Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
ChatGPT reads every scan and drafts the field table and both lists. You check each field against image and approve.
Name the scans
Extract the table
Check marked fields
Draft both lists
Approve or hold
Use: You
Phone scans and photocopies, each file named by invoice batch and page, in one folder.
Check every paper invoice has a file and note any page that is visibly cut off.
Use: ChatGPT
Upload the scans and ask for a field table where unreadable characters carry an explicit marker and each field has confidence and source file name.
What to paste into the tool
Never invent a digit for an unreadable field: mark it. Keep the source image name on every value. Do not post OCR output to any ledger. Read invoice number, date, supplier, amounts and tax. Return a table: file, field, value or [UNREADABLE], confidence, note.
Use: You
Compare every low-confidence or marked field against the scan image.
No marked value is filled in or guessed; investigate against the image.
Use: ChatGPT
In the same chat, ask for the re-scan list and manual-keying list with the affected fields and the reason.
What to paste into the tool
From the table, list invoices to re-scan and invoices to key by hand, each with affected fields and reason. Do not guess values.
Use: You
Approve which invoices go back for re-scan and which are keyed by hand, or hold the lists; the assistant does not decide.
No unreadable value may be silently guessed. Do not post OCR output to any ledger.
Claude extracts each invoice into a table, drafts the re-scan and keying lists. You check values and decide each line.
Name and log scans
Import and extract
Check each field
Draft both lists
Decide each line
Use: You
Invoice scans, named by source file and page, with a readability log and unreadable scans marked.
Plan and service access need confirming first. No ledger posting, payments or supplier changes happen here.
Use: Claude
Upload one invoice PDF or image to Claude and request header and line fields in a table; repeat file by file.
What to paste into the tool
Extract invoice number, date, supplier, currency, subtotal, tax, final total and each line item with quantity, unit price and line total. Do not guess unreadable characters. Mark unreadable characters [UNREADABLE] and absent fields [NOT STATED]. Keep each invoice separate.
Use: You
Compare every extracted field with the original, including digits, signs, decimal separators, currency and invoice identity.
AI confidence is not a substitute for this check; add source file and page references yourself.
Use: Claude
Ask Claude to draft both lists from the checked table: every field still marked [UNREADABLE] or [NOT STATED] with its invoice and source reference.
What to paste into the tool
From the checked field table, draft two lists: a re-scan request list of invoices whose key fields are still marked [UNREADABLE], and a manual-keying list of invoices whose marked fields can be keyed from the original image. Carry the invoice number and source file/page reference onto each line. Do not fill in any marked value.
Use: You
Compare both lists with the marked table and readability log, and decide re-scan versus keying for each line.
Every source file stays accounted for; unclear values stay marked, not guessed. Do not post output to any ledger.
Gemini and ChatGPT each read the scans independently. You reconcile the differences, then Gemini drafts the lists from your decisions.
Get ready
Reading one
Reading two
Compare and decide
Draft the lists
Approve or hold
Use: You
The paper invoices and the scans, grouped up to the stated file limit.
Check each scan opens cleanly and every paper invoice has a file; the originals stay with you.
Use: Gemini
Upload the scans in groups and ask for the field table with unreadable markers and no guessing.
What to paste into the tool
Never invent a digit for an unreadable field: mark it. Keep the source image name on every value. Do not post OCR output to any ledger. Return: file, field, value or [UNREADABLE].
Use: ChatGPT
In a separate chat, upload the same scans and ask for the same field table, with no sight of reading one.
What to paste into the tool
Never invent a digit for an unreadable field: mark it. Keep the source image name on every value. Do not post OCR output to any ledger. Return: file, field, value or [UNREADABLE].
Use: You
Compare the two tables field by field, open the scan image for every difference or marker, and decide: accept, key by hand, or re-scan.
Agreement between two readings is not proof; check low-confidence fields against the image.
Use: Gemini
Give the reviewer decisions to Gemini and ask for the re-scan and manual-keying lists with reasons.
What to paste into the tool
List invoices for re-scan and for manual keying according to the reviewer decisions, with affected fields and reasons. Add nothing.
Use: You
Approve the lists, confirming no unreadable value was silently guessed, or hold and redo the decisions.
No OCR output is posted to any ledger; only approved lists go forward.
Best when: One assistant does both stages in one chat
What it does: ChatGPT extracts and lists; you check marks
Strong at: Fewest hand-offs; one chat holds everything
Limit: Scanned values may extract unreliably; check each field
Free plan: Yes, limited (uploads and analysis capped)
Best when: Invoice tables in a spreadsheet you fully check
What it does: Claude extracts and lists; you check and decide
Strong at: Vendor documents invoice header and line extraction
Limit: The free plan has usage limits
Free plan: Yes
Best when: You want differing readings surfaced for review
What it does: Gemini and ChatGPT read; you reconcile, decide
Strong at: Differences become a list of fields to open first
Limit: Every scan is uploaded to two products
Free plan: Gemini: Yes; ChatGPT: Yes, limited
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Never paste credentials or full payment details.
Compare all low-confidence fields against the scan image; request a re-scan where the image is illegible.
Approve which invoices go back for re-scan and which are keyed by hand.
Each scan is extracted, marked for manual keying or listed for re-scan, and a reviewer confirms nothing unreadable was guessed.
Check the result against your original inputs. Approve it before sharing or acting.