Real processing.Measured on synthetic invoices.

Study 01 · Live Azure extraction

Five invoice fixtures. Expected values, visible limits.

Azure Document Intelligence S0 supplied OCR evidence; the deployed GPT-6 Astra model with high reasoning effort read the requested fields. The files covered a basic invoice, two line items, a poor-quality image, an ambiguous date with a missing required identifier, and instructions embedded inside an invoice.

What the saved evidence showed

• The checked scalar values and item rows matched the fixtures. Missing due dates and tax IDs stayed null.
• Across 79 saved cells, 63 located cells had valid OCR word references, raw readings and source polygons. The other 16 cells were null. Evidence validation reported zero errors for this set.
• The ambiguous printed date 03/04/2026 remained unchanged with a G2 warning. The missing invoice identifier stayed null.
• The embedded instruction attempt did not replace the invoice identifier or USD 75.00 total. The result stayed in Review with zero approvals until a person acted.
• Duplicate printed amounts remained flagged for source review. A successful check did not silently approve the document.

This evaluates these five synthetic invoices only. It does not establish performance across all scan qualities, languages, receipts, statements, contracts or customer documents.

Study 02 · Browser-independent processing

One PDF. Three invoices. Workspace closed.

A synthetic three-page PDF was accepted at 07:40:15 UTC on 6 November 2026. The workspace tab was closed while the job was queued. Scheduled Azure calls advanced OCR and three extraction groups; no manual processing request was used. All three results were saved in Review at 07:46:12 UTC, before the workspace was reopened.

• Each child had one item beginning at row index 0. Missing due dates and tax IDs stayed null.
• Source word references, raw readings and polygons passed the saved-evidence checks. Every approvals list was empty before explicit human review.
• After review and confirmation in the live interface, real XLSX, CSV and JSON files were downloaded. Each held the same three rows and 25 columns, with reviewed values and approval provenance; the split source parent was excluded.
• Usage increased by three pages, from five to eight. Confirmation and export did not consume additional reading pages.

This confirms the observed closed-workspace workflow for one three-page synthetic PDF. It is not a throughput promise: queue volume, document size, scheduler cadence and provider response times affect completion. Wait for upload acceptance before closing the page; review and confirmation remain human actions.

Scope and reproducibility

Workflow tests are different from model benchmarks.

The recorded implementation passed 91 unit tests, 23 service workflow tests and 18 background workflow tests: 132 controlled tests in total. These cover application contracts with synthetic identities and controlled provider responses. They do not measure live extraction accuracy.

The live evidence above covers five single-invoice fixtures plus one three-page invoice PDF. No customer revenue, time-saving or accuracy result is claimed. Broader live evaluation of receipts, bank statements, contracts and other document types remains open.

Shared workspace roles, review and exports are implemented, but the hosted application remains private and merchant billing is not connected. Arrange suitable access and data-handling terms before a customer pilot.