Financial statements in, checked numbers out.
Drop in a P&L, balance sheet, cash flow or vendor invoice. The agent pulls out every number, maps it to a chart of accounts and checks that the totals still tie. Anything it isn't sure about waits for a person to review before it goes to Google Sheets.
- 1ParsePDF, Excel, CSV or scan
- 2Extract & mapClaude + chart of accounts
- 3ValidateDo the totals tie?
- 4ReviewHuman fixes what's flagged
- 5PushGoogle Sheets / ERP
Northwind Supply Co. is a made-up company. The three files use different units and formats but tie to each other, just like a real close package.
Files are processed in memory for this one request and never saved. This is a public demo, though, so please don't upload real confidential statements.
Measuring accuracy
Every sample comes with an answer key. Each run is scored field by field: exact values, plus the right account for each line. In a real engagement the answer key is 20–30 of your own statements, and every prompt or model change gets re-scored against it before it ships.
Why the checks aren't AI
The model reads the document, and plain arithmetic checks the result: assets = liabilities + equity, cash rolls forward, invoice lines add up to the total. If a number was misread, the totals stop tying, so the line gets flagged instead of slipping into your books.
Human in the loop
A line goes to review if the model's confidence is below 80%, it has no account, or it breaks a check. Nothing can be pushed until each flagged line is approved or corrected. Every run is logged with its timings, token counts and outcomes.