Standard Copilot chat is built for speed. Microsoft says so itself: it is "optimized for speed and efficiency and is great for quicker tasks like summarizing emails or drafting short replies." Two agents in the same app are built for the opposite job. They take longer, think harder and return something you can hand to a colleague.
They are Researcher and Analyst. This guide explains what each does, when to use which, what is confirmed about licensing, and the one limit we could not verify. If you are new to the product, start with our beginner's guide to Microsoft 365 Copilot.
TL;DR: Researcher vs Analyst
| Question | Researcher | Analyst |
|---|---|---|
| What it does | Multi-step research, cited report | Turns data into insights and charts |
| Typical input | A question, plus your scope | A spreadsheet or table |
| Sources | Web and work content you can access | Your data, plus work and web data |
| Visible work | Citations and clarifying questions | Python code you can inspect |
| Best for | Briefs, market scans, client reports | Forecasts, trends, joined sheets |
| Not for | Quick replies | Anything you cannot check |
| Licensing | Microsoft 365 Copilot licence announced; current limits unconfirmed | Same |
What does the Researcher agent do?
Microsoft's Researcher agent documentation says it handles complex, multi-step research tasks and delivers comprehensive, source-cited reports. It "draws insights from both the web and your work content - including files, emails, meetings, and chats you have access to." You find it under Agents in Chat in the Microsoft Copilot app.
The page lays out when to pick it:
| Use this when you... | Choose |
|---|---|
| Need a quick summary, a short reply, or lightweight brainstorming | Standard Copilot chat |
| Need deeper reasoning across multiple sources (web and work) and a shareable report | Researcher |
| Want citations and a clean structure with headings, bullets and visuals | Researcher |
| Have a back-and-forth task without heavy analysis | Standard Copilot chat |
Microsoft frames the slower pace as a design choice: Researcher "deliberately spends more time retrieving and analyzing to write a more in-depth analysis." The final answer can include charts and graphs, organised sections and cited sources.
Microsoft's launch post from March 2025 adds history. It said Researcher combines OpenAI's deep research model with Copilot's orchestration and search, and can use third-party data through connectors such as Salesforce, ServiceNow and Confluence. Its examples were a go-to-market strategy, finding product opportunities from trends and internal data, and a client quarterly report. Models change over time, so treat that as launch-era detail, not a current specification.
An agent cube with binoculars facing a web of pages, showing the Copilot Researcher agent gathering sources from work content and the web
How do you get a good Researcher result?
Microsoft's tips are three, and they are the right three:
- Be specific. Clear, detailed questions lead to more focused and reliable results.
- Define your scope. Say whether you want workplace resources, the web or both.
- Answer the clarifying questions. Microsoft says this phase is important to steer the research and keep the report on target.
Prompt: "Research how our three main competitors price onboarding and implementation. Use the web and /CompetitorNotes.docx. Exclude anything older than 2025. Return a one-page report with a comparison table, sources, and the three biggest unknowns."
Our guide to writing Copilot prompts explains the goal, context, source and expectations pattern that makes this prompt work.
What does the Analyst agent do?
Microsoft describes Analyst as an agent that can "transform complex data into clear insights and visualizations" inside Copilot Chat. The March 2025 launch post said it "thinks like a skilled data scientist," uses chain-of-thought reasoning, works through problems iteratively, and can run Python for complex data queries. You can see the code as it runs and check its work. Examples were combining several spreadsheets into a demand forecast, a customer purchasing visualisation and a revenue projection.
The launch post said Analyst was built on OpenAI's o3-mini reasoning model. That detail dates from 2025, and we did not find a current model statement, so do not rely on it.
A green thread following four beads beneath a magnifying lens, showing Copilot Analyst code and steps you can inspect to check its work
The visible code is the feature to use. A chart without its method is a screenshot. A chart with its Python is something you can audit. Ask for it every time.
Prompt: "Using the attached sales workbook, find the three regions with the largest quarter-on-quarter decline. Show the Python you used. List any rows you excluded and why. Suggest one chart."
Our guide to Copilot in Excel covers the in-sheet route. Choose Analyst when the question spans several files or needs more than a formula.
We did not find official file-size or format limits for Analyst on the pages we read. A Microsoft Tech Community post shows adding it from the Agent Store and uploading an Excel file, but lists no limits. Test with your largest real file before you promise a team it works.
When should you use Researcher, Analyst, chat or Cowork?
| Task | Best fit |
|---|---|
| Summarise this email thread | Standard chat |
| Draft a reply | Standard chat |
| A cited brief on a market or topic | Researcher |
| Combine three sheets and forecast demand | Analyst |
| Build a weekly report across files and send it for approval | Copilot Cowork |
| Run something on a schedule | Copilot Studio |
Researcher and Analyst work inside a single request that you start. They are not standing automations. Our piece on repeating work in Microsoft 365 Copilot explains where the line sits and why scheduled work belongs in Copilot Studio.
Are Researcher and Analyst generally available, and what do they cost?
This is the part with the most noise, so here is what we could verify.
- Availability. Microsoft's agents in Microsoft 365 page lists both as generally available inside Microsoft 365 Copilot Chat. It gives no date. The March 2025 launch post carries a banner saying both are now generally available. A Collab365 article reports that five agents, including Researcher and Analyst, reached general availability on June 26, 2026, with Project Manager still in public preview. We could not match that date to a Microsoft page, so we do not rely on it.
- Licence. The launch plan was Microsoft 365 Copilot licence holders. Thurrott's general-availability report says anyone with that licence can use them. The Learn and adoption pages we read do not mention licensing or limits at all. Our pricing and licensing guide covers the add-on price and the free Copilot Chat tier.
- Query limit. Thurrott reported 25 combined Researcher and Analyst queries per month for licensed users at general availability, and 37 languages for Researcher against eight for Analyst. That is 2025 reporting. We found no current Microsoft page that states a cap, so we cannot say whether it still applies.
- Ignite 2026. Microsoft's Office at Ignite 2026 post did not mention either agent.
The honest advice: before you roll Researcher out to a team, run a pilot, count how many runs a user actually gets, and ask your Microsoft account team to confirm the current limit in writing.
What are the risks?
Wrong sources. A cited report is not a verified report. Open the citations for any number or quote that will leave your team. Microsoft's own training says AI-generated content can be inaccurate and sources are provided for review.
Permission leaks. Microsoft says Researcher respects the permissions you already have. That is good, but if a user can open a file they should not, Researcher can use it. This is why admins should review oversharing before a wide rollout. See our Copilot readiness checklist.
Prompt injection. Researcher reads emails, files and web pages. Any of them can carry instructions meant for the model. We explain a documented case in the Word document AI worm and the ChatMate remote prompt execution research. Keep a human approval on anything that acts.
Web leakage. Microsoft's Purview DLP documentation says a policy can block Copilot and Cowork from using external web search when a prompt contains sensitive information types. Ask your admin whether that rule is on before you paste regulated data into a research prompt.
A short pilot plan
- Pick three questions your team already answers by hand each month.
- Run each in Researcher with a defined scope. Record time, citations opened, and errors.
- Pick two spreadsheet questions. Run each in Analyst, and read the Python.
- Compare with the manual result. Keep a use only if the checked output wins.
- Write the prompts into a shared library.
If you want a guided run through Copilot, Cowork and agents on your own work, our Microsoft Copilot and Copilot Studio workshop runs live on December 20 and 21, 2026.
Related reading
- What is Microsoft 365 Copilot? A beginner's guide
- Microsoft 365 Copilot pricing and licensing
- Copilot Cowork explained
- Microsoft Copilot in Excel
- How to write prompts for Microsoft 365 Copilot
- Repeat work in Microsoft 365 Copilot
- Agent Builder vs Copilot Studio
- Copilot readiness for admins
- Official: What is Researcher agent in Microsoft Copilot?, Agents in Microsoft 365
Product behaviour reflects Microsoft documentation as of November 19, 2026. Model names, the 25-query limit and language counts come from 2025 launch coverage and were not confirmed on a current Microsoft page. Check your tenant and your Microsoft account team before you plan around them.
