Quick answer
Researcher handles complex, multi step questions that need evidence gathered from several places. It combines a deep research model with Copilot's ability to search both the open web and your own work content: files, emails, meetings and chats you already have access to. It takes minutes rather than seconds, and returns something closer to a briefing than an answer.
Analyst works on data. Give it a messy spreadsheet or an export and it will interpret it, find patterns, produce visualisations and explain what it sees in plain language. It is aimed at the moment when you have numbers and no clear picture of what they say.
Both are generally available to Microsoft 365 Copilot users. Both are worth the wait they impose, and neither removes the need to check the result.
Key takeaways
- Researcher is for questions with several moving parts, not quick lookups.
- Analyst is for turning raw data into an interpretation you can act on.
- Researcher reads your work content as well as the web, which is what makes it different.
- Expect minutes, not seconds. That is the point, not a fault.
- Both produce confident narratives, so verify anything specific before it travels.
- The quality of the answer tracks the quality of the question far more than with ordinary chat.
- Bad source data defeats Analyst completely, so tidy the sheet first.
Most people use Copilot as one thing: a box you type into. That undersells what is actually available. Microsoft 365 Copilot now ships two specialist agents, Researcher and Analyst, built for work that ordinary chat handles badly. Knowing which one to reach for, and which questions are worth their time, is the difference between a tool you use occasionally and one that removes a recurring afternoon from your week.
What Researcher is genuinely for
The questions worth giving it
Researcher earns its keep on questions you would otherwise assign to a person for half a day. Things like: what has changed in our sector in the last quarter and what does it mean for our positioning. Or: pull together everything we have said to this client across email and meetings, plus what is publicly known about them, and tell me where the relationship stands.
That second example is the one people underestimate. Because Researcher can draw on your own Microsoft 365 content alongside the web, it can answer questions that no external tool can touch. Nothing else you have access to knows both the market and your own last eleven months of correspondence.
The questions to keep away from it
Anything with a single factual answer. Asking Researcher for a date, a definition or a number wastes several minutes producing an essay where ordinary Copilot chat would have replied instantly. Reach for it when the answer needs assembling rather than retrieving.
How to phrase the request
Researcher responds to structure. A vague question gets a broad, shallow briefing. The pattern that works:
- State the decision the research is feeding. "I am deciding whether to bid for this contract."
- Name the specific things you need to know, as a list.
- Say what you already believe, so it can challenge or confirm it.
- Name the sources you trust and any you want excluded.
- Specify the output shape: a one page brief, a comparison table, five bullet points with sources.
That takes two minutes to write and roughly triples the usefulness of what comes back.
Reading the output properly
Researcher writes well, which is a hazard. A fluent briefing feels authoritative regardless of whether the underlying evidence is strong. Read it in this order: first the sources it used, then the claims that carry a number or a date, then the narrative. If a claim matters and has no source you can open, treat it as a hypothesis rather than a finding.
What Analyst is genuinely for
The moment it fits
You have an export from a system nobody enjoys using. Sales by region for eighteen months, or ticket volumes by category, or attendance across a programme. You need to know what it says before Thursday. Historically that meant a pivot table, several charts and an hour of squinting.
Analyst takes the data, works through it and explains what it found in language you can put in front of people who do not read spreadsheets. It will surface outliers, trends and relationships you had not asked about, which is often where the value sits.
Where it falls over
Analyst inherits every flaw in the source. Duplicate rows, inconsistent category names, dates stored as text, a column that changed meaning halfway through the year: none of these announce themselves. The tool produces a confident analysis of whatever it was given.
Ten minutes spent checking that your categories are consistent and your dates are real dates will do more for the quality of the output than any amount of prompt refinement.
Asking it the right thing
The weak request is "analyse this data." The strong request names the decision and the shape of the answer.
- "Which three regions are underperforming against the same period last year, and by how much?"
- "Is the drop in November explained by fewer customers or smaller orders?"
- "Show me the categories where volume rose but satisfaction fell."
Each of these has a checkable answer. That matters, because a checkable answer is one you can defend in a meeting.
Choosing between them
| Situation | Reach for | Why |
|---|---|---|
| Need evidence assembled from many places | Researcher | It searches the web and your work content together |
| Have a spreadsheet and no clear story | Analyst | It interprets and visualises rather than summarising |
| Need one fact quickly | Ordinary Copilot chat | Both agents are slower by design |
| Preparing for a client or bid decision | Researcher first, then Analyst | Context before numbers |
| Monthly reporting pack | Analyst | Repeatable structure, objective standard |
Making them part of how the team works
The common failure is that one enthusiast uses these agents brilliantly and nobody else knows they exist. Three things fix that.
Publish the good prompts. When someone writes a request that produces an excellent briefing, that request is an asset. Keep a shared list of the ones that work for your recurring questions. This is the single highest return habit available to a team using Copilot.
Agree what gets verified. Researcher output that informs an internal discussion needs a lighter check than output going into a tender response. Say which is which before anyone is under time pressure.
Give it the boring recurring work first. The monthly market summary. The quarterly data pack. Repeated tasks are where you notice the saving and where the standard is easiest to define.
Frequently asked questions
Do Researcher and Analyst cost extra?
They are part of the Microsoft 365 Copilot experience rather than separately purchased add ons, though exactly what your organisation sees depends on its licensing. Longer running agentic work elsewhere in Microsoft 365 does carry usage based costs, which is worth understanding separately.
Can Researcher see files I do not have permission to open?
No. It works within your existing Microsoft 365 permissions, so it can reach the same content you could find yourself. That is reassuring at an individual level and precisely why organisations should tidy up overly broad sharing before rolling Copilot out widely.
How long should I expect to wait?
Minutes rather than seconds for Researcher, depending on the breadth of the question. That feels slow if you are expecting chat. It is the correct trade for work that would otherwise take a person half a day.
Is Analyst better than doing it myself in Excel?
It is faster at the first pass and better at spotting things you were not looking for. It is not a substitute for understanding your own data. The strongest pattern is Analyst for the initial read, then your own check of anything that will drive a decision.
What if the answer is wrong?
Assume it might be, and design for that. Open the citations, sanity check the numbers against something you already know, and never let an unverified figure become a commitment. An agent that saves you four hours and costs you one hour of checking is still an excellent trade.
Practise this with support
For practical work with bounded agents, see Copilot Cowork and Agent Mode in practice. For organisation-wide ownership and controls, see the corporate workshop Governing AI agents with Agent 365.
Author note
Written for AI Vision Consulting, a practical AI training and automation company based in Newcastle upon Tyne and serving UK organisations. We focus on getting real use out of the tools organisations already pay for, with clear habits for checking anything that matters.
Sources
Related articles
More practical guides from AI Vision Consulting.
