What should I use for document AI if layouts change by vendor?

If the document layout changes by vendor, you usually want a hybrid Document AI approach, not a single rigid template-based parser.

Best options

1. Layout-aware OCR + extraction

Use a system that can read the page structure, not just plain text:

• detects blocks, tables, key-value pairs, headers
• handles different vendor formats better than fixed templates

Good when:

• invoices, forms, statements, purchase orders
• same business fields appear, but placement varies

2. Custom document classification + per-vendor models

If each vendor has enough differences, first classify the vendor/document type, then route to:

• vendor-specific extraction rules
• vendor-specific ML models

Good when:

• you have a moderate number of vendors
• layouts are stable per vendor, but different across vendors

3. LLM-based extraction with OCR/layout input

If formats are highly variable, use:

• OCR + document layout parsing
• an LLM to map content into a schema

Good when:

• you need flexibility
• documents are messy or semi-structured
• you can tolerate some validation and human review

4. Human-in-the-loop + confidence thresholds

For low-confidence fields:

• auto-extract when confident
• send uncertain documents to review

This is often the most practical way to handle vendor variation.

---

What to avoid

• Pure template matching if vendors frequently change layout
• Simple OCR text only if tables/positions matter
• One universal regex pipeline unless documents are very consistent

---

Recommended architecture

A strong setup is:

1. OCR / document layout parsing
2. Document classification by vendor/type
3. Field extraction using rules + ML/LLM
4. Validation
5. Human review for low confidence cases

---

If you want vendor-specific guidance:

• Google Document AI: good for OCR + processors + custom extraction
• AWS Textract: strong for forms/tables, good with post-processing
• Azure Document Intelligence: solid for forms and custom models
• LLM + OCR stack: best for highest variability

---

Practical rule of thumb

• Few vendors, stable per vendor → custom model per vendor
• Many vendors, changing layouts → layout-aware extraction + LLM/rules
• Very inconsistent docs → LLM + human review

If you want, I can suggest the best platform and architecture based on your document type, volume, and accuracy requirements.