Every operations team has a version of this problem.
You process the same type of document regularly — invoices, purchase orders, patient intake forms, insurance claims. Each one contains the same set of fields you need to extract: a date, a name, an amount, a reference number. And every time, someone is either manually copying those fields into a spreadsheet, or running a one-off conversion and hunting through the output for the data they need.
This works at low volume. At 50 documents a day, it becomes a full-time job. At 200 a day, you are hiring for a role that should not exist.
The extraction flow approach
An extraction flow is a saved configuration that tells PaperAI exactly what data to pull from a document and how to process it. Instead of configuring each conversion manually, you define the rules once and reuse them. This is the core of AI data extraction at scale.
Here is what a flow includes:
- A custom AI prompt tuned to the document type (generated by AI from your sample documents, then editable)
- Extraction field definitions — the specific data points you need from every document
- An AI model selection — which model to use and at what temperature
- Auto-approve rules — a confidence threshold above which documents skip human review entirely
The key difference between ad-hoc processing and flow-based processing is codified decision-making. When the rules are in the system, every document gets the same treatment regardless of who uploads it.
How the creation wizard works
Creating a flow takes about five minutes. PaperAI's wizard walks you through three steps.
Step 1: Upload samples. Drop in one to four sample documents of the same type. If you are building a flow for vendor invoices, upload four representative invoices — ideally from different vendors so the AI sees layout variation.
Step 2: AI analysis. PaperAI's AI examines your samples and generates a suggested configuration: a name, a description, a conversion prompt, and a list of extraction fields it detected in the documents. For an invoice, it might suggest fields like invoice_number, vendor_name, invoice_date, due_date, line_items, subtotal, tax, and total.
Step 3: Review and create. You review everything the AI suggested and edit as needed. Add fields the AI missed, remove ones you do not need, adjust the prompt, pick a model tier, set the auto-approve threshold. Then click Create.
After that, every document processed through this flow follows the same rules automatically.
Extraction fields
Each field has three properties:
- Label — A descriptive name like "Invoice Date" (auto-converted to a slug:
invoice_date) - Type — What kind of data to expect: string, number, date, currency, boolean, or array
- Required or optional — Whether the AI should flag the document if this field cannot be extracted
When a document is processed through a flow with extraction enabled, the results include a structured grid of field names and their extracted values. This is the data you actually need — not raw Markdown, but the specific data points your downstream systems consume.
Auto-approve with confidence thresholds
Every conversion produces a confidence score. When you enable auto-approve on a flow, documents scoring above your threshold (configurable from 50% to 100%) are automatically approved and skip the review queue.
This is how you scale from processing 50 documents a day to 500 without adding headcount. The AI handles the clear cases. Your team reviews only the edge cases where the AI is uncertain.
A practical starting point: set the threshold at 85-90%. Monitor the auto-approved documents for a week. If quality is consistent, you can raise it to 95%. If you are catching errors in auto-approved documents, lower it.
Auto-approve is available on Business plans and above.
Real example: accounts payable
A five-person AP team processing 400 invoices per month from 80 vendors.
Before flows: Each person uploads invoices individually, picks their own model, manually checks the output, and copies extracted data into a spreadsheet. Three people use different settings. Error rate is unknown because nobody measures the same things.
With a flow: The AP manager creates an "Invoice Processing" flow:
- Nine extraction fields (invoice number, vendor, dates, line items, amounts)
- Premium model (handles tables reliably)
- Auto-approve at 90% confidence
Now every team member uploads invoices to the same flow. Same fields extracted. Same model. Same quality bar. New hires produce the same output quality on day one. When someone goes on vacation, nothing changes.
The team went from 33 hours per month of manual processing to roughly 16 hours — and the output is measurably more consistent.
Flow limits
The number of flows you can create depends on your plan:
| Plan | Max Flows | |------|-----------| | Starter | 1 | | Pro | 5 | | Business | 25 | | Scale | Unlimited |
Start with your highest-volume document type. One well-configured flow is worth more than five rough ones.
Getting started
If you process the same document type more than 20 times per month, it should be a flow. The setup takes five minutes. The time savings compound every day.
Head to the Flows page in your dashboard, click New Flow, and upload a few samples. The AI does the heavy lifting.
For a deeper comparison of flow-based versus ad-hoc processing, see why rules-based automation beats ad-hoc processing.
Related resources
- AI data extraction — how PaperAI extracts structured fields from any document
- Invoice processing use cases — the most common extraction flow for finance teams