How good is Claude in Excel?

The Spreadsheet Problem: Complex Workbooks AI Tools Keep Getting Wrong

If you've ever inherited a fifteen-tab workbook with no documentation and formulas that reference cells three sheets away, you know the fear of pressing Enter on anything. That's the real test for AI in Excel — and it's where most tools fall apart.

The hard part isn't writing a VLOOKUP. It's tracing formula dependencies across tabs, auditing a fragile model built by someone who left the company years ago, and changing an assumption without silently breaking a chain of calculations. As one DataCamp tutorial puts it, Excel behaves more like a structured programming system than people admit — cells, formulas, and dependencies reward tools that can trace logic, not just generate text.

Here's why tool choice matters so much. AIMultiple's independent benchmark of 15 AI Excel tools found a startling spread: Claude for Excel scored 95% accuracy on a 40-question workbook benchmark across four real-world use cases, but results fell off a cliff from there — Quadratic, R2 Copilot, and Paradigm managed roughly 75%, Formula Bot just 15%, one tool hit 5%, and ExcelAIBot scored 0%. Picking the wrong assistant isn't a minor efficiency loss; it's actively wrong numbers in your model.

Even the best tools still make mistakes. Hands-on testing by Artificial Corner found Claude's first-draft financial model contained formula-linking errors — increasing headcount didn't increase staff costs — plus a missing growth assumption, requiring a 10–15 minute human audit. AIMultiple warns bluntly that tools "still return wrong values," and ICAEW guidance cited in a Claude vs. ChatGPT comparison recommends treating AI formulas as a draft, not a finished answer.

The practical takeaways before you trust any AI with your workbooks:

• Complex, multi-tab reasoning is where tools diverge most — benchmark spreads range from 95% down to 0%.
• First drafts routinely contain broken formula links and missing assumptions.
• Every credible source insists on human review before trusting output.
• The new skill isn't building the model — it's auditing the logic and interpreting the output.

That last point reframes the whole question. The value isn't in the build; it's in the audit. And if you extend your assistant with community-built agent skills for spreadsheet work, remember that published research found roughly 26% of skills carry at least one vulnerability — SkillzBench is building free security scanning and sandbox benchmarking precisely so you can verify a skill before it touches your data.

Where Claude Actually Excels in Excel: The Evidence

Ask an AI to write a VLOOKUP and you'll get roughly the same answer everywhere. Ask it to trace a broken formula across five tabs in an inherited financial model, and the differences become stark — that's where Claude for Excel separates itself from the pack.

In an independent benchmark of 15 AI Excel tools, Claude for Excel scored 95% workbook accuracy, the second-highest result behind GPT for Excel's 97.5% — well ahead of tools like Quadratic, R2 Copilot, and Paradigm at roughly 75%, and Formula Bot at just 15%. The test covered a 20-question finance benchmark (DCF, PV, FV) and a 40-question workbook assessment across four real-world use cases, with a 1% error margin. Notably, the researchers still warn that "tools still return wrong values, as our benchmark shows, so users need to check the output."

What gives Claude an edge on complex work is how it reasons before it acts. Unlike ChatGPT, which writes the formula directly with a plain-language explanation, Claude highlights the specific cells it references, explains its logic first, and flags assumptions about your data structure. Every response includes cell-level citations, so you can trace exactly where each number came from. DataCamp's testing confirms it never silently edits a file — proposed changes appear as before/after previews, and nothing updates without approval.

Hands-on testing bears this out on real tasks:

• Building a usable dashboard from a 6,000+ row Olympic medals dataset, using Python scripts for insights that would be hard to recreate manually in Excel
• Tracing formula dependencies across tabs and auditing multi-step financial models — tasks where it reportedly outperforms ChatGPT for Excel
• Explaining nested formulas and navigating large, undocumented workbooks, which DataCamp's author names as its best use cases

That said, the honest picture includes the tie: on everyday formulas like VLOOKUP and IF statements, Claude and ChatGPT perform similarly, so you're not gaining much on simple asks. And Claude's first-draft financial model in one hands-on test contained formula-linking errors — increasing headcount didn't raise staff costs — requiring 10–15 minutes of human auditing to fix.

That audit step is the real takeaway. As the testers put it, "the value isn't in the build; it's in the audit." Claude's transparency makes that audit faster and more trustworthy, which is exactly why it suits models and analyses that must withstand scrutiny.

For readers extending Claude's Excel abilities with agent skills — community packages promising Power Query ETL or plain-English spreadsheet Q&A — the same verification mindset applies. Published research found roughly 26% of agent skills carry at least one vulnerability, and SkillzBench is building a free directory with security scanning and sandbox benchmarking to help you separate useful skills from risky ones before they touch your workflow.

Claude's Real Limitations: What It Can't Do (and What It Costs)

For all its strengths, Claude for Excel comes with constraints that are consistent across every credible evaluation — and they matter for anyone planning a workflow around it. The first is cost: there is no free tier. The add-in requires a paid Claude plan, starting with Pro at $20/month and scaling through Max (from $100/month), Team, and Enterprise tiers, so teams without existing subscriptions simply cannot use it, according to a comparative review of AI Excel tools.

Platform coverage is another hard limit. Claude for Excel works only in Microsoft Excel — there is no Google Sheets support — which rules it out for organizations that live in Sheets. And while it handles formulas and multi-tab reasoning well, it offers no native VBA or macro execution (code must be generated in chat and pasted manually) and no native Power Query support, per DataCamp's hands-on tutorial.

Large datasets expose further friction. Each interaction processes the full workbook with multiple tool calls, which consumes usage limits quickly, and there is no bulk row-by-row processing for big datasets — so high-volume, repetitive tasks are a poor fit for the add-in, as the same review documents. DataCamp also flags a security consideration worth taking seriously: prompt-injection risk from untrusted spreadsheets, where hidden cells, comments, imported text, or linked external sources can carry instructions the model might follow.

The evidence on accuracy itself is genuinely contested, and readers should weigh it fairly. AIMultiple's independent benchmark — a 20-question finance test plus a 40-question workbook benchmark with a 1% error margin — scored Claude for Excel at 95% workbook accuracy, second of 15 tools tested. But a competitor vendor's own July 2026 benchmark of 20 finance tasks claimed Claude scored just 66.3%, behind its own product. That figure is explicitly a vendor claim with no neutral corroboration, and the vendor is itself a rival tool — so treat both numbers as pieces of evidence, not settled fact.

The documented limitations, in brief:

• No free tier — Pro ($20/month) minimum, with Max from $100/month
• Excel-only: no Google Sheets support
• No native VBA, macros, or Power Query execution
• No bulk row-by-row processing; fast usage-limit consumption
• Prompt-injection exposure from untrusted workbook content

One practical note on extending Claude's capabilities: community agent skills promise Power Query ETL, advanced formula guidance, and plain-English spreadsheet Q&A — but these are largely unverified marketing claims, and one popular Excel skill carried a single GitHub star. With published research finding roughly 26% of agent skills carrying at least one vulnerability, verification matters. That's the gap SkillzBench is building for: a free directory with security scanning, version pinning, and sandbox benchmarking that publishes each skill's Lift separately from its trust score. If you want to test and equip skills for your Excel agents safely, join the SkillzBench waitlist to be first through the door.

The Audit-First Workflow: How to Use Claude in Excel Safely

The most important number in Claude's Excel story isn't its 95% benchmark accuracy — it's the 10–15 minutes a human spent fixing a first draft. That's what hands-on testing found when Claude built a financial model: the output looked polished, but formula links were broken (raising headcount didn't raise staff costs) and a growth assumption was missing entirely.

The lesson isn't that Claude is unreliable. It's that the workflow matters more than the tool. As that same testing concluded, "the value isn't in the build; it's in the audit" — your job shifts from writing formulas to verifying logic. An audit-first workflow makes that shift concrete.

Start with the guardrails Claude already gives you. Keep Ask Before Edits mode switched on; testing confirms nothing updates without approval, and Claude shows proposed changes before applying them. Then demand what makes Claude distinctive: cell-level citations that trace every number back to a specific cell. If a response doesn't reference the cells it used, ask again.

The "ask why" workflow looks like this:

• Ask before edits stays ON — review every proposed change rather than letting Claude write directly.
• Demand cell-level citations so you can trace each figure to its source cell.
• Ask Claude to explain its reasoning before it writes a formula — a documented strength over tools that just output the formula.
• Treat every output as a draft, per ICAEW guidance that AI-generated formulas "can be inconsistent and occasionally overconfident."
• Test the model's logic: change an input and verify the downstream numbers actually move.

That last step is exactly what caught the formula-linking errors in testing. A model that doesn't respond correctly to changed assumptions is broken, no matter how clean it looks. Budget 10–15 minutes for this kind of audit on any substantial build — plan for it the way you'd plan for review of a junior analyst's work.

This caution is echoed everywhere. Independent benchmarking warns that even top-scoring tools still return wrong values, so users need to check the output. DataCamp's testing agrees Claude "can still be wrong" and shouldn't run unattended. A 95% accuracy score means roughly one in twenty workbook answers has a problem — and in a financial model, one broken link can silently corrupt everything downstream.

The same audit-first discipline applies if you extend Claude with agent skills for spreadsheet work. Community skills promise Power Query automation and plain-English data analysis, but published research found roughly 26% carry at least one vulnerability. SkillzBench is building free security scanning and sandbox benchmarking precisely so you can verify a skill before it touches your workbooks — the same trust-but-verify mindset you should bring to Claude's own output.

Used this way, Claude becomes a fast, transparent first-drafter with an auditor attached: you.

Extending Claude with Agent Skills — and Why Verification Matters

Agent skills promise to fill the gaps in Claude for Excel — no native Power Query, no bulk processing, no Google Sheets — and the marketplace has responded enthusiastically. An "Excel Analysis skill" claims structured guidance for dynamic pivot tables, advanced formulas like XLOOKUP and INDEX/MATCH, and full ETL automation via Power Query and M code. A separate "CSV & Excel Data Analyst" skill advertises plain-English spreadsheet Q&A with a roughly five-minute setup, so users never write a VLOOKUP again.

Here is the uncomfortable part: these are vendor marketing claims, not verified performance evidence. The Excel Analysis skill listing had a single GitHub star — minimal community validation for a package promising enterprise-grade ETL. And the risk isn't just that a skill might not help. Published research found that roughly 26% of agent skills carry at least one vulnerability, meaning about one in four instruction packages you install could expose your agent to security problems.

This matters more in Excel than in most domains. Spreadsheets reward tools that trace logic and explain what's happening, and Claude's standout traits — cell-level citations, reasoning shown before formulas are written, nothing updating without approval — exist precisely because spreadsheet errors are expensive. An unverified skill bolted onto that careful system can undermine the transparency you're paying for. It's the same lesson every hands-on evaluation of Claude for Excel reached: first drafts fail. A financial model built by Claude required a 10–15 minute audit to catch broken formula links, and the value, as one tutorial put it, "isn't in the build; it's in the audit." Skills deserve the same skepticism.

So how do you find Excel skills that actually help? This is the problem SkillzBench is being built to solve. Its approach covers the full lifecycle of skill adoption:

• Discover and dedupe — continuously crawling GitHub so community Excel skills are findable in one place, not scattered across marketplaces.
• Verify before you trust — pinning every skill version to a commit SHA, running license checks and security scans, and assigning a 0–100 trust score from security, quality, license, maintenance, and community signals.
• Benchmark in a sandbox — testing every skill with the same model, with and without the skill, and publishing the Lift in percentage points so you can see what a skill actually adds.
• Equip per agent — one free MCP server plus per-agent loadouts, so your Excel agent only sees the skills and tools you chose for it.
• Stay current — tracking upstream changes, re-scanning, and re-benchmarking so a skill that was safe last month doesn't silently become a liability.

Everything on SkillzBench is free — the directory, the MCP connector, benchmarks, and alerts — and money never touches scores. Lift is always shown separately from ratings, and advertising revenue never influences listings. The platform is pre-launch and actively being built, so the waitlist is the way to get in early.

If you're extending Claude for Excel with community skills, treat them like Claude's own first drafts: promising, occasionally wrong, and always worth verifying before they touch your workbook. Join the SkillzBench waitlist to be first in line when verified, benchmarked Excel skills are ready to browse.