Capital Resources.
A US commercial lender, and the paperwork that arrives before every decision. Borrower documents used to be opened one at a time and typed into a spreadsheet; now they are read the moment they land, and the officer confirms instead of transcribing.
* Not real data. Names, figures and identifiers are invented, and anything under NDA is stripped from the capture.
Underwriting starts with documents – financial statements, tax returns, bank records – and they arrive as PDFs in a shared mailbox. Each one was opened, read, and its figures typed into a workbook by hand. The lending ratios were then computed in that workbook, and the compliance picture for a borrower lived across several of them. Nothing was wrong with the numbers. What was wrong was that a person's afternoon went into moving them.
A pipeline that watches the mailbox through Microsoft Graph and picks up each borrower PDF as it lands. Text is read where the document has text and OCR is run where it does not; an agent built on LangGraph lifts the figures out, computes the lending ratios from them, and puts the result in front of the officer as a proposal rather than a fact. The officer confirms, corrects or rejects – that click is the audit trail. Compliance status became a live matrix inside the Microsoft Teams app the team already had open, so there is no second system to log into.
The whole thing runs in the client's own Microsoft 365 tenant and Azure subscription. Documents never leave it; the models are called from inside it. The network rules are the client's, and atopos holds no copy of anything.