AI automation

How to automate invoice data entry with AI into Google Sheets

By Sauravjit Singh · Updated 2026-11-13 · 6 min read

Short answer: You can automate invoice data entry by sending each bill (PDF or photo) to an AI model that reads it and returns fixed fields such as GSTIN, invoice number, taxable value and tax split, then logging them in a Google Sheet. A validation step should flag totals that don't add up. Your accountant still reconciles.

Typing purchase bills into a sheet or accounting software is slow, error-prone and exactly the kind of work AI now does reasonably well. Modern AI models can read a clear PDF or a photographed bill and pull out the fields you ask for, without separate OCR software.

This guide sets up a no-code pipeline for Indian GST invoices: which columns to capture, the extraction prompt, a check that catches wrong totals, and a clear boundary on what the automation should not do.

Which GST invoice fields should your sheet capture?

Capture the fields your accountant needs to match bills later, plus a few for checking the AI's work. Here is a practical column map for purchase bills.

Sheet columnWhat it holdsWhy it matters
supplier_nameVendor's name as printedMatching and search
supplier_gstinVendor's 15-character GSTINNeeded for matching with GST records; the first two digits are the state code
buyer_gstinYour GSTIN as printed on the billCatches bills billed to the wrong entity
invoice_no / invoice_dateAs printed; date stored as YYYY-MM-DDDuplicate checks; avoids 03/04 confusion
place_of_supplyState named on the invoiceHelps explain which tax type applies
hsn_sacHSN (goods) or SAC (services) codes, comma-separatedClassification your accountant checks
taxable_valueTotal before tax, in ₹Base for tax checks
cgst / sgst / igstEach tax amount, in ₹, zero if absentGenerally CGST+SGST for intra-state and IGST for inter-state supplies
total_amountInvoice total, in ₹Validation and payment
file_link / statusLink to the original file; OK or CHECKEvery row traceable to its source

GST rules have exceptions and change over time. This column map is a data-entry aid, not tax advice; confirm the fields and treatment with your CA or the official GST portal.

The pipeline, step by step

The simplest version uses a Google Drive folder as the inbox and Make or Zapier to run the rest.

  1. Create a Drive folder called 'Bills – to process'. Everyone drops PDFs and photos there, or you forward vendor emails with attachments into it using a Gmail filter plus an automation.
  2. Trigger: 'new file in folder' in Make or Zapier.
  3. Send the file to an AI model that accepts PDFs or images (OpenAI, Gemini or Claude modules, or a dedicated document-extraction service) with the extraction prompt below.
  4. Parse the JSON reply into fields.
  5. Run the validation checks (next section) in a formatter or a few sheet formulas.
  6. Add a row to the 'Purchase register – draft' sheet with all fields, the file link and a status.
  7. Move the file to 'Bills – processed' so nothing is read twice.

If most bills come by email, you can skip the folder and trigger on new Gmail messages with a label such as 'vendor-bills', applied by a Gmail filter on known supplier addresses. The automation then saves the attachment to Drive for the record and continues from step three. Keep the folder route as well for paper bills scanned in the office, so both paths end in the same sheet with the same columns.

If the AI step is new to you, our guide on how to use ChatGPT in Zapier and Make explains the module setup and why fixed JSON output matters.

Worked example: the extraction prompt

Write the prompt so the model fills fields and never invents them.

  • You extract data from Indian GST purchase invoices. Return JSON only with these keys: supplier_name, supplier_gstin, buyer_gstin, invoice_no, invoice_date (YYYY-MM-DD), place_of_supply, hsn_sac (array), taxable_value, cgst, sgst, igst, total_amount, confidence (high, medium or low).
  • Amounts are numbers in rupees with no commas or symbols. If a tax is not shown, use 0. If any other field is missing or unreadable, use null. Do not guess or calculate missing values. Copy GSTINs exactly as printed.

Say a stationery supplier's bill to a Pune CA firm shows taxable value ₹12,400, CGST ₹1,116, SGST ₹1,116 and total ₹14,632. A good run returns those four numbers, the two GSTINs and the HSN code, with confidence high.

Critique of a typical first attempt: without the 'do not calculate' line, models sometimes fill a blurred total by adding the parts. That looks helpful and hides the very errors you want to catch. Without the date format rule, '04/05/2026' may come back as April or May. Both lines earn their place.

Add a validation step that flags totals that don't add up

Never trust extracted numbers blindly. A few simple checks catch most problems, and anything that fails is marked CHECK for a person.

  • Arithmetic: taxable_value + cgst + sgst + igst should equal total_amount within about ₹1 for rounding. Round-off lines and extra charges can cause legitimate differences, so flag rather than reject.
  • Tax type: a bill should normally show either CGST and SGST together or IGST, not both. Flag mixed or one-sided cases for review.
  • GSTIN shape: 15 characters. Flag anything shorter, longer or null. (A formula can check length; a CA can verify validity on the GST portal.)
  • Buyer GSTIN: flag if it doesn't match your own.
  • Duplicates: flag if the same supplier GSTIN and invoice number already exist in the sheet.
  • Low confidence: anything the model marked medium or low goes to CHECK.

Using the example above: 12,400 + 1,116 + 1,116 = 14,632, so the row is OK. If the AI had read the total as ₹14,362 (two digits swapped, a classic photo misread), the difference of ₹270 would flag it immediately.

What about WhatsApp photos and messy scans?

In many Indian offices, a large share of bills arrive as WhatsApp photos from the delivery person or a field team. These work, but the quality of the photo decides the quality of the data.

The easiest route is to have staff save or forward bill photos to the shared Drive folder or a dedicated email address rather than building a WhatsApp bot. That keeps the pipeline simple.

  • Ask for flat, well-lit, full-page photos; folded corners often hide the total.
  • Thermal-paper receipts fade; photograph them on the day.
  • Handwritten kachha bills are hit and miss. Expect more CHECK rows.
  • Multi-page invoices should be one PDF, not three photos.

Expect the CHECK rate to be higher in the first fortnight while staff get used to taking usable photos. Share two examples of a good and a bad photo in the team group; that small step usually does more for accuracy than changing the prompt.

What does it cost, and is my data safe?

For a small business processing a few hundred bills a month, the AI cost is usually modest, often in the range of a few hundred rupees a month as of November 2026, depending on the model and page count. Your Make or Zapier plan adds its own cost; see our Zapier vs Make cost comparison for how usage is counted.

Invoices contain business and sometimes personal data. Use a business account for every connection, check each provider's data-use terms, and read our checklist on whether it is safe to connect Gmail and Drive to automation tools before switching this on.

Where this stops and your CA starts

This automation is a data-entry helper. It does not decide input tax credit eligibility, file returns or reconcile against GSTR-2B. Those depend on rules that change and on facts the AI cannot see. Keep the sheet as a draft register that your accountant reviews, and consult your CA on anything tax-related.

When the pipeline stops working, it is usually a renamed folder, an expired connection or a vendor whose bill layout changed. Our guide to why Zapier and Make automations fail covers the fixes, and the no-code automation overview shows what else in the accounts office can be automated. Document-extraction workflows like this one are a project in our Applied AI Professional course.

Frequently asked questions

Can AI read invoices into Google Sheets without OCR software?

Yes. Many current AI models read PDFs and images directly and return structured fields. For very high volumes or poor scans, a dedicated document-extraction service may be more consistent.

How accurate is AI invoice extraction?

On clear, printed bills it is usually good, but not perfect. Photos, faded receipts and handwritten bills cause more errors, which is why the validation checks and a human review of flagged rows matter.

Can this automation file my GST returns?

No, and it shouldn't. It prepares a draft register of bill data. Filing, input tax credit and reconciliation should be handled by your CA or accountant using the official GST systems.

Can I send the data to Excel or Tally instead of Google Sheets?

Excel on OneDrive is supported by Zapier and Make. Tally usually needs a connector or import file; check what your Tally partner supports before building around it.

Is this suitable for a CA office or accounts team?

It suits teams that type many purchase bills, as long as every row stays traceable to its file and someone reviews CHECK rows. Spettro School's AI course covers building this type of extraction workflow without coding.

Want to learn this with a trainer?

Leave your details and our counsellor will call you within one working day with batch dates, fees and the right course for you.

  • Free counselling call, no obligation to join
  • Weekday and weekend batches, online or offline
  • Small batches of 15 to 20 students

Or chat on WhatsApp

Enter a 10-digit mobile number.

We use your details only to contact you about Spettro School courses.

Chat on WhatsApp