Supplier Invoice Intake Agent
Manual invoice keying removed
Invoices arrive as email attachments from several dozen suppliers, and every one of them has its own layout. Somebody opens each PDF, finds the reference, types the totals into the accounting system and files the attachment. It is not difficult work. It is just work that never stops arriving, and it lands on whoever is least able to say no to it.
This study illustrates the kind of engagement we run rather than documenting one client. It carries no client metrics or quotes.
The challenge
Document extraction is the easy half. The hard half is what happens when the model is unsure: a scanned invoice with a smudged total, a supplier who changed their template last month, a line that could match two open purchase orders. An agent that guesses in those cases costs more than the manual process it replaced, because a wrong figure entered confidently is only found later, usually by the supplier.
Our approach
- 1
Scoped the agent to one inbox and one document type. Anything that is not a supplier invoice is left alone rather than handled badly.
- 2
Grounded extraction in the client's own purchase order records, so every invoice is read against what was actually ordered rather than in isolation.
- 3
Set an explicit confidence threshold with a review queue behind it. Below the line, the invoice goes to a person with the fields pre-filled and the uncertain ones flagged, which is faster than starting from the PDF.
- 4
Logged every run: what was read, what was matched, what was escalated and why, so a disputed entry can be traced back to the document it came from.
The outcome
Clean invoices flow into the accounting system without anyone opening the attachment.
Ambiguous ones surface as a short review queue instead of hiding inside a full inbox.
Every posted entry has a run record behind it, so a query from a supplier is answered from the log rather than from memory.
Have a project in mind?