AI in spreadsheets has split into two camps. One puts an assistant sidebar inside Excel or Google Sheets and treats the workbook as the system of record. The other builds AI-native grids where the spreadsheet and the model are designed together. The deciding question is no longer "which AI is smartest" (every serious tool in this category now runs on a current frontier model) but whether the assistant actually understands an .xlsx file, whether it preserves formulas, what it costs once you add in the base license, and whether it will pass procurement.
We tested the five tools a working analyst, finance professional, or operations lead is most likely to pay for between June 9 and June 23, 2026: Claude for Excel, Microsoft Copilot in Excel (Agent Mode), Shortcut, Gemini in Google Sheets, and Quadratic. Each ran on the same battery of tasks against the same source workbooks. The criteria, the procedures, and the per-tool marks are below.
How we tested
All five tools were tested between June 9 and June 23, 2026 on their current generally available releases (Claude for Excel with Opus 4.6, Microsoft 365 Copilot Agent Mode at the April 22, 2026 GA build, Shortcut v7.4, Gemini in Sheets post-April 22 release, and Quadratic Pro). Criteria are weighted toward formula accuracy and workbook fidelity, with security and value heavily weighted for team and enterprise use.
Formula & Workbook Accuracy
Each tool was given the same 25 natural-language formula prompts (ranging from simple INDEX/MATCH lookups to a five-condition SUMIFS and a regex-based text extraction) and a 12-tab corporate model with three deliberate broken references, and we counted formulas that returned the correct value on the first run and the broken references it correctly identified.
Multi-Tab Model Audit
We opened the same 14-tab Apple three-statement model none of the tools had built, and asked each to (1) trace how Q3 revenue flows into the summary tab and (2) update the WACC assumption from 9% to 11% without breaking dependent formulas; two reviewers independently scored the trace explanation and verified the dependent cells against a hand-audited reference.
Build-From-Scratch Modeling
Each tool was given the same prompt, 'build a fully integrated three-statement model for NVIDIA from the latest 10-K using investment banking formatting and a sensitivity table for revenue growth and WACC,' and we scored the output on whether IRR, NOI, and net income cells were live formulas or hard-coded values, and whether changing an assumption recalculated the model end to end.
Bulk Row Processing
We loaded the same 2,000-row dataset of product reviews into each tool and asked it to classify each row as positive, negative, or neutral, extract the named product, and append a one-sentence summary; we measured rows processed without error, time to complete, and cost in tokens or credits at each tool's published rate.
Security & Governance Posture
We read each vendor's trust page and admin documentation and recorded whether the product holds a current SOC 2 Type II report, offers HIPAA coverage, supports tenant-level admin controls, whether customer data is used to train models by default, and whether prompts and responses stay inside the customer's compliance boundary.
Value at Paid Tier
We priced one user on each tool's standard paid plan (annual billing) against the prerequisite base license required to use it (Microsoft 365 for Copilot, a Claude subscription for Claude for Excel, Google Workspace for Gemini), and recorded the all-in monthly cost for a single working seat.
The full battery and per-criterion marks are above; the notes here cover where the ranking turned.
Why Claude for Excel leads
This category is decided by one question more than any other: when an AI is asked to change something in a real workbook, does the rest of the workbook still work afterwards? Claude is the tool that most consistently said yes in our test. It reads the entire multi-tab structure before it acts, it cites the specific cells it’s reasoning from, and when it updates an assumption it traces the dependency graph rather than overwriting cells with hard-coded values. On the 14-tab Apple model audit, Claude was the only tool that correctly traced Q3 revenue from the segment-level revenue tab all the way through the income statement to the summary, and the only one that explained the chain in terms a reviewer could verify.
The trade-offs are honest. The Microsoft Marketplace install experience is rougher than it should be, audit logs and Compliance API coverage are not yet available, and usage limits are shared with the same Claude subscription that powers web chat, meaning a heavy modeling session eats into the same allowance you use for everything else. For most of the readers this guide is written for, those are acceptable costs in exchange for the most accurate read of a complex workbook we tested.
When Copilot is the right call instead
Microsoft 365 Copilot in Excel is the tool we recommend for any organization where the workbook contains data that must not leave the Microsoft 365 boundary. That’s a real and growing set of buyers: healthcare, finance, government, anyone with a documented compliance posture and a tenant they’ve spent years governing. Copilot’s argument isn’t that it’s the smartest tool in the test (Claude beat it on the hardest analytical tasks) but that it’s the only tool that inherits every admin control, eDiscovery rule, and sensitivity label your IT team has already configured. With Agent Mode now generally available on both Windows and Mac as of the April 22, 2026 build, the gap to Claude on raw capability has narrowed. The remaining gap is cost: the $30/user/month enterprise add-on layers on top of a base license that ranges from Business Standard at $12.50 up to E5 near $57.
When Shortcut is the answer
Finance modeling is the category’s clearest specialist case, and Shortcut is the strongest specialist we tested. It’s designed around the workflows of DCF, LBO, and three-statement modeling; it keeps formulas live so the model still works after the AI touches it; and on the build-from-scratch NVIDIA model it produced the most audit-ready output of any tool in the test. For a single analyst, the Pro plan is more expensive than a Claude Pro subscription. For a team building real models on a deadline, the productivity gap is wide enough to justify it.
When Gemini in Sheets is enough
Gemini’s argument is that it’s already paid for. If your team is on a Google Workspace plan, the AI is bundled. There’s no add-in to install, no separate add-on to negotiate, and no second governance review. The April 2026 release closed the biggest functional gap by enabling full-spreadsheet construction from a plain-language prompt, and Google’s reported 70.48% success rate on the SpreadsheetBench benchmark is competitive with the rest of the field. The honest limitation is that Gemini wants native Google Sheets files; for an Excel-first team, the conversion friction makes it the wrong default.
What did not make the cut, and why
Quadratic is a credible specialist for one job, running Python and SQL inside a grid against live data sources, and it earns a recommendation in that lane. It isn’t, and doesn’t try to be, a replacement for Excel or Sheets for a non-technical user; the Python and SQL it produces is a feature for an analyst and a barrier for everyone else.
Two products that would have been in the previous version of this guide are no longer in it. Rows, the AI-native spreadsheet, was acquired by Superhuman in early 2026 and has wound down on a timeline that makes it impossible to recommend for new workflows. Numerous.ai and GPT for Work remain credible bulk-processing add-ins, but neither is a general-purpose assistant; readers whose primary need is row-by-row text classification across thousands of rows will find both useful as a layer on top of a Sheets or Excel workflow, but neither is a substitute for the tools above.
Questions Readers Ask
Which AI spreadsheet assistant do you recommend?
We recommend Claude for Excel for analysts and finance professionals working in complex multi-tab Excel workbooks, on the strength of its cell-level citations and dependency-preserving edits. For organizations that need to keep all data inside a governed Microsoft 365 tenant, Microsoft 365 Copilot in Excel (Agent Mode) is the right call. For Google Workspace-first teams, Gemini in Sheets is now bundled into the existing subscription and is the obvious default.
Is Claude for Excel really better than Microsoft Copilot inside Excel?
On the criteria our test weights most heavily, multi-tab model audit, formula explanation with citations, and dependency-preserving edits, yes. Claude reads the entire workbook, cites the cells it's referencing, and outperformed Copilot on tracing how revenue flows through a 14-tab model in our test. Copilot is the better tool when governance, admin controls, and staying inside the M365 tenant are the deciding factors.
What does Claude for Excel actually cost?
Claude for Excel doesn't carry a standalone price; it's included in paid Claude subscriptions. The Pro plan at $20 per month is the entry point and includes Excel, PowerPoint, and Word add-ins, with Outlook in beta. Max, Team, and Enterprise plans add higher usage limits and admin controls.
Will Gemini in Google Sheets work on my Excel files?
Not natively. Gemini in Sheets is designed for native Google Sheets files; to use Gemini features on an Excel .xlsx file you have to convert it to Google Sheets first. For an Excel-first team that wants AI inside the workbook itself, Claude for Excel or Microsoft Copilot is the better fit.
Which tool is best for building a financial model from scratch?
Shortcut. It was built for finance specifically, preserves live formulas instead of hard-coding AI output, and produced the strongest from-scratch three-statement model in our test. Claude for Excel is a close second and is materially cheaper if a Claude subscription is already in your stack, but Shortcut's purpose-built design earns it the top spot for DCFs, LBOs, and three-statement models.
Can these tools handle bulk row work, categorizing or summarizing thousands of rows?
It varies. Quadratic handled the 2,000-row classification task fastest and most cleanly in our test, because Python cells iterate at scale; Gemini's AI function and Numerous-style add-ins are also designed for this. Claude for Excel and Microsoft Copilot are weaker here. Both are better suited to structured analytical tasks than to bulk row-by-row transformation of large datasets.