In-house project (internal tool)
Puenjer Books
An in-house, AI-supported bookkeeping system for a Spanish S.L.: reading receipts, classifying them by chart of accounts, reconciling the bank and preparing tax returns.
6
tax forms prepared automatically
01 · Starting point
Standard accounting software does not take over the actual work: receipts still have to be matched by hand, tax types checked and discrepancies hunted down. Migrating from the old system also revealed errors in the existing data – wrong currencies and wrong tax types.
02 · Solution
- 1 A chain of capture, classification, posting, reconciliation, review and tax preparation under the Spanish chart of accounts (PGC Pymes).
- 2 AI reads and classifies receipts – posting is done by a deterministic engine, so every entry stays traceable.
- 3 Uncertain cases stop at a confidence threshold and go to a human. Every step is written to an audit log, and closed periods are locked.
- 4 A two-stage review agent: code checks the maths, the language model judges plausibility.
03 · What I learned
AI classifies, code decides
Language models are strong at reading and categorising. The binding result – here the journal entry – is produced by deterministic logic. That makes the system auditable.
Judge every case on its own
Per-supplier defaults are convenient but dangerous: suppliers change how they are taxed. Every receipt is therefore assessed individually.
Clear limits to automation
Open tax questions go explicitly to the tax adviser. A good AI tool knows when something is not its call.