Where does the time actually go in the invoicing process?
In most operations, invoicing is not a single process but a series of small decisions. An invoice arrives as a PDF or email. Someone reads the amount, looks up the corresponding order in the ERP, checks whether the quantity is correct, requests approval if needed, and then posts it. With twenty invoices a day that is manageable. With a hundred or more it becomes a full-time job. The bottleneck is rarely the posting itself, but the finding, comparing and deciding. Automation that only speeds up the posting misses the point.
What document AI concretely solves
Document AI, such as what Bonsai builds through Dottle, takes over the reading and recognition of invoices. The system extracts structure from unstructured documents: supplier name, invoice number, line items, amounts, VAT. It then automatically matches that data against open orders or delivery notes in your existing system. Invoices that are fully correct pass through without anyone needing to look at them. Invoices with a discrepancy, a missing order number or a price difference are routed to a team member with a clear flag: this does not match because of reason X. The team member decides. This shifts the work from manual re-entry to handling genuine exceptions.
Where organisations get stuck during implementation
Reading invoices is the easiest part. The real obstacles lie elsewhere. First: suppliers that do not use consistent invoice formats. A regular supplier always sends the same layout. An ad-hoc supplier sends something different every time. Document AI is strong on patterns, but with highly variable input it requires more human oversight. Second: integrations with the ERP. If the ERP has no reliable API, or if the order data is messy, the system cannot match accurately. Automation makes poor data visible, but does not fix it. Third: approval workflows. Who is authorised to approve what? If those rules are not defined, you build a system that in practice still ends up on someone's desk waiting for a signature.
When automating the invoicing process is not the right fit
Low volumes are the most common reason not to automate. If you process fifteen invoices a week with a fixed set of suppliers and a straightforward accounting package, a structured manual workflow is likely cheaper and more stable. Automation pays for itself at volume, with diverse input and where the cost of manual errors is high. Beyond that: if your ERP is due for replacement and you are already considering rebuilding your core system, it is sometimes smarter to integrate invoicing into that new system rather than adding a separate layer on top of an ageing foundation.
Two paths: an AI layer or rebuilding the system
At Bonsai, clients choose between two directions. The first is an AI Worker on top of the existing system: document AI that reads invoices, matches them and flags exceptions, without replacing the ERP or accounting environment. This is the fastest route and well suited when core systems are still functional but manual work is consuming too much time. The second direction is a Digital Twin: rebuilding the core system itself, with invoicing, order management and approval workflows built in as part of a coherent whole. This is more relevant when current systems fall structurally short or when you want to break the dependency on expensive software vendors. The choice depends on how deep the problem runs, not on which solution looks more appealing.