What Is Excel AI? A Plain-English Guide for 2026

Why Spreadsheets Still Eat Your Time — and Why AI Help Was Overhyped

You know the feeling: it's 2 PM, you've been wrestling with a VLOOKUP for an hour, and the formatting on your report still looks like it was designed in 1998. Spreadsheet work eats time in ways that are hard to explain to anyone who hasn't lived it — manual formula writing, data cleanup, conditional formatting, and the endless cycle of fixing one cell only to break another.

The promise of "AI in Excel" arrived with loud marketing and high expectations. But the term itself has become a source of confusion. Older tools — Flash Fill suggestions, Ideas, even the App Skills feature — get lumped together with Microsoft 365 Copilot in Excel, which is the current product most people mean when they say "Excel AI" today. That confusion matters because feature drift has left the landscape littered with stale tutorials.

App Skills were removed from Excel by late February 2026, with Microsoft redirecting users toward its "Analyst" tool for deeper analysis. Many YouTube walkthroughs and blog guides from 2024 still teach features that no longer exist. Buyers who tested those old workflows often came away thinking Excel AI was a toy — when in reality, they were testing a discontinued product layer.

A few reasons that skepticism lingers:

• Microsoft itself warns that Copilot "can make mistakes, misinterpret information, or produce inaccurate outcomes," which doesn't inspire confidence for finance or legal work
• Editing only works within the currently open workbook — Copilot can't pull from other files, emails, or enterprise systems during a session
• Perpetual Office licenses (Office 2019 or 2024) don't include Copilot at all, so users on older setups never even see the new features
• Pricing varies by plan and has shifted — Copilot Pro launched at $20/month for consumers, while business tiers now run around $18–$25 per user per month depending on commitment

The gap between expectation and reality is real, but it's narrowing fast. Copilot now operates in three modes — Edit, Plan, and Chat — and behaves less like a chatbot and more like an agent that builds a step-by-step plan, executes it, and reviews the result before handing it back. That shift toward agentic workflows is the bigger story, and it's where the broader ecosystem of agent skills comes in. Anyone building repeatable spreadsheet automations can browse and benchmark free agent skills through SkillzBench's pre-launch directory — every skill version-pinned to a commit SHA, security-scanned, and sandbox-tested before you equip it. Join the SkillzBench waitlist to get early access when it goes live.

What Excel AI Actually Is: Copilot's Three Modes and Core Capabilities

Most people picture "Excel AI" as a chatbot hovering next to their spreadsheet. The reality is more useful — and more structured — than that.

Microsoft defines it plainly: AI in Excel uses Microsoft 365 Copilot to speed up tasks like adding columns and formulas, formatting tables, and surfacing insights about your data. It works with Excel's native tools — tables, charts, PivotTables, formulas — rather than replacing them.

Copilot runs in three modes, and the mode you pick changes what happens to your file:

• Edit mode (the default) — Copilot edits the workbook directly, handling complex multi-step tasks like reshaping data, merging sheets, and building multi-element reports.
• Plan mode — Copilot drafts a structured step-by-step plan for your review before it touches anything.
• Chat mode — Copilot analyzes data and returns insights without modifying the workbook at all.

Under the hood, the workflow is agentic: Copilot analyzes your request, builds a plan, executes it in the workbook, then reviews whether the output matches your intent. Microsoft's documentation notes you can watch its reasoning live in the pane and hit Stop at any time.

What can it actually do? The capability list is broad: edit worksheets (insert values, apply conditional formatting, data validation, borders, styles), generate formulas and calculations across multiple sheets, summarize data, and create charts, PivotTables, and shapes. It also builds dashboards and templates from scratch, highlights and filters data, identifies trends and outliers, and can search the web with citations.

Two 2026 additions matter for anyone evaluating the tool now. Copilot gained a brand kit skill that applies approved company colors, fonts, and logos to tables and charts, plus Power BI grounding that pulls governed metrics directly into the grid while respecting row-level security — both rolled out mid-2026. It also supports custom skills for repeatable tasks, which connects to the broader shift toward agent skills that can automate the rest of a workflow.

If you're building out that kind of automation, SkillzBench is building a free directory of agent skills — every one security-scanned, version-pinned, and sandbox-benchmarked before you equip it. The MCP connector and benchmarks are in development now, and the waitlist is open.

What's New in 2026: Brand Kits, Power BI Grounding, and the Excel Agent

If you last looked at Excel AI in 2024, the July–August 2026 rollouts will feel like a different product. Microsoft spent the summer turning Copilot from a helpful chat pane into something that formats like a designer, reasons like an analyst, and acts like an agent.

The headline additions are two new design skills. A brand kit skill rolled out in July 2026 applies an organization's approved colors, fonts, and logos to tables and charts, so reports come out on-brand without anyone manually restyling them. A theme design skill follows in August 2026, applying coordinated palettes and visual hierarchy that Microsoft describes as turning raw data into a clean, readable report source.

For business users, the bigger change is Power BI grounding. Rolling out in August 2026, it lets Copilot ground its Excel analysis in governed Power BI data while respecting row-level security, bringing trusted business metrics directly into the grid. In plain terms: no more exporting from Power BI, pasting into Excel, and hoping the numbers still reconcile.

Other summer updates fill long-standing gaps:

• Expanded workbook availability (August 2026): Copilot now works in more cloud-saved files, including workbooks where AutoSave is switched off on Windows and Mac.
• Excel agent in Copilot Chat (July 2026): @mention the Excel agent to create spreadsheets without leaving the Copilot app.
• New models (July 2026): OpenAI's GPT-5.6 and Anthropic's Claude Sonnet 5 are now in Microsoft 365 Copilot, delivering stronger reasoning for complex, multi-step tasks, including deeper analysis in Excel.

That last bullet matters more than it looks. Excel AI has shifted from "chat beside a spreadsheet" toward an agentic workflow — Copilot analyzes your task, creates a step-by-step plan, makes changes live in the workbook, and reviews whether the result matches your intent, with a Stop button to pause at any time source. Microsoft's own documentation confirms users can watch Copilot's reasoning in the pane as it works source.

This agentic direction is the broader trend across AI tools, and it's where skills come in. Copilot already supports custom skills for repeatable tasks, and agent skills — instruction packages that teach agents like Claude Code, Codex, Cursor, and Gemini CLI new capabilities — are proliferating across GitHub. The catch: published research found roughly 26% of agent skills carry at least one vulnerability, so quality and security vary wildly. That's the problem SkillzBench is building against: a free directory that continuously discovers skills from GitHub, pins every version to a commit SHA, security-scans and quality-judges each one, and benchmarks it in a sandbox to publish a Lift score in percentage points — with everything free, from the directory to the MCP connector.

CTA: Want agent skills that are actually proven? Join the SkillzBench waitlist — every skill security-scanned, version-pinned, and benchmarked before you equip it. Free.

Social proof: Skills, proven.

The Honest Fine Print: Costs, Licensing, and Real Limitations

Excel AI sounds impressive in demos, but the fine print determines whether it fits your budget and your risk tolerance. Here's what Microsoft doesn't put on the landing page.

Access isn't automatic. Copilot in Excel requires an active Microsoft 365 subscription plus a separate Copilot license that your IT admin must assign. If you bought a perpetual license — Office 2019 or Office 2024 — you're out of luck; those don't include Copilot at all, according to licensing details from SF Magazine.

Pricing adds up quickly. Copilot Pro runs $20/month for home and student users, while enterprise Copilot 365 costs $360 per user per year. More recent pricing data puts the business tier at roughly $18/user/month with an annual commitment — or $25.20/user/month if you pay monthly — on top of a qualifying Microsoft 365 plan. Budget for two line items, not one.

Then there's the accuracy question, which Microsoft addresses surprisingly directly. The company's own guidance states that Copilot "can make mistakes, misinterpret information, or produce inaccurate outcomes", and advises against relying on it for sensitive decisions in finance, legal, or medical contexts without human review. That's not boilerplate — it's a genuine caution for anyone feeding it a P&L.

Two practical limitations matter most:

• Editing works only within the currently open workbook — Copilot can't reach other files, emails, or enterprise data during an editing session.
• Edits are saved automatically as they happen, so Microsoft recommends working on a copy of critical or sensitive workbooks.
• Copilot struggles with messy inherited files, multi-sheet reconciliation, and formal financial model review — it shines on clean tables with one or two clear actions.

The verification mindset applies beyond Excel. As Copilot shifts toward agentic workflows — planning, executing, and reviewing its own work — the same "trust but verify" discipline extends to the broader ecosystem of agent skills, where published research has found that roughly 26% of third-party skills carry at least one vulnerability. SkillzBench is building a free answer to that: a directory where every skill gets security-scanned, version-pinned, and sandbox-benchmarked before you equip it.

The takeaway: Excel AI is worth the license fee for the right tasks, but treat it like a fast, occasionally overconfident junior analyst. Give it clean data, review its output, and never let it touch your only copy of a sensitive file.

Getting Started: Best-Fit Use Cases and Where Agent Skills Fit In

Knowing where Excel AI shines — and where it stumbles — is the difference between saving an hour a week and wasting an afternoon babysitting a confused chatbot. The pattern is consistent: Copilot performs best when a task starts with a clean table and ends in one or two clear actions, and it struggles with the messy, inherited files most teams actually have.

Where Copilot in Excel works best: formula generation, charts and PivotTables, quick summaries, and light edits like highlighting, sorting, and filtering. These map directly to Microsoft's own capability list — Copilot can generate formulas, summarize data, and identify insights like trends and outliers, all while you watch its reasoning live and hit Stop at any point.

Where it falls short:

• Messy inherited files with inconsistent formatting and merged cells
• Multi-sheet reconciliation across workbooks — editing only works within the currently open file
• Formal financial modeling review, where Microsoft itself warns Copilot "can make mistakes" and advises human verification for finance, legal, or medical decisions

That accuracy caveat matters more than most tutorials admit. Microsoft recommends working on a copy of critical workbooks, since edits save automatically. The practical rule: give Copilot clean input and a narrow ask, and it delivers.

For repeatable work, Microsoft's documentation notes Copilot supports custom skills — and 2026 added concrete examples, like the brand kit skill that applies your organization's approved colors, fonts, and logos automatically, rolled out in July 2026 alongside a theme design skill.

This points to a broader shift: Excel AI now runs as an agent workflow — plan, execute, review — and agent skills are becoming the standard way to package repeatable instructions for tools like Claude Code, Cursor, and Codex. The catch is trust: published research found roughly 26% of agent skills carry at least one vulnerability, so grabbing one off GitHub is a gamble.

That's the gap SkillzBench is building to close. Currently in pre-launch, it's creating a free directory of agent skills where every skill is security-scanned, version-pinned to a commit, license-checked, and benchmarked in a sandbox — with and without the skill — before you equip it. A free MCP connector then links any agent to the skills you choose, which is handy for spreadsheet-adjacent workflows like data cleanup, report formatting, and CSV transformation that fall outside Copilot's clean-table comfort zone. Everything — directory, benchmarks, connector, alerts — is free.

If you want agent skills that are actually proven rather than just popular, join the SkillzBench waitlist and get notified at launch.