What does AI invoice processing actually do?
An incoming invoice contains fixed fields: supplier, invoice number, lines with description, quantity and amount, VAT, total. AI invoice processing reads those fields automatically, even when every supplier uses a different invoice format. Dottle, Bonsai's document AI, does this without requiring a template to be configured per supplier. The extracted data is then compared with the corresponding purchase order in your system: do the lines, quantities and prices match? Only when everything matches, or when a flagged discrepancy is found, does the proposal land with the accountant. They approve with one click, or adjust it if something is off. The human remains responsible; the groundwork is automated.
How does matching with the purchase order work?
The two-way or three-way match principle is nothing new in accounts payable, but executing it manually takes time. Dottle automatically compares invoice lines with the purchase order and, where available, the goods receipt confirmation. Discrepancies above a configured threshold are flagged and submitted for review. Minor deviations within the accepted margin are processed automatically. The result is a posting proposal ready in your accounting package or ERP: correct general ledger account, cost centre, VAT code. The accountant only needs to verify the proposal, not rebuild it from scratch.
When is automated invoice processing a good fit?
Dottle works best when you regularly receive invoices from a fixed group of suppliers, when you have an ERP or accounting package that can be integrated, and when invoices arrive digitally (PDF via email or a procurement portal). Companies in logistics, trade and manufacturing that process tens to hundreds of incoming invoices per month see immediate results: less manual work, fewer data entry errors and a shorter lead time from invoice to payment.
When is it not a good idea?
Being honest about the limitations is part of the approach. If your suppliers send unstructured or messy invoices, deliver invoices on paper without a scanning process, or if no ERP integration is available, then automation becomes difficult. Dottle does not learn new formats independently over time. The quality of the output depends on the quality of the input. If the source data is structurally poor, that problem needs to be solved first. Automation on top of chaos only makes the chaos faster.
What does the implementation look like?
We start with a brief intake: which invoice formats come in, which system sits on the receiving end, what is the current lead time and where do errors occur? Based on that, we build the integration and configure the matching logic. The accountant then works in a straightforward approval screen: invoice on the left, proposal on the right, discrepancies highlighted. No new system to learn, but noticeably less manual work. Ownership of the integration and configuration stays with the client.