"Which AI model should I use?"
It's the most common question we get at PaperAI. And the honest answer is: it depends on what you're processing.
There is no single best model. Anyone who tells you otherwise is either selling you something or hasn't tested enough document types. The right model depends on document quality, layout complexity, handwriting vs. typed text, and how much you're willing to spend per page.
Here's how to think about it.
The two tiers
PaperAI offers 5 AI models via Azure OpenAI — GPT-4o Mini, Mistral Document AI, GPT-5.4 Mini, GPT-4o, and GPT-5 Chat. They fall into two practical tiers based on cost and capability.
Standard models (2-5 credits/page)
These are your workhorses for clean, predictable documents. The Standard tier includes GPT-4o Mini, Mistral Document AI, and GPT-5.4 Mini.
Use them for:
- Typed invoices with standard layouts
- Receipts from major POS systems
- Digital-native PDFs (not scanned)
- Forms with clearly printed text and consistent structure
- Scanned invoices (typical office scanner quality)
- Multi-page contracts with varied formatting
- Documents with headers, footers, tables, and body text mixed together
- Forms that have been filled in with a mix of typed and selected/checked fields
Standard models handle clean documents with 95%+ accuracy on most fields. They're fast, affordable, and perfectly adequate when the source material is reasonably clean. They're also good at understanding document structure — they can parse a table even when the grid lines are faint, handle rotated pages and skewed scans, and deal with stamps, signatures overlapping text, and coffee stains.
For most business document processing, standard models hit the sweet spot of cost and accuracy.
Where they struggle:
- Handwritten notes and forms
- Historical documents with faded or damaged text
- Complex multi-column layouts (academic papers, newspapers)
- Documents in multiple languages on the same page
- Low-quality scans (under 200 DPI, poor contrast)
Premium models (8-10 credits/page)
Reserve these for the hard stuff. The Premium tier includes GPT-4o and GPT-5 Chat.
Use them for:
- Handwritten notes and forms
- Historical documents with faded or damaged text
- Complex multi-column layouts (academic papers, newspapers)
- Documents in multiple languages on the same page
- Low-quality scans (under 200 DPI, poor contrast)
Premium models are multimodal. They don't just run OCR and then parse text. They see the entire page as an image and reason about what they're looking at. This matters enormously for handwriting, where traditional character recognition fails but visual understanding succeeds.
The accuracy improvement over standard models on difficult documents can be 20-40 percentage points. That's the difference between usable output and garbage.
The decision framework
Here's how to pick the right model in practice.
Step 1: Categorize your documents
Before you touch a model selector, sort your documents into groups:
- Group A: Clean, typed, consistent layout. You get the same invoice format from the same vendor every month.
- Group B: Typed but messy. Different layouts, varying scan quality, occasional handwritten annotations.
- Group C: Handwritten, damaged, complex, or unusual.
Most organizations find that 60-70% of their documents fall into Group A, 20-30% into Group B, and 5-10% into Group C.
Step 2: Start with standard models
Run a test batch of 20-30 documents from each group through a standard model. Review the results field by field.
For Group A documents, you'll likely see 95%+ accuracy. Done. Keep using standard models for these.
For Group B, check the fields that matter most. If the vendor name, date, and total amount are correct but a line item description is slightly off, that might be acceptable. Define your accuracy threshold before you test — not after.
For Group C, standard models will probably produce poor results. That's expected. Move to Step 3.
Step 3: Upgrade selectively
Only move to a more expensive model for the document groups where the standard model falls short. This is the key insight that saves money.
Run the same Group B and Group C test batches through a premium model. Compare the results. Only use premium models where the accuracy improvement justifies the extra credits.
Step 4: Set up per-type routing
This is where PaperAI's Smart Flows become important.
A Smart Flow lets you define different processing configurations for different document types. You can set:
- Model selection per document category
- Extraction fields specific to that document type
- Confidence thresholds for auto-approval
- Review requirements
So your clean vendor invoices run through a standard model at 2-5 credits/page, and your handwritten inspection notes get a premium model at 8-10 credits/page.
You're not overpaying for easy documents. You're not underpaying for hard ones.
Common mistakes
Using the most expensive model for everything. We see this constantly. Someone tests a premium model, gets great results, and uses it for all documents — including the clean typed invoices that a standard model handles perfectly. They burn through credits faster than necessary.
Judging accuracy on too few test documents. Twenty documents is a minimum. If your documents have high variance (different vendors, different formats), test at least 50. Accuracy on 5 cherry-picked samples tells you nothing.
Ignoring confidence scores. Every field PaperAI extracts comes with a confidence score. A field extracted at 98% confidence almost never needs review. A field at 72% usually does. Use these scores to set smart review thresholds instead of reviewing every single field.
Not re-evaluating. Models improve. A standard model that struggled with a document type six months ago might handle it fine today. Re-run your test batches every quarter with newer models. You might be able to downgrade some of your processing and save credits.
A practical example
A property management company processes three document types:
- Lease agreements (typed, consistent template): Standard model, 3 credits/page. 200 pages/month = 600 credits.
- Maintenance invoices (various vendors, scanned): Standard model, 5 credits/page. 150 pages/month = 750 credits.
- Inspection reports (handwritten notes + photos): Premium model, 9 credits/page. 40 pages/month = 360 credits.
Total: 1,710 credits/month.
If they used the premium model for everything: 390 pages x 9 credits = 3,510 credits/month. That's roughly double the cost for no meaningful accuracy improvement on 90% of their documents. See pricing plans for how credits map to each plan tier.
The bottom line
Cost-efficiency in document processing isn't about finding the cheapest model. It's about matching the right model to each document type.
Start cheap. Measure accuracy. Upgrade only where the numbers justify it. Use Smart Flows to automate the routing.
That's the framework. It's not complicated, but it requires a bit of upfront testing. The payoff is a processing pipeline that's both accurate and economical.
Need help figuring out the right model mix for your documents? Reach out at hello@paperaiapp.com.
Related resources
- Features overview — explore all 5 AI models and processing capabilities
- Pricing plans — understand how credits map to different model tiers
- Why you need more than one AI model — the strategic case for multi-model processing
- Credit-based pricing for document AI — how model choice affects cost