What is the actual problem you want to solve?
Invoice processing takes time because people read documents, type over data and then manually check whether the amounts match the order or the delivery. At ten invoices a day that is manageable. At a hundred or more it becomes a full-time task for someone who has better things to do. The problem is not the approval itself, but the typing and the preparation: reading invoice lines, matching them with purchase orders, flagging discrepancies. That preparation work is exactly where AI invoice automation adds value, provided the preconditions are right.
Where does AI invoice automation work best?
It works best when incoming volume is high enough to recognise patterns, when suppliers use reasonably consistent formats and when there is a clear reference to match on, such as a purchase order number or contract number. In that situation a well-configured system picks up the invoice, reads the relevant fields, matches them with the purchasing or order data in the ERP and presents discrepancies to a team member. The person approves or rejects; the AI does the lookup work. Think of distribution companies, wholesalers or manufacturing businesses with established suppliers and a standardised purchasing process.
Where does everyone get stuck?
Most projects stall on three points. First: supplier formats that vary every time. A PDF from one supplier has a clean table; from the next it is a scanned receipt from 2008. AI is good at recognising structure, but chaos has its limits. Second: missing references. If an invoice arrives without a purchase order number, or with a proprietary invoice number that means nothing, the system cannot match without additional logic or human input. Third: exceptions that fall outside the model. Credit notes, splits across multiple cost centres, invoices in foreign currencies or invoices for services without a concrete delivery: these always take more effort than the standard cases. A realistic estimate: in most operations, 60 to 80 percent of invoices can be automated; the rest will always need human attention. That is fine, as long as you set expectations accordingly.
When is AI invoice automation not the right fit?
If you process fewer than a few dozen invoices per week, the return on a fully automated pipeline is low. Implementation takes time and maintenance; the saving is marginal. In that case a smart upload tool with manual review is faster to set up and cheaper to maintain. Also, if your ERP data is messy, has missing master data or uses outdated item codes, automation backfires: the system matches on incorrect data and creates more work than it removes. AI invoice automation is not a fix for a poorly structured accounting process; it accelerates what already works, it does not repair what is broken.
How does Bonsai approach this?
Bonsai builds AI invoice automation in two forms. For organisations that want to replace or rebuild their core system, invoice processing is embedded in the Digital Twin: a domain-specific system designed from the ground up with AI as part of the core. For organisations that want to keep their current ERP, we build an AI Worker that runs on top of the existing system, reads documents, matches them and passes exceptions to the right person. In both cases the principle is the same: the person approves, the AI does the typing. Dottle, our product for document processing, is the engine behind reading the invoices themselves. The system becomes the property of the client; no subscription, no lock-in.