Researcher: Your Deep-Work Agent

A regional manager wants a two-page brief by Thursday. It must compare three onboarding options, use last quarter's pilot data, and explain how other companies approach the same problem. That's more than a lookup. It's a compact research project, which is where Researcher is useful.

Use Researcher when the answer needs a trail

Researcher handles complex, multistep investigations in Microsoft 365 Copilot. It searches the organizational content you're permitted to see and the public web, then builds a structured report with citations. Use it when several sources need to be compared, disagreements need explaining, or a recommendation needs evidence behind it. If one file and one good paragraph will settle the question, Copilot Chat is the lighter option.

Grounding sets the evidence boundary for the answer. Work grounding covers permitted organizational material such as projects, policies, decisions, and measurements. Web grounding brings in public material such as industry context, competitors, and standards. Researcher can use both, but they answer different questions. An industry trend can't establish what happened in your organization, and one internal pilot can't establish an industry rule. Ask for separate internal findings and external context so the recommendation doesn't borrow confidence from evidence that belongs elsewhere.

Build the brief around the decision

A topic alone gives Researcher too much room to guess. Start with the four elements from The Anatomy of a Great Prompt, then add the boundaries a research task needs.

State the goal, including the research action and the decision it will inform. Add context about your role, the audience, and why they need the report. Name the source material, including which work files Researcher may use and whether the web is in scope. Set expectations for the report structure, citations, and treatment of gaps. Then define the research-specific limits: scope for the population, product, or geography, a time boundary for the reporting period or "as of" date, and an evidence rule for missing, uncertain, or conflicting information.

Be precise about dates. If pilot results cover April through June and manager interviews took place on July 2, label the interviews as post-period qualitative evidence. Otherwise the report may fold two periods together and make the measurement carry more than it can support. Give web sources a cutoff date for the same reason. With the decision, sources, scope, period, and evidence rule written down, Researcher has less to infer and you have a clear basis for reviewing the result.

Check the claim, not the citation

A citation tells you where a statement came from. It doesn't tell you whether Researcher read the source correctly. A report can link to the right spreadsheet and still describe its numbers incorrectly. For any figure that may affect the decision, compare the wording of the claim with the calculation and population in the source.

Record the claim, source, numerator, denominator, population, and date. Suppose 48 of 60 survey respondents rated a pilot favorably, while 120 people took part in the pilot. "80% of respondents were favorable" is supported because 48 divided by 60 is 80%. "80% of participants were favorable" isn't supported. That wording quietly changes the denominator from 60 respondents to 120 participants, two-thirds of whom didn't answer the survey. The percentage stayed put, but the claim became much broader.

Keep conflicting figures separate. A January forecast of 72% adoption and a June survey in which 46 of 100 respondents express interest were produced at different times, with different methods and populations. Document both, then investigate the uncertainty most likely to affect the decision. Averaging them into "59%" invents precision. The newer number isn't automatically more relevant, just as a cited claim isn't automatically verified.

Spend each query on a defined gap

Researcher work has four stages: scope the question and evidence boundary, draft the first report, refine unsupported claims and conflicts, then finalize from verified evidence. Microsoft's Generate reports with AI research agents Applied Skill assesses the same workflow, so this practice also prepares you for that credential.

The first report is rarely the last one, and refinement uses a limited allowance. Researcher permits a maximum of 25 queries per user per month. Don't rerun the entire report whenever one section needs attention. Ask for one correction at a time, such as recalculating percentages, comparing two conflicting sources without combining them, or rewriting the recommendation from verified rows only. A focused follow-up spends one query on one gap. A vague "try again" spends the same query and may disturb sections that were already sound. If two follow-ups leave the same problem in place, return to the original brief. Its scope or evidence rule probably needs work.

Advanced controls narrow the search

Once the basic loop is second nature, Researcher offers four controls for harder questions. They solve different problems, and none of them turns an unsupported statement into a fact. Your confidence still comes from the source record.

A few caveats worth knowing. Claude only appears when an administrator has enabled Anthropic as a subprocessor, so its absence is a tenant setting, not a failing on your part. Don't burn queries hunting for it. Model Council's consensus is not proof: several models can repeat the same wrong assumption, so the disagreement record matters more than the vote. Critique produces findings. A human still decides what to do with them, and each one remains unresolved until you revise the claim or explicitly keep it as an open question. Computer Use is Frontier-gated and read-only: give it a stopping condition, and never let it sign in, submit forms, or buy anything.