Join me exploring Microsoft Copilot through each letter of the alphabet. This week, R is for Researcher agent.

I’ve mentioned this one a few times now but it deserves its own spotlight, so here we go. 

A quick reminder: what is Researcher agent?

Standard Copilot is great for a quick summary. Researcher is what happens when Copilot puts on a scuba suit and goes deeper diving. 

It’s one of the built-in agents you get with a paid Copilot licence, and it’s designed for “deeper reasoning”. You’ll give it a brief, it will go away, do some proper legwork across your internal work data and the web, and come back with a structured, cited report.

Think less “answer my quick question” and more “go and actually research this for me and write it up.”

You can find it under ‘more agents’ on the left hand nav: 

What I actually wanted to do with it

In my last post, Q is for Questions, I talked about end-result thinking. When using tools like this it’s important to know what you actually want before you start prompting. So let me practise what I preach.

I co-run a quarterly Microsoft Copilot and Cloud User Group, and we’ve got two goals we’re stuck on: reaching genuine end users outside our existing networks, and getting more women through the door. 

What I wanted was a proper, practical plan for doing that. Not a generic “here are some tips” list, but something grounded in what’s working for comparable groups.

Set it off on the right path 

Before it gets started, Researcher will often check in with you. It’ll ask clarifying questions to confirm the scope. It will also give you options for the results you get back. Do you want a short report 1-5 pages? Or a long one with 5+ pages? You can steer this.

This matters more than it sounds. A quick question doesn’t need a twelve-page report, and a big strategic piece of work is badly served by three bullet points. Telling Researcher the depth you’re looking for up front is really just end-result thinking again; being clear about what “done” looks like before it starts.

It will also ask other questions about the response “do you care more about x over y?”, “is your main concern a, b, c, or a mix of all three?”. This helps you further hone the results you get back.

The bit I want you to steal: Model Council

As usual, Microsoft has buried a really useful feature under a dropdown somewhere – it’s called Model Council and it is worth finding!! (Today you’ll need to be on the frontier programme for this.) 

Quick recap: Copilot doesn’t have its own AI brain. It borrows models from OpenAI (GPT) and Anthropic (Claude). When you’re using Researcher, look out for the ‘Auto’ dropdown in the top right that lets you choose which model handles your query: Auto, Claude, or Model Council.

Worth noting – you need to make this selection before you input your prompt! 

Model Council uses both GPT and Claude models in parallel. You get a side by side comparison of responses from each, plus an easy to digest summary of where GPT and Claude agreed, and where they differed.

When it kicks off you can literally see the models reasoning next to each other: 

If you’re using AI generally, I’d say it’s best practice to compare responses between tools as standard. Remember that AIs are built by different companies with different access to different data (and biases) – so if you want a well-rounded picture (who doesn’t?) you need to be comparing responses between tools. Model Council in researcher agent is a great shortcut to doing that for you without having to flip between each tool yourself. 

For my user group plan, both models strongly agreed on the big rocks: get off LinkedIn-only and onto Meetup and Eventbrite, use explicitly inclusive language (“women and first-timers especially welcome”), feature a female speaker, run a beginner track alongside the technical stuff, and partner with women-in-tech communities. When two different AIs, built by two different companies, trained differently, both land firmly on the same advice, that’s a decent signal it’s worth listening to.

But the disagreements were arguably more useful. One leaned towards launching every channel at once in an aggressive week-one sprint; the other preferred phasing it out over 30-day blocks. They differed on whether to pay for Meetup premium or stick to free tools, and on how much to prioritise an ongoing community chat between events. None of those has a single “right” answer, those are judgement calls. But having them flagged as the open questions meant I knew exactly where I needed to apply my own brain and discuss with co-organisers, rather than just accepting one confident-sounding plan.

Did it actually help? 

There’s nothing more frustrating than when the output of an AI tool is either so generic you have to go off and do your own research anyway, or takes you so long to get to the end result you wanted that you’d have been better off not using it in the first place. This was different though. 

In LBC (Life Before Copilot) I know I would have set off on an expedition hunting around Google, ending up down various rat holes looking at other user groups, with 52 browser tabs open, and at best a couple of half-baked new ideas. Researcher did all the trawling for me and handed back something structured, with the comparable groups already identified and the tactics already pulled out.

Was every single detail perfect? No. But as a way to go from “blank page and a vague goal” to “a real plan I can pressure-test and act on,” it did exactly what I hoped.

TLDR

Researcher is Copilot in a scuba suit – diving deeper. Give it a proper brief and it’ll go away, research across your work data and the web, and come back with a structured, cited report. You control how deep it goes, so match the depth to what you actually need. 

If you take one thing from this post: use Model Council. It runs your query through both GPT and Claude and shows you where they agree (trust it) and where they differ (apply your own judgement). It’s a built-in second opinion, and comparing AI tools is something you should be doing anyway. Just remember to check what comes back. It’ll be a brilliant starting point, not the finished job.

Have you used Researcher for anything meaty yet? And has anyone else played with Model Council? Did the agreements and disagreements between the models change how much you trusted the output? Let me know in the comments.

Useful links:

Get started with Researcher in Microsoft 365 Copilot Researcher in Microsoft 365 Copilot (where Model Council lives) https://support.microsoft.com/en-us/microsoft-365-copilot/get-started-with-researcher-in-microsoft-365-copilot 

And, since it came up — if you’re ever in London and fancy hanging out with other Copilot-curious folk: https://mccug.co.uk/