Can ChatGPT extract invoice data? What it reads and what you still check

Yes: ChatGPT can read an invoice you upload and return the supplier, dates and amounts as a list or a table, and OpenAI lists PDF among the formats ChatGPT accepts. What it does not give you is a reliable sign of which values are wrong: OpenAI's help centre says the model may express high confidence even in incorrect answers. Every value it returns needs the checks below before it reaches your books.

How to use a general assistant on an invoice

1. Check the account first. Personal and business plans treat your data differently (see below). On the ChatGPT Free plan the limit is 3 file uploads a day.
2. Ask for a fixed list of fields: supplier name, supplier VAT number, invoice number, invoice date, due date, net amount, VAT rate, VAT amount, total and currency.
3. Tell it to leave a field empty when the value is not printed, and to quote the text each value came from. Anthropic's guide to reducing hallucinations recommends both: explicit permission to say "I don't know" and, for long documents, answers grounded in word-for-word quotes.
4. Ask for dates as year, month and day (2026-04-03), so a swapped day and month is easier to spot.
5. Run the six checks below, then copy the confirmed values into your spreadsheet or accounting software yourself.

Why read values go wrong

• Characters that look alike. 0 and O, 1 and l, 5 and S, 8 and B, and a comma or a dot as the decimal mark: 1.250,00 and 1,250.00 are the same amount in two conventions. Microsoft's Document Intelligence documentation advises improving the quality of the input documents when confidence in the read text is low.
• Several amounts on one page. An invoice can show a subtotal, tax, a discount, the amount due, the amount paid and a prior balance; Amazon Textract treats each as a separate standard field. Taking the wrong one gives a plausible, wrong total.
• Dates in different orders. 03/04/2026 is 3 April for a supplier that writes the day first and 4 March for one that writes the month first.
• Multi-page invoices. The total often sits on the last page, and a table row can break across two pages; Microsoft notes that such a row can get a lower confidence than rows that are not split.
• A value that is not on the page at all. OpenAI's help centre calls responses that are not factually accurate hallucinations, and its examples include incorrect facts, fabricated quotes and overconfident answers. An invoice without a due date can still come back with one.

Six checks that catch a wrong or invented number

These work on the output of any tool, from a chat assistant to a dedicated extraction service, and each can be done in a spreadsheet with the invoice open next to it.

1. The net amount plus tax equals the total. If not, at least one of the three values was misread. Check the amount due separately when the invoice shows a deposit or an earlier payment.
2. The tax amount matches the rate. Multiply the net amount by the printed VAT rate, line by line where several rates apply: under EU rules a full VAT invoice shows the VAT rate applied, the amount payable and a breakdown of the VAT amount by rate or exemption.
3. The supplier's VAT number is valid. Compare it with earlier invoices from the same supplier, then check an EU number in the European Commission's VIES service. Your Europe explains that an invalid result means the number is not registered in the relevant national database, and that authorities do not give out the name and address behind a number, only confirm whether a given name and address match it.
4. The invoice number has not been seen before for that supplier. EU rules require a unique sequential number on every full invoice, so the same supplier and number twice means a duplicate or a misread. Also look for the same supplier and amount close together, which catches a copy saved under another file name.
5. The due date is after the invoice date. If not, the day and month were probably swapped or another date, such as the delivery date, was taken. Where payment terms are printed, the gap should match them.
6. A person confirms before anything is posted. OpenAI asks users to verify data from ChatGPT, and Microsoft describes confidence as a way to decide whether to accept a value automatically or flag it for human review. No value should be booked or paid until someone has compared it with the document.

What dedicated extraction tools add

Dedicated tools read the same documents but build the review and the hand-off around the reading. From their own documentation:

• A confidence value per field. Amazon Textract returns a confidence score with each value it detects on an invoice, including the total, the tax and the vendor's VAT number. Azure Document Intelligence returns field confidence between 0 and 1. Both are services called from your own software, which decides what happens to low scores.
• A validation screen. In Rossum, which Coupa has acquired, the validation screen gives a field a grey tick when the AI has validated it, for example because its confidence passed the threshold, and a green tick once a person has checked it; to find the problem fields, users are told to check those that have no tick. See also Rossum alternatives.
• Export. Parseur lists Excel, CSV and JSON export and Google Sheets on every plan, including the free one. Rossum's site lists integrations with SAP, Coupa, NetSuite, Workday and Microsoft Dynamics.
• Published prices per page. Textract's invoice and receipt analysis costs USD 0.01 per page in US West (Oregon) for the first million pages a month, with 100 free pages a month for three months for new AWS customers. Parseur's free plan covers 20 pages a month and its Micro plan 100 pages for USD 49 a month, excluding VAT; every page of a PDF counts.

Before you upload company invoices to any AI service

Invoices carry supplier names, bank details and prices. Ask three questions, and read the answer on the provider's own pages for the plan you use.

• Is the data used for training? On personal ChatGPT Free, Plus and Pro accounts, data sharing is enabled by default and can be turned off with the setting "Improve the model for everyone"; by default OpenAI does not train on content from ChatGPT Business, Enterprise or Edu workspaces or the API. By default Anthropic does not use inputs or outputs from its commercial products, such as Claude for Work and the API, to train its models; on Claude Free, Pro and Max it uses chats if you choose to allow it or when a chat is flagged for safety review. A conversation rated with the thumbs up or down button in Claude is kept for up to five years and may be used for training, on business plans too.
• Where is it processed? OpenAI's API offers data residency controls to eligible customers, with Europe (EEA and Switzerland) among the regions. Anthropic's API documentation lists "us" and "global" as the only inference locations, "global" being the default, and "us" as the only workspace storage location at present.
• Who can see it? In ChatGPT Business, workspace admins can view, export and delete members' conversations, and OpenAI's access is limited to authorised employees (engineering support, abuse investigations, legal compliance) and to contractors bound by confidentiality who review for abuse. Anthropic deletes API inputs and outputs within 30 days unless a listed exception applies, such as enforcing its Usage Policy or a legal duty.

Where Arhivix fits

Arhivix is a document system. Documents arrive by upload, scanner, phone camera or a connected company email inbox; Arhivix recognises the document type and proposes key fields (supplier, amount, net amount, VAT, due date, invoice number), and every read field is a proposal that a person confirms. It warns about possible duplicates (the same invoice number, the same supplier and amount, or the same file), and a person decides. The document stays filed and searchable: content search is in the paid plans, while the free plan searches file names. The data processing agreement forbids using customer data to train AI models, and files are hosted on AWS in the EU (Frankfurt and Dublin).

Posting invoices to your accounts stays in your accounting software, with you or your accountant; other software connects to Arhivix through the REST API and signed webhooks in every paid plan. The guide to invoice data extraction software sets out the kinds of tool. Pricing: a free plan called Arhivix Lite, then paid plans priced per company, not per user, with 5, 10 or 20 user accounts.