Enterprise Document Extraction Pipeline with Reducto, Textract, and AACFlow

Claude Vision bridges the gap. When Textract returns a jumbled mess and Reducto cannot parse the structure, Claude can look at the document image and reason about what it means. It understands context: "this column is probably VAT because the header says НДС and the numbers sum to 20% of the subtotal." It also handles multilingual documents without special configuration.

The correct architecture is not a single tool — it is a routing layer that sends each document to the right tool based on detected characteristics.

The AACFlow Pipeline Architecture

In AACFlow, you build this as a sequential workflow with conditional branches. Here is the structure of the invoice processing pipeline:

Stage 1: Ingest and classify. A file upload trigger receives the document. A Claude Vision block analyzes the first page: Is it machine-generated or scanned? What language? Does it contain tables? What document type does it appear to be? This classification takes roughly one second and costs fractions of a cent.

Stage 2: Primary extraction. Based on classification:

• Machine-generated PDF with known structure → Reducto structured extraction
• Scanned document with good quality → AWS Textract with table detection enabled
• Poor quality scan or handwritten content → Claude Vision with a detailed extraction prompt
• Mixed content (typed form with handwritten fields) → Textract for typed sections, Claude Vision for handwritten fields

Stage 3: Validation. The extracted data goes through a validation block. This is a function block with custom JavaScript that checks business rules: Does the invoice total equal line items plus VAT? Is the date within a reasonable range? Are all required fields present? Does the supplier tax ID format match the expected pattern for the country?

Stage 4: AI enrichment. A Claude Sonnet block takes the validated extraction and enriches it. For invoice processing: normalize the supplier name against your vendor master data, classify the expense category, flag anything unusual (a line item description that does not match the PO, a total that differs from the quoted amount by more than 2%).

Stage 5: Storage. Validated, enriched data goes to two places: the structured fields into your database via a PostgreSQL block, and the original document plus extraction result into the AACFlow knowledge base for future reference and audit.

Building the Invoice Extraction Workflow

Here is a concrete example of the Reducto extraction configuration for invoice processing:

Reducto returns this structure populated from the document. For a clean, machine-generated invoice from a major ERP system, you get 100% field coverage in under two seconds. The extraction is deterministic — run the same document ten times, get the same result ten times.

For the Textract path, AACFlow calls the AWS Textract block with AnalyzeDocument using the FORMS and TABLES feature types. The raw Textract response is then passed to a function block that maps the key-value pairs and table rows to your target schema. This mapping layer is where you encode supplier-specific rules: "For invoices from Supplier X, the invoice number is always in the top-right cell of the first table, not in the form fields."

Handling Edge Cases That Break Naive Pipelines

Low-quality scans. When Textract confidence scores average below 70%, AACFlow routes to a different path: first, an image preprocessing step (contrast enhancement, deskewing via a Python function block), then re-extraction. If confidence is still low after preprocessing, the document goes to a human review queue — a Slack notification with the document and a form for manual field entry.

Multilingual documents. Documents that mix languages are common in international trade. A supplier in Germany might send an invoice with the header in German, the line items in English (product codes and descriptions), and the payment terms referencing Russian bank details. Textract handles multi-language detection automatically. For Reducto, you specify the expected language in the extraction config. For the AI enrichment step, Claude handles this natively — no special configuration needed.

Template variations. Even the same supplier may change their invoice template after an ERP upgrade. The AACFlow workflow handles this by storing validated extraction results in the knowledge base. When an invoice fails validation (a field in an unexpected location), the system checks whether recent successful invoices from the same supplier can inform a schema update. If there is a pattern, it proposes a template update.

Duplicate detection. Invoices get submitted twice. The AACFlow workflow computes a fuzzy hash of key fields (supplier tax ID + invoice number + total amount + date) and checks against recently processed documents. A match with high similarity triggers a duplicate warning rather than a full rejection — because sometimes a corrected invoice replaces an original and shares most fields.

Results: 97% Field Accuracy on Real Invoice Data

The pipeline described above, running on a corpus of 12,000 invoices from 340 different suppliers across 8 countries, achieves 97.3% field-level accuracy on the 50 fields we track. The breakdown:

• Machine-generated PDFs (60% of volume): 99.8% accuracy via Reducto
• Good-quality scans (28% of volume): 96.1% accuracy via Textract + mapping
• Poor-quality scans routed through Claude Vision (12% of volume): 91.4% accuracy, with 4% going to human review

The human review queue processes roughly 480 documents per month. Without the pipeline, the team was manually processing 12,000 documents per month. That is a 96% reduction in manual work.

Processing cost per document: approximately $0.04, broken down as $0.01 for Textract or Reducto extraction, $0.02 for Claude enrichment, and $0.01 for AACFlow execution overhead. At 12,000 documents per month, that is $480 in processing costs, compared to the labor cost of manual processing.

Setting Up in AACFlow

The complete workflow is available as a template in the AACFlow template library under "Enterprise Document Processing." You can deploy it in under ten minutes:

1. Connect your AWS account for Textract access
2. Add your Reducto API key
3. Add your Anthropic API key for Claude Vision and enrichment
4. Configure your target schema (the fields you want to extract)
5. Set up your validation rules in the function block
6. Connect your PostgreSQL database or configure knowledge base storage

The workflow includes pre-built error handling, the human review routing, and a monitoring dashboard that tracks extraction accuracy per supplier and document type over time.

Document extraction is not a solved problem. But with the right tool for each document type, a validation layer that catches errors before they enter your systems, and an orchestration platform that ties it all together, you can get to production-grade accuracy without building custom infrastructure from scratch.

— Alexandr Chibilyaev