Extract Invoice and Receipt Data into a Spreadsheet

Tier: Core · Hands-on time: ~30 minutes · Tools: ChatGPT, Claude, Google Sheets

Manual invoice entry is the purest form of office drudgery: high volume, zero judgment, and one transposed digit away from a payment dispute. Models read documents well enough to do the typing — but "well enough" is not "always right", which is why this recipe's centrepiece is a reconciliation column that makes every extraction self-auditing.

When to use — and when not to

• Expense claims and card reconciliationsa month of receipts that must become rows before anyone gets reimbursed.
• Small-batch invoice entrya few dozen supplier invoices to log into a tracker or import into accounting software.
• Historical cleanupa folder of old PDFs that finance wants summarised into a searchable sheet.

And when not to:

• Your company already runs OCR-equipped AP software. If invoices flow through an accounts-payable system with extraction built in, use that pipeline — it is audited, this is not.
• Handwritten or badly photographed receipts. A blurry thermal-paper photo defeats the model quietly — it will guess. Retype those few by hand.
• Payment execution. This recipe produces a data sheet, not a payment run. Approvals and bank details stay entirely in your controlled systems; never let extracted data trigger a payment without the normal checks.

The recipe

1. 1.

   Batch the documents and define your columns (5 minutes). Decide the exact columns once: file name, vendor, invoice/receipt no., date, currency, subtotal, tax, total, category, line items. Work in batches of five to ten documents — small enough to check, large enough to be worth it. Blank out or crop anything the sheet doesn't need, especially bank account details.

2. 2.

   Extract with a strict schema and an uncertainty rule. Upload the batch to ChatGPT or Claude with this prompt:

   Copy

   Extract data from each attached invoice/receipt into ONE table with
   EXACTLY these columns:
   
   file name | vendor | doc number | date (YYYY-MM-DD) | currency |
   subtotal | tax | total | line items (item: amount; ...)
   
   Rules:
   - Copy values EXACTLY as printed. Do not round, convert currencies,
     or normalise vendor names.
   - If a field is missing or unreadable, write UNCLEAR — never guess.
     A wrong number is far worse than a blank.
   - One row per document. Dates reformatted to YYYY-MM-DD only when
     the printed date is unambiguous; otherwise UNCLEAR.
   - After the table, list any document where the printed total does
     not equal subtotal + tax, or where line items do not sum to the
     subtotal, with the amounts you read.
   
   Output the table as tab-separated values so I can paste it into
   Google Sheets.
   

   A good output ends like this:

   Copy

   acme-0312.pdf  Acme Supplies Ltd  INV-88213  2026-03-12  USD
   412.50  33.00  445.50  toner: 180.00; paper A4 x10: 232.50
   
   Checks:
   - officehub-receipt.jpg: line items sum to 96.00 but printed
     subtotal reads 99.00 — possible misread of item 2, flagging.
   - parkfee.jpg: date UNCLEAR (print faded).
   

3. 3.

   Paste into Sheets and add the reconciliation column (5 minutes). Paste the TSV into Google Sheets. Add one column: =IF(ROUND(F2+G2,2)=ROUND(H2,2),"OK","MISMATCH") — subtotal plus tax must equal total. Add a second if you extracted line items: their sum versus subtotal. These formulas are the proof layer: the model read three numbers per document that must be arithmetically consistent, so a misread almost always surfaces as a MISMATCH.

4. 4.

   Check every MISMATCH and UNCLEAR against the source (8 minutes). Open only the flagged documents — that is the whole time saving. A MISMATCH means either the model misread a digit (fix the cell) or the document itself is wrong (flag it to the vendor or finance — you just caught a real error, which happens more often than you'd expect). Spot-check two unflagged rows too; consistency checks catch most errors, not all.

5. 5.

   Standardise and hand off. One last pass in the chat if needed: "Here is my vendor column — map each to this master vendor list, exact matches only, UNCLEAR otherwise." Then the sheet is ready for the tracker, the expense claim, or an import. Keep the source files attached or linked — an extraction without its source is an assertion, not a record.

Make it yours

• Claude for big batches: it handles many PDFs per conversation and is disciplined about UNCLEAR when told. For fifty-plus documents, run batches of ten and keep the same conversation so formatting stays identical across batches.
• ChatGPT for photos and mixed bags: strong OCR on phone photos of paper receipts. Ask it to also output a "confidence" note per row — low-confidence rows join your checking pile even if arithmetic passes.
• Google Sheets as the audit layer: beyond the reconciliation formulas, add conditional formatting on MISMATCH/UNCLEAR and a pivot by vendor-month. Duplicate doc number detection (=COUNTIF(C:C,C2)>1) catches the same invoice extracted twice — a paste-twice error no model will notice for you.

Quality check

• Reconciliation column shows OK on every rowor the mismatches were resolved against the source document.
• Every UNCLEAR was resolved by a human, not deleted.
• Two clean rows spot-checked against their PDFs, including one line-item breakdown.
• No duplicate document numbers unless the documents genuinely repeat.
• Currency column is populatedmixed-currency batches silently summed in one currency is the classic downstream disaster.