Our own books: AI reads the purchase bills, our checks decide, a person approves

1. 01

   Every bill keyed by hand.

   Vendor, GSTIN, bill number, tax split and expense account were all entered into Zoho Books manually.

2. 02

   Bills arrive in every shape.

   PDFs, scans, phone photos, marketplace invoices and foreign software invoices.

3. 03

   Manual entry is not error-free.

   When we compared 15 bills our team had already posted with the documents, 4 entries differed: two by a few paise of rounding, one by its date, and one with the vendor's GSTIN recorded against the wrong state registration.

4. 04

   No draft status in the API.

   The Zoho Books API creates bills as open, with no draft, so the review has to happen before posting, not after.

1. 1

   Reading.

   An AI model (Anthropic's Claude) reads a PDF or image into a fixed set of fields: vendor, GSTINs, bill number, dates, each line with its HSN or SAC code and tax, totals, round-off and TDS withheld. It leaves a value blank rather than guess, and flags anything hard to read.

2. 2

   Checks, in plain code.

   Billed to our GSTIN; vendor GSTIN format and check digit; vendor matched to our Zoho Books records by GSTIN, then PAN, then name; no duplicate bill number across all vendor records; IGST or CGST and SGST according to the two states; lines, tax, round-off and TDS adding up to the printed total; bill date; reverse charge for foreign online services. A bill that fails a check cannot be approved.

3. 3

   Numbers are never the AI's.

   Amounts come back exactly as printed and our code does every sum. In testing, the AI's free-text note once got a TDS calculation wrong while the extracted figures were right, which is why the checks never use its notes.

4. 4

   Review beside the original.

   The review page shows the bill next to the extracted fields and the check results, with the expense account, TDS section and tax codes prefilled from the same vendor's previous bill. Every field can be corrected and re-checked; Approve and Reject are recorded.

5. 5

   Posting with a read-back.

   On approval the tool refreshes the Zoho Books data, re-checks for a duplicate, creates the bill through the Zoho Books API, attaches the original, then reads the bill back and compares Zoho's total with the printed total.

• On the 15 bills checked against the printed documents, the reading matched every key field: vendor, GSTINs, bill number, date, taxable value, tax split and total.

• 11 of the 15 agreed exactly with what our team had posted by hand; the 4 differences were in the manual entries.

• On 21 October 2026 the first live bill (a foreign software invoice under reverse charge) was posted through the tool, with Zoho's total equal to the printed total and the original attached.

• Average AI reading cost: under 2 US cents a bill.

• Not yet measured: time saved per bill and the share of entries needing correction. We will measure both over a month of use on our own books before offering the tool to clients.

• Our Senior Manager, Implementation & Automation Lead: design and build, using AI-assisted development tools.
• A Chartered Accountant director: review.

Hosting on Zoho Catalyst with a Zoho WorkDrive inbox, after testing.