Copilot Agent Mode and Claude in Excel can both turn a raw sales table into an executive dashboard, but they work very differently. In my test Claude built the whole dashboard from one prompt without asking a question, while Copilot proposed a plan, asked about the time period, and needed one nudge about blank charts before it was finished.
I ran both on the same workbook, a table named sales data with about 15,000 rows of global sales, with the same prompt: it opened with “You are a senior enterprise BI analyst” and said the table would feed an executive dashboard inside Excel. I recorded this in February 2026. Since then Agent Mode has been folded into Copilot in Excel, and Anthropic’s add-in has become Claude for Excel and left beta, so the working styles below are from my test and the product facts are current.
How do Copilot and Claude compare in Excel?
| Copilot in Excel (Agent Mode) | Claude for Excel | |
|---|---|---|
| How you get it | Built into Excel with a Microsoft 365 Copilot licence or Microsoft 365 Premium | Add-in from Microsoft AppSource, signed in with a Claude Pro, Max, Team or Enterprise account |
| Models | Auto, Claude or GPT | A curated set of Claude models (Opus 4.5 in my test) |
| First move in my test | Proposed a dashboard blueprint and asked about the time lens | Started building straight away |
| Errors in my test | Fixed its own errors; first dashboard had blank charts until I flagged them | One “worksheet not found” error, fixed without help |
| Output in my test | Separate Dashboard and Model sheets | One Executive Dashboard sheet, charts needed moving into place |
| Links back to the workbook | Explained what it built and fixed | Clickable links from the chat to the tabs and tables it referenced |
| Admin and audit | Runs under your Microsoft 365 Copilot licence and admin settings | Not in Enterprise audit logs; chat history stored locally in the browser |
Claude got to a finished first draft faster in my test, and Copilot is the one I’d use when the workbook lives in SharePoint and the data needs to stay under the Microsoft 365 controls you already run.
Does it ask questions first or just build?
Claude read the table in chunks of 500 rows, then built a KPI banner, summary tables, top 10 countries and sales reps, region and year trends, and four charts. I’d set it to accept all edits, so it never stopped to ask, and about 90 seconds in the dashboard sheet was already filling up.
Copilot came back with a blueprint instead (a dashboard sheet, a calculation sheet and the executive questions the columns could answer) and asked what time lens to use. I said the last 12 months, rolling, and to go ahead. It reasoned through 31 steps and added Dashboard and Model sheets.
When you already know the reporting period, Claude’s approach saves you a round trip. When the person who asked for the dashboard hasn’t decided it, Copilot’s question about the time lens is worth the extra turn, because the period changes every number on the page.
How did each one handle errors?
Both hit errors and recovered on their own. Claude referenced a worksheet name that didn’t exist, fixed the reference and carried on. Copilot ran into a range it couldn’t access and adjusted its approach, and you can expand its reasoning to watch each fix.
Copilot reported the dashboard as complete, but the charts were empty placeholders. When I told it they were blank, it checked the data types, rebuilt them and explained what it had fixed. Claude’s charts had data first time, though they landed on top of each other and I had to drag them into position.
Which one gave better executive insights?
Both picked up the cost of discounting. I asked each one what would stand out to an executive audience. Claude said revenue growth had flatlined, called out discount leakage and named APAC as the only region growing. When I asked for the three biggest changes over time, it pointed to new customer acquisition falling from 2019 to 2025 and a 2021 rebound led by APAC and EMEA.
Copilot came back with strong revenue scale, discounts costing the business, a healthy forward signal and balanced regional revenue. Asked what it would simplify for a board or CFO, it said to take the helper data off the dashboard and cut the KPI tiles to three or four board-grade metrics.
On regions they disagreed. Claude saw one growth engine and Copilot saw balance, partly because Copilot was working on a rolling 12 months and Claude on the full history. Before either summary goes to a board, check which period it covers and compare the claim with the numbers on the dashboard.
What’s changed since this test?
In Copilot you no longer turn on Agent Mode from the Tools button. Select Home, then Copilot, and editing is on by default, with Plan (Copilot proposes an approach for you to review first) and Chat only as the other modes. The workbook needs to be on OneDrive or SharePoint. The model switcher offers Auto, Claude and GPT while editing, and using Claude needs your admin to enable Anthropic as a Microsoft sub-processor in the Microsoft 365 admin centre.
Claude for Excel runs on Excel for the web, Windows and Mac, and admins can deploy it from Settings, Integrated apps in the Microsoft 365 admin centre. Answers carry cell-level citations, and it warns you before overwriting existing data. Its activity isn’t in Enterprise audit logs, although Enterprise organisations using Anthropic’s Compliance API do get these sessions.
Which should you use for Excel dashboards?
I’d start with where the data lives and who’s accountable for it. If your sales data sits in SharePoint and your security team has already done the work on Copilot’s permissions and policies, Copilot in Excel works inside all of that, and you can run it on a Claude model anyway. A separate add-in with its own account, audit trail and retention rules is a conversation to have with your compliance people first.
Claude for Excel makes sense for an analyst or finance team that already has Claude Enterprise, works in workbooks all day and wants the agent to get on with it. Its citations from the chat into the workbook help when you’re checking its numbers.
Whichever you pick, don’t judge it on one prompt. My best results from Copilot came on the third and fourth turn, after it had asked its question and I’d pushed back on the blank charts. Check every figure before it goes anywhere near a board pack. Claude for Excel isn’t recommended for audit-critical calculations without verification.
If you’re weighing the two vendors more broadly, I compare Claude skills with Copilot agents separately. For dashboards built outside a workbook, see how Copilot Cowork builds a dashboard from raw data or a live Copilot dashboard over a Microsoft List.
Frequently asked questions
Can Copilot in Excel use Claude models? Yes. When Copilot is editing a workbook, the model switcher offers Auto, Claude and GPT, and it may be tucked into the settings menu. It needs a commercial Microsoft 365 Copilot or Microsoft 365 Premium subscription, and closing Excel resets it to the default model.
Do I need a Copilot licence to use Claude for Excel? No. Claude for Excel uses your Claude account. Organisations that route AI through Azure AI Foundry, Amazon Bedrock or Google Cloud Vertex AI can deploy it without individual Claude accounts, and the models then come from that platform.
What can’t Claude for Excel do? It doesn’t support Excel data tables or macros and VBA, and it won’t run on Excel 2016 or 2019 perpetual licences, iPad or Android. Anthropic also warns against using it on untrusted spreadsheets, because files from outside can carry hidden instructions.
