Invoice OCR software in India: what to check before buying

Good invoice OCR software for India does more than read text. It pulls the supplier GSTIN, invoice number, date, HSN codes, taxable value and the CGST, SGST or IGST split from every bill, checks that the numbers add up, and sends anything doubtful to a person before it reaches Tally. Before you buy, test it on a few hundred of your own bills, bad phone photos included, not on the vendor's demo set.

This checklist is for owners, accounts heads and plant managers who receive hundreds of supplier bills a month and still type them in.

Which fields should it pull from a GST bill?

Ask the vendor for the output schema, not the brochure. For a purchase bill the minimum is:

On HSN: since 1 April 2021 a supplier with turnover above ₹5 crore in the preceding financial year must show a 6 digit HSN or SAC, and one up to ₹5 crore must show 4 digits on B2B invoices, according to the GST Council's press release on Notification 78/2020. A tool that reads HSN as free text and never checks its length lets a misread digit through.

On GSTIN: the GST portal's Search Taxpayer page shows a GSTIN's legal name and whether it is active, cancelled or suspended, without logging in. Checking each new supplier once is cheap.

Header fields or line items: which do you need?

Header fields sit at the top and bottom of the bill: supplier, number, date, totals. Most tools read them well because every bill has them in some form. Line items are the table in the middle: description, HSN, quantity, rate, discount, tax rate, amount. They are much harder. Tables run over pages, columns change between suppliers, and a thermal paper bill has no ruled lines at all.

Decide which you need before you compare tools. If your books take one purchase voucher per bill against a ledger, header fields with the tax split are enough. If you track stock, match bills to purchase orders or check rates against a contract, you need line items, and you should test on your longest bills. Our comparison of OCR and intelligent document processing explains why a table needs more than text recognition.

Can it read multi-page, photographed and handwritten bills?

Ask for results on the bills that actually arrive:

• Multi-page PDFs. Does it know that page 2 continues the table on page 1, and that page 3 is a different invoice in the same scan?
• Phone photos. Shadows, folds and skew. Tesseract's documentation says it works best on images of at least 300 dpi, and a quick photo of a crumpled bill is a long way from a flatbed scan. Ask whether the tool straightens images and whether it says when a photo is too poor to read, rather than guessing.
• Handwritten bills. Many small suppliers still write cash bills by hand. Handwriting, stamps over the total and faint carbon copies all lower accuracy. The useful question is whether low confidence fields are flagged for a person.

What can the e-invoice QR code do for you?

Suppliers who must e-invoice report each B2B invoice to the Invoice Registration Portal and print the signed QR code it returns (see e-invoicing in India). According to the West Bengal commercial tax department's e-invoice guide, that QR code carries the supplier and recipient GSTINs, the invoice number and date, the invoice value, the number of line items, the HSN of the main item, the IRN and the IRN date.

Software that decodes the QR code gets those header fields exactly, with no OCR, and can cross-check what it read from the printed page. A mismatch between the QR value and the printed total deserves a person's attention. Line items are not in the QR code, so OCR still does that part. Ask each vendor whether it reads the QR at all.

What happens to doubtful bills, and how do they reach Tally?

No tool is right on every field of every bill. What matters is what happens when it is unsure. Look for:

1. A confidence score per field, not one score per document.
2. A review screen with the bill image beside the extracted fields and the doubtful field highlighted.
3. Rules that always send a bill to a person: a new supplier, a total above a limit, a failed tax check.
4. A log of who approved what, and when.

Then the export. TallyPrime Release 4.0, released on 27 November 2023, added import of masters and transactions from Excel with field mapping, and TallyPrime also accepts XML requests over HTTP, so checked bills can arrive as purchase vouchers on the right ledgers. Ask whether the tool creates a missing supplier ledger or stops and asks. Our Tally integration page shows how these imports work.

We have built an OCR pipeline of this kind for invoices, IDs and forms, with a person reviewing the doubtful fields. It is the pattern our OCR and document intelligence service follows.

How do you measure accuracy on your own bills?

A vendor demo uses clean invoices. Run your own test:

1. Pick 200 to 300 past bills that look like a normal month: your main suppliers, the odd ones, phone photos, multi-page PDFs, a few handwritten.
2. Have your accountant enter the correct values for the fields you need. That is your answer key.
3. Run every tool you are considering on the same set.
4. Score per field: how many GSTINs, totals and tax amounts came out exactly right, and how many wrong values were flagged rather than passed silently.
5. Time how long a person spends in the review queue per bill.

A wrong total marked as confident costs more than a blank field, so weigh the silent errors most.

If your team types in supplier bills every day and you want to know what a pipeline would read correctly from your own documents, send us a sample and we will test it. Book a consultation.