Ready-Made AgentsIntermediate14 min

Analyst and Coach Agents

Two ready-made agents, one discipline. Analyst turns your files into charts and totals you can reconcile. Coach agents help you practise and draft, and both leave the verifying and the deciding to you.

What you'll learn

  • Attach a file and give Analyst a prompt you can check
  • Independently reconcile an Analyst total instead of trusting the chart
  • Use Coach agents for practice while keeping every decision with a person
  • Tell a supplied fact from a generated suggestion in any agent's output

Before you start

On this page

One colleague sends you a 5,000-row CSV and asks "what's going on with March," while another wants to rehearse a difficult update to leadership. Microsoft 365 Copilot has a ready-made agent for each job: Analyst for the file and the Coach family for the rehearsal. In both cases, the agent helps with the work while you remain responsible for the result.

Different jobs, same checking habit

Analyst and the Coach agents do very different things. Analyst runs Python over your files and returns tables, charts, and totals. A Coach agent works with a prompt, draft, or plan. Yet both can produce an answer that looks sound while hiding a bad assumption. Analyst may calculate the wrong column perfectly. A coach may name an owner nobody assigned. You won't necessarily get an error message.

Separate the material you supplied from anything the agent generated. Then verify every part that could influence a decision. That habit runs through the rest of this lesson.

Know what Analyst is doing

Analyst is a generally available Microsoft 365 Copilot agent included with a Microsoft 365 Copilot license. Attach a file and ask a question in plain language. Its code interpreter runs Python against the data and returns tables, charts, or grouped summaries. You don't write the code, but you still need to check the calculation. Code can run cleanly against the wrong field.

The documented supported file types show the range of material you can provide:

File type Good first task
Excel workbook Summarize sales, expenses, or survey results
CSV / TSV Analyze a table exported from another system
Word document Extract and organize information from a report
PowerPoint Compare or summarize slide content
PDF Analyze a supplied report
TXT / UTF-8 Analyze logs, notes, or delimited text

For numbers, structured files like Excel, CSV, and TSV are the easiest starting point because their rows and columns are explicit. For a document, ask Analyst to identify the relevant content first, before you request any calculation. Keep a simple four-part model in mind while you work: the input (your file and question), the method (the Python Analyst chooses), the output (its tables and charts), and the verification (your checks against the source). Analyst owns the first three. You own the fourth. Skip it and the first three don't add up to anything you can use.

Inspect before you calculate· copilot-chat
Bad example

Look at the attached sales file and tell me if anything is wrong.

Good example

I've attached a sales file. Before calculating anything, inspect it and report: the column names and count, the apparent data type of each column, the number of data rows, any blank cells and where they are, any duplicate rows, any values that look invalid, and how you interpreted the Month column (text, date, or other). Don't invent missing values, and don't interpret business performance yet.

Why this works: A structure-and-quality readout you can check against the file you meant to attach, so you catch a wrong file or a misparsed date before it poisons every later chart.

Ask for a method you can inspect

"Make a chart from this file" leaves Analyst to choose the measure, grouping, calculation, and treatment of missing values. A checkable prompt names those choices. Specify the numeric field, grouping fields, aggregation such as sum or average, any filter, the output format, units, and missing-value rule. "Create a line graph of the sum of Revenue by Month, ordered chronologically, with the y-axis in dollars" gives you something concrete to compare with the result.

Ask for the summary table behind every chart. Then reconcile it yourself. Monthly revenue of 26,750, 38,900, and 41,250 must total 106,900. A regional breakdown of the same rows should reach that same figure. If it doesn't reach 106,900, rows may have been dropped or counted twice.

Some clients show Analyst's generated Python. Read it as you would a colleague's spreadsheet formula. Check the file, columns, date parsing, aggregation, and chart labels. If you can't see the code, ask Analyst to explain its method and show the rows it used. Keep the interpretation within the data as well. Higher revenue in one region establishes a difference. It doesn't establish a cause. "West led because its team ran a better campaign" still needs campaign data in the file.

A tidy chart can hide the wrong method

A bar chart looks equally convincing whether it plots a sum, an average, or a count. Never accept a visual on its appearance: ask for the table beneath it and reconcile the totals to a number you calculated yourself.

A chart you can reconcile· copilot-chat
Bad example

Make a useful chart from this sales file and explain the trend.

Good example

Using only the attached sales file, calculate the sum of Revenue by Month and create a line graph with Month on the x-axis, ordered chronologically, and the y-axis in dollars. Below the chart, show the summary table you used and a grand total. State how any missing or invalid Revenue values were handled. Then break the same total down by Region and confirm both breakdowns reconcile to the same grand total.

Why this works: A chart plus the table and grand total behind it, with month and region breakdowns that agree, so you're trusting a reconciled number, not a nice-looking line.

Coach agents help you prepare

Microsoft's directory of agents built by Microsoft lists five Coach agents: Prompt Coach, Idea Coach, Writing Coach, Learning Coach, and Career Coach. Their role is preparation, reflection, practice, and feedback. Consequential decisions remain with people.

Consider a manager deciding whether an employee should receive a formal warning. A Coach agent must not make that determination. The qualified manager and the HR process own it. Once those people have set the process and boundaries, a coach can help the manager rehearse the conversation. The distinction matters in less sensitive work too. Coaching can improve your preparation without taking on your authority or accountability.

Sort statements in a coach response into three groups. A supplied fact came from your material. A generated suggestion is new wording, an idea, a recommendation, or a proposed action. A verification gap is information the response needs but the source didn't provide. A generated suggestion may be useful, but it carries a different evidence status. "Suggested owner: Jordan, subject to confirmation" is a proposal. "Jordan is assigned to update the docs" claims an assignment. If no source supports it, reject the sentence or mark it for verification.

A proposed owner is not an assignment

When a coach agent hands you an owner, a date, or a cause that wasn't in your source, treat it as a suggestion to confirm, never as a fact. Label it, or mark it for verification, before it travels into a real plan.

Turn a vague ask into a real prompt· copilot-chat
Bad example

Turn “Help with our product launch” into a better prompt.

Good example

Turn this rough request into a complete prompt for a one-page launch-readiness briefing for product leadership: "Help with our product launch." The eventual assistant should use a launch checklist and status note as its only factual sources, require sections for Status, Risks, Owners, Decisions, and Next Steps, and mark absent information for verification. First list up to three missing inputs you'd need, then give me the revised prompt.

Why this works: A structured prompt naming task, audience, sources, sections, and an uncertainty rule, plus the gaps to fill. It's a better instruction, not evidence about your launch.

Availability isn't capability

A name in a directory is not a promise that an agent is in your tenant or that it behaves the way its name suggests. Keep three ideas separate: an agent can be listed in Microsoft's directory, available when you can open it in your tenant, and documented when a source describes what it does. Prompt Coach, for example, has a documented sample use. Learning Coach and Career Coach are directory-listed, and their current availability and behavior are best confirmed in your own tenant rather than assumed from the name.

Your success criteria don't depend on the agent at all. If a coach agent isn't available, run the same prompt in Copilot Chat and label the output as a fallback. The exercise still works, because what you're testing is your prompt, your fact-audit, and your labels. Judge the response against the output you asked for and the evidence behind it, never against what the agent's name implies it should be able to do. Use the same discipline in the next lesson when you decide which agent a job needs.

Revise without inventing facts· copilot-chat
Bad example

Rewrite this launch announcement so it sounds clearer and more complete.

Good example

Review the announcement below for an audience of affected employees. The goal is to make the required action clearer. Identify anything unclear or missing, then give me a revision that preserves every supplied fact and invents no date, owner, deadline, cause, or support channel. Mark any necessary missing detail for verification. Announcement: "The launch date has changed. Two launch checks remain open. Employees should continue their assigned preparation tasks. The project owner will publish the confirmed date on the launch-status page."

Why this works: A clearer announcement whose every fact still traces to the source, with genuine gaps flagged rather than filled in. Wording is generated. Dates and owners must be earned.

Try it yourself

Grade the coach's advice

Coaching output is a set of suggestions to judge, not instructions to follow. Practise the judgment on five recommendations a Coach agent produced for a review update about a support-hours pilot. The situation: demand evidence uses mismatched measures, capacity is limited, the cost ceiling is missing, and one data value is unknown.

  1. 01

    Read the five recommendations: (1) State the evidence supports a limited proposal, not automatic launch. (2) Promise satisfaction will rise because most survey respondents favor later hours. (3) Ask the cost owner to set an approved ceiling before launch. (4) Describe the 70-case forecast as proven weekly demand. (5) Propose a human review after cost, approval, and missing-data questions are resolved.

  2. 02

    Label each one Accept, Adapt, or Reject, with a one-line reason grounded in evidence or policy.

    Hint: Two of them make promises the evidence doesn't support. A preference isn't proven satisfaction, and a forecast isn't proven demand.

  3. 03

    For every Reject, write the corrected version that keeps the claim inside the evidence.

  4. 04

    Name the human owner and the review point your final plan routes the decision to.

A graded list where recommendations 2 and 4 are rejected or corrected, and the pilot decision stays with a named person.

Key takeaways

  • Analyst runs Python over uploaded files to produce charts and totals, but you own the verification.
  • Specify the measure, aggregation, grouping, filter, output, and units so Analyst's result is checkable.
  • Reconcile a chart's grouped totals to an overall total you calculate yourself before trusting it.
  • Coach agents support practice and feedback. Consequential decisions stay with people.
  • Sort every response into supplied facts, generated suggestions, and verification gaps. A proposed owner is not an assignment.

Check your understanding

  1. 1. Which of these can you hand to Analyst for analysis?

  2. 2. Analyst returns a bar chart of revenue by region. Before you use it, what's the strongest check?

  3. 3. A workbook has merged cells, two columns both named Status, values mixing dollars and thousands of dollars, and undocumented blanks. What should happen before Analyst is used?

  4. 4. A manager asks a Coach agent to decide whether an employee should receive a formal warning. What's the appropriate response?

  5. 5. A retrospective note names no owner, but the Coach agent's summary says "Jordan is assigned to update the documentation." How should you treat that sentence?

Frequently asked questions

Terms used in this lesson

Analyst
A Microsoft 365 Copilot agent that analyzes uploaded files with a Python code interpreter and returns tables, charts, and summaries.
code interpreter
An execution environment that lets an AI system run code, here Python, against the data you supply for a task.
pivot-style summary
A compact cross-tabulation that groups one field into rows, another into columns, and aggregates a numeric measure in the cells.
coach agent
A Microsoft-built agent that supports preparation, practice, and feedback without taking over the decision.
generated suggestion
A new idea, wording, or proposed action an agent produced, distinct from a supplied fact, and to be confirmed before it drives a decision.

Further reading