Can AI Agents Actually Handle Excel? What Skills Can and Cannot Do in 2026

The Excel Agent Hype vs. A 0%-97.5% Reality

If you ask "can AI agents handle Excel?" in 2026, the honest answer is a shrug followed by a spreadsheet of test scores. Independent testing of 15 AI Excel tools found accuracy ranging from 0% to 97.5% — and the researchers' conclusion was blunt: tools still return wrong values, so users need to check the output.

That spread is the real story. In the same benchmark, one tool scored 97.5% on workbook questions while another scored literally zero. Formula Bot hit 65% on finance calculations but only 15% on workbook questions. The reason for the divergence is simple: the tests measure different skills, and tools excel at different ones. Calculation accuracy and explanation quality don't even reliably co-occur.

So "can it do Excel?" is the wrong question. The useful version has three parts:

• Which task? Formula creation, pivot tables, and data cleaning score well across tools; bulk row-by-row transformations and complex multi-step analysis trip up most of them.
• Which tool? A purpose-built finance agent and a general assistant are not interchangeable, even when marketing copy sounds identical.
• Verified how? Vendor benchmarks abound — Shortcut claims 77.3% vs. Claude for Excel's 66.3% on finance tasks; PwC claims 3x accuracy on enterprise workbooks — but these are self-reported, not neutral tests.

The failure modes observed in independent testing are worth internalizing before you trust an agent with your month-end close. One tool returned only the first two years of a six-year financial plan. Another required a fourth chat attempt to produce consistent results. Still another couldn't use figures stated in the chat unless they also lived inside the spreadsheet.

This matters because adoption is racing ahead of verification. Some 79% of companies claim to have adopted AI agents, and two-thirds of those report measurable productivity gains. Enterprise usage is shifting from asking models for outputs to delegating multi-step workflows, per OpenAI's enterprise report. Meanwhile, the agent-skills ecosystem that extends these tools is growing fast — and largely unvetted. Published research found roughly 26% of agent skills carry at least one vulnerability.

That's the gap SkillzBench is being built to close: a free directory that verifies, security-scans, and benchmarks agent skills, publishing each skill's measured Lift so you can see what it actually adds before it touches your workflow. The tools will keep improving. The verification habit is yours to build either way.

What AI Agents Genuinely Do Well in Excel (And Where They Fall Apart)

AI agents in Excel are no longer guessing at formulas from a chat window — they read workbooks, make edits, and explain what they changed. But independent testing shows the gap between what agents promise and what they deliver is enormous.

In June 2026, analysts ran a hands-on benchmark of 15 AI Excel tools and found accuracy ranging from 0% to 97.5% — and even top performers "still return wrong values," so users need to check the output. That single finding shapes everything else: agents are genuinely useful, but verification is not optional.

Where agents genuinely shine is well-documented. According to independent benchmark testing, the strongest tools handle formula creation, pivot tables, charts, and multi-tab workbook comprehension — GPT for Excel scored 97.5% and Claude for Excel 95% on workbook questions. Data cleaning is another real strength, with vendors reporting reductions of up to 80% in cleaning time. Plain-language Q&A about your spreadsheet also works well in most modern tools.

The weak spots are just as clear:

• Bulk row-by-row processing at scale — most mainstream agents, including Copilot, Claude for Excel, and the ChatGPT add-in, cannot reliably process thousands of rows one at a time.
• Complex multi-step analysis — Gemini's own documentation notes that multi-step requests can return incomplete or inconsistent results.
• Very large workbooks — enterprise spreadsheets with dense formulas and hidden dependencies were historically "beyond the reach of even the most advanced models."
• Usage quotas and context limits — recurring constraints across nearly every tool tested.

Failure modes in testing were telling. One tool answered only the first two years of a six-year financial plan; another needed four attempts to respond consistently. Calculation accuracy and explanation quality don't always co-occur in the same tool, so a strong performer on one task may fail on another.

This is exactly why vetting matters before you trust any agent — or any agent skill — with real spreadsheets. Published research found roughly 26% of agent skills carry at least one vulnerability, and most published benchmark claims (like Shortcut's 77.3% vs. Claude's 66.3%) come from vendors testing their own products. SkillzBench is building a free directory that verifies, security-scans, and benchmarks skills in a sandbox so you can see the measured Lift each one delivers before it touches your workbooks.

The honest 2026 picture: agents handle a meaningful share of Excel work well, but no single agent excels at everything, and every output deserves a human check.

ctaText: Agent skills can automate the rest of your spreadsheet workflow — browse and test free ones on SkillzBench. socialProofText: SkillzBench is free, independent, and never lets money touch scores — benchmarks and security scans included.

Why Vendor Benchmarks Keep Lying to You

Every vendor benchmark you read about AI and Excel was run by someone selling the tool. That is not a conspiracy — it is just how the market currently works, and it explains why the numbers you see never agree with each other.

The conflict is stark. Shortcut's July 2026 benchmark reported 77.3% accuracy against Claude for Excel's 66.3% on twenty finance tasks, a result the covering comparison explicitly flags as a vendor claim rather than a neutral test. PwC's headline figure — up to three times higher accuracy on enterprise-grade spreadsheets — comes from internal research that no outside party has verified. Neither claim is false. Neither is independently confirmed either.

Then there is the deeper problem: the tests measure different skills. Independent testing of fifteen AI Excel tools found accuracy ranging from 0% to 97.5%, and the analysts behind it note that scores diverge precisely because the tests measure different things. Formula Bot scored 65% on finance calculations but only 15% on workbook questions. ExcelAIBot scored zero. A tool that wins a formula-generation benchmark may collapse on multi-sheet comprehension, and a tool that explains its reasoning well may still return wrong values.

• A high calculation score does not predict explanation quality, and vice versa
• Workbook comprehension and row-by-row bulk processing are separate capabilities
• Vendor rankings shift depending on who is publishing them
• Small sample sizes (20 to 40 questions) make single benchmarks fragile

The rule that follows is simple: match the tool to your task, then test on your own workbooks. No published benchmark can tell you whether an agent handles your messy, half-documented, twelve-tab model — only your own workbook can.

That same discipline applies beyond spreadsheets. Agent skills — the SKILL.md packages that extend Claude Code, Codex, Cursor, and Gemini CLI — are proliferating without independent verification, and published research found roughly 26% carry at least one vulnerability. SkillzBench is building a free directory that verifies, security-scans, and benchmarks skills so you can see the Lift each one actually delivers before you trust it.

Your 2026 Excel AI Playbook: Test, Verify, Then Delegate

The gap between "my agent works great on the demo" and "my agent works on quarter-end books" is where most AI spreadsheet projects die. Independent testing of 15 AI Excel tools found accuracy ranging from 0% to 97.5% — and the testers' own conclusion was blunt: tools still return wrong values, so users need to check the output (AIMultiple's benchmark).

Adoption momentum is real. According to PwC's AI agent survey, 79% of companies claim agent adoption, and two-thirds of those report measurable productivity gains. But claiming adoption is not the same as proving capability on your workbooks. Here is how to close that gap.

Test on your actual workbook types first. A tool that scores 97.5% on workbook questions may still stumble on your specific structure — one benchmarked tool returned only the first two years of a six-year financial plan. Run every candidate agent on a representative sample: a messy multi-tab file, a dense formula model, a routine cleaning task. Score results yourself before committing.

Build human review into every agent-edited workflow. This is not optional overhead; it is the difference between delegation and liability. The failure modes are documented and predictable:

• Partial or truncated answers that look complete at a glance
• Misunderstood questions on multi-step requests
• Inconsistent results requiring multiple attempts on the same prompt
• Context limits and usage quotas cutting off large workbooks mid-task

Verify bulk-scale capability specifically. Most mainstream agents cannot reliably process thousands of rows row-by-row (one comparison notes only a handful handle bulk scale). If your workload is mass transformation, test with your real row counts, not a 50-row sample. Vendor claims like 1,000 rows per minute are exactly that — vendor claims.

Treat enterprise-workbook claims as unverified until proven on your data. PwC's frontier agent claims up to 3x accuracy on enterprise-grade spreadsheets, but that figure is internal research, not independent verification (PwC's announcement). The same discipline applies to any benchmark a vendor publishes about itself. Test first, trust second, delegate third.

If you are extending agents like Claude Code or Codex into spreadsheet workflows with agent skills, the same rule applies — published research found roughly 26% of skills carry at least one vulnerability, so vetting before deployment matters. SkillzBench is building a free directory that security-scans and benchmarks skills so you can see exactly what each one delivers. Verification is not a step in the workflow — it is the workflow.

Leveling Up: Agent Skills for Spreadsheet Workflows — Vetted, Not Vibed

Once an AI agent has written a formula or cleaned a column, the next question is how to make that capability repeatable. That's where agent skills come in: SKILL.md instruction packages that teach tools like Claude Code, Codex, Cursor, and Gemini CLI to handle recurring spreadsheet workflows the same way every time, instead of re-prompting from scratch.

The timing fits a real shift. OpenAI's enterprise research describes a transition "from asking models for outputs to delegating complex, multi-step workflows," and notes that 75% of surveyed workers now complete new technical tasks, including spreadsheet analysis and automation. Skills are how that delegation gets packaged and reused.

But there's a catch that most skills hype ignores. Published research found that roughly 26% of agent skills carry at least one vulnerability. Installing an unvetted skill that reads — and edits — your financial workbooks is a different risk profile than trying a chatbot once. Independent testing of AI Excel tools already shows why skepticism is warranted: accuracy across 15 tools ranged from 0% to 97.5%, and tools "still return wrong values," according to AIMultiple's benchmark. A skill that amplifies a weak or unsafe agent just industrializes the problem.

Before trusting any skill with your data, look for evidence on three fronts:

• Security — has the skill been scanned for vulnerabilities, and is the version pinned to a specific commit so you know exactly what you installed?
• Provenance — who wrote it, under what license, and is it actively maintained as the upstream repo changes?
• Measured impact — does the skill actually improve results on spreadsheet tasks with the same model, tested with and without it?

That last point matters most. A skill should earn its place with a measurable Lift — a concrete improvement in percentage points — not with a persuasive README. Calculation accuracy and explanation quality don't always co-occur in AI spreadsheet tools, so a skill that helps one task may quietly hurt another.

This is the gap SkillzBench is being built to close: a free, independent directory that continuously discovers skills from GitHub, security-scans and license-checks each one, benchmarks every skill in a sandbox with and without it installed, and publishes the Lift it delivers. A free MCP connector then lets you equip any agent with only the vetted skills you choose — and the platform re-scans and re-benchmarks as upstream code changes, so yesterday's safe skill can't silently become tomorrow's liability.

Everything on SkillzBench — the directory, the scanning, the benchmarks, the MCP connector, the alerts — is free, with trust scores kept strictly separate from any advertising. Spreadsheet work sits at the heart of business decisions; the skills that touch it should be vetted, not vibed.