Purchase bill entry is the slowest part of most Indian back offices. Sales invoices are generated by software and arrive structured. Purchase bills arrive as paper, scans, email PDFs and WhatsApp photographs, in whatever format each supplier happens to use.
What OCR actually reads from a purchase bill
Optical character recognition converts an image into text. The harder part is extraction: deciding which piece of text is the invoice number and which is the purchase order number when both appear in the same corner.
| Invoice number | High | Confused with PO or challan number | Invoice date | High | DD/MM vs MM/DD ambiguity | Supplier GSTIN | High | 15-character format is self-checking | Taxable value | Medium | Multiple subtotals on multi-page bills | CGST / SGST / IGST split | Medium | Rate columns misaligned in scans | Line items and HSN | Low | Merged cells, rotated scans, handwriting |
A practical rule: trust OCR for header fields, verify tax amounts, and review line items manually when input tax credit depends on them.
The scan quality that actually matters
Why GSTIN extraction matters more than the rest
A supplier GSTIN is 15 characters with a checksum, so it can be validated automatically. Once it is confirmed, the bill can be matched against GSTR-2B and the credit position becomes verifiable. See the GSTR-2B reconciliation guide for the matching step.
A review workflow that stays fast
- Capture the bill and let extraction fill the header fields.
- Confirm GSTIN and invoice number against the document on screen.
- Check the tax split, since this is where extraction most often drifts.
- Assign the expense or purchase ledger, reusing supplier defaults.
- Post the entry and keep the source image attached for audit.
The attachment matters as much as the data. During assessment, the ability to show the source document next to the ledger entry is what makes the credit defensible.
What to do with handwritten bills
Handwriting recognition remains unreliable for figures, and a misread digit in a tax amount is worse than no extraction at all. For handwritten bills, use capture for the image and audit trail, and enter the amounts manually.
Where extraction fits in the purchase cycle
Extraction is one step in a longer chain, and its value depends on what happens either side of it. Dropping it into an otherwise manual process produces less benefit than expected.
- Bill arrives by email, WhatsApp, courier or hand delivery.
- Capture: scan or forward it into one place rather than several inboxes.
- Extraction: header fields and tax amounts are read from the document.
- Review: a person confirms GSTIN, invoice number and tax split.
- Coding: expense or purchase ledger assigned, using supplier defaults.
- Posting: entry created with the source image attached.
- Matching: the invoice is reconciled against GSTR-2B when it appears.
The most common failure is step two. Bills scattered across three inboxes and a WhatsApp group cannot be processed reliably no matter how good the extraction is.
Supplier templates improve accuracy faster than better OCR
Most businesses receive bills from a stable set of suppliers, each using a consistent layout. Once the system learns where a given supplier prints its invoice number and tax block, accuracy on that supplier rises sharply.
Measuring accuracy honestly
Vendors quote accuracy percentages that are rarely comparable. A figure of ninety-nine percent at character level can still mean a meaningful share of bills carry at least one wrong field.
| Character accuracy | Raw OCR quality | One wrong digit fails the whole field | Field accuracy | Per-field correctness | Treats invoice number and paise equally | Straight-through rate | Bills needing no correction | The number that actually matters |
Straight-through rate is the honest measure: the proportion of bills that reach posting without a human changing anything. Track it monthly and judge any tool on that.
Retention, storage and what an assessment expects
Capture is also a record-keeping exercise. GST law requires records to be retained for a defined period from the due date of the annual return, and the source document is what supports a credit claim under scrutiny.
That last point is regularly overlooked. A business that can export its ledger but not its attachments has kept only half of what an assessment will ask for.
When manual entry is still the right answer
Automation has a floor below which it costs more than it saves. Extraction is not worth configuring for a business receiving a handful of bills a month from two suppliers.
Coding rules do more work than extraction
Extraction reads the document. Coding decides which ledger the expense hits, and that decision is what makes the entry useful for reporting. Most of the residual manual effort after automation sits here.
Blocked credits need catching early
Certain input credits are blocked regardless of a valid invoice, and identifying them at entry is far cheaper than reversing them after a claim.
Common categories include motor vehicles below the specified seating capacity, food and beverages, club memberships, works contract services for immovable property, and goods lost, stolen or written off. Coding these to a separate ledger at entry means the return preparation never has to hunt for them.
What good looks like after six months
A business that has run this properly for two quarters should be able to answer these without opening a filing cabinet.
- How many purchase bills were received last month, and how many are still unposted.
- The straight-through rate, and which suppliers drag it down.
- Total input credit available against total claimed, with the difference explained.
- Which suppliers have unmatched invoices in GSTR-2B.
- How quickly a specific bill from eight months ago can be produced on screen.
Frequently asked questions
Is OCR accurate enough to post entries without review?
Not for tax amounts. Header fields are usually reliable, but input tax credit depends on exact figures, so a human confirmation step should stay in the workflow.
Does this replace GSTR-2B reconciliation?
No. Extraction gets the bill into the books. Reconciliation confirms the supplier actually filed it. Both steps are needed before the credit is safe to claim.
What file format should suppliers be asked for?
A digitally generated PDF. It carries a real text layer, so extraction is near-perfect and no scanning step is involved.
