Document Intelligence: read receipts, invoices and IDs safely

Concepts

Document Intelligence turns a file (PDF, photo or scan) into checked values. It reads the cheap way first (the PDF text layer), asks a vision model only when it must, checks the answer with fixed rules, and sends doubtful documents to a person. It never posts a business transaction on its own: it hands values on as a draft, a workflow or a callback.

Piece	What it does
Flow	A versioned pipeline you design in Studio: Document in, Read text, Classify, Extract fields, Validate, Human review, Output
Preset	A ready-made document type: RECEIPT, INVOICE, ID_DOCUMENT, BUSINESS_CARD
Run	One document read through a flow, with values, a confidence per field and a list of issues
Review	The original beside the values; a person corrects, approves or rejects

Design a flow

1. Open Studio, AI Studio, Document Intelligence.
2. Choose New flow; the starter pipeline reads a receipt.
3. Click a step to set it up. In Extract fields name the fields, their type and the ERP field each maps to. In Validate add rules (required, pattern, range, totals add up, date not in the future, not expired, duplicate check).
4. Set the flow settings: review threshold, largest file, most pages, how long runs are kept, learn from corrections.
5. Save draft, Test with a file, then Publish.

Use it

From the API (document.read is the same call as a step in the Integration Designer):

POST /api/v1/ai/documents/runs
{ "preset": "RECEIPT", "fileName": "taxi.jpg", "mediaType": "image/jpeg", "data": "<base64>" }

The answer holds status, needs_review, result_json.values, confidence_json and issues_json. Documents with NEEDS_REVIEW wait in Runs & review.

Limits and safety

• Each tenant uses its own AI profile document-intelligence (falls back to default). With no model set up, people see "not set up" and type the details instead.
• A file read before is answered from a cache of its SHA-256.
• Permission resource documentintelligence: view, run, review, configure.