Can AI Agents Write Excel Formulas?

The Formula Gap: Why Excel Users Are Drowning in Syntax They Never Memorized

Open almost any laptop in any office and there's a good chance Excel is already running. It's the most widely used business tool on the planet, and most of what people do in it every week is repetitive in a way that's genuinely automatable.

The problem isn't laziness. It's that Excel ships with hundreds of formulas, and even the best power users don't know every single one. So a task that should take a minute becomes twenty minutes of Google searches, Stack Overflow tabs, and half-remembered syntax. That gap between this should be automated and I don't know how to automate it is exactly where frustration lives.

And the manual work most people do is fragile. Linking an income statement to a balance sheet to a cash flow statement is where most financial models break. One circular reference and the whole thing spirals into #REF errors that take hours to untangle. The work is repetitive, but it's also brittle enough that one wrong click costs an afternoon.

That brittleness is what pushes people toward AI help. According to a recent breakdown of Claude Skills for Excel, the most painful spreadsheet tasks aren't exotic — they're the same joins, lookups, and cross-sheet references that nobody wants to rebuild from scratch each month.

The repetitive work that fills a typical week includes:

• VLOOKUPs and INDEX-MATCH across messy sheets
• Nested IF statements for categorizing transactions
• SUMIFS and COUNTIFS for reporting
• Cross-sheet links between statements
• Cleaning and reshaping imported data

None of this is glamorous. All of it is the kind of work where syntax memorization becomes the bottleneck, not the actual thinking. That's the formula gap — and it's the reason a generation of office workers is quietly hoping the machine can just write the formula for them.

Yes, AI Agents Can Write Formulas — The Evidence Across Eight Sources

The short answer is yes — and it's no longer a fringe capability. Across all eight sources examined, every one independently confirms that AI agents can generate Excel formulas from plain-language descriptions, from Microsoft's built-in Copilot features to third-party tools and open-source agent frameworks.

Microsoft's own Formula AI demo makes the case concretely. When asked to calculate the total Trailhawk and Roadhawk bikes available in European warehouses, the model generated a SumIfs formula, correctly identified Dublin, Berlin, London, Paris, and Madrid as European cities, and arrived at a verified total of 845. Users simply describe what they want without knowing the formula name, and Copilot suggests the right one — with automatic completion triggered the moment you type "=". Microsoft's official Excel AI page confirms formula generation among its supported features.

A three-agent benchmark put the capability to a practical test, and all three tools successfully wrote and applied a formula to flag returned orders. Formula-writing took 9 seconds for the fastest tool, 34 seconds for Claude in Excel, and 44 seconds for Copilot. The same benchmark found these agents handling VLOOKUP, INDEX-MATCH, nested IF statements, and array formulas from plain English — no syntax memorization required.

But there's an important distinction between AI functions and true agents. As one hands-on tutorial puts it, "a cell function can only do one thing per cell. An agent can plan, execute, and refine data across an entire spreadsheet." Microsoft's =COPILOT() function blurs that line — it lives in Excel's calculation engine, updates automatically when source data changes, and can be nested inside other formulas. Agent Mode goes further, building and iterating on workbooks until outcomes are verified.

Where the evidence is weaker is on accuracy claims. Vendor-reported figures — such as a specialized spreadsheet agent achieving 96.1% accuracy versus OpenAI's Code Interpreter at 75.3%, or top tools reaching roughly 75% accuracy on complex financial calculations — come without cited primary sources, so treat them as claims rather than established facts. The practical takeaway: AI formulas are drafts to verify, not finished work.

For readers who want to push formula work into full automated workflows, agent skills are the natural next step — instruction packages that give agents structured, repeatable Excel capabilities like the Anthropic XLSX skill covering spreadsheets, formulas, and financial models. Quality varies, though: published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is building a free directory that benchmarks agent skills in a sandbox and publishes the Lift in percentage points, so you can see which Excel skills actually help — the waitlist is live at skillzbench.com.

Speed, Scale, and Accuracy: Where AI Formula Writers Shine and Where They Break

Speed matters when you're staring at a 10,000-row spreadsheet, and AI formula writers deliver — but inconsistently. In a three-tool benchmark, formula writing ranged from 9 seconds (GPT for Excel) to 44 seconds (Copilot), with Claude in Excel landing at 34 seconds.

The real gap shows up at scale. The same benchmark found GPT for Excel completed 100-row, 1,000-row, and 10,000-row bulk tests without issue. Claude in Excel hit rate limits and failed the 100-row, 1,000-row, and 500-row tests, never even starting the 10,000-row run. The benchmark's author flagged that combination — rate limits plus no progress indicators — as "a dealbreaker for production workloads."

Consistency checking tells a similar story. GPT for Excel finished in 16 seconds; Claude took 26 seconds; Copilot needed 4 minutes 46 seconds — roughly 18× slower than the fastest tool. On the full 11-test benchmark, GPT for Excel finished in under 7 minutes while Copilot regularly took 1–5 minutes per task, often 5×–10× slower.

Accuracy claims deserve more skepticism. Vendor-published figures include a specialized spreadsheet agent at 96.1% accuracy with GPT-4 versus 75.3% for OpenAI's Code Interpreter, and top AI Excel tools at roughly 75% accuracy on complex financial calculations. None of these figures have independent verification. Treat them as marketing claims, not benchmarks.

The practical takeaway:

• Match the tool to the job size — confirm row limits and progress visibility before committing to bulk work.
• Treat every AI-generated formula as a draft. Verify outputs, especially on complex logic.
• Watch for silent failures. One tool "finished fastest" on a bulk test by applying text formulas instead of generating real content — "worse than producing no result."
• Cap agent loops and validate tool calls if you're building your own. Models hallucinate parameter names; try/catch wrappers let them self-correct.

Agent skills extend this further — the Anthropic XLSX skill covers spreadsheets, formulas, and financial models, while community Office MCP servers expose dozens of spreadsheet-specific tools. But skill quality varies; published research cited by SkillzBench found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory that benchmarks skills in a sandbox and publishes the Lift in percentage points, so users can see which Excel and office skills actually help before installing them.

From One-Off Formulas to Full Workflows: How Agent Skills Extend the Capability

Writing a single Excel formula is one thing. Running a full spreadsheet workflow — merging workbooks, executing macros, building a three-statement financial model — is where agent skills start to change the economics of office work.

The ecosystem around this is already substantial. Anthropic publishes an official XLSX skill that handles "Excel spreadsheets, formulas, financial models," and the broader community has built 136+ Office Skills and 39 MCP tools, seven of them spreadsheet-specific. Claude Skills demonstrate two practical approaches: live control via xlwings (running VBA macros, live dashboards) and file manipulation via openpyxl (bulk merging and normalization). A routine that takes "two hours of clicking" can collapse to "thirty seconds," per one hands-on report.

The shift matters because it moves agents beyond cell-level assistance into multi-step autonomous work. Microsoft's Agent Mode "brings advanced agentic AI reasoning to your open files" and iterates until outcomes verify. Where traditional automation — macros, VBA, Power Query, Office Scripts — follows rigid rules and breaks when data structures shift, agents "reason about tasks instead of just following fixed rules," according to nexos.ai.

But skill quality is uneven. Published research found roughly 26% of agent skills carry at least one vulnerability, so installation without verification is a real risk. A few categories of Excel skill tend to deliver the most lift:

• Financial modeling skills — three-statement models, scenario analysis, DCF templates
• Data merging and normalization — bulk openpyxl workflows across many files
• Live Excel control — xlwings-based macros, dashboards, and Office Script execution
• Formula generation and audit — natural-language to formula plus consistency checking
• Pivot table and reporting skills — automated summarization across recurring reports

This is exactly the gap SkillzBench is building to close. The platform is being designed as a free, independent directory that security-scans and sandbox-benchmarks every skill, publishing the Lift so you can see which Excel skills actually move the needle before you install them. A free MCP connector will let any agent load only the skills you select. The directory, MCP, benchmarks, and alerts are all planned to remain free, supported by advertising that never touches scores.

Skills, proven — that's the standard the space needs. Join the SkillzBench waitlist to get early access when the directory and benchmark results go live.

Your Action Plan: Getting Reliable AI-Written Formulas Without the Risk

Knowing AI agents can write Excel formulas is one thing; getting reliable results from them is another. Here is a practical playbook for using them well — whether you're a business user or building the agent yourself.

Start with plain-language generation for routine formulas, but treat every output as a draft. In a three-tool benchmark, all tested agents successfully wrote and applied a formula to flag returned orders, so the capability is real. But with vendor-claimed accuracy of roughly 75% on complex financial calculations, verification is non-negotiable — spot-check the logic against a few rows you already know the answer to.

Keep traditional methods where they win. Per guidance on Excel automation, plain formulas remain best for in-cell calculations, macros for fixed predictable tasks, and Power Query for stable recurring pipelines. AI agents earn their keep when work goes beyond math into interpreting, transforming, or producing new outputs — especially with messy or changing data.

For bulk work, confirm row limits and progress visibility before committing. The benchmark showed one tool completing 10,000-row tests in under 12 minutes while another hit rate limits and failed at 100 rows — and reviewers called missing progress indicators "a dealbreaker" for production workloads.

If you're building spreadsheet agents, engineer for failure from day one:

• Cap agent loops — the reference implementation limits agents to 10 steps so they can't spiral.
• Validate tool calls with try/catch and return errors to the model — per the hands-on tutorial, "most models will fix their own mistakes on the second attempt if you tell them what went wrong."
• Batch writes and show visible progress so long jobs stay observable.
• Use larger models (70B+, GPT-4o, or Claude) for complex multi-step reasoning across 50+ rows.

Agent skills can extend formula work into full workflows — packages like the Anthropic XLSX skill cover spreadsheets, formulas, and financial models, and community MCP servers offer dozens of Office tools. But skill quality varies widely, and published research has found roughly 26% of skills carry at least one vulnerability. That's why SkillzBench is building a free directory that verifies, security-scans, and benchmarks agent skills in a sandbox — so you can see which ones actually help before you install them.

Ready to put this into practice? Browse and test free, verified Excel agent skills on SkillzBench — skills, proven.