The shoebox on the shelf above the dryer has not had its lid sit flat in two years.
Open it. A Bunnings receipt for a hose fitting, faded to nothing. Two clipper servicing invoices. A vaccination certificate handed across the counter in 2024, never scanned. An SD card from a phone she no longer owns. A bank statement, still sealed.
That is the shoebox you can see. Two more have no lid at all: the laptop folder called Photos, and the spreadsheet three people have typed into since 2022.
The direct answer. To organise business files with AI, you point a tool that runs on your own computer at a folder, describe in plain English what the files are and where a copy of each one should end up, and let it read every file and do the filing. It is not an upload and it is not a chat window. The rule that makes it safe is that it works on copies: it reads your originals, writes new files somewhere else, and never moves or deletes anything.
Take Steph, a two-groomer salon in Sydney's Inner West. She is a worked example we follow through this series rather than a real client's books. Around 4,000 photos, four years deep. Roughly 200 receipts, loose. One spreadsheet.
Three jobs below, in order of bravery. The photos. The receipts. The spreadsheet nobody maintains.
This is the job that stops it being a chat window
Everything in the two levels below this one happens in a conversation. You type, it types back, you copy the useful part out. Genuinely useful, and still a conversation.
This one is different, and the difference is what people are describing when they say the penny dropped. You write one instruction. It opens 4,000 files, works out what each one is, prints what it is doing into the terminal line by line, and hands you a log at the end. You are in the room for it. You are not the one typing.
We are deliberately not telling you how long it takes, and we are not quoting you files per minute either. That depends on your machine and your files, and anyone putting a figure on it has invented it. What matters is who is sitting there while it happens. By hand those hours are yours; this way something that is not you spends them.
This post assumes you have read what Claude Code actually is and have already given it one harmless job.
The one rule for organising business files with AI: work on copies
Every instruction below repeats the same clause in the middle of it. Copy, never move. Do not rename anything in the source folder. Do not delete anything, on any run.
Here is why, in the only terms that matter. A sort that goes wrong on copies costs you 10 minutes: delete the new folder, run it again with a better instruction. A sort that goes wrong on your originals can cost you your records, four years of before-and-afters or the receipts behind a tax return. Same tool, same afternoon, two completely different worst cases.
So the instruction has to say it, and you have to check it did. Count the source folder before, count it after, and the two numbers match or something is wrong.
A tool that reads a folder is also a tool that can misread one, which is why every job below carries its own one-minute check. Why it gets things wrong, and the 10-second check that catches it explains the mechanism.
What might be in a folder you have not looked inside
Before you point anything at a folder, open it and look. Just enough to know what kind of files are in there, because a folder nobody has opened in three years is where a scanned bank statement or somebody's driver licence ends up.
Whatever it reads goes into its context, and that is the same act as showing it to the model. So the list you already know applies here, with the folder standing in for the paste.
Never paste into an AI chat:
And never let it:
The full list, and the Australian law behind each line, is in what you must never paste into an AI chatbot. For a mixed folder, pull the sensitive files out first and run the job on what is left.
What this tool may do on your computer without asking you first is a setting. What to allow it to do and what never to is the only post here that covers those settings, and it belongs before all three jobs.
Job one: organise thousands of photos so you can find one dog
A phone gives you 400 files called IMG_4821.JPG and nothing else. Your booking export knows which dog was on the table at 10:45 on Wednesday. The whole job is putting those two facts next to each other.
Copy the photos off the phone into one folder, download your appointments export covering the same dates, then type this.
Three things in there do the real work. The count at both ends, so a lost file cannot hide. The _ask-me folder, because two dogs finishing within a few minutes of each other is normal and a guess there is worse than a question. And matching on the clock and the diary rather than the picture: it reads a timestamp, it does not identify a Groodle.
Your one-minute check. Open one dog's new folder, open one photo, and confirm that really is that dog on that date. Then count the source folder and confirm the number is what it was. Those two catch the only failures that matter: the wrong dog, and a lost file.
The count on the phone in her apron pocket is a different number from the 4,000 on the laptop, and turning phone photos into a feed is what the sorted folders make possible.
Job two: the receipts, and the honest bit about tax records
Same shape, higher stakes, one extra rule: it must never invent a number.
Now the part this job needs and the photo job does not. A renamed file is still the same receipt. Tidier, findable, and legally exactly what it was before. So do not delete an original because a neater copy exists somewhere else, and do not clear out the shoebox on the strength of a CSV. Keep the originals until your bookkeeper confirms the list, then for as long as your accountant tells you to.
It will also refuse the bit you might most want done. It records whether GST is printed on the receipt. It does not work out your GST position, your claimable amount, or which expense bucket a receipt belongs in. That is not timidity, that is the correct answer.
Your one-minute check. Pick two rows in the CSV, open the filed copy of each, and compare the supplier, the date, and the total against what is printed on the receipt. Check the cents. Totals are where the mistakes are.
Job three: the spreadsheet nobody maintains
Four years, three people typing, no rules. Two columns for phone numbers. Dates in three formats. The same client entered twice because somebody spelled the surname differently. Nobody owns it and everybody relies on it.
The braver-looking job is the safest, because of how you word it.
Deduplicated does not mean it deletes rows. It finds the pairs and puts them in front of you, and you decide which one survives. That is the whole reason this is safe to run on the file your business depends on.
Your one-minute check. Sort by the duplicate column, read the top 10 pairs, and confirm each really is one client and not two families with the same surname. Then compare the row count of the copy against the original. Same number, or it did something you did not ask for.
When a job needs one more thing installed
Expect this once, on your first PDF. We hit it ourselves the first time we ran the receipt job. Reading a PDF can need a small free piece of software a Mac does not come with, and when it is missing the job stops and names what it needs rather than quietly returning a blank total. Two ways out, both a minute: photograph the receipt and put the photo in the folder instead, or install the missing piece once and never see the message again.
That is a snag, not a fault. A tool that fails loudly is worth far more than one that fails politely. The failure you can see costs a minute. The blank total you did not notice ends up in your bookkeeper's list.
Three times is the sign it should be a skill
Read back over the three instructions. Count the folder. Copy, never move. Set aside anything uncertain, with the reason. Verify at the end. Print a log.
You have now written the same job three times, from memory, slightly differently each time. That is the moment to stop typing it out and write it down once. A written-down job is a skill, and teaching it a job once so it runs forever is the next post in this level.
If you would rather not start from a blank file, the free pet business skill library later in this series is 12 of them written for a salon, and two are the photo job and the receipt job above, with the counting and the checks already written in.
What this actually reclaims for you
Steph's two admin evenings, two hours each, are already spoken for, and none of that time is the shoebox. The shoebox is the job that never gets done at all, until the week her accountant asks and it eats a Sunday.
Say the photos are four hours by hand, the receipts three, and the spreadsheet two. Nine hours, once a year, done badly under time pressure. Against that: 20 minutes writing three instructions and reading three logs.
The honest version is not nine hours back every week. You get them back once, and the second time round the job costs 20 minutes instead of nine hours, which is the argument for writing it down rather than remembering it. Making it repeat itself on a Monday morning is where that goes next.
There is also something that is not measured in hours. The next time an owner asks whether you have a photo of Moose from March, the answer takes under a minute, because there is a folder with his name on it.
Job one only works if your diary can hand you an appointments export with the dog, the date, and the time on it. If yours cannot do that, the sorting has nothing to match against. Start your free trial and download the past appointments CSV before you sort a single photo.