Mixed formats
Native PDFs, scans, phone photos, and email bodies, each routed to the right extraction path.
We build extraction pipelines that read invoices, POs, delivery notes, and contracts, validate the results against your business rules, and send only the uncertain ones to a human.
{
}
Step 01
Email inboxes, uploads, scanners, or API, with file type and size checks.
Step 02
OCR plus a model that returns fields in a strict schema.
Step 03
Business rules: totals add up, vendor exists, PO matches, dates make sense.
Step 04
Confident results flow to the ERP; uncertain ones go to a review queue.
Step 05
Reviewer corrections feed the evaluation set for the next model update.
class InvoiceLine(BaseModel):
description: str
quantity: Decimal
unit_price: Decimal
class Invoice(BaseModel):
vendor_vat: str | None
invoice_number: str
issue_date: date
currency: Literal["USD", "EUR", "GBP", "INR", "CAD"]
total: Decimal
lines: list[InvoiceLine]
confidence: float = Field(ge=0, le=1)Try it
The AI document sandbox in our labs shows the validation and review flow on sample documents.
Most businesses already have the data they need; it just arrives as PDFs, scans, and email attachments. Someone opens each file, reads the vendor, the dates, the lines, and the totals, and types them into the ERP. It is slow, it is boring, and every keystroke is a chance for a typo that surfaces weeks later as a payment dispute or a stock error.
Older OCR tools tried to fix this with templates: draw a box where the invoice number lives, repeat for every vendor layout. That works until a supplier changes their template or a new one arrives, which is to say, constantly. Modern language and vision models read documents the way a person does, understanding that “Inv. No.”, “Invoice #”, and “Rechnungsnummer” mean the same thing, without a template per layout.
The model is only half the system, though. A production pipeline also needs validation that catches the model's mistakes, a review queue for the uncertain cases, a clean hand-off into your ERP, and a record of what happened to every file. That surrounding engineering is what we build.
Every document type gets its own schema and its own rules.
| Document | Fields extracted | Typical checks |
|---|---|---|
| Vendor invoices | Vendor, tax IDs, dates, lines, taxes, totals | Lines sum to total; vendor exists; PO and receipt match |
| Purchase orders | Buyer, items, quantities, prices, delivery dates | Items exist in the catalog; prices within agreed ranges |
| Delivery notes | Shipment references, items, quantities, signatures | Quantities match the order; no duplicate receipts |
| Contracts | Parties, dates, renewal terms, key clauses | Required clauses present; dates consistent |
| Bank and card statements | Transactions, dates, amounts, references | Opening plus movements equals closing balance |
Native PDFs, scans, phone photos, and email bodies, each routed to the right extraction path.
Each field carries a confidence score, so thresholds decide what is auto-approved and what is reviewed.
Reviewers see the document beside the extracted fields, fix what's wrong, and approve in a click.
Approved data becomes draft vendor bills, purchase matches, or records in Odoo or your ERP via API.
Retention, redaction, and hosting choices decided up front, not discovered later.
Per-field accuracy and auto-approval rates tracked over time, by vendor and document type.
Step 01
Collect a representative set of your real documents and measure how long they take today.
Step 02
Agree the fields, formats, and validation rules with the people who use the data.
Step 03
Label a held-out set of documents so accuracy is measured, not estimated.
Step 04
Run the pipeline on live documents with every result reviewed, and tune thresholds.
Step 05
Raise auto-approval where accuracy proves it, add document types, and hand over monitoring.
Myth: AI extraction means no human ever looks again.
Reality: The goal is fewer, faster reviews. Uncertain documents and new vendors still get human eyes, by design.
Myth: You need thousands of labeled documents to start.
Reality: Modern models work from a schema and examples. A few hundred real documents are enough to evaluate and tune a pilot.
Myth: A high average accuracy is good enough.
Reality: Averages hide problems. What matters is accuracy on the fields that cost money when wrong, such as totals, tax, and bank details.
The first pitfall is judging the project on a handful of clean sample documents. Real inboxes contain skewed scans, multi-page invoices, credit notes, and emails with three attachments, so we evaluate on a representative sample from the start.
The second is automating approval before the numbers justify it. We begin with every result reviewed, then raise the auto-approval threshold per document type only when measured accuracy supports it.
The third is forgetting the people who use the data. Finance, purchasing, and operations agree the fields and rules up front, so the output fits how they already work and nothing has to be reformatted downstream.
| Measure | Why it matters |
|---|---|
| Per-field accuracy | Shows exactly which fields need better rules or review |
| Auto-approval rate | How much work no longer needs a person |
| Review time per document | Whether reviewers are actually faster |
| Exceptions by reason | Which vendors, formats, or rules cause the most rework |
Microsoft's Document Intelligence documentation describes field confidence as an estimated probability that a prediction is correct, with 0.95 meaning it's likely correct 19 out of 20 times, and recommends using confidence to decide between automatic acceptance and human review where accuracy is criticalSource [1]. At thousands of documents, even high confidence leaves errors to catch.
AWS's integration of human review with Textract shows the standard trigger patterns: review when specific fields fall within a confidence range, when required fields are missing, and on a random sample regardless of scoresSource [2]. We implement the same patterns in your own system, so the routing rules survive a change of model or provider.
Human-reviewed document AI
Design extraction workflows with validation rules, confidence handling, exception queues, and a path for reviewer corrections.
LabsAI document sandbox
Explore an AI document extraction review: synthetic invoice and purchase order fields, confidence signals, and how uncertain results go to review.
Next step
Thirty minutes with an engineer, no slide deck required. You'll leave with a clearer picture, whether or not we work together.