Sort a pile of receipts into a clean expense table

What you will build

A monthly Receipt Desk: a way to turn a pile of receipt photos, PDFs or pasted receipt text into one clean expense table with date, supplier, description, category, amount, payment method and receipt number. Anything the AI cannot read is marked UNREADABLE instead of guessed. Likely duplicates are listed separately. Totals are calculated by your spreadsheet, not by the AI.

You end up with a table you can paste into your spreadsheet or hand to your accountant, and a routine that takes the same shape every month.

Before you start

• A batch of receipts: paper, phone photos, PDFs or e-receipts from email. Start with one month.
• Your expense categories. Ask your accountant or bookkeeper for the list they use, so your table matches their work.
• Claude, ChatGPT or Gemini. Claude's help page lists JPEG, PNG, GIF and WebP images and PDF files as supported uploads, and suggests images of at least 1000 x 1000 pixels. See Upload files to Claude.
• A spreadsheet: Google Sheets or Excel.
• About 30 minutes for the first month.

How it works

Sorting receipts is slow because every receipt looks different. AI tools can read most printed receipts from a clear photo and pull out the fields you need. But they can misread faded print, confuse a total with a subtotal, or count the same receipt twice if you photographed it twice.

So the Desk is built around three safety rules. Copy amounts exactly as printed, never calculate or round. Write UNREADABLE instead of guessing. Let the spreadsheet do every total. Then you check the table against the pile with two simple counts before it goes anywhere. This is bookkeeping help, not tax advice. Your accountant decides what is claimable and how records must be kept.

1. Agree your categories with your accountant.
2. Photograph and prepare receipts safely.
3. Run the extraction prompt in small batches.
4. Check the table against the pile.
5. Fix UNREADABLE cells and duplicates from the originals.
6. Move the table into your spreadsheet and total it there.
7. File the originals and save the known-good routine.

Step 1: Agree your categories with your accountant

Categories are where AI sorting goes wrong most quietly. If you let the AI invent categories, you get a different list every month. Use the list your accountant already works with, and add one line of meaning to each.

EXPENSE CATEGORIES (from our accountant)
Stock / materials: [what counts, e.g. car shampoo, wax, microfibre cloths]
Transport: [petrol, tolls, parking]
Equipment: [items above a certain value, as our accountant defines]
Utilities: [phone, internet, electricity]
Rental: [shop or storage rental]
Marketing: [ads, printing, boosting posts]
Meals: [only if our accountant says to track them]
Other: [anything that does not fit; the owner decides later]

When in doubt, the AI should put an expense in "Other" with a note, not force it into a category.

Check

• The category list came from your accountant or bookkeeper.
• Each category has a one-line meaning.
• "Other" exists for anything unclear.

Step 2: Photograph and prepare receipts safely

A clear photo prevents most errors. Lay each receipt flat in good light, fill the frame, and avoid shadows and folds. Faded thermal receipts are hard for anyone to read, so photograph them as soon as you get them.

Before uploading:

• Crop or cover card numbers, bank account numbers and IC numbers.
• Crop out anything unrelated in the background.
• Keep one receipt per photo where you can. If several are in one photo, make sure none overlap.
• Name photos by date if your phone allows, or put each month's photos in one album or folder.

For e-receipts in email, save them as PDFs or copy the text. For receipts that staff sent you on WhatsApp, save the images to one folder first.

Check

• Every photo is sharp and the total is readable by eye.
• No card, account or IC numbers are visible.
• One month's receipts are in one place.

Step 3: Run the extraction prompt in small batches

Upload a batch of about ten to fifteen receipts at a time. Smaller batches are easier to check and keep each chat focused.

Sort these receipts into one table with these columns:
date (YYYY-MM-DD) | supplier | description | category | amount_RM | payment_method | receipt_number | file

Categories to use (only these): [paste your category list with meanings]

Rules:
- Copy the final amount paid exactly as printed. Do not round, add up or calculate tax.
- If a receipt shows both a subtotal and a total, use the total paid and note "subtotal also shown".
- If any field is unreadable, write UNREADABLE in that cell. Do not guess.
- If the category is unclear, use Other and add a short note.
- "file" = the photo or file name the row came from.
- After the table, list any receipts that look like duplicates (same supplier, date and amount), and say why.
- Do not total the table. I will total it in my spreadsheet.

Return the table as CSV so I can paste it into a spreadsheet.

RECEIPTS:
[upload photos or PDFs, or paste receipt text]

The "file" column is your link back to the original. If a number looks wrong later, you know exactly which photo to open.

Check

• Every row has a file name.
• UNREADABLE appears where the photo is unclear, not a guessed value.
• Possible duplicates are listed separately, not silently removed.

Step 4: Check the table against the pile

Two quick counts catch most problems.

PILE CHECK
1. Count: number of receipts uploaded = number of rows in the table (+ duplicates listed)?
2. Money: add up the receipts yourself (calculator or quick sum) and compare with the spreadsheet total from Step 6.
3. Spot-check: pick three rows at random, open their files, compare date, supplier and amount.
4. Categories: scan the category column for anything that looks wrong.

If the count does not match, find the missing or extra receipt before you go further. A missing row is worse than an ugly table.

Check

• Row count matches receipt count.
• Three random rows match their receipts exactly.
• Nothing is in a category that surprises you.

Step 5: Fix UNREADABLE cells and duplicates from the originals

Go through every UNREADABLE cell and fill it from the paper receipt or the original photo, zoomed in. If you truly cannot read it, ask the supplier for a copy or note it for your accountant. Do not let the AI try again with a guess.

For the duplicate list, open both files. Sometimes you photographed the same receipt twice. Sometimes you really bought the same thing twice on the same day. Decide each one yourself and delete only true duplicates.

FIX LOG
Row | field | was | now | source (paper / photo zoom / supplier copy)

Check

• No UNREADABLE cell is left without a note.
• Each duplicate was decided by looking at both originals.
• The fix log shows what you changed and why.

Step 6: Move the table into your spreadsheet and total it there

Paste the CSV into a monthly tab in your spreadsheet. Then build totals with formulas, not typed numbers:

SPREADSHEET SETUP
Tab name: [YYYY-MM]
Total per category: SUMIF over the category column
Month total: SUM of the amount column
Check cell: Month total = sum of category totals (should be TRUE)
Notes column: anything your accountant should know

The check cell is a small safety net. If the month total and the sum of categories do not match, a row has a typo in its category name.

If you use an accounting app, follow its import format instead. The AI can reshape the CSV columns to match, as long as it copies values exactly.

Check

• Every total comes from a formula.
• The check cell shows the totals agree.
• The month tab is named and dated.

Step 7: File the originals and save the known-good routine

The table does not replace the receipts. Keep the originals the way your accountant asks: paper in a monthly envelope, photos in a monthly folder, or both. Name the folder the same way as the spreadsheet tab.

When a month goes smoothly, save the prompt, category list and checks as your known-good routine. Next month you repeat the same steps with the same prompt.

MONTHLY RECEIPT ROUTINE
Week 1 of next month:
1. Collect all receipts for last month in one place.
2. Crop private numbers.
3. Run the prompt in batches of 10 to 15.
4. Pile check, fix UNREADABLE and duplicates.
5. Paste into the month tab, check totals.
6. File originals. Send the tab to the accountant.

Check

• Originals are filed by month.
• The prompt and categories are saved as a known-good copy.
• The routine has a fixed week in the month.

Worked example

This example is made up for teaching. The business, suppliers and amounts are invented.

A mobile car detailing business in Cheras collects receipts in the van's glove box and in a WhatsApp chat where the owner's helper sends photos. For September, the owner gathered 28 receipts: petrol, tolls, car shampoo and wax from a supplier in Kajang, microfibre cloths, a pressure washer part and a phone bill.

The owner photographed the paper receipts on the van seat, cropped a card number on two of them, and uploaded them in two batches of 14. The first batch came back with 13 rows, one UNREADABLE amount on a faded petrol receipt, and one duplicate pair: two receipts from the same shop with the same date and amount, RM 46.50 (made up).

The pile check: 14 uploaded, 13 rows plus 1 duplicate. The owner opened both duplicate photos. Same receipt, photographed twice. One row deleted. For the faded petrol receipt, the owner found the transaction in the fuel app history and filled the amount by hand, noting "from app history" in the fix log.

One row was wrong: a pressure washer nozzle landed in "Stock / materials". The owner's accountant treats tools as "Equipment". The owner added to the category meaning: "Equipment: tools and machine parts, including pressure washer parts." Batch two was run with the updated list and sorted a hose connector correctly.

In the spreadsheet, SUMIF gave per-category totals and the check cell showed TRUE. The owner sent the September tab to the accountant and put the receipts in an envelope marked September.

Common mistakes and fixes

• The AI rounds or adds tax → "copy the final amount paid exactly as printed; do not calculate".
• A faded receipt gets a guessed amount → require UNREADABLE and fill it yourself from the original or another record.
• The same receipt is counted twice → ask for a duplicates list and decide each pair by opening both files.
• Categories change every month → paste the accountant's list with meanings and allow only those.
• The AI's total does not match yours → never use the AI's total; use spreadsheet formulas and a check cell.
• Card numbers appear in uploaded photos → crop before uploading, every time.
• Receipts pile up for months → set a fixed week each month and run small batches.

Take it further

• Weekly report: feed the month's expense totals into A weekly business report you check instead of write, or a monthly version of it.
• Project: keep the category list and prompt as Project files, so each month starts the same way. See Set up a Claude Project as your business brain.
• Skill: if you run this every month, package the prompt and categories as a Claude Skill. See Write your first Claude Skill for quotes or enquiry replies for the method.
• Monthly habit: first week of the month, same steps, same prompt.
• Related: Before you paste anything into AI: a two-minute safety check and Check the AI's answer before a customer sees it.

Quick checklist

• Categories from your accountant, with meanings.
• Clear photos, private numbers cropped.
• Batches of 10 to 15 receipts.
• Amounts copied exactly; UNREADABLE instead of guesses.
• Duplicates listed, decided by opening both files.
• Row count matches receipt count.
• Totals by spreadsheet formula, with a check cell.
• Originals filed by month.
• Known-good prompt saved for next month.