| You get | A cleaned copy, a list of every change, and a list of what it wasn't sure about |
|---|---|
| Stays with you | Your original file, deleting anything, and merges it isn't sure of |
| You need | The spreadsheet (or a regular export), an agent, a computer |
| Rough cost | Your AI plan; usually little more, because the cleaning runs as code |
| Skip it if | It's a one-off: upload it to ChatGPT or Claude instead |
1What it looks like handed off
A customer list with the same person three times. Product names spelled four ways. Dates as "3/10", "10 March" and "2026-03-10" in one column. Handed off, you get a clean copy and two short lists: what it changed, and what it left for you because it wasn't sure. Your original file isn't touched.
It's most useful when the mess comes back: a weekly export from your till, booking system or website forms that needs the same fixing every time.
| Change | Rows | |
|---|---|---|
| Merged "J. Patel", "Jay Patel" and "jay patel" (same email address). | 3 → 1 | Done |
| Phone numbers written as (918) 555-0142 and 918.555.0142 put in one format. | 212 | Done |
| Dates in three formats changed to 2026-03-10 style. | 87 | Done |
| "Maria Lopez" twice, with different emails and addresses. | 2 | Left for you |
| Rows with no name and no email. | 6 | Flagged, not deleted |
Original: customers.xlsx, unchanged. Clean copy: customers-clean.xlsx.
2The job, end to end
- You drop the file in a folder, or the agent picks up the weekly export where it lands.
- It makes a copy. All the work happens on the copy.
- It looks at the columns and writes down what it plans to fix: formats, spellings, duplicates.
- It writes a short program to make those fixes and runs it, so every row is treated the same way and the result can be checked. It doesn't retype your data.
- It counts rows before and after, and lists every change, with the uncertain ones set aside for you.
- You read the lists, decide the uncertain ones, and use the clean copy. Next week, the same rules run on the new export.
3What stays with you
- Your original
- It's never edited. If the clean copy is wrong, you've lost nothing.
- Deleting rows
- It flags empty or junk rows; it doesn't remove them unless you say so.
- Merges it isn't sure of
- Two people with the same name but different details are listed for you, not merged.
- Anything that leaves the folder
- Uploading the clean list to your email tool or CRM is a step you take. Out of the box, it asks before it spends money or sends anything.
4When it gets something wrong
- It merges two different people
- The change list shows every merge with the rows it combined, so you can spot one that's wrong and undo it from the original.
- It "fixes" something that wasn't broken
- An unusual but correct spelling, a product code that looks like a date. Tell it, and add the rule to its instructions.
- Rows go missing
- The before-and-after counts are there to catch this. If they don't add up with the change list, don't use the copy.
- The export changes shape
- If your system adds or renames a column, the weekly run stops and tells you instead of guessing. On upkeep, I update the rules.
5What it costs
| Item | What it means |
|---|---|
| Your AI plan | Because the cleaning runs as code, the AI reads a sample of your rows and writes the rules, not every row. A subscription covers it; on a pay-as-you-go key, my rough estimate is cents to a few dollars per file. |
| The agent | Claude Code and Codex are good at this kind of work and come with their subscriptions; OpenClaw and Hermes are free and open source. |
| A computer | Any computer that can run the agent. It only needs to stay on for a weekly run. |
| Your time | A few minutes to read the change list and decide the uncertain rows. |
Doing it yourself
- Install an agent such as Claude Code or Codex, and make a folder with a copy of the spreadsheet.
- Tell it what clean means for you: which columns, which formats, what counts as the same customer or product.
- Ask it to write a script that makes the fixes on a copy, counts rows before and after, and writes a change list and an unsure list.
- Check the change list against the original for the first few runs.
- For a regular export, keep the script and schedule it, or ask the agent to rerun it each week.
Budget an hour or two for one file (a rough guide); a weekly routine takes a little longer to get right.
6When not to bother
- It's a one-off. Upload the file to ChatGPT or Claude, describe what clean means, and download the result. Check the row counts the same way.
- Only exact duplicates or stray spaces. Excel's Remove Duplicates and Google Sheets' data cleanup tools do it in seconds.
- The mess starts at the source. If a form lets people type dates any way they like, fixing the form beats cleaning up after it every week.
If a $20 subscription does it, I'll tell you. An agent on its own computer is worth it when the same export needs the same cleaning every week, and you'd rather not think about it.
7Want this running on your own computer?
I'll set it up, and keep it running if you like. On a Mac mini, a Windows PC or Omarchy, with the agent you use or want: OpenClaw, Hermes, Claude Code, Codex or another. Out of the box, it asks before it spends money or sends anything. Tell me about the job and I'll quote it. If a $20 subscription does it, I'll tell you. You keep the computer, the accounts and the setup.
Tyler Mayberry, who runs Cirlet (formerly Animas)
Cirlet sets up assistants and teaches teams at independent insurance agencies nationwide. Monthly live training and lifetime learning-library access are included. Open the team training library.