Can AI Clean Up Messy Spreadsheet Data? A Practical Test Guide

The Messy Spreadsheet Problem — And Why It Costs More Than You Think

If you've ever inherited a spreadsheet with broken formulas, dates formatted three different ways, and duplicate rows hiding in plain sight, you know the drill: hours of find-and-replace, cross-checking totals, and hoping you didn't miss something that breaks the reconciliation downstream. Manual data cleaning is tedious, error-prone, and expensive in ways that are easy to underestimate.

The price tag is well documented. Research citing Gartner puts the average annual cost of poor data quality at $12.9 million per organization. That figure reflects more than wasted time — it includes bad decisions built on inconsistent records, duplicate entries inflating reports, and reconciliation headaches that ripple through finance and operations.

The mess itself usually follows a familiar pattern:

• Broken or inconsistent formulas that silently produce wrong numbers
• Mixed formats — dates, names, and currencies entered differently across rows
• Duplicates and near-duplicates that resist simple removal
• Cross-row integrity problems, like values that make no sense in context
• Reconciliation gaps when cleaned data doesn't match control totals

The obvious question is whether AI can take this off your plate reliably — or whether it's another round of hype. The early evidence is genuinely encouraging, but nuanced. In a Journal of Accountancy study, an Excel-integrated AI agent cleaned, validated, and reconciled 59,157 rows of messy general ledger data in roughly 10 minutes — work the authors note would take many hours manually. Crucially, that result depended on a custom instruction file (a "skill") giving the agent procedural knowledge, plus a human approving the cleaning plan before anything changed.

Hands-on testing tells a more cautionary side. When a reviewer gave the same broken spreadsheet to ChatGPT, Claude, and Gemini, all three fixed formulas and formats but none detected a block of shuffled data — the kind of context-level error that matters most. As the tester put it, "fixing the formulas is the easy part."

That's the tension this guide explores: AI is dramatically faster at the mechanical work, but data intent and subtle integrity issues still need a human. And since quality varies wildly between tools and agent skills — published research found roughly 26% of agent skills carry at least one vulnerability — knowing which ones to trust is part of the problem. It's exactly why SkillzBench is being built: to independently test and benchmark agent skills so you can find reliable ones, free.

What the Evidence Says: AI Cleans Spreadsheets — With Caveats

AI spreadsheet cleaning works — until it doesn't. The strongest evidence for what's possible comes from a Journal of Accountancy study, where an Excel-integrated AI agent cleaned, validated, and reconciled 59,157 rows of messy general ledger data in roughly 10 minutes — a task the authors note would take many hours to complete manually.

Two ingredients made that result possible. First, the researchers built a custom "skill" file: a text package of instructions giving the agent procedural cleaning knowledge, analogous to handing a staff accountant a written procedure manual. ChatGPT for Excel ships only three generic built-in skills and none for data cleaning, so the custom skill filled the gap. Second, the workflow enforced human approval before execution — the agent presented a documented cleaning plan in a separate worksheet, worked from a copy of the data, and logged every change in a Modifications_Log audit trail.

The caveats are just as instructive. In a hands-on comparison, a first-person test of ChatGPT, Claude, and Gemini on a deliberately broken spreadsheet found that none of the three detected a block of shuffled word-count data — a cross-row integrity problem that made one article appear to take 40 hours to draft. The tester's verdict: fixing formulas is the easy part; the real question is whether AI understands what the data means.

The pattern across both studies is consistent:

• AI excels at formula fixes and format standardization — the rule-shaped work that tools like Power Query already handle well.
• Agent-level tools can reason across rows, like Claude inferring a missing value of 4 through cross-row pattern-matching.
• All three frontier models missed shuffled data, showing AI still struggles with data intent.
• Human-in-the-loop validation "reduces, but doesn't eliminate, the risk of inappropriate or unauthorized data alterations," per the Journal of Accountancy authors.

There's also a transparency gap worth noting. The study's authors emphasize that the real challenge "lies not solely in data cleaning but in ensuring that the process is transparent, auditable, and appropriate for professional use" — and the strongest results came from a custom skill built specifically for the job. That matters because the agent-skills ecosystem is growing fast, but quality is uneven: published research has found roughly 26% of skills carry at least one vulnerability. SkillzBench is being built to address exactly that — security-scanning, sandbox benchmarking, and publishing each skill's Lift so you know what actually works before you connect it to your agent.

The takeaway: AI can clean your spreadsheets dramatically faster than manual work, but treat every output as a draft that a human still signs off on.

The Skills Gap: Why Instructions Files Are the Real Breakthrough

The most important breakthrough in AI spreadsheet cleaning isn't a smarter model — it's a text file. When researchers at the Journal of Accountancy got an Excel agent to clean, validate, and reconcile 59,157 rows of messy general ledger data in about 10 minutes, the secret ingredient was a custom "skill": a package of written instructions that gave a generic agent the procedural knowledge of an experienced data cleaner.

Think of a skill as a procedure manual for your AI. ChatGPT for Excel ships with three built-in skills covering corporate finance, financial planning, and analysis — but none for data cleaning. So the researchers wrote their own, encoding rules like working only from a copy, logging every change in a Modifications_Log, and hitting 100% accuracy on a small control dataset before touching real data. That discipline is what separated a reliable cleanup from a confident guess.

The catch is that skills vary wildly in quality and safety. Published research has found that roughly 26% of agent skills carry at least one vulnerability, and most people installing them have no way to tell a well-tested procedure file from a risky one. The same governance concern shows up in the cleaning-tool market: analysts warn that AI-assisted cleanup "introduce[s] governance considerations around reproducibility and personally identifiable information handling," and urge teams to always inspect and test AI suggestions before applying them to critical pipelines.

A trustworthy skill for spreadsheet work should demonstrate:

• Security scanning and license checks before you install it
• Version pinning, so a skill that worked yesterday doesn't silently change overnight
• Sandbox benchmarking — tested with the same model, with and without the skill, so you can see the actual Lift

That's the gap SkillzBench is being built to close: a free, independent directory that continuously crawls skills from GitHub, verifies each one, benchmarks it in a sandbox, and connects your agent through a free MCP tool so it only sees the skills you've chosen. The Journal of Accountancy authors put it well: the real challenge "lies not solely in data cleaning but in ensuring that the process is transparent, auditable, and appropriate for professional use." Skills, proven — not just installed — are what turn a generic Excel agent into a data cleaner you can actually trust.

Your Practical Test Protocol: How to Try AI Cleaning Safely

Don't trust an AI agent with your production spreadsheet on day one. The strongest evidence for AI cleaning — an Excel agent that processed 59,157 rows of messy general ledger data in about 10 minutes — came from a workflow built entirely around controlled testing and human checkpoints, and that's the protocol worth copying.

Start by building a small control dataset with known flaws and a pre-calculated clean output. In the Journal of Accountancy study, researchers ran the agent against this control set and iterated until it achieved 100% accuracy before touching anything larger. Verify that agent-generated formulas actually trace back to source cells, not just produce plausible-looking numbers.

Before any execution, require the agent to present a documented cleaning plan in a separate worksheet and give formal approval. The study's authors note this human-in-the-loop checkpoint "reduces, but doesn't eliminate, the risk of inappropriate or unauthorized data alterations" — which is exactly why the remaining safeguards matter.

• Never clean the original file. Work from a copy so raw data stays preserved no matter what the agent does.
• Demand a documented audit trail — the validated workflow logs every change in a Modifications_Log with plain-English explanations of each modification.
• Reconcile cleaned data to control totals so you can prove nothing was silently dropped or duplicated.
• Prefer tools that write cleaned values next to originals with changed cells highlighted, so you compare before and after instead of trusting a black box.

Matching the tool type to the job matters too. In-place tools that clean inside Excel or Google Sheets suit one-off rescues, while standalone platforms with repeatable pipelines, scheduled runs, and audit trails fit recurring cleaning work. And don't assume "AI" means better: many products are rule engines with a fresh coat of marketing, and free baselines like Power Query and OpenRefine remain benchmarks any AI tool has to beat before it earns a subscription.

One caution from hands-on testing: all three frontier models tested (ChatGPT, Claude, Gemini) missed a block of shuffled word-count data, so treat AI output as a draft for human review, not a finished product. That review burden shrinks dramatically when your agent runs a well-built skill file — the text package of instructions that drove the study's strong results. Since published research found roughly 26% of community-shared agent skills carry at least one vulnerability, it pays to use skills that have been security-scanned and benchmarked rather than grabbed blind from a repo. That verification problem is exactly what we're building SkillzBench to solve, with free benchmarking and a free MCP connector so your agent only sees the skills you choose.

Run this protocol end to end on a copy of your own messiest file, and you'll know within an hour whether AI cleaning deserves a place in your workflow.

Where to Go Next: Equipping Your Agent With Vetted Skills

The clearest lesson from every test we've examined is that AI agents don't fail because they can't clean — they fail because they don't know how you want it cleaned. The strongest result in this space, a Journal of Accountancy study where an agent cleaned and reconciled 59,157 rows in about 10 minutes, succeeded for one specific reason: researchers wrote a custom skill file giving the agent step-by-step procedural knowledge. The agent wasn't smarter — it was better instructed.

That finding points directly at the next step in this workflow: instead of hand-writing instructions every time, equip your agent with vetted skills. Skills — instruction packages that work across Claude Code, Codex, Cursor, and Gemini CLI — encode proven procedures like "work from a copy," "log every modification," and "reconcile against control totals." The Journal of Accountancy team targeted 100% accuracy on a small control dataset before letting the agent near real data, and that discipline is exactly what a good skill encodes.

The catch is that skills vary wildly in quality and safety. Published research found roughly 26% of skills carry at least one vulnerability, and nothing in the current ecosystem tells you which ones are safe, which actually help, and which quietly break your workflow. That's the gap SkillzBench is being built to close.

Here's what SkillzBench will offer, all free:

• Discovery — continuous crawling and deduplication of skills from GitHub, so you see the full landscape, not just what surfaces first.
• Verification — every version pinned to a commit SHA, then validated, license-checked, security-scanned, and quality-judged.
• Benchmarking — every skill tested in a sandbox with the same model, with and without the skill, publishing the Lift in percentage points.
• Equipping — a free MCP connector and per-agent loadouts, so each agent only sees the skills you've chosen for it.

Two design choices matter for spreadsheet work specifically. Lift is always shown separately from ratings, so a skill that sounds impressive can't hide mediocre measured results. And because money never touches scores, the benchmark numbers stay comparable to the vendor-neutral discipline practitioners already recommend — comparing tools against honest baselines rather than marketing claims.

SkillzBench is pre-launch and being built right now. Join the waitlist at skillzbench.com to get notified when the directory, security scanning, benchmarks, and MCP connector go live — and start equipping your agent with skills that are proven, not just published.