Anthropic rolled out Claude in Excel to Pro subscribers last week, expanding access beyond the Max, Team, and Enterprise users who've had it since the October 2025 research preview. The add-in runs in a sidebar, reads your entire workbook, and explains what it's doing with cell-level citations. It's powered by Claude Opus 4.5.
The move puts Anthropic directly in Microsoft's territory. And yes, Microsoft has its own Copilot features in Excel, including an "Agent Mode" launched last month. But Anthropic's pitch is different: this isn't a general-purpose assistant; it's built specifically for financial work.
What the tool actually does
Claude in Excel can query multi-tab workbooks, update assumptions without breaking formulas, debug errors, and build models from templates. The November update added pivot tables, charts, and file uploads. Recent improvements include drag-and-drop for multiple files, protection against overwriting existing cells, and auto-compaction for longer sessions.
The keyboard shortcut (Control+Option+C on Mac, Control+Alt+C on Windows) opens the sidebar fast enough that you won't resent using it.
What it can't do yet: macros, VBA, data validation rules. Anthropic says those are coming. For now, it's read-analyze-modify territory, not full automation.
Six workflows worth testing
I spent a few hours with the tool. Here's what worked, what didn't, and what finance people should actually try.
1. Inheriting someone else's model
Every analyst has inherited a budget model that looks like it was built by a madman. Nested IFs referencing cells across eight tabs, no documentation, circular reference warnings that nobody remembers how to silence.
The prompt that works: "Walk me through how the revenue forecast in tab 'Projections' gets calculated. What cells does it depend on?"
Claude traces the dependencies and explains them in plain language, with clickable citations that jump to the relevant cells. This is genuinely useful. I gave it a 47-tab consolidation model, and it correctly identified that Q3 revenue in the summary was pulling from a named range that referenced a hidden tab. Finding that manually would have taken twenty minutes.
What didn't work: Asking it to "explain everything" about a complex model. The responses got long and vague. Narrow questions beat open-ended ones.
2. Sensitivity analysis without touching the structure
Finance people constantly run scenarios: "What if revenue grows 2% slower? What if COGS increases by 50 basis points?"
The prompt that works: "Increase the revenue growth rate in C12 from 8% to 10% and show me what happens to EBITDA in the Summary tab."
Claude changes the input, tells you exactly what it changed, and reports the downstream effect. It preserves all formulas. The change log is useful for audit trails, though it doesn't persist between sessions (a gap Anthropic should fix).
Practical example for FP&A: You're prepping for a board meeting and need three scenarios. Instead of manually building each one, you describe them: "Create a downside case where revenue grows at 5% instead of 8%, and OpEx stays flat instead of increasing with headcount." Claude adjusts the assumptions and walks you through what changed.
3. Debugging formula errors
The tool is surprisingly good at tracing #REF!, #VALUE!, and circular reference errors. Better than clicking through cells manually.
The prompt that works: "Cell G145 shows #VALUE!. What's causing it?"
Claude identified that a VLOOKUP was returning text into a cell that fed a SUM formula. It suggested IFERROR wrappers and SUMIF alternatives. Not revolutionary, but faster than debugging by hand.
4. Building DCF models from templates
Anthropic has been pushing pre-built "Agent Skills" for financial tasks, including discounted cash flow models. You can upload a 10-K or provide assumptions, and Claude will populate a template.
What worked: Starting from Anthropic's DCF skill and asking Claude to populate it with assumptions: "Use 8% revenue growth, 35% gross margin, WACC of 10%, and a 3% terminal growth rate. Build a 5-year projection."
The output was reasonable, not something I'd send to a client without review, but a solid first draft. The sensitivity tables and scenario toggles were correctly wired.
What didn't work: Expecting it to make judgment calls. Claude populated the numbers I gave it but didn't push back on unrealistic assumptions. If you tell it to use 25% revenue growth for a mature company, it will. The skill is execution, not strategy.
5. Comparable company analysis
Another pre-built skill. Give Claude a list of companies and metrics, and it'll build a comp table with valuation multiples.
Practical example for investment banking: You're updating a pitch book and need to refresh the comp set. Upload the existing table, tell Claude which companies to keep, and ask it to recalculate the multiples using updated data. It won't pull live data (that requires the financial data connectors Anthropic announced for Enterprise), but it'll restructure and recalculate what you give it.
The formatting was cleaner than I expected. Still needed manual tweaks for client-facing work.
6. Variance analysis on actuals vs. budget
This is bread-and-butter FP&A work, and Claude handles it decently.
The prompt that works: "Compare actual Q4 results in tab 'Actuals' to the budget in tab 'Budget'. Show me the top 5 variance drivers by absolute dollar amount."
Claude identifies where the numbers diverge and ranks them. It doesn't know why sales in the Northeast region underperformed, but it finds the discrepancy fast.
What would make this better: Integration with ERP exports. Right now you have to upload the files yourself. Anthropic's connectors (S&P Capital IQ, Daloopa, etc.) are Enterprise-only.
The security question
Anthropic's help documentation includes a warning about prompt injection attacks, where malicious instructions hidden in spreadsheet content could trick the AI into taking unintended actions. It's a real risk when you're working with files from external parties.
The practical advice: don't open random Excel files from strangers and assume Claude will behave. Same hygiene you'd apply to macros.
Also worth noting: chat history isn't saved between sessions. And for Team/Enterprise plans, the Excel add-in doesn't inherit custom data retention settings or appear in audit logs. Compliance teams should know this before rolling it out.
Is it better than Copilot?
Different tools, different approaches. Microsoft's Copilot is tightly integrated with the Office ecosystem and can do things Claude can't, like in-cell formulas and broader Excel automation. But Claude's strength is understanding financial modeling conventions and explaining complex workbooks without making you feel like you're prompting a generic assistant.
For finance-specific work, Claude is more useful. For general Excel tasks, Copilot probably wins on integration. Most power users will end up with both.
The bottom line
Claude in Excel isn't going to replace your financial analyst. But it will save them time on the tedious parts: inheriting unfamiliar models, running sensitivity scenarios, debugging formula spaghetti. The cell-level citations are genuinely helpful. The pre-built skills for DCF and comp analysis are solid starting points.
Pro plans are $20/month. If you spend real time in Excel, it's worth testing. Just review everything before it goes to anyone important.
The add-in is available now from the Microsoft Marketplace.
