Most of this track so far has used standard Copilot Chat. This lesson introduces two built-in specialists that live one level deeper: agents Microsoft built and ships with Copilot itself, for the two jobs that a quick chat answer genuinely isn't built for.
Two agents, one job each
Researcher handles multi-step research. Give it a real question and it gathers from both your work content, meaning files, emails, meetings, and chats you have access to, and the public web, then hands back a structured report with headings, visuals, cited sources, key findings, and suggested next steps. It's built for exactly the kind of question a quick chat answer would flatten: one that needs several sources weighed against each other before an answer is worth trusting.
Analyst handles data. Point it at spreadsheets, Excel files or CSVs, and data already indexed through Microsoft Graph, and it writes and runs actual Python code to turn fragmented numbers into forecasts, demand planning, charts, and graphs. The code it writes and runs is visible to you in real time, so you're not taking its analysis on faith.
Neither is a general-purpose chat you nudge line by line. Both take on a defined job, work at it for a while, and come back with something finished, which is the same shift in posture Claude Cowork makes: you stop steering every step and start reviewing a result.
Opening Researcher
Find it under Agents in Copilot Chat and select Researcher. Ask a specific, detailed question rather than a vague one, since the detail is what it researches against. It may ask clarifying questions before it starts, including letting you set the scope explicitly: your work content, the public web, or both. Answer those, because they genuinely shape what comes back.
Response time is part of the deal. A simple question typically takes under five minutes; a genuinely complex one can run ten to forty-five minutes, since Researcher is spending that time retrieving and weighing sources rather than returning the first plausible answer. That's a real trade-off against standard chat's speed, not a bug to route around.
A worked example: "How has our competitor's pricing for the enterprise tier changed over the last two quarters, and what does that suggest for our Q1 renewal conversations?" is the kind of question worth Researcher's time. "What's today's date" or "summarise this one email" isn't, standard chat answers those in seconds and a report format would only slow you down.
Opening Analyst
Find it the same way, under Agents, and select Analyst. Attach the file you want examined, an Excel workbook or CSV is the common case, or point it at data already indexed in your tenant. Ask for what you actually want: a trend, a forecast, a breakdown by category. Analyst writes Python to answer it and shows you that code alongside the result, so if a number looks off, you can actually trace how it was produced instead of just trusting the output.
Say you've got three months of scattered expense spreadsheets, one per region, with slightly different column names. Ask Analyst to combine them and chart total spend by category, and it writes the code to reconcile the columns, runs it, and hands back both the chart and the logic it used to build it. That visible working is the whole point: a forecast you can't check is just a confident-sounding guess.
What they can't do
Both are first-party agents Microsoft built and ships as-is: unlike the agent you built in lesson 4.2, you can't open either one in Agent Builder and customise its instructions or knowledge sources. Researcher can't take an image as input, and it can't yet generate a PowerPoint or PDF directly, so a report you want polished into a deck still needs that last step done by hand. For a job that's really about a spreadsheet rather than open-ended research, Microsoft's own guidance points you to Analyst instead of Researcher.
The query budget
Here's a concrete constraint neither Claude Skills nor Cowork has in quite this shape: Researcher and Analyst share a combined cap of twenty-five queries per user per month. That's not a soft guideline, it's a hard limit tied to your licence, so it's worth saving these two for questions that genuinely need the depth, rather than routing routine work through them out of habit. Standard Copilot Chat has no such cap.
Safety and data boundaries
Both agents run inside the Microsoft 365 commercial data processing boundary and inherit the suite's existing security, privacy, and compliance commitments. Neither is trained on your tenant's data, and neither sends your files or attachments to OpenAI. Researcher's web access is governed by a single tenant-level toggle, on or off, with no setting to allow or block individual sites, and turning it off means Researcher works from your permitted content alone. Your organisation's admins can disable either agent entirely, though Researcher stays reachable under Tools in Copilot Chat even when agents more broadly are turned off, and you personally can't unpin or remove it yourself.
Where it earns its keep against Cowork
It's worth being honest about where this genuinely differs from Claude Cowork, not just renamed. Cowork connects to a folder on your own computer and reads, writes, and builds files in place there, a spreadsheet with live formulas, a deck, a finished document, landing directly where you work. Researcher and Analyst don't reach into a local folder at all: they work against your tenant's content and whatever you attach in the chat, and their output is a report or an analysis inside that chat, not a file written back to a folder on your machine.
So reach for Researcher when the job is genuinely research: weighing several sources, work content and the web together, into something citable and shareable. Reach for Analyst when the job is a real dataset that needs actual computation, not just a glance. For everything Cowork covers instead, building and editing real files in place on your own computer, Copilot simply doesn't have a direct equivalent yet.
That's a genuine gap worth naming plainly rather than glossing over. If what you actually need is "turn this folder of receipts into an expense report, saved back to this folder," Researcher and Analyst won't do that end to end the way Cowork does: Analyst can do the computation on data you attach, but it won't write a finished workbook back into a folder on your computer for you. Knowing which of the two problems you actually have, research-and-synthesize versus build-and-save-a-file, is most of the battle in picking the right tool.
Where this goes next
You've now met two of Copilot's purpose-built specialists. The next lesson, Connectors, widens what any of them, Researcher, Analyst, or an agent you built yourself, can actually reach by linking Copilot to more of your other tools.
Try it
Pick a real question you'd normally answer by digging through a few documents yourself, something with no single obvious source. Open Researcher, set the scope to your work content plus the web, and ask it. When the report comes back, check one cited source yourself: does it actually say what the report claims it says?