Invoice data entry is the most reliably profitable AI project we run, because the before-state is so clearly wasteful: a person retyping numbers that already exist in a PDF.
Real accuracy
On clean PDF invoices, field-level accuracy above 95% is normal. On a phone photo of a crumpled bill under a ceiling fan, it drops. Anyone quoting you a single accuracy figure without seeing your documents is guessing.
This is why every pipeline we build has a confidence threshold and a human review queue. The goal is not zero humans — it is turning three hours of typing into ten minutes of checking.
The Tally piece
Tally accepts structured XML over its HTTP interface, which is how we write entries in. GST fields, ledger mapping and voucher types all need configuring against your existing chart of accounts, and that configuration is most of the project.
What to check before starting
- Roughly how many bills per month, and in what formats
- Whether your ledger naming is consistent
- Who will own the review queue
- Whether you need it in Tally, or whether a clean spreadsheet would do