On this page
Can an AI agent do spreadsheet analysis?
Partly, as an assistant you check. An AI agent for spreadsheet analysis can summarize a dataset, add formula columns, build PivotTables and charts, flag outliers and trace why a formula breaks. Microsoft's Copilot in Excel and Anthropic's Claude for Excel both do this inside the workbook, according to their documentation.
Neither vendor tells you to trust the output unchecked. Microsoft's FAQ says Copilot "can sometimes make mistakes" and tells users to "review, edit, and verify anything Copilot creates." Anthropic's documentation says Claude for Excel is not recommended for final client deliverables without human review or for audit-critical calculations without verification.
So hire it the way you would use a fast junior analyst: it does the first pass, and someone who knows the numbers checks it before anyone acts on it.
What the agent does and what stays with you
Agent
0of 10
Agent with review
8of 10
- Summarize a dataset and answer questions about it
- Add calculated columns, summary rows and lookups
- Build PivotTables and charts
- Clean messy data before analysis
- Explain formulas and trace how a number is derived
- Find the cause of formula errors
- Summarize themes in text columns such as survey comments
- Produce a recurring analysis on a schedule
Human
2of 10
- Define the question, the metrics and which data is authoritative
- Decide what the result means and act on it
10 tasks: 0 agent, 8 agent with review, 2 human
| Task | Who does it | Why |
|---|---|---|
| Summarize a dataset and answer questions about it | Agent with review | Microsoft documents Copilot in Excel returning summaries, trends and outliers as charts, PivotTables or text, and says to verify anything it creates. |
| Add calculated columns, summary rows and lookups | Agent with review | Microsoft's documentation shows Copilot adding formula columns, summary rows and XLOOKUP-based lookup columns and explaining each formula. |
| Build PivotTables and charts | Agent with review | Copilot in Excel creates charts and PivotTables on request, and Claude for Excel can edit pivot tables, according to each vendor's documentation. |
| Clean messy data before analysis | Agent with review | Anthropic ships a starter skill for cleaning messy spreadsheet data in a range, sheet or whole file; a person checks what was changed. |
| Explain formulas and trace how a number is derived | Agent with review | Claude for Excel answers with cell-level citations, and Copilot in Excel explains the formula in a selected cell, so the reader can follow the trail. |
| Find the cause of formula errors | Agent with review | Anthropic documents Claude for Excel locating the root cause of errors such as #REF! and #DIV/0 and suggesting fixes. |
| Summarize themes in text columns such as survey comments | Agent with review | Microsoft documents Copilot summarizing feedback into themes and sentiment, with superscript references to the source rows for checking. |
| Produce a recurring analysis on a schedule | Agent with review | Microsoft reports Pamarco's sales director receives an automatically generated monthly board sales analysis and reviews and refines the narrative. |
| Define the question, the metrics and which data is authoritative | Human | No source reviewed for this page shows an agent setting metric definitions; Microsoft advises naming the columns to analyze in the prompt. |
| Decide what the result means and act on it | Human | Microsoft advises against using Copilot for decisions in sensitive areas such as finance, legal or medical topics. |
The agent handles the mechanical part of analysis. Microsoft's documentation shows Copilot in Excel answering questions such as "Which region had the highest and lowest sales?", adding a profit-margin column, building lookups with XLOOKUP and summarizing free-text feedback into themes. Copilot has three modes: edit changes the workbook directly, plan proposes steps first, and chat keeps answers in the chat pane. Claude for Excel reads multi-tab workbooks, answers with cell-level citations you can click, updates assumptions while keeping formulas intact, and can sort, filter, edit pivot tables and apply conditional formatting, according to Anthropic.
What stays with you is the part that decides whether the answer is right: which question matters, how a metric is defined, which data source is authoritative, and what to do with the result.
Production use exists, reported by the vendors. In a September 2026 customer story, Microsoft says Pamarco, a manufacturer with about 225 employees, uses Copilot Cowork for scheduled daily, weekly and monthly reports, with results going to Excel, Power BI and other outputs. A sales director who spent about five hours a month preparing sales analysis for the board now receives an automatically generated presentation on the sixth working day and reviews and refines the narrative. Anthropic's May 2026 announcement quotes Citadel saying its analysts use Claude for Excel to build and update coverage models, Bain & Company saying its teams build initial versions of complex models faster, and ServiceNow saying the add-in is being rapidly adopted there. These quotes give no time savings or error rates.
A worked example: one sales export
This is an illustration, not a customer story. Assume, for this example, that you export 4,800 order lines for October from your CRM, that the accounting system shows October revenue of $612,400, and that you ask: "Revenue by region, top 10 customers, and any unusual discounts."
The agent adds a PivotTable by region, a top-10 table and a column flagging discounts more than two standard deviations from the mean. Its summary says October revenue was $618,900 and the Northeast grew fastest.
The analyst checks the control total first. $618,900 is $6,500 above the accounting figure. She asks the agent which rows make up the difference. It finds three orders marked as test accounts and one order that was refunded in October but still shows as closed in the CRM. She tells the agent to exclude test accounts and net refunds, as her written definition says. The total now ties to $612,400. Only then does she read the regional story and the discount flags, two of which turn out to be approved contract prices. The agent did the first pass; the analyst decided which numbers were true.
What you need to provide
Give the agent a copy of the workbook, not the master. Its access should be edit rights on that copy and nothing else: no write access to the source CRM, accounting system or shared master file. If the data comes from a database, export it through a read-only account.
Name the data the agent will see. Your workbook may contain customer names, employee salaries or financial results. Anthropic's documentation says Claude for Excel sends conversations to Anthropic, including the content it reads from the workbook, and that the add-in does not inherit custom data retention settings your organization may have set. For Copilot, Microsoft points business users to its data, privacy and security documentation for Microsoft 365 Copilot. Check these terms before you hand over sensitive files.
Write down the definitions an analyst carries in their head. Microsoft's documentation says naming the columns you want analyzed produces more accurate results. Claude for Excel has a persistent Instructions field for conventions such as number formats, and Microsoft's documentation describes Copilot in Excel rules: a visible worksheet named .Rules whose guidelines Copilot follows for that workbook. Put metric definitions there, and always include one control total you already trust.
On licenses: Claude for Excel is available on Claude's Pro, Max, Team and Enterprise plans. Copilot in Excel requires one of several Microsoft 365 or Microsoft Copilot subscriptions listed in Microsoft's FAQ. Prices were not researched for this page; check the vendors' current pricing pages.
Job brief template
Use this as the brief for a vendor, an operator or your own team. Replace the bracketed values with your own.
Copy-ready brief
- Goal
- Within [2] business days of receiving the workbook, every question in the brief has an answer that ties to the source data, with the formulas and steps visible in the workbook.
- Deliverable
- A new sheet or workbook in the shared folder containing the answer tables and charts, formulas left in place (not pasted as values), and a short note listing each step taken, each assumption made and any rows excluded.
- Inputs
- A copy of the workbook or export, with edit rights to that copy only; the original stays read-only for the agent
- The questions to answer, in writing, for example: revenue by region and month, top 10 customers, outliers in discount rate
- Metric definitions: what counts as revenue, which date column to use, how refunds and test rows are treated
- One or more control totals you already trust, such as total revenue for the period from the accounting system
- The output format you want: new sheet, PivotTable, chart or short written summary
- Acceptance criteria
- The total of the analyzed dataset matches the control total named in the brief to the cent, or the difference is explained row by row.
- Every figure in the written summary points to a cell or range in the workbook.
- Excluded or changed rows are listed with the reason; no source data is overwritten.
- On a spot check of [5] figures recalculated by hand, all [5] match.
- Running the same request again on the same file gives the same numbers.
- Systems
- spreadsheet tool (for example Microsoft Excel with Copilot or the Claude for Excel add-in)
- shared file storage with version history (for example OneDrive or SharePoint)
- source system the data comes from, such as an accounting system or CRM export
- Constraints
- The workbook may hold customer, employee or financial data; use only files your organization allows in the AI tool and check the vendor's data-retention terms first.
- Only use workbooks from trusted sources; Anthropic warns that external files can contain hidden instructions.
- The agent works on a copy; it has no write access to the master file or to source systems.
- Recalculation must be set to Automatic for Copilot edits to work.
- Volume
- Example: 2–10 workbooks a month, each up to a few tens of thousands of rows
- Cadence
- Example: ad hoc requests plus one recurring monthly analysis
- Human checkpoints
- Metric definitions: the analyst or manager who owns the numbers writes them down before the work starts.
- Control totals: a person checks the reconciliation before reading any other result.
- Changes to data: a person reviews every excluded, edited or deduplicated row.
- Use of results: the owner of the decision reads the analysis and the step log before it is shared or acted on.
A control-total check can catch discrepancies before you rely on the analysis. Spreadsheet errors are quiet: a wrong filter or a duplicated row produces a clean-looking table. A number that ties to a figure you already trust catches errors of this kind before anything else is read.
Risks and human checkpoints
The agent answers with fluent, confident language and a wrong number.
How you notice it
Check the dataset total against a control total and hand-recalculate a few figures every time; Microsoft's FAQ says Copilot's content can be inaccurate and tells users to verify it.
The agent uses the wrong column or definition, such as order date instead of invoice date.
How you notice it
Name the columns and definitions in the request and check the step log; Microsoft's documentation says naming the columns produces more accurate results.
The agent changes or overwrites source data while cleaning it.
How you notice it
Work on a copy and ask the agent what it changed; Copilot can summarize its own edits from the workbook's change history, and Version History restores earlier versions.
A workbook from outside your company carries hidden instructions that make the agent leak or alter data.
How you notice it
Use only trusted files and read each confirmation prompt; Anthropic documents that testing found Claude for Excel could be manipulated this way and asks users to confirm risky operations.
The analysis needs a feature the agent cannot handle, such as macros or data tables, and it works around it silently.
How you notice it
Before the trial, check the vendor's list of unsupported features; Anthropic lists data tables, macros and VBA as unsupported in Claude for Excel.
Human checkpoints
- Metric definitions: the analyst or manager who owns the numbers writes them down before the work starts.
- Control totals: a person checks the reconciliation before reading any other result.
- Changes to data: a person reviews every excluded, edited or deduplicated row.
- Use of results: the owner of the decision reads the analysis and the step log before it is shared or acted on.
Two product facts shape the checkpoints. Copilot in Excel can only report changes that are in the workbook's change history, which Excel keeps for up to 365 days, so keep files in a location with version history. And Anthropic warns that files from external sources, such as downloaded templates, vendor files and data imports, can contain hidden instructions; when Claude proposes a risky operation it asks you to confirm, and those confirmations deserve a careful read.
How to run a trial
Run the agent next to your current process on work you have already done, so you know the right answers.
- Pick three to five past analyses where the result was checked: a monthly sales summary, a survey readout, a cost breakdown.
- Give the agent copies of the original files and the same questions, with your metric definitions and one control total each.
- Compare its results with the checked versions. Count figures that match, figures that are wrong, and wrong figures it presented with confidence.
- Ask the same question twice on the same file and compare the numbers.
- Plant one problem, such as a duplicated block of rows or a test account, and see whether the agent finds it or includes it.
- Time the review. The agent's speed matters less than how long a person needs to check its work.
Pick three to five past analyses
Results already checked: a monthly sales summary, a survey readout, a cost breakdown
Give the agent copies of the original files
The same questions, your metric definitions and one control total each
Compare its results with the checked versions
Count figures that match, figures that are wrong, and wrong figures presented with confidence
Ask the same question twice
On the same file, then compare the numbers
Plant one problem
A duplicated block of rows or a test account: does the agent find it or include it
Time the review
How long a person needs to check the work matters more than the agent's speed
If the results tie out and review time is acceptable, move to live requests on copies, keep the control-total check, and widen the scope one type of analysis at a time.
For structuring the test, see how to run a paid trial with an AI agent, and for the brief, how to write a job description for an AI agent. Related pages: reporting and BI, data engineering, FP&A and forecasting, the data analyst role and the data overview.
Questions buyers ask
Is the agent's analysis accurate? Neither vendor publishes an accuracy rate on customer workbooks. Microsoft describes testing Copilot in Excel against public and internal benchmarks, including a finance benchmark built with outside finance professionals, and says benchmark runs gate releases, but the post cited here gives no scores. Measure accuracy yourself against analyses you have already checked.
Can it work with macros and very complex models? Anthropic lists macros, VBA and data tables as unsupported in Claude for Excel. Copilot in Excel only edits workbooks when calculation is set to Automatic. Check these limits against your files before the trial.
Will it change my data without asking? In Copilot's edit mode, Copilot changes the workbook directly; plan mode proposes a plan before editing, and chat mode keeps responses in the chat. Claude for Excel warns before overwriting existing data. Working on a copy helps protect the master workbook from direct edits.
Sources
- Use Claude for Excel
Anthropic (Claude documentation)PrimaryAccessed Nov 14, 2026
- Frequently asked questions about Copilot in Excel
Microsoft SupportPrimaryAccessed Nov 14, 2026
- Get data insights with Copilot in Excel
Microsoft SupportPrimaryAccessed Nov 14, 2026
- Review the change history of your workbook with Copilot in Excel
Microsoft SupportPrimaryAccessed Nov 14, 2026
- Create guidelines for a workbook with Copilot in Excel rules
Microsoft SupportPrimaryAccessed Nov 14, 2026
- Pamarco accelerates finance operations with autonomous agents and Copilot Cowork
MicrosoftVendorAccessed Nov 14, 2026
- Collaborate with Claude across Excel, PowerPoint, Word and Outlook
AnthropicVendorAccessed Nov 14, 2026
- Advancing Claude for Excel and PowerPoint
AnthropicVendorAccessed Nov 14, 2026
- An Inside Look at Copilot in Excel
Microsoft Tech Community (Excel Blog)VendorAccessed Nov 14, 2026
What we could not verify
- Every production report on this page (Pamarco, Citadel, Bain & Company, ServiceNow) is published by the software vendor involved. No independent study of agent accuracy on business spreadsheet analysis in production was found.
- No accuracy rate for Copilot in Excel or Claude for Excel on customer workbooks is published. Microsoft describes its internal benchmarks but does not publish scores in the post cited here.
- Agent pricing is not shown. Both products cited here come with subscription plans (Claude Pro, Max, Team and Enterprise; Microsoft 365 and Copilot licenses), but per-seat prices were not researched to a sourced range for this page, so it has no cost comparison.
- No source was found showing an agent choosing the business question, defining metrics or deciding what to do with the result, so those tasks are marked human.
- The Citadel, Bain & Company and ServiceNow statements are short quotes in an Anthropic product announcement; they give no volumes, time savings or error rates.
Related
- AI agents for data analysis and BI: what to hand over
- Can AI replace data analysts? A task-by-task answer
- AI agents for business intelligence and reporting
- AI agents for data engineering: what they do today
- AI agent for FP&A and forecasting: what it can do
- AI agent job description: how to write one, with template
- How to test an AI agent before buying: run a paid trial