How can AI work in Excel?

Why 'AI in Excel' Is So Confusing: One Label, Five Different Products

Search for "AI in Excel" and you'll get Microsoft Copilot, a $10 formula bot, a $6 spreadsheet app, and a $180/year workbook agent — all wearing the same label. That's not a search problem; it's a market that genuinely contains five different product categories pretending to be one thing.

The confusion is real because the demand is real. According to the Stanford HAI 2025 AI Index Report, 78% of organizations reported using AI in 2024, up from 55% the year before. Meanwhile, 96% of FP&A teams still use Excel for planning, and Acuity Training research cited by the BBC found two-thirds of office workers touch Excel at least once every hour. AI wants in, and Excel is where the work already lives.

The problem is that "Excel AI" spans wildly different products at wildly different prices — from $6/user/month (Rows AI) to $30/user/month (Microsoft 365 Copilot for larger customers). Industry analysis identifies at least five distinct categories:

• Embedded Copilot — Microsoft's native assistant, including its in-cell COPILOT function and Agent Mode for multi-step tasks
• Third-party add-ins — Claude for Excel, ChatGPT for Excel, and others that bolt onto your existing workbook
• =AI() formula functions — tools like Numerous.ai that put AI inside cells as formulas
• AI-native spreadsheet apps — Rows and Quadratic, built with AI as a core feature
• Workbook agents — tools like Deckary that plan, execute, and verify multi-step tasks across a file

As one vendor analysis bluntly puts it: "You need to check each vendor directly instead of assuming 'AI for Excel' means the same thing everywhere." A pricing comparison, a privacy policy, and a capability check all change depending on which of the five categories you're actually looking at.

There's also a quieter layer worth knowing about: both Microsoft and Anthropic now use "skills" — reusable SKILL.md instruction packages — as a core integration mechanism inside Excel, from Copilot's custom finance skills to Claude's shareable one-click workflows. That's why SkillzBench is building a free directory that security-scans, version-pins, and sandbox-benchmarks agent skills before you trust them with workbook data.

This article is your decoder: what each category actually does, what it costs, where it breaks, and which one fits your workflow.

The Five Ways AI Actually Gets Inside Your Workbook

"Excel AI" is not one thing. Search for it and you'll find at least five distinct product categories, which is why the market feels so muddled — as one vendor analysis puts it, you need to check each vendor directly instead of assuming the label means the same thing everywhere.

1. Microsoft 365 Copilot. The native option embeds AI directly in Excel, including an in-cell COPILOT function and Agent Mode, which according to industry reporting, plans, executes, and validates multi-step tasks from natural-language prompts. Its Plan mode drafts actions and asks you to approve or edit before executing, with AI edits tracked alongside human edits in the Show Changes pane, per Computerworld.

2. Third-party add-ins. Claude for Excel and ChatGPT for Excel bolt external AI onto your existing workbook. Anthropic's Claude for Excel offers shareable one-click "Skills" workflows, such as variance analyses, that teams save and distribute org-wide.

3. Formula-based =AI() functions. Tools like Numerous.ai let you write AI calls as spreadsheet formulas, so a cell can summarize text or classify data just like a VLOOKUP. This is the lightest-touch integration — no new interface to learn.

4. AI-native spreadsheet apps. Rows and Quadratic rebuild the spreadsheet around AI from the ground up, with AI assistance built into the grid itself rather than layered on top.

5. Full workbook agents. Tools like Deckary plan and execute entire multi-step workflows — building tables, writing formulas, producing charts — rather than answering single prompts.

• Rows AI: from $6/user/month billed annually
• Numerous.ai: $10/month billed annually
• Melder: $50/month Pro — and claims it never stores your Excel data outside your own workbook
• Microsoft 365 Copilot: $30/user/month for larger customers ($21 for organizations under 300 employees)

Privacy policies diverge sharply: ChatGPT for Excel may store some data logs with OpenAI for 30 days for safety purposes, while Melder claims local-only retention. Verify each vendor before deploying on sensitive workbooks.

Notably, both Microsoft and Anthropic now use SKILL.md-style instruction packages inside Excel — the same format used by coding agents like Claude Code and Codex. SkillzBench is building a free directory that security-scans, version-pins, and sandbox-benchmarks those skills, since published research has found roughly 26% carry at least one vulnerability. Skills, proven — before they touch your data.

Trust, Oversight, and the Rise of Excel 'Skills'

The most important feature in Excel AI isn't the model — it's the "are you sure?" button. As AI assistants gain the power to restructure entire workbooks, the market is racing to build guardrails around them, and that race says a lot about how much you should trust the output.

The Stanford HAI 2025 AI Index Report is blunt: AI models "often fail to reliably solve logic tasks even when provably correct solutions exist," limiting their effectiveness in high-stakes settings where precision is critical. A financial model is exactly such a setting. And the problem predates AI — Panko's 15 years of spreadsheet research found that spreadsheet errors are "both common and non-trivial." Layer a fallible model on top of an error-prone medium, and human review stops being optional.

That's why the industry is converging on the same safety architecture. Microsoft's Copilot now includes a Plan mode that drafts a list of intended actions and asks users to approve, edit, or answer clarifying questions before executing anything, with AI edits appearing alongside human edits in the Show Changes pane, according to Computerworld's coverage of the update. Agent Mode similarly surfaces each step of a multi-step task for review before touching your data.

When evaluating any Excel AI tool, prioritize these capabilities:

• Plan-then-execute workflows that show intended actions before changes are made
• Ask-before-write controls on destructive edits like deletions or overwrites
• Change tracking that distinguishes AI edits from human edits
• Clear vendor data policies — some tools may retain logs for 30 days

The oversight story now has a companion trend: reusable instruction packages called skills. Microsoft's new Skills feature lets users define repeatable processes — building a discounted cash flow, preparing a variance analysis — saved as a SKILL.md file in OneDrive, with partners like LSEG and Ramp building their own. Anthropic introduced parallel Skills for Claude for Excel: shareable one-click workflows teams can distribute org-wide.

The appeal is obvious. As Microsoft's VP for Excel Brian Jones put it, a skill guides Copilot through the steps "instead of starting from scratch each time," producing output that is easier to review, reuse, and trust. But skills are instruction packages, and instruction packages can be wrong — or worse. Published research has found roughly 26% of AI agent skills carry at least one vulnerability.

That's the gap SkillzBench is being built to close: a free directory of agent skills with security scanning, version-pinned validation to a commit SHA, and sandboxed benchmarking that tests every skill with the same model, with and without it, before it ever touches your workbook data. Skills, proven — the same standard Excel's oversight features are applying to the model, applied to the layer on top.

Should You Augment Excel or Replace It? A Practical Decision Framework

Should your AI strategy treat Excel as a problem to eliminate or a surface to enhance? The industry genuinely disagrees, and the honest answer depends on what your teams actually do with spreadsheets.

On one side, the case for replacement is real. Prof Mark Whitehorn of Dundee University argues that Excel's decentralized data makes it hard to "extract it for broader analysis and to fuel AI," and Telus moved hundreds of staff off Excel expecting to save roughly C$42m (~$30m) per year. Small departments hoard data in macros, and as Whitehorn puts it, "It isn't your data, it's the company's data."

On the other side sits a stubborn fact: 96% of FP&A teams still use Excel for planning, and two-thirds of office workers open it at least once an hour. The a16z analysis resolves the tension by separating use cases: routine "mini software" workflows may indeed be replaced, but for analytical work like financial models, "the spreadsheet is not storage. It is the reasoning surface" — which makes in-grid augmentation the better fit.

That suggests a practical decision framework:

• Replace the workflow, not the tool — if Excel is being used as a data pipeline (macros, manual entry, handoffs), move that data to a system AI can reach.
• Augment when the grid is the thinking — modeling, variance analysis, and ad-hoc exploration benefit from embedded Copilot-style assistants, plan-mode oversight, and reusable skills.
• Match the integration method to the job: in-cell functions for transformations, add-ins for conversational analysis, full workbook agents for multi-step tasks.
• Verify vendor data policies before deploying on sensitive workbooks — retention terms vary widely, from 30-day log storage to claims of no external storage at all.

Your next steps are concrete. Audit which workflows are storage versus reasoning. Pick the integration method that matches each. Then, for agentic setups, agent skills — the same SKILL.md-style packages Microsoft and Anthropic now use inside Excel — can automate the rest of the workflow. SkillzBench is building a free, security-scanned directory where you can browse, verify, and test those skills before trusting them with your data.