How to do Excel work with AI?

The Excel Bottleneck: Why Manual Spreadsheet Work No Longer Scales

It's 4 p.m. on a Friday, and you're still cleaning a 5,000-row export — trimming spaces, fixing dates, rewriting the same nested IF formula for the twentieth time. If that sounds familiar, you're not inefficient; the manual spreadsheet workflow itself has stopped scaling.

The scale of the problem is bigger than most people realize. Roughly 85% of company data is unstructured, which is exactly the kind of messy, inconsistent data that makes Excel work so slow, according to automation research. Every report you build starts with hours of prep before a single insight appears.

The automatable share of the work is enormous, too. Up to 80% of finance departments' transactional work could be automated, and employees themselves estimate they could save 240 hours per year by automating repetitive tasks — business leaders put the figure even higher, at 360 hours. Payment firm iPSL offers a glimpse of what's possible: by automating its Excel-based reporting, it cut budget and forecast reporting from six weeks to ten days, a 66% acceleration.

So AI is clearly the answer — but which AI, for which task? That's where most people get stuck. The market has split into three categories, and they solve genuinely different problems:

• Native assistants built into Excel, like Microsoft Copilot, which operate in Edit, Plan, and Chat modes and are well-suited to interactive work — building charts, PivotTables, and summaries with your data staying editable throughout.
• Dedicated add-ins designed for bulk processing, which vendor testing suggests dramatically outperform built-in assistants on large tasks like cleaning or translating thousands of rows.
• Standalone platforms such as KNIME or Tableau, where analysis happens outside the spreadsheet entirely.

The tension is that picking the wrong category wastes the very hours you're trying to reclaim. Microsoft's own Copilot documentation shows it's designed for guided, step-by-step work — not row-by-row throughput on a million cells. Meanwhile, most head-to-head performance comparisons come from vendors selling their own tools, so the numbers deserve healthy skepticism.

There's also a trust gap on the human side. Finance and accounting staff are famously hesitant about automation — only 66% view it positively beforehand — yet that jumps to 89% after they actually use the tools. The resistance isn't to AI; it's to unverified AI doing critical work. That same instinct matters if you go beyond plugins into agent-based automation, where published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory and MCP connector that security-scan, version-pin, and benchmark every skill before it reaches your agent — the waitlist is live now.

The bottleneck is real, the hours are quantifiable, and the tools exist. The rest of this guide shows how to match them correctly.

Three Paths to AI-Powered Excel: Native Assistants, Add-ins, and Bulk Processors

Not every AI Excel tool does the same job — and picking the wrong category for your workload is the fastest way to waste both time and budget. Roughly 85% of AI users say adoption accelerated because AI showed up inside tools they already use, so the good news is you rarely have to leave Excel at all.

Path 1: Native assistants for guided, interactive work. Microsoft's own documentation positions Copilot as a hands-on collaborator operating in three modes: Edit (changes the workbook directly), Plan (proposes a structured plan before acting), and Chat (analyzes without modifying). Its strengths are charts with labeled axes, PivotTables, conditional formatting, and anomaly detection — with you watching the reasoning live and a Stop button always available.

Path 2: Dedicated add-ins for bulk operations. Here the gap widens dramatically. Vendor testing reports Copilot is capped at roughly 100 function calculations per 10 minutes, while bulk engines like GPT for Work claim 200,000–1 million rows per single run — figures that are vendor-reported, not independently verified. Alternatives include GPTExcel ($6.99/month paid tier, 20+ million formulas created), Kutools ($49/user with 300+ features), and Claude for Excel (from $20/month via Claude Pro).

Path 3: Free built-in options. Power Query ships free inside Excel and handles serious data cleaning without any subscription — a meaningful point when cost is the top adoption barrier for 51% of businesses.

Whichever path you choose, apply four selection criteria before spending anything:

• Works where your data lives — no upload, copy-back, and reformat detours through a separate app.
• Handles bulk data — thousands of rows in one run, not a hundred cells at a time.
• Goes beyond formulas — pivots, charts, and plain-language analysis, since "a SUM formula is not really data analysis."
• Offers real model choice — check which models are selectable on your actual plan, not the marketing page.

Pricing spans the full spectrum: Copilot runs $30/user/month, GPT for Work installs free then starts at $25/month, and several tools — including the ChatGPT add-in and GPTExcel's 4-requests-per-day tier — cost nothing to start.

If your ambitions extend past plugins into agent-driven automation, the same vetting discipline applies: published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory and MCP connector that pins versions, security-scans, and benchmarks every skill, so you can test before you trust.

Match the Tool to the Task: Formulas, Data Cleaning, and Analysis in Practice

The single biggest mistake teams make when adopting AI in Excel is buying one tool and expecting it to do everything. The research points to a different approach: match the tool to the workload, because the AI that excels at building a chart will choke on cleaning 10,000 rows.

For guided, interactive work — charts with labeled axes, PivotTables, reshaping data, merging sheets, and chat-based summaries — Microsoft's native Copilot is well-suited. According to Microsoft's official documentation, Copilot operates in three modes (Edit, Plan, and Chat), creates a step-by-step plan before acting, and keeps every modification under your control with a Stop button. That design makes it ideal for beginners and one-off tasks, not heavy throughput.

Bulk work is a different story. Vendor testing suggests Copilot handles roughly 100 function calculations per 10 minutes, while bulk-capable add-ins like GPT for Work claim 200,000 to 1 million rows per run — a vendor-reported figure, but the directional gap is consistent across sources. For cleaning thousands of rows, translating reviews, or categorizing transactions, dedicated add-ins are the practical choice.

Where AI adds the most value, though, is neither of those. Analysis — forecasting, variance analysis, and data prep — is where the payoff compounds. Automation statistics show 28% of CFOs already use AI for automated forecasting, with another 39% planning to adopt within the year. And up to 80% of finance departments' transactional work could be automated. Payment firm iPSL cut budget and forecast reporting from 6 weeks to 10 days using automated Excel tooling — a 66% acceleration.

Before you spend anything, one critical caveat: almost every head-to-head benchmark you'll read is vendor-reported. Shortcut's claimed 77.3% accuracy versus Claude for Excel's 66.3% across 20 finance tasks is explicitly "a vendor claim, not a neutral test" — the same applies to speed comparisons published by any tool's own blog. Pilot tools on representative samples of your own spreadsheets instead.

A practical starting sequence:

• Start with free tiers: Power Query (built into Excel), GPTExcel's free plan (4 requests/day), GPT for Work (free to install), or ChatGPT's Excel add-in, which works on all plans including Free.
• Test each candidate on a real, messy sample of your own data — not the vendor's demo file.
• Check four criteria: works inside Excel without upload/copy-back, handles bulk data, goes beyond formulas, and lets you choose which AI model you actually get on your plan.
• Only then consider paying — Copilot runs $30/user/month, and costs are the top AI adoption barrier for 51% of businesses.

If you extend this into agent-driven automation beyond plugins, verify the agent skills you use first — published research found roughly 26% of published skills carry at least one vulnerability. SkillzBench is building a free directory that security-scans, benchmarks, and pins every skill by version, with a free MCP connector to attach vetted skills to your agents; the waitlist is open now.

Beyond Plugins: Automating the Whole Excel Workflow with Verified Agent Skills

Plugins get you far, but the next step in AI-assisted Excel work is handing the entire workflow — data cleaning, analysis, reporting — to an AI agent that runs it end-to-end. Agent skills are the mechanism making that possible.

Agent skills are instruction packages (SKILL.md files) that teach agents like Claude Code, Codex, Cursor, and Gemini CLI how to perform specific tasks reliably. Instead of writing a formula or asking Copilot to reshape one sheet at a time, you load a skill and the agent executes a whole process. Microsoft is already moving this direction — Copilot's "custom skills" let users define how it completes repeatable processes, and its Agent Mode went GA on Windows in January 2026, according to Microsoft's official documentation.

The enterprise plumbing is maturing too. In February 2026, Workato launched production-ready MCP servers that let agents like Claude and ChatGPT integrate securely with enterprise systems, and market research shows AI automation expanding from rule-based tasks into full cognitive workflows. MCP is emerging as the standard way to connect agents to the tools they need — including your spreadsheets.

There's a catch, and it's a serious one. Skills are published by anyone, and published research found that roughly 26% of agent skills carry at least one vulnerability. Installing an unverified skill into an agent that touches your financial data is a risk most teams haven't thought through yet.

Before installing any skill, check for:

• A pinned version tied to a specific commit, so you know exactly what code you're running
• A security scan and license check, not just a star count on GitHub
• Evidence the skill actually improves results — tested with the same model, with and without the skill
• Ongoing monitoring, since upstream changes can introduce new risks after you install

That's the gap SkillzBench is being built to close. It's a free, independent directory of agent skills that continuously crawls and dedupes skills from GitHub, pins every version to a commit SHA, runs security scans and license checks, and benchmarks each skill in a sandbox — publishing the performance Lift in percentage points separately from its 0–100 trust score. A free MCP connector lets you toggle exactly which skills and tools each agent can see. Everything stays free; revenue comes from blog advertising, never from scores or rankings.

If you've followed the earlier steps — picking the right plugin for bulk versus interactive work, testing on your own data — agent skills are the natural next layer. And because the platform is pre-launch, joining the SkillzBench waitlist is how you get access the moment the directory, scanner, and benchmarks go live. Your Excel workflow can be automated end-to-end; the only question is whether the skills doing it have been verified first.