Claude's Excel Plugin or Built-In Copilot: Which Should Your Team Use?

Teams already running on Microsoft 365 now have two distinct ways to put an AI assistant inside the spreadsheet grid: Copilot built into Excel and the Claude add-in (often called the Claude Excel plugin). Both can read cells, write formulas and reshape sheets without making you copy tables into a browser tab, which leads managers to ask whether they need one, the other or both.

In Microsoft 365 workplaces, people usually reach for Copilot in Excel, often for data cleanup or report building. Choose the Claude add-in when your analysts need heavier multi-tab modeling, formula debugging or clickable cell-level citations on an existing Claude plan. Paying for both is generally only worth it when Copilot is already bundled into your Microsoft license and a specific team has an advanced modeling or assumption-testing workflow that Claude handles better; otherwise, pick one.

The two Excel AI tools at a glance

Prices, license tiers and platform rules below were checked on November 9, 2026, from Microsoft's Copilot business pricing, Microsoft's Copilot overview and Anthropic's Excel add-in documentation. Annual seat figures are billed yearly in US dollars.

Before assigning seats, check which Microsoft tier your users actually see inside the app. In Microsoft's license options and overview, Copilot Chat (Basic) is the free web-grounded chat for Entra work accounts and does not include the assistant inside Word, Excel, PowerPoint or OneNote. Microsoft 365 Copilot (Basic) is also unlicensed, but gives users standard, capacity-based access inside those desktop and web apps without automatic Microsoft Graph grounding.

Because standard access varies with service capacity and tenant configuration, do not oversell unlicensed in-app features to your staff before testing what your tenant exposes. Microsoft 365 Copilot (Premium) is the paid add-on license that unlocks priority access, full in-app editing and automatic Microsoft Graph and Work IQ grounding. For a role-by-role breakdown of those tiers, see our Copilot Chat vs Microsoft 365 Copilot guide.

Three similarly named options people mix up

Vendor naming around these products overlaps enough that buyers and IT admins regularly talk past each other on discovery calls. Keep these three distinct items straight before you scope a purchase or submit an IT ticket:

• The Claude add-in for spreadsheets: Anthropic's task-pane tool inside Excel that connects to a paid Claude Pro, Max, Team or Enterprise account (or your own Bedrock, Vertex AI or Foundry endpoint).
• The "Claude for Microsoft 365" AppSource listing: The single marketplace package IT deploys in the Microsoft 365 Admin Center, which covers Excel alongside Word, PowerPoint and Outlook.
• Claude models inside Microsoft's own assistant: A model-switcher option inside Copilot in Excel that lets paid commercial Copilot or Microsoft 365 Premium users pick an Anthropic model instead of an OpenAI GPT model for the active session, provided an enterprise admin has enabled Anthropic as a Microsoft subprocessor.

Two everyday prompts that work in either tool

Most day-to-day spreadsheet requests do not require an intricate financial model. They start with a raw export from another software tool that needs to be cleaned up for a database import or summarized for a quarterly review. Both of the prompts below work well inside Excel regardless of which vendor's pane you open.

When a messy website export has to be formatted before importing it into a database, we use a reference tab in the same workbook and run this prompt:

Clean and normalize the data in this sheet so it matches the format in the Example sheet

Giving the assistant a concrete Example tab in the same file works better than writing a paragraph of formatting rules. In practice, that single instruction clears out blank rows, recombines data that split across columns (such as street addresses or city and state fields) and converts null numerical values to zeros so the database import does not fail.

For recurring issue reporting, you can point the assistant at a raw export of support logs and ask for both the categorization and the written takeaway at once:

Using this data, make me a document identifying the top categories of support tickets received in the last quarter along with best options to proactively reduce them.

Where the Claude add-in pulls ahead (and where it stops)

Anthropic built its Excel add-in for analysts and operators who live inside multi-tab models. When you ask how a forecast or margin line is calculated, its answers include clickable cell-level citations that jump straight to the referenced cell. You can ask it to flex a discount or growth rate while keeping formula relationships intact so downstream cells recalculate cleanly, trace the root cause of #REF! or #DIV/0! errors, or build and populate three-statement and LBO templates from scratch.

The add-in also handles native spreadsheet operations such as sorting, filtering, editing PivotTables, applying conditional formatting and creating data validation dropdowns. Teams on Claude can pull external context directly into a workbook through connectors like S&P Global, LSEG and Daloopa, apply saved Skills via / commands, and set persistent Excel-only formatting preferences under Settings > Instructions. Conversation context auto-compacts during long sessions, warns before overwriting existing data, and can coordinate with open Word, PowerPoint and Outlook files when "Let Claude work across files" is enabled (on by default for Pro and Max, off by default for Team and Enterprise).

The boundaries are just as important to document for your team:

• Not for unreviewed final work: Anthropic explicitly does not recommend the add-in for final client deliverables without human review, audit-critical calculations without verification, or models holding highly sensitive or regulated data without proper controls.
• Unsupported Excel features: It cannot create or edit Excel Data Tables, and it does not support macros or VBA operations.
• Platform exclusions: It requires SharedRuntime support and does not run on Excel 2016 or 2019 perpetual licenses, iPad, Android or older Microsoft 365 desktop builds below 16.0.13127.20296 on Windows or version 16.46 on Mac.
• Prompt injection risk on external files: Downloaded templates, vendor spreadsheets and raw data imports can hide instructions in cells that trick the add-in into extracting data or modifying records. Only run it on trusted workbooks, and read confirmation prompts carefully before approving risky actions.
• Governance gaps to plan around: Add-in activity does not inherit custom organization data retention settings and is not recorded in standard Claude Enterprise audit logs, though sessions are captured if you enable the Enterprise Compliance API or route telemetry to a custom OpenTelemetry collector.

Where Copilot built into Excel fits best

In workplaces already standardized on Microsoft 365, Copilot in Excel is usually the first stop for data cleanup, ad-hoc tables and report building. It opens in three modes on Windows, Mac and the web: edit mode (the default) makes direct changes in the workbook, plan mode drafts a step-by-step approach for your confirmation before touching cells, and chat mode keeps analysis and answers inside the conversation pane without altering the sheet.

Inside the grid, it creates and edits Excel tables, charts, PivotTables and shapes with editable links to source data, writes formulas across sheets, highlights outliers and applies sorting, filtering and conditional formatting. It can also pull information from the web with source citations, import data from other Excel files stored on OneDrive, SharePoint or your computer, and record repeatable tasks as custom skills when you demonstrate them in a workbook. On a paid Premium license, it connects to your wider organizational context through Microsoft Graph and Work IQ.

Keep Microsoft's operational constraints in mind during training:

• File state and calculation rules: Workbook editing only works when Excel's Calculation Options are set to Automatic. If a file is saved in an older or strict format (such as Strict Open XML Spreadsheet) instead of .xlsx, or if it is checked out in SharePoint, desktop Excel returns an "Unsupported file state" error. If a SharePoint library requires check-out for editing, desktop users on Windows and Mac must open the file in Excel for the web to use the assistant.
• Live coauthoring visibility: Any direct edit saved to a shared workbook is immediately visible to everyone else in a coauthoring session, so staff should use version history or Undo when testing broad changes.
• No independent admin kill-switch: According to Microsoft's Excel FAQ, Copilot in Excel is a Microsoft Copilot Chat tool and cannot be turned off independently in the admin center; whatever access policies you set for Copilot Chat apply to Excel as well.

Workbook size limits: check used cells, not row counts

Neither tool has a simple row-count cutoff, and treating them as if they do is the fastest way to freeze a user's desktop app. Always check a workbook's total used cells—calculated as the last row multiplied by the last column on each worksheet, added across all tabs—before running an analysis prompt. Empty cells that carry stray formatting still count toward that range.

In Anthropic's performance documentation, the add-in reads up to 2,000 populated cells per request and steps through larger sheets in batches. Workbooks under 1 million total used cells across all sheets tend to work well; files between 1 million and 5 million used cells carry higher crash or timeout risk and need narrower prompts that name a specific tab and cell range; and anything over 5 million used cells is use-at-your-own-risk. A single worksheet above 5 million cells also hits Office's own per-range read limit and can fail outright, while Excel on the web enforces a 5 MB per-request cap (and browser workbook limits up to 100 MB), meaning heavy workbooks belong on 64-bit desktop Excel.

On Microsoft's side, an earlier version of the support FAQ (article ID 7a13758f, which previously stated that Copilot worked with Excel tables up to two million cells while actions like formula columns, highlighting, sorting and filtering had no volume limit but could take 30 seconds or more) has since been updated. As verified on November 9, 2026, Microsoft's current English and international FAQ no longer publishes a fixed cell-count cap, noting only that editing requires Automatic calculation mode and a modern .xlsx file state. Even without a published number on the current FAQ page, large ranges and heavy formulas still bog down the single-threaded Excel process, so staff should check used-cell counts first.

The Lilac Spreadsheet Fit Test

Before buying add-on seats or approving an AppSource deployment request, run the team's actual spreadsheets through the Lilac Spreadsheet Fit Test. Answer these six questions in order:

1. Who owns and edits the workbook? If a broad group of operations, sales or service staff need quick help cleaning exports or building summary tables inside shared files, start with Copilot in Excel. If a small finance, FP&A or analytics group owns complex workbooks and already uses Claude for long-document analysis, point that group to the Claude add-in.
2. Is the file a multi-tab financial model or a flat data dump? Flat exports, basic PivotTables and quick charts fit Microsoft's edit and plan modes well. Multi-tab models where someone needs clickable cell citations, assumption flexing that preserves downstream formulas, or #REF! debugging across tabs favor Anthropic's add-in.
3. Does the task need Microsoft 365 tenant grounding or external financial connectors? If the spreadsheet work depends on pulling context from Outlook threads, Teams meetings and SharePoint lists via Microsoft Graph and Work IQ, only a paid Microsoft Copilot seat does that automatically. If the model pulls from S&P Global, LSEG, Daloopa or custom MCP servers—or only needs the open Word, PowerPoint and Outlook windows on the user's desktop—the Claude add-in fits.
4. Which licenses are already paid for? Check whether your Microsoft 365 tenant already exposes Microsoft 365 Copilot (Basic) in Excel or whether the team already holds paid Copilot or Claude Team seats. Paying for both is generally only justified when Copilot in Excel is already bundled into a user's Microsoft license and they have an advanced modeling or assumption-work use case that Claude handles better.
5. What are your audit, retention and file-state rules? If compliance requires every prompt to flow through Microsoft Purview eDiscovery and standard tenant retention—or if your team works on iPad or in SharePoint libraries that require file check-out—stay with Copilot built into Excel. Remember that Anthropic's add-in does not inherit custom Claude org retention or standard Enterprise audit logs unless you wire up the Enterprise Compliance API or a custom OpenTelemetry collector.
6. What does Ctrl+End show on the workbook's largest tab? Multiply the last row by the last column across every tab before testing. If total used cells cross 1 million, clean up stray empty-cell formatting or copy the target tab into a lighter workbook before judging either tool's speed.

To compare the broader company-level contracts behind these two tools, read our Microsoft 365 Copilot vs Claude Enterprise comparison and our breakdown of Claude Team vs Claude Enterprise on Microsoft 365.

Rollout notes and documented admin gotchas for a pilot

If the Fit Test points to a pilot of Anthropic's add-in alongside or instead of Copilot in Excel, hand your IT lead or MSP two documented Microsoft 365 rollout risks up front. We have not hit these two failure modes directly in client rollouts, but both are documented by Microsoft and Anthropic and are worth preventing before your pilot group logs in.

1. Office Store access and the Entra PIM deployment bug

To deploy the add-in centrally, an admin goes to the Microsoft 365 Admin Center > Settings > Org settings > User owned apps and services and turns on "Let users access the Office Store", then opens Settings > Integrated apps > Add-ins, searches Microsoft AppSource for "Claude for Microsoft 365" and deploys it to the organization or a pilot group. If your security rules keep the Office Store disabled, IT can download Anthropic's custom Excel manifest XML file (https://pivot.claude.ai/manifest-excel.xml) and upload the custom manifest directly in Integrated apps instead.

Watch out for a documented Microsoft bug (tracking ID 11126536): if your administrators activate their admin roles through Microsoft Entra Privileged Identity Management (PIM), the Integrated apps page may fail to recognize the PIM-activated role and block the deployment. Have IT run the deployment from an admin account where the required role is permanently assigned rather than PIM-eligible, as noted in Anthropic's Excel deployment docs.

2. Missing chat history and skills after Windows M365 updates 2608 and 2609

According to Anthropic's Excel documentation and data storage guide, the September and October 2026 Microsoft 365 updates on Windows (versions 2608 and 2609) can cause Office to open the Claude add-in against a new, empty WebView2 storage folder. Users see first-run onboarding screens and appear to have lost their chat history and uploaded skills.

Anthropic notes that the earlier data is still on the device inside %LOCALAPPDATA%\Microsoft\Office\16.0\Wef\webview2—so tell IT and users not to clear the Office add-in cache, which permanently deletes that folder. Instead, have users back up uploaded skills to their Claude account via the cloud icon (Save to claude.ai), download important chat transcripts via More options > Download Transcript, and have IT follow Anthropic's PowerShell steps to copy the earlier .leveldb and .blob IndexedDB folders into the active WebView2 directory.

3. Train the pilot group on size habits and human review

Technical setup is only half of a spreadsheet pilot. In our AI training sessions, we teach staff to close extra workbooks so they do not compete for the same 50-batch Office command queue, wait for each step to finish without clicking into the grid, and verify every formula and assumption change before sending a file to a client or executive.