Can AI read scanned GST invoices? Yes, field by field. Make the arithmetic prove the result.
Short answer. Yes, with checks. OCR engines and vision-language models can read a scanned tax invoice and return the supplier, GSTIN, invoice number, date, HSN codes, taxable value, tax and total as structured data. They misread digits, so the output should never be booked unchecked. Two cheap tests catch most errors: the lines must add up to the total, and the GSTIN must pass its format and checksum rules. If the supplier issues e-invoices, which have been required since 1 August 2023 for aggregate turnover above ₹5 crore, the invoice carries a digitally signed QR code holding the key fields, and reading that QR code beats reading the printed text.
01
What AI can extract, and how to check each field
Field
Typical misreads
Cheap check
Supplier and recipient GSTIN
Letters and digits that look alike, such as O and 0, or 1 and I.
A GSTIN has 15 characters: a two-digit state code, the holder’s PAN, a registration number, the letter Z and a check character. Validate the format and checksum, then match it to your vendor list.
Invoice number and date
Slashes and hyphens dropped, day and month swapped.
Numbers in one series should not repeat, and the date should fall in the period.
HSN or SAC code
One wrong digit changes the rate.
Compare with the code list and the item description.
Taxable value and tax
Decimal-point slips and Indian digit grouping such as 1,23,456.00.
Tax should equal rate times taxable value, line by line.
Invoice total
The last line misread.
Lines plus tax must add up to the printed total.
02
Read the QR code when there is one
Under the e-invoicing system the invoice registration portal returns a signed QR code that the supplier may print on the invoice. The official FAQ lists what it holds: the supplier’s GSTIN, the recipient’s GSTIN if applicable, the invoice number, the date, the invoice value (taxable value and gross tax), the number of line items, the HSN code of the main item, which is the line with the highest taxable value, and the unique invoice reference number. The content is digitally signed, so tampering makes verification fail, and anyone can verify it with the offline app on the portal.
E-invoicing has applied since 1 August 2023 to taxpayers with aggregate turnover above ₹5 crore in any preceding financial year from 2017-18, under Notification 10/2023 Central Tax. Thresholds have been lowered several times, so check the current notification. The QR code covers header fields only, not every line, so you still need the lines from the image.
03
Tools that read scans
Tool
Kind
Licence
Worth knowing
Tesseract
OCR engine
Apache 2.0
Works best at 300 dpi or more. Hindi data is available. Phone photos need deskewing and noise clean-up.
PaddleOCR
OCR toolkit
Apache 2.0
Says it supports 100+ languages and converts PDFs and images into structured data.
Qwen2.5-VL-7B
Vision-language model
Apache 2.0
Its card says it supports structured outputs for scans of invoices, forms and tables.
Docling
Document converter
MIT
Handles scanned PDFs with OCR and can run locally, including in air-gapped environments.
04
What a person must still check
Any invoice that fails the arithmetic or the GSTIN check. Send it to a person, never to the ledger.
A sample of invoices that pass. Passing the checks is not proof that a digit was not swapped for another that also fits.
Handwritten corrections, stamps and overwritten amounts, which OCR often skips.
Duplicates: the same invoice scanned twice, or an original and a copy.
Anything on which a tax claim depends. The accountant signs off, not the tool.
05
How an office would set this up
Scan at 300 dpi or more, straight and flat, in colour or grey, rather than photographing at an angle.
Read the QR code first. Use OCR for the fields and lines it does not cover.
Have the model return the fields as JSON that follows a fixed schema. Ollama and vLLM can both constrain output to a JSON schema.
Run the arithmetic and GSTIN checks automatically, and route failures to a review queue.
Keep a link from every extracted value back to the page image so a reviewer can see the source.
Review a random 5 per cent of passed invoices each month and track the error rate.
06
What ANVE does today
ANVE has a Bills function that reads GST bills, and its vision models read scanned documents with OCR. Both run today in ANVE’s development build. No sealed unit exists yet, and ANVE publishes no accuracy figures for bills.
07
Questions people ask
Can AI read handwritten invoices?
Handwriting is a harder, separate problem from printed text. Test on your own documents and expect to review more of them. The question on handwritten documents goes further.
Does a photo from a phone work as well as a scan?
Often not. Tesseract’s documentation advises at least 300 dpi and describes fixes for skew, noise and borders, which phone photos tend to need.
Can I trust the totals the AI reads?
Only after the arithmetic check passes: the lines and tax must add up to the printed total.
Can the e-invoice QR code be checked offline?
Yes. The official FAQ says anyone can verify the signed QR code using the offline app provided on the invoice registration portal.
08
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
