The Copilot Analyst agent: what it does, the data it uses, and where it fails
Analyst is the data half of the built-in Microsoft 365 Copilot agents. Where Researcher writes a brief, Analyst works with numbers and tables. This page covers what it is built to do, the data it needs, the failure modes to expect, and how it differs from its sibling. It is the second of our two deep dives; the shortlist is at the best Copilot agents. For the concept, see what a Copilot agent is.
How to read this deep dive
We have not run a scored, reproducible lab test of Analyst, and this page does not pretend otherwise. The material here is drawn from Microsoft's documentation and public first-hand write-ups, with the specific sources named. The worked example is an example prompt, not a transcript of a test. Where a limit is described, it is described as plan-dependent rather than quoted, because Microsoft changes these figures.
What Analyst is built for
Analyst is a reasoning agent for data. You give it a dataset and a question, and it explores the data, works through the analysis, and explains what it found and how. Microsoft pairs it with Researcher as the two advanced reasoning agents in Microsoft 365 Copilot; the distinction is that Analyst is about working with data rather than gathering written sources. The details are in Microsoft's getting started with agents material and the agents overview.
The useful framing is exploration. Analyst is good at the first pass over a dataset: what is in here, what stands out, what does this segment look like, what changed. That first pass is often the slow part of analysis, which is where the time saving comes from.
The data it needs
Analyst is only as good as the data you hand it, and it is happiest with clean, tabular data. Before you give it anything, make sure:
- There is one table. One dataset per sheet or file, with a single header row.
- The headers are unambiguous. "Revenue" beats "Rev.", "Q3 2026" beats "Q3".
- The columns are consistent. No merged cells, no mixed units, no text where a number belongs.
- It can read the format. A supported table format beats a scanned image every time.
- You have permission. It works on what you can access and no more.
A worked example
A task shape Analyst handles well is a reconciliation. For example:
"Here is a monthly sales sheet and a list of invoiced accounts. Find the accounts that appear in one file but not the other. Group the differences by region, show the totals, and explain the steps you took so I can check them."
Asking for the steps is the part that matters. It turns a black-box answer into something you can audit, and it makes the agent's assumptions visible before you act on the totals.
Where it fails
- Messy data. Merged cells, hidden rows, and inconsistent headers are the most common cause of a wrong answer.
- A wrong assumption carried through. If it misreads what a column means, the whole analysis follows the mistake and still looks confident.
- Silent data limits. Very large data may be sampled or truncated, and the analysis may not flag it.
- It is not the record. Its output is an analysis to verify, not a system of record. The maintained model should stay in a sheet you control.
- It is capped. Usage is limited on eligible plans, and the allowance is shared with Researcher.
- It is plan-dependent. Whether you have Analyst depends on your licensing and your organization's configuration.
Analyst and Researcher
They are meant to be used together and they are easy to mix up. Researcher gathers and reasons across sources and returns a written brief. Analyst works with numbers and returns an analysis. If the task is "what do these numbers show?", use Analyst. If it is "what does the evidence say?", use the Copilot Researcher agent. The two share a usage allowance, so choosing correctly also protects the budget.
Getting better results from Analyst
- Give it the smallest clean table that answers the question.
- State the units, the time period, and what a row represents.
- Ask it to show its steps so you can audit the method.
- Split a multi-part question into separate passes.
- Verify the totals you will act on against the source data.
If your question is really about fitting data work into a working day, rather than about this one agent, see Copilot agents for work.
Analyst agent: common questions
What does the Copilot Analyst agent do?
Analyst is built for data. It explores a dataset, summarizes it, segments it, and explains its working, returning a structured analysis rather than a single number. It reasons over the data it is given and shows the steps behind its answer.
What data does Analyst need?
It does its best work on clean, tabular data with clear headers, ideally one table per file or sheet. Messy headers, merged cells, and data split across many files make it less reliable. Give it the smallest clean dataset that answers the question.
Where does Analyst fail?
It can misread messy or ambiguous data, it can carry a wrong assumption through a confident-looking analysis, and it can miss a limitation in the data. Check its numbers and its method against the source before you use them.
What is the difference between Analyst and Researcher?
Analyst works with data and numbers; Researcher gathers and reasons across written sources. Analyst answers "what do these numbers show?", Researcher answers "what does the evidence say?". They are designed to be paired and they share a usage allowance.
Can Analyst replace a spreadsheet?
No. It is an analysis assistant, not the record. It is excellent for exploring and explaining, but the model you control and audit should stay in your spreadsheet. Use Analyst to get to an answer faster, then keep the maintained version yourself.
Is the Analyst agent available on my plan?
Availability depends on your Microsoft 365 plan and licensing, and usage is capped. The cap applies to Analyst and Researcher together. Check your own plan and Microsoft documentation, because both change.
TopCopilotAgents is independent and not affiliated with Microsoft. General information only; confirm current behavior and limits with Microsoft.