question budget. Blow off the clarifying questions and you’ve spent the runtime to get a long version of a chat answer.
- Analyst doesn’t know your data is dirty. It has no opinion about whether your export is trustworthy; it analyzes what’s there. Every conclusion is conditional on data quality you have to verify yourself.
- Both inherit the retrieval rules. Security trimming, indexing lag, stale-document ranking — everything from how Copilot actually answers still applies. Researcher reasoning beautifully over the 2022 version of your policy is still wrong.
Start this week
One run each. For Researcher: a real decision you’re facing, briefed with scope/sources/format, clarifying questions answered like they’re the prompt — because they are. For Analyst: a dataset you already know well, so you can judge its code review against your own ground truth. You’ll know within twenty minutes which of your recurring tasks these two should take off your plate — and you’ve already paid for them.