I’m testing the Payables Agent in Business Central (US cloud) and I’m trying to understand how it actually “learns” from history.
I created multiple test invoices (10+) for the same vendor using the exact same Word template, exported to PDF each time. The only differences were invoice number, date, and sometimes quantities. The line descriptions were identical across all invoices.
All invoices were:
- Sent to the configured AP Agent email
- Created as Purchase Document Drafts by the Agent
- Reviewed manually
- Coded consistently to the same GL accounts (in one test)
- Coded consistently to the same Item (in another test)
- Posted
The Agent correctly identifies the vendor header every time, so OCR and vendor matching are working fine.
However, even after 10+ identical posted invoices, the Agent still does not automatically populate either:
- the GL account (in the GL test), or
- the Item No. (in the Item test)
It leaves the Type and No. blank in the draft.
We are aware of Text-to-Account and vendor default accounts/items. We intentionally did not use those because the goal was to test whether the Agent truly learns from processing history. If it only works when exact text-to-account mappings or vendor defaults are configured, then it’s effectively doing OCR plus lookup rather than intelligent learning from prior postings.
Has anyone observed the Agent actually converging and reliably auto-populating GL or Item after repeated identical postings without using Text-to-Account or vendor defaults? Is there documentation that explains the threshold or learning behavior?
Trying to understand whether this is expected design behavior or a limitation in the current release.
