How can I use AI to help me in Excel?

The Excel Bottleneck: Why Manual Spreadsheet Work Still Slows You Down

It's 2026, and most knowledge workers still spend their mornings doing the same thing: clicking through spreadsheet cells by hand. A 2025 Alteryx survey found that 76% of data professionals still rely on spreadsheets as their primary data-cleaning tool — the very people whose job it is to wrangle data haven't automated the wrangling.

The bottleneck isn't a lack of awareness. AI adoption is now mainstream — 88% of organizations use AI in at least one business function — yet the day-to-day grind remains stubbornly manual: fixing broken formulas, deduplicating customer lists, standardizing inconsistent entries, and rebuilding the same monthly report for the third time.

Leaders feel the cost of this gap. In a 2025 IBM study, 43% of COOs ranked data quality issues as their top data priority, and Gartner predicted that poor data quality would contribute to at least 30% of generative AI projects being abandoned by the end of 2025. When the data underneath is hand-cleaned, every AI initiative built on top inherits the errors.

The irony is that the tools have moved well beyond simple formula generation. Modern Excel AI can now:

• Write, explain, and fix formulas from plain-language descriptions
• Clean messy columns — deduplicating rows, standardizing formats, and filling missing values
• Classify and tag rows in bulk by category or sentiment
• Build pivot tables and charts, then surface patterns and insights from the data

Microsoft itself positions Copilot in Excel as a way to speed up adding columns and formulas, formatting tables, and coming up with insights about your data. OpenAI's ChatGPT for Excel add-in lets you describe what you need in plain language and updates live models directly in the workbook.

So why does the manual work persist? Partly because the hype hasn't matched reliability — independent testing has caught even flagship assistants freezing mid-task or failing to apply changes across thousands of rows. Partly because it's hard to know which tools actually deliver versus which just write a decent SUM formula. As one practitioner put it: "If all it does is write a SUM formula, it is not really doing data analysis."

The gap between AI's promise and spreadsheet reality comes down to choosing and verifying the right tools for the specific job — which is exactly why we're building SkillzBench: a free directory where agent skills for spreadsheet workflows are tested and benchmarked so you can see which ones actually help before trusting them with your data. The waitlist is open now.

Two Kinds of AI for Excel: Agents That Edit vs. Tools That Suggest

Pick the wrong type of AI spreadsheet helper and you'll either overpay for power you don't use or spend hours pasting formulas by hand. The Excel AI market has quietly split into two camps, and understanding that split is the single most useful decision you can make before spending a dollar.

The first camp is agents that edit your workbook directly. Microsoft Copilot, Claude for Excel, and ChatGPT for Excel live inside your file: they add columns, write formulas, format tables, and surface insights without you touching a cell. Microsoft positions Copilot as an accelerator for exactly these tasks, per its official product page, while OpenAI's add-in creates and updates live Excel models from plain-language requests.

The second camp is standalone formula generators like Formula Bot, Ajelix, and GPTExcel. They hand you formulas or scripts to paste in yourself — as one detailed comparison puts it, none of them can edit your file, and that is the single biggest difference between the two groups.

The trade-off comes down to cost versus effort:

• Agents cost more but save more time. Copilot runs $30/user/month, Claude for Excel starts at $20/month, and ChatGPT for Excel is included with ChatGPT plans.
• Standalone tools are cheaper to start. GPTExcel's paid tier is $9/month — the lowest in the category — but the work of applying results stays with you.
• Reliability varies even among agents: an independent five-task test found Copilot froze mid-task and never finished one error check.

Then there's the differentiator most people miss: bulk processing. According to capability testing across major tools, most built-in assistants — Copilot, Claude for Excel, the ChatGPT add-in — cannot reliably apply prompts across thousands of rows, such as classifying or tagging entire columns. They suit structured, step-by-step tasks, not large-scale transformations.

That matters because spreadsheets remain the workhorse: a 2025 Alteryx survey found 76% of data professionals still rely on them as their primary cleaning tool. If your job involves 50,000 rows of sentiment tagging, verify bulk capacity before you buy anything.

One way to close that gap is to pair your spreadsheet tool with agent skills that automate the surrounding workflow — SkillzBench is building a free directory and MCP connector for exactly that, so you can browse and test skills for feeding cleaned data into downstream agents.

What AI Can Actually Do in Excel Today — and Where It Still Fails

AI in Excel has quietly gone from party trick to daily workhorse — but the gap between marketing claims and real performance is where most users get burned. Knowing both sides helps you use it well.

Today's tools genuinely handle a broad slice of spreadsheet work. Microsoft positions Copilot as an accelerator for adding columns and formulas, formatting tables, and surfacing insights, while OpenAI's ChatGPT add-in creates and updates live models directly inside the workbook for analysis, reporting, and budgeting, per its official announcement. AI cleaning also matters because spreadsheets aren't going anywhere: a 2025 Alteryx survey found 76% of data professionals still rely on spreadsheets as their primary cleaning tool (OvalEdge's analysis).

What AI can realistically do right now:

• Generate and explain formulas from plain-language descriptions
• Clean data — dedupe, standardize, and fill missing values at scale
• Categorize rows or tag sentiment across columns
• Build pivot tables and charts from raw data
• Surface patterns and insights you might otherwise miss

Now the honest part. OpenAI itself admits that complex formulas or edge cases may still require manual refinement, and outputs occasionally need cleanup to match your formatting conventions. Independent testing adds weight: one test across five finance tasks found Copilot froze mid-task and never finished the error check, returning a broken result on the final analysis. Reddit users called its agent mode "gimmicky in its current state."

Bulk work is another weak spot. Most built-in assistants cannot reliably apply prompts across thousands of rows — classifying or tagging entire columns remains beyond them. Only a handful of tools handle large-scale transformations, so verify bulk capability before committing to a workflow.

Treat vendor benchmarks skeptically, too. Headline numbers like 77.3% accuracy or 16-second task times come from the vendors' own tests, not neutral third parties — the sources themselves flag these as claims rather than established facts. That absence of independent verification is exactly the problem SkillzBench is built to solve in the agent-skills space: benchmarking every skill in a sandbox, with and without it, so Lift scores reflect real testing rather than marketing. If you want to extend your Excel workflow with agent skills, you'll be able to browse and test free ones on SkillzBench once it launches.

The takeaway: use AI for formulas, cleaning, and first-pass insights — but always review the output, especially on complex logic and full-column transformations.

Your AI-in-Excel Playbook: Prompts, Reviews, and Data Safety

Knowing the pitfalls is only half the battle — the other half is a repeatable routine that keeps AI helpful without letting it quietly corrupt your numbers. The difference between an AI assistant that saves you hours and one that wastes them usually comes down to four habits you control.

First, write specific prompts. A recurring complaint in practitioner reviews is that vague prompts produce confidently wrong answers — the same dataset yields shallow categorization or deep trend analysis depending entirely on how narrowly you scope the request. Instead of "analyze my sales data," try "classify column D by product line, then show month-over-month revenue change for Q3 using only rows where status equals 'closed.'"

Second, always review complex formulas. OpenAI itself admits that complex formulas and edge cases may still require manual refinement, and outputs may need cleanup to match your formatting conventions. Prefer tools with cell-level citations — Claude for Excel, for example, lets you trace exactly where each number or insight comes from, and ChatGPT for Excel links answers to the specific cells it references.

Third, verify bulk capability before committing to a tool for large jobs. Most built-in assistants cannot reliably apply prompts across thousands of rows, while purpose-built options handle up to 1 million rows per Excel run.

• Confirm the tool supports bulk processing if you need to classify or tag entire columns
• Test on fake or dummy numbers before feeding in real data
• Check for ISO 27001 or SOC 2 certifications and an explicit "data will not train AI models" statement before paying

That last point matters more than most users realize. Community warnings note that free tools are often the ones that keep what you feed them, so protect your data before you pay, not after. The stakes are real: a 2025 IBM study found 43% of COOs rank data quality as their top priority, and Gartner predicts at least 30% of generative AI projects will be abandoned by end of 2025, with poor data quality a key factor.

Finally, remember that AI strengthens governance rather than replacing it — AI should flag issues; humans should approve them, especially for sensitive datasets. If you want to extend this workflow beyond the spreadsheet — feeding cleaned data into downstream agents or automating the next step — SkillzBench is building a free directory and MCP connector for exactly that kind of agent skill discovery, with security scanning and benchmarking built in.

Beyond the Workbook: Extending Your Excel Workflow with Agent Skills

A cleaned, well-structured spreadsheet is rarely the end of the story. Once your data is tidy, the real value often comes from what happens next: feeding those rows into reports, dashboards, forecasts, or automated workflows that act on the numbers without you touching them again.

This is where agent skills come in. Unlike an AI assistant working inside your workbook, agent skills are instruction packages that extend AI agents — tools like Claude Code, Codex, Cursor, and Gemini CLI — to handle tasks beyond the sheet itself. A skill might take the customer segments you just tagged in Excel and push them into a CRM, generate a summary report, or trigger follow-up automations. The demand is real: nearly 3 in 4 companies plan to deploy agentic AI within two years, according to industry adoption statistics, up from just 23% today.

But there's a catch. Agent skills are everywhere, and nobody can tell which ones actually help — and which are risky. Published research found that roughly 26% of skills carry at least one vulnerability. Installing an unvetted skill to handle your freshly cleaned financial data is exactly the kind of shortcut that creates problems later.

That's the gap SkillzBench is being built to close. It's a free, independent directory of AI agent skills, currently in pre-launch with a live waitlist, designed around a few core pillars:

• Discover — continuously crawling and deduplicating skills from GitHub so you can find spreadsheet-adjacent skills in one place.
• Verify — pinning every skill to a specific commit SHA, then running license checks, security scans, and quality reviews.
• Benchmark — testing each skill in a sandbox with the same model, with and without the skill, and publishing the measured Lift in percentage points.
• Equip — one free MCP connector that links any agent to only the skills and tools you explicitly toggle on, so your agent never sees more than you chose.

Transparency is the guiding principle. Every skill gets a trust score from 0 to 100 based on security, quality, license, maintenance, and community signals — with Lift always shown separately, and money never touching the scores. SkillzBench is free across the board, supported by ads in its blog content rather than user fees.

The same skepticism you'd apply to an Excel AI tool applies here. As one independent comparison of Excel AI tools warns, free tools are often the ones that keep the data you feed them — so verifying what any automation touches, and how it's been tested, matters before connecting it to real business data.

If your Excel workflow now ends with clean data waiting for its next step, agent skills can automate the rest. Join the SkillzBench waitlist to be first in line when the directory and MCP connector go live.