Copilot’s Researcher Agent: Deep, Source-Cited Reports From Your Work Data

What the Researcher Agent Actually Does

The Copilot Researcher Agent is a built-in agent in the Microsoft 365 Copilot app that handles multi-step research and returns a structured, source-cited report rather than a chat reply. It draws on the web and, at work, on the files, emails, meetings and chats you already have access to. Microsoft’s own framing is blunt about the trade-off: it takes longer on purpose, because the extra processing time is what produces the depth.

The practical difference from ordinary Microsoft Copilot chat isn’t intelligence, it’s scope and patience.

What comes back is a document, not a wall of chat text. Expect headed sections, visuals or charts where the data supports them, and a citation attached to each factual claim so a sceptical reader can follow it back. Simple questions finish in a few minutes; genuinely broad ones take considerably longer, which is why the run feels stalled the first time you try it.

One naming note worth having, because it saves a confused search: Deep Research has been retired, and the Copilot Researcher Agent is now the in-depth research experience. If a colleague mentions Deep Research, they mean this.

Who Can Use Copilot Researcher Agent

The Copilot Researcher Agent needs a paid Copilot licence. According to Microsoft’s official Researcher documentation, it’s available to Microsoft 365 Premium subscribers and to business and enterprise users with a Microsoft 365 Copilot add-on licence. Free Copilot Chat doesn’t include it, no matter how much you prompt at it.

This is where most articles get it wrong, so here are the four lanes as they stand in 2026:

The business add-on carries a promotional rate through 30 September 2026, and both business tiers sit on top of a qualifying Microsoft 365 plan rather than replacing it. Check the live pricing page before you build a budget case — this is the fastest-moving fact on the page.

That cap matters more than the price. Microsoft’s Researcher agent FAQ on Microsoft Learn puts it at a maximum of 25 queries per user per month, shared across Researcher and Analyst. There’s no counter in the interface, so you’re tracking it yourself.

Researcher with Computer Use

There’s a second tier worth knowing about, because it answers a question people ask constantly: what happens when the source sits behind a login? Microsoft’s Frontier programme includes a version of the Copilot Researcher Agent with Computer Use, which works inside a virtual environment and can navigate sites the ordinary run can’t reach.

You watch it work — the browser or terminal it’s using, and a preview of what it’s doing — and you confirm actions like signing in to a site when it asks. Your admin controls whether it’s available at all and which sites it may visit, so this is a conversation with IT rather than a setting you flip yourself.

The Scope-Spend-Audit Method

The Scope-Spend-Audit Method is a three-stage habit for metered research tools: scope the question and sources before you type, spend the run on a question worth 25th of your month, then audit the citations before the report leaves your hands. It’s the approach we teach at PromptPeakAI for any AI tool where runs are limited and the output carries your name.

Scoping is not the same as writing a long prompt. It’s four decisions: what will this report be used to decide, which pool of sources answers it, what period counts, and what shape the output needs to take. Get those wrong and you’ve burned a run producing something beautifully formatted and beside the point.

Running Your First Research Report

Here’s the sequence, start to finish. Budget about fifteen minutes for the first one and three or four for every run after.

Sign in to the Microsoft 365 Copilot app and look in the left navigation for Agents. Researcher sits there alongside Analyst. If you can’t see it, that’s a licence or admin question rather than a bug — the troubleshooting table below covers both.

Decide whether this run should read your work content, the web, or both, and say so in the prompt itself. A report that quietly mixes an internal deal note with a public press release, without you having asked for both, is a report you’ll have to untangle later.

Full-shape means role, scope, sources, and output format in one go. This is the same structure behind the AI Execution Loop, and it matters more here than in ordinary chat, because you don’t get to iterate cheaply.

Act as a senior analyst preparing a briefing for [who will read this].

Research question: [the one decision this report has to inform].

Sources: use [work content only / the web only / both]. From my work content, focus on [SharePoint folder, mailbox, or Teams channel] between [start date] and [end date]. From the web, prioritise [type of source, e.g. official announcements and filings].

Produce: an executive summary of no more than 150 words, then [3-5] sections covering [the sub-questions], then a short list of what the evidence does not tell us.

Cite a source for every factual claim. Where you are inferring rather than reporting, say so in the sentence.

The Copilot Researcher Agent often comes back with questions before it starts. Answer them properly — this is the cheapest steering you will get, and saying “go ahead” out of impatience is how a run drifts. Then leave it alone. A pause of a minute or two is the tool working, not the browser hanging.

The difference a scoped prompt makes is not subtle:

Four Prompts for Real Work

These four cover the jobs professionals actually hand to the Copilot Researcher Agent. Swap the bracketed parts and they’re ready to run.

Act as a market analyst preparing a pre-meeting brief for our leadership team.

Using the web, research [Competitor]'s public announcements, pricing changes and positioning shifts between [start date] and [end date]. Then cross-reference against customer feedback in my work content from the same period.

Output: a 150-word executive summary, a table of their moves against ours, the three objections our customers raised most often, and one open question we cannot answer from available evidence. Cite every claim.

Act as a procurement analyst.

Compare the vendor proposals in [folder or mailbox] against our internal requirements document, [document name]. Use the web only to check whether each vendor's stated certifications are current.

Output: a comparison table covering scope, pricing structure, support terms and compliance gaps, followed by the two risks I should raise before signing anything. Cite the file or page behind every row.

Project post-mortems are where Researcher earns its licence fee, because the raw material is spread across months of meetings and threads. If you mostly need the meeting side of that, turning transcripts into structured minutes is the lighter-weight job.

Act as a project analyst reconstructing what happened on [project name].

Using my work content only, review meeting recaps and email threads about [project name] between [start date] and [end date].

Output: a chronological timeline of decisions, scope changes and escalations, with the date, the decision and who owned it. Cite the meeting or email subject for each entry. Where the record is ambiguous or a gap exists, list it separately rather than filling it in.

Act as a compliance analyst.

Using the web, find changes to [regulation or policy area] affecting [country or state] since [date]. Then review our internal policy documents in [folder] and identify which sections are now out of step.

Output: a table of each change, the internal section affected, and the severity of the gap. Add suggested revised wording as a clearly labelled draft. Cite the official source for every regulatory claim, and note that this is a draft for legal review, not legal advice.

How to Audit the Citations

A citation proves a source was consulted. It does not prove the claim reflects what that source says — and the report is going out under your name, not Microsoft’s. Anyone who has read about how AI tools produce confident wrong answers knows the failure mode; the citation layer makes it easier to catch, not impossible to have.

Work through the report in this order, which takes about ten minutes:

1. Every number, every date. Open the cited source and confirm both. These are what people quote back at you.
2. Anything that surprised you. A surprising claim is either your best finding or the error. Check it first.
3. Internal citations especially. Confirm the email or file actually says what the sentence claims, and that it isn’t a superseded draft.
4. Recency on web sources. Check the publication date on anything time-sensitive.
5. Anything uncited. If a sentence carries no source, either find one or cut it.

Let a second model review it

The Copilot Researcher Agent can run more than one model over the same question, and the setting sits at the top of the Researcher screen. Auto is selected by default and is the right choice most of the time.

Where your admin has enabled Anthropic’s models, two other options appear. Critique has the report drafted by a GPT model and then reviewed by Claude, which tightens the structure and pushes key statements towards clearer sourcing. Model Council runs the same question through multiple deep-reasoning models so you can compare what each one surfaced — useful when the stakes justify the extra time, overkill for a routine brief. Neither replaces your own audit; they just mean fewer obvious gaps reach it.

Turning the Report Into Deliverables

A report stuck in a chat thread helps nobody. The Copilot Researcher Agent’s output can move into Copilot Pages, and from there into the formats people actually asked for. Pages sits between the chat and the finished file — it’s where you edit, cut and reorder before anything gets exported.

The export routes, as they stand right now:

That third row is the one to read twice. Microsoft’s support documentation currently states that the previously available option to convert a Copilot Page into a PowerPoint presentation is unavailable — so the deck route today goes through a Word file, which Copilot in PowerPoint can build a presentation from. This particular feature has moved before and will likely move again, so check the Create menu before you promise anyone a deck.

Once the draft is in Word, the tidying is ordinary document work — Copilot in Word handles the rewriting and formatting pass. Keep the citations in the document until the final version is approved; they’re the reason anyone will trust the thing.

What Researcher Can’t Do

The Copilot Researcher Agent only sees what you can already see. It respects your existing permissions, so it will never surface a file you couldn’t open yourself — which means “no documents found” is usually a permissions or storage story, not a failure of the agent.

That permission boundary is also the answer to the question everyone asks quietly:

There’s also the question of which agent you should have opened in the first place. Numbers-heavy work belongs to Analyst, not Researcher — if your real question is about a spreadsheet, Copilot in Excel and the Analyst agent are the better starting point.

Fixing Common Copilot Researcher Agent Problems

Four Copilot Researcher Agent problems come up often enough to be worth naming, and none of them mean the tool is broken.

If a report comes back thin, the instinct is to rerun it. Resist that for a moment — a thin report is almost always a scoping problem, and rerunning the same vague question costs you a second slot for the same result. Rewrite the scope first.

Where This Fits Your Week

The Copilot Researcher Agent is not an everyday tool, and treating it like one is the fastest way to hit the cap by the tenth of the month. Save it for the questions that genuinely span your inbox, your files and the outside world:

• The board paper or quarterly review that pulls from six months of scattered material
• The vendor or platform decision where getting it wrong is expensive
• The post-mortem nobody wants to write because the record is spread across twenty meetings
• The compliance check where a missed update has consequences

Everything smaller belongs in ordinary Copilot Chat, which costs you nothing from the monthly allowance.

If you want to see how the same job plays out in other ecosystems, ChatGPT’s Deep Research mode and Gemini Deep Research follow a similar shape without the work-content grounding. And when the task is really about pulling several documents into one brief rather than researching anything new, synthesising multiple documents is the cheaper move. The method travels: scope it, spend the run, audit what comes back.