What AI is replacing Excel?

The Spreadsheet Problem: Why AI Is Knocking on Excel's Door

Excel turned 40 recently, and it's still everywhere: research cited by the BBC finds two-thirds of office workers use it at least once an hour, and Microsoft claims monthly usage has grown consistently for six years. Yet the same spreadsheet habits that made Excel indispensable are now colliding with an uncomfortable reality — fragmented, undocumented workbooks are exactly what AI-driven workflows can't digest.

Among tech leaders, the BBC reports, Excel is viewed "at best, a blocker to smoother digital workflows and AI, at worst, an accident waiting to happen." Spreadsheet data scattered across personal files can't be centrally controlled or extracted to fuel machine learning systems. As Prof Mark Whitehorn of Dundee University puts it, spreadsheets are poorly documented — "the guy who wrote the macros has gone and the people in the department don't know how to run them."

The risks aren't theoretical. Recent high-profile failures show what happens when organizations lean on workbooks for critical processes:

• Health New Zealand used an Excel spreadsheet as its primary financial data file, causing discrepancies and errors.
• UK anesthetist recruitment was "plunged into chaos" in 2023 by spreadsheet confusion.
• The Afghan data scandal began with a shared Excel file exposing sensitive information.

Meanwhile, AI is already inside the spreadsheet. Tools like Microsoft Copilot for Excel write formulas, build pivot tables, clean data, and surface insights in plain language — and independent testing shows a 40-route logistics task dropping from roughly 45 minutes manually to 10 minutes AI-assisted. Some organizations are going further: Canadian telecom Telus moved hundreds of staff off Excel entirely and expects to save C$42m a year.

So the real question isn't "will AI replace Excel?" — the evidence says full replacement isn't happening yet. The sharper question is which spreadsheet tasks AI is already taking over, and what that means for the skills you'll still need to verify the output. As one comparison bluntly noted: "If you cannot read a pivot table, you will not know when an AI-generated one is wrong." That verification mindset extends beyond Excel too — if you're wiring AI agents into your workflows, SkillzBench is building a free directory that security-scans and benchmarks agent skills before you trust them with your data.

What AI Is Actually Replacing: Formulas, Pivots, Cleaning, and Insights

The question isn't whether AI will delete Excel from your desktop — it's which spreadsheet jobs AI has already taken off your plate. According to one tool-by-tool comparison, AI now handles three distinct categories of work: in-cell generation, data interpretation, and full spreadsheet creation.

In-cell generation means describing what you want in plain language and getting formulas, VBA macros, or Apps Script back. Data interpretation covers summarizing ranges, spotting trends and outliers, and answering plain-language questions about your data. Full creation goes further: a single prompt can produce a working spreadsheet from scratch. Microsoft positions Copilot the same way, promising to speed up "certain spreadsheet tasks, like adding columns and formulas, formatting tables, and coming up with insights about your data."

The time savings are real. A worked logistics example cut a 40-route anomaly-flagging task from roughly 45 minutes manually to about 10 minutes AI-assisted. But the tool you pick matters more than most buyers realize, because of one technical divide:

• Microsoft Copilot — $30/user/month add-on; strong on structured, step-by-step tasks
• Claude for Excel — available on Pro ($20/mo) and up, with cell-level citations for verification
• ChatGPT for Excel & Sheets — works on every plan, including Free
• GPT for Work — from $25/month, built for bulk processing

The key differentiator is bulk processing. Per independent tool testing, most built-in assistants — Copilot, Claude for Excel, the ChatGPT add-in, Gemini — cannot reliably apply prompts across thousands of rows. GPT for Work, by contrast, processes up to 200,000 rows in Google Sheets and 1 million in Excel per run, at roughly 1,000 rows per minute.

That gap matters because verification remains the human's job. As one reviewer put it bluntly: "If you cannot read a pivot table, you will not know when an AI-generated one is wrong." Hands-on testing found AI tools exhibit gullibility bias, don't question flawed premises, and produce divergent results from vague prompts. Anthropic's own internal analytics accuracy was just 21% before significant process investment — and 95% after.

The same verification principle applies when you push AI beyond the spreadsheet into agent workflows. SkillzBench is building free, independent benchmarking and security scanning for AI agent skills, so you can test whether a skill actually helps before trusting it with your data. Because in spreadsheets and in agents alike, the formula is the same: generate fast, verify always.

The Counter-Trend: Companies Replacing Excel Entirely to Enable AI

While Microsoft insists Excel usage keeps climbing, some organizations are ripping it out entirely — and their reasons have everything to do with AI. The logic is simple: fragmented spreadsheet data cannot be centrally controlled or extracted to fuel AI and machine learning systems.

The BBC's reporting captures the stakes bluntly: among tech leaders, Excel is viewed as "at best, a blocker to smoother digital workflows and AI, at worst, an accident waiting to happen." The BBC's case studies show what replacement looks like in practice.

Canadian telecom Telus moved hundreds of staff off Excel onto a custom planning system, explicitly to enable AI and automation. The payoff: an expected C$42m per year in savings from eliminating misaligned capital. Director Moutie Wali was blunt about how: "You have to force it by not allowing the spreadsheet to coexist with your [new] applications."

The pattern repeats at smaller scale. A UK charity replaced its Excel-based accounts with an AI-enabled accounting package, saving over £6,000 a year and eliminating the need for a bookkeeper role. The failures driving this are well documented:

• Health New Zealand used an Excel file as its "primary data file" for financial performance management, causing discrepancies and errors
• UK anesthetist recruitment was "plunged into chaos" in 2023 by spreadsheet confusion
• The Afghan data scandal originated from sharing an Excel spreadsheet

The experts explain why spreadsheets block organizational-scale AI. Tom Wilkie, CTO of Grafana, still calls Excel "just a really good tool" for small datasets and quick charts — but a blocker for AI-driven workflows at scale. Prof Mark Whitehorn, emeritus professor of analytics at Dundee University, points to documentation rot: "the guy who wrote the macros has gone and the people in the department don't know how to run them." His reminder to spreadsheet owners: "It isn't your data, it's the company's data."

Here is the tension worth holding: both narratives are true. Microsoft claims Excel monthly usage has grown consistently for six years, and two-thirds of office workers use it at least hourly. Meanwhile, organizations that centralize their data are finding AI works far better without spreadsheets in the middle.

For teams building AI workflows beyond the spreadsheet, agent skills are increasingly the bridge — and SkillzBench is building a free directory and MCP connector to discover, verify, and benchmark them so you know which ones actually help.

The Verification Gap: Why AI Spreadsheet Output Still Needs Humans

Here's the uncomfortable truth about AI spreadsheet tools: the same hands-on testing that shows impressive time savings also shows accuracy can collapse without human verification. When Anthropic analyzed its own AI analytics workflow, internal accuracy was just 21% before a significant investment in process — and 95% only after. The tool didn't change; the verification did.

This matters because AI analysis tools don't calculate your results directly. They generate and run code (mainly Python) in response to your questions, which means every answer is only as reliable as the generated code. A detailed hands-on test of these tools concluded bluntly: "Never rely on analysis without code." If you can't read what the AI produced, you can't check it.

The testing also documented a pattern the author calls "gullibility bias" — AI tools rarely question flawed premises in your prompts. Vague questions produced divergent interpretations across different tools, and conversation history quietly biased results. In other words, the AI doesn't push back; it confidently executes whatever you asked, errors included.

Vendor benchmarks deserve equal skepticism. Shortcut's July 2026 benchmark of 20 finance tasks reported 77.3% accuracy versus 66.3% for Claude for Excel — but that figure comes from Shortcut itself, and reviewers flagged it explicitly as "a vendor claim, not a neutral test." When accuracy numbers come from the company selling the tool, treat them as marketing until independently confirmed.

So where does AI genuinely save time? In bounded, checkable tasks. A worked example from independent tool testing found that flagging anomalies across 40 logistics routes dropped from roughly 45 minutes manually to about 10 minutes AI-assisted. That's a real, meaningful gain — but it works because the output is small enough to verify quickly.

The verification checklist is straightforward:

• Understand the structure of your data before asking AI to interpret it.
• Spot-check AI-generated formulas and pivot tables against known results.
• Ask precise questions — vague prompts invite divergent, unreliable answers.
• Be suspicious when a vendor publishes its own accuracy benchmark.

The bottom line echoes the testing consensus: "No AI tool in this comparison replaces spreadsheet fundamentals. If you cannot read a pivot table, you will not know when an AI-generated one is wrong." A formula generated in ten seconds can still be wrong in a way that takes an hour to trace back through a report.

This verification-first mindset extends beyond spreadsheets to the broader agent ecosystem. Research has found roughly 26% of published agent skills carry at least one vulnerability, which is why SkillzBench is building independent sandbox benchmarking and security scanning for agent skills — so you can see tested Lift numbers before trusting any skill in your workflow. The principle is the same everywhere: verify first, then automate.

Your Action Plan: Choosing AI Spreadsheet Tools and Extending to Agent Skills

Choosing an AI spreadsheet tool is less about which chatbot sounds smartest and more about four practical questions. Get those right, and the tool fits your workflow instead of fighting it.

First, ask whether the AI works where your data lives. According to hands-on tool comparisons, most built-in assistants — Copilot, Claude for Excel, the ChatGPT add-in, Gemini — cannot reliably apply prompts across thousands of rows. If your workbooks run long, that constraint matters more than any feature list.

Second, test bulk handling with your own files, not vendor demos. GPT for Work processes up to 1,000 rows per minute and handles up to 1 million rows in Excel per run — roughly 100 times what a person can do by hand, per the same comparison. Tools that cap out on small ranges will disappoint quickly.

Third, look beyond formula generation. The tasks AI genuinely takes over include pivot tables, data cleaning, pattern detection, and plain-language Q&A over your data, as Coursiv's tool comparison outlines. Fourth, check model choice and transparency: cell-level citations and logged cell changes beat a slick interface when you need to audit what the AI did.

A quick checklist before you commit budget:

• Does it work natively in Excel or Google Sheets, or require exporting data elsewhere?
• Can it process your typical dataset size in one run?
• Does it do analysis and interpretation, or only write formulas?
• Can you switch models, or are you locked to one vendor's?

Remember the verification lesson, too: independent testing found AI analysis accuracy depends heavily on process discipline — Anthropic's internal accuracy sat at 21% before process investment and 95% after, per one journalist's hands-on testing. Treat AI output as a draft, not a verdict.

Increasingly, spreadsheet AI doesn't stop at the grid. Claude for Excel's live data integrations — pulling from sources like S&P Global, LSEG, and Moody's — run on MCP connectors, the same protocol that lets AI agents reach external tools and skills. If your automation ambitions extend past the spreadsheet into agent workflows, SkillzBench's free directory and MCP connector let you browse and test agent skills — with the security scanning and with/without-skill benchmarking SkillzBench is building, since published research found roughly 26% of skills carry at least one vulnerability. Verify first, then equip your agents.