Optimizing PDF/Image Analysis for Visual Compliance using AI Vision

Help me Build my Workflow
,
1

Hi everyone,

I’m looking for the best way to build an n8n workflow that manages the quality control and compliance of legal documents. The main goal is to analyze documents (PDFs or images) page-by-page to verify the visual presence of mandatory elements (signatures, dates, checkboxes).

The Desired Workflow

  1. Trigger: New document uploaded to an S3 bucket.

  2. Step 1: Pre-Analysis (OCR/AI Vision):

    • I need a solution to analyze the input.

    • It must extract data and, crucially, perform a visual recognition check (e.g., “Is there ink in this area? Is there a checkmark in this box?”, “Is there a signature, even a digital one, in this rectangle?”).

  3. Step 2: Quality & Compliance Check:

    • Verify that every page is legible (visual quality control).

    • For each page, check the presence of required elements (Ink/Digital Signatures, Handwritten Location/Date, Completed Checkboxes).

    • The result should be a structured JSON object for each page (similar to the output of an AI Assistant).

  4. Step 3: Final Action:

    • If the result is “Conforming,” move the file.

    • If the result is “Non-Conforming,” send an email notification with the detailed JSON output.

My Key Questions (Seeking Advice)

  1. Visual Recognition Service: What is the best node or service to integrate for the visual recognition of signatures/graphic marks?

  2. Feasibility: Do you think this overall process is truly feasible and robust within an n8n environment?

  3. PDF Handling: What is the most efficient way to handle a multi-page PDF input and analyze it page-by-page within the n8n workflow?

Thanks in advance for any suggestions or examples!

2

If I was doing something like this I would possibly steer away from using AI and rely on parsing to avoid sending confidential documents to an AI.

With AI however you could use the OpenAI analyze image node, with a structured prompt to find handwriting or specifically analyse where a signature box is supposed to be.

You could also potentially use the information extractor node and define it to try and extract text out of the boxes - it should still be able to extract legible handwriting from a flat PDF document. Illegible handwriting may not be able to be extracted and then you have whether it can be read or not.

The extract PDF node is also your best bet with the added option of separate pages and JSON output. This will enable you to split pages at this node then push it out with a split node. Then use code nodes to determine legibility.

I have built something similar, but to extract key information in a compliant field without any AI involvement.

Some signed PDF documents also have metadata you can extract at the extract node - so this might give you a check to see if a document has actually been signed. The legibility part might be hit or miss, with or without AI. Probably worth testing these nodes with your documents first to see what you get out.