Use case · Documents and finance
A folder of scanned PDFs becomes one Excel sheet, every night
A scheduled workflow picks up every new file in the folder, sends it to an IntelliDocs stream described in one sentence ("delivery notes; extract supplier, delivery date, PO number, items with quantity") and appends the result to the sheet. Fields that fail validation are flagged in a review column and posted to a channel; the sheet is the source of truth by the next morning.
- Who it is for
- Operations and back-office teams with a scanner folder, an SFTP drop or a SharePoint library that fills with PDFs and images — delivery notes, intake forms, receipts, lab reports, archived contracts — and a spreadsheet someone fills in by hand.
- Systems involved
- SharePoint / OneDrive / Google Drive / SFTPIntelliDocsExcel 365 / Google SheetsTeams / Slack review
How the run goes
5 steps · logged per run
- Trigger01
Nightly, or new file
Folder, SharePoint, SFTP
- Document02
IntelliDocs extracts
Fields you described, handwriting included
- Agent03
Validate and normalize
Dates, decimals, duplicates
- System04
Row appended
Excel 365 / Google Sheets
- Person05
Review flagged rows
Only field that fails validations
How it runs, step by step
- 1
The trigger is a schedule or a file event on the folder. Files already processed are remembered by hash, so re-running never duplicates rows.
- 2
Each file goes to a Document Stream. The stream description is the whole configuration; when the forms change, the description changes. Photos from phones and skewed scans are handled by the OCR step; handwritten fields are read.
- 3
A validation step normalizes dates (dd/mm vs mm/dd by locale), decimal commas, currency symbols and reference formats, and recomputes any totals the document carries.
- 4
Rows are appended to Excel 365 or Google Sheets with the source file link, run ID and a confidence flag per field. Flagged rows are posted to a Teams or Slack channel for a quick look; corrected values are written back by the reviewer in the sheet.
- 5
For high volumes the sheet becomes a table in AI Workflows or a database, and Power BI or Looker reads from there.
Limits — what this does not do
- 20 pages per file and Latin scripts; long or non-Latin documents are split or excluded.
- A spreadsheet is a fine destination to 50,000 rows; past that, use a table or database step.
- Handwriting accuracy depends on the handwriting; the review column exists for a reason.
Questions about this process
OCR gives you text; this gives you fields. The model reads the page and returns the structure you described, including a confidence per field.
Yes, with the review flag doing the work on bad handwriting. We measure it on your forms in the sandbox.
No. It fills the spreadsheet or table your DMS doesn't; the files stay where they are, linked from each row.
Related use cases
Build this on your systems in a guided sandbox.
Tell us the process; after a 30-minute call the sandbox is ready from 24 hours, and we build the first run with your team.