Most companies asking about AI want a chatbot. Almost none of them have a problem a chatbot solves. The return is somewhere much less interesting: documents arriving in formats you did not choose, being retyped by people.
Supplier invoices, prescriptions, admission forms, delivery challans. These arrive as scans and photographs, in inconsistent layouts, and someone keys them in. It is exactly the work language models are now genuinely good at, and exactly the work nobody wants to do.
The thing that makes it safe is not model quality, it is the approval gate. Every extraction carries a confidence score. Below a threshold, or above a value limit, it routes to a person. The agent never silently posts a large payment.
You also need an evaluation set before you start. Fifty documents with correct answers, run after every prompt change. Without it you are guessing whether last week's improvement broke something, and with document work you will not notice for a month.
At one distributor we took roughly 80% of invoices hands-free. The three people doing data entry moved to supplier reconciliation, which had been on the list for two years and never had anyone to do it.
