The problem
An assistant can understand an email and suggest a useful action. That does not confirm the order, book the delivery, check the invoice or match the incoming payment.
The approach
I automated the order process in an operating company with AI, from incoming order through to bank reconciliation. The process connects unstructured inputs with the corresponding ERP records and continues to follow open steps.
The workflow first reads the current ERP state. It matches the input to a case and checks deterministic rules. Where interpretation is needed, a model produces structured classification, not permission to post a transaction.
What this demonstrates
ERP automation rarely fails because a model cannot understand a sentence. It fails because ownership is unclear, data is inconsistent or exceptions were never designed as part of the process.
Where it applies
Good candidates are processes where people copy information from emails or documents into the ERP, repeat the same checks or manually chase open transactions.
The specific implementation is not a transferable package. ERP configuration, data models, roles and actual working practices differ from one company to another.