Re-keying supplier invoices into Tally or an ERP is slow, repetitive and easy to get wrong. Document AI can read the PDFs and photos for you, but the projects that disappoint usually fail on preparation, not on the technology: samples that do not represent reality, fields nobody defined, and no plan for the invoices the system is unsure about.
This checklist covers what to gather and decide before any extraction work starts. It is written for finance heads and accounts managers in India who receive GST invoices in many formats and want clean data in their books with fewer manual corrections.
Quick answer: Before you automate invoice data entry, collect a representative set of real invoices with their correct values, list each field you need and where it goes in your accounting system, write validation rules such as GSTIN checks, totals and purchase order matching, decide who reviews flagged invoices, and agree how accuracy will be measured before go-live.
1. Build a representative sample set
The sample set becomes your test set, so it has to look like your real inbox, not like your neatest supplier.
- Every common format: digital PDFs, scanned copies, phone photos, multi-page invoices and invoices with handwritten notes or stamps.
- Every major supplier: plus a mix of occasional suppliers, because those are where layouts surprise you.
- The difficult ones: faded thermal prints, skewed scans, invoices in Hindi or with mixed languages.
- Known correct values: for each sample, record what was actually posted in your books. Without this "ground truth", accuracy cannot be measured.
Keep the sample in a restricted folder. Invoices can contain personal details of proprietors and bank account numbers, so treat them like any other confidential record.
2. Define the fields and where each one goes
Write down every field you need, how it will be checked and where it will be posted. This table is a starting point for a typical purchase invoice.
| Field | Check to apply | Where it goes |
|---|---|---|
| Supplier GSTIN | 15-character format and match with your supplier master | Party ledger |
| Invoice number and date | No duplicate for the same supplier in the same financial year | Voucher reference and date |
| Place of supply | Decides IGST or CGST plus SGST | Tax ledgers |
| Line items, HSN or SAC, quantity, rate | Quantity and rate within purchase order tolerance | Stock items or expense ledgers |
| Taxable value, tax amounts, total | Line totals add up; tax matches rate within rounding | Voucher amounts |
| Purchase order number | Open purchase order exists for this supplier | Matching record |
| IRN and QR code, if present | Decoded values agree with printed values | Stored with the voucher |
Leave out fields nobody uses. Every extra field adds tuning and review effort. If you are unsure which fields matter downstream, our intelligent document processing team usually works this out with your accounts and audit staff in the first workshop.
3. Write validation rules before anyone writes code
Extraction tells you what the invoice says. Validation tells you whether it is right. Most of the value comes from these rules.
- Format checks: a GSTIN starts with a two-digit state code followed by the supplier's PAN and ends with a check character, so a misread digit is often detectable. Dates must fall in an open financial year.
- Arithmetic checks: line totals must add to the taxable value, and tax must equal rate times value within a rounding tolerance.
- Master data checks: the supplier must exist, and its GSTIN must match your master. You can confirm a GSTIN's status with the Search Taxpayer facility on the GST portal.
- Matching checks: invoice against purchase order, and against the goods receipt note where you use one (a three-way match).
- Risk checks: changed bank details, a new supplier with a large first invoice, or the same invoice number arriving twice.
4. Decide the human review loop
No extraction system is right every time, so design the review step deliberately rather than treating it as a failure.
- Set confidence thresholds per field. Amounts and GSTIN deserve stricter thresholds than a description line.
- Name the reviewers and how quickly flagged invoices should be cleared, especially near GST return dates.
- Make sure the review screen shows the invoice image next to the extracted value, so a correction takes seconds.
- Feed corrections back into tuning, so the same supplier layout stops being flagged.
- Keep approval before posting until accuracy on your own invoices has been proven over a few weeks.
5. Plan the posting into Tally or your ERP
Clean data is only useful if it lands correctly. Agree the posting design with whoever owns your books.
- Voucher type and ledgers: which purchase voucher type, which ledgers for each tax rate, and how cost centres are chosen.
- Name mapping: suppliers print their names in many ways, so a mapping table from GSTIN to ledger avoids duplicates.
- Stock item mapping: supplier item descriptions rarely match your item names, so maintain a mapping per supplier.
- A test company first: post into a copy of your books before touching the live company.
- An audit trail: every voucher should link back to the source image and record who approved it.
6. Agree how accuracy will be measured
Agree the measures before the pilot so the result is not argued about afterwards.
- Field-level accuracy: the share of individual fields extracted correctly, reported separately for critical fields.
- Document-level accuracy: the share of invoices where every required field is correct.
- Straight-through rate: the share of invoices posted without any human correction.
- Held-out testing: measure on samples that were not used for tuning, and compare with your current manual error rate.
How EthicsComputer can help
Our Intelligent Document Processing service begins with exactly this sampling and field design, then tunes extraction field by field on your own invoices and connects a review screen and posting into Tally, SAP, Zoho or your own system. A proof of concept typically takes 3 to 4 weeks. If invoice entry is one step in a longer approval chain, we can automate the rest with business process automation, and day-to-day books can be supported by our accounting and bookkeeping team.
How many supplier invoices does your team key in each month?
Tell us your invoice volume, formats and accounting system, and we will suggest how to test extraction on your own samples.
About the author
Vinay Grover founded EthicsComputer in Rohtak, Haryana in 2018 and leads its cybersecurity, software and AI work for clients across India and abroad. EthicsComputer is ISO 9001, ISO/IEC 27001 and ISO 20000-1 certified.
