How We Used Google Apps Script and Gemini AI to Parse PDFs and Automate Invoice Processing

Key Takeaways

• We built an AI-powered automation tool using Google Apps Script and Gemini 2.5 Flash-Lite that parses PDFs, validates data, and submits invoices automatically.

• The solution saved our client ~1 FTE per week and eliminated manual data entry errors.

• Choosing the right LLM matters:

  
  

  • U.S. models like GPT-4o and Gemini deliver excellent accuracy but come at a higher cost.

  • Chinese LLMs like DeepSeek, Qwen, and Yi-Lightning provide similar structured parsing at 10×–50× lower cost.

  
  

• Leveraging cost-efficient AI automation can give companies a significant competitive edge.

Introduction

Invoice processing used to be one of the biggest hidden productivity drains. Finance teams spent hours every week opening PDFs, copying invoice data into spreadsheets, validating line items, and manually submitting records into their ERP systems.

We recently helped a client replace this tedious process with an AI-powered PDF automation tool built using Google Apps Script and Gemini AI.

Results:

• 1 full-time equivalent (FTE) saved every week

• Zero manual entry errors

• PDF-to-submission in seconds

The Problem

Our client received hundreds of PDF invoices weekly. The finance team spent hours:

• Opening each PDF manually

• Copying vendor names, invoice totals, and line items into spreadsheets

• Validating against POs and vendor IDs

• Submitting invoices into their ERP system

This workflow was repetitive, error-prone, and expensive.

The Solution: AI-Powered Automation

We built a serverless workflow using Google Apps Script and Gemini 2.5 Flash-Lite that handles everything from PDF parsing to ERP submission.

1) Smart PDF Ingestion

Invoices are automatically collected from:

• Email inboxes

• Google Drive watch folders

• Cloud storage integrations

2) AI-Powered PDF Parsing

Using Gemini’s multimodal capability, the tool reads PDFs directly and returns structured JSON with:

• Invoice numbers and dates

• Vendor details

• Line items, tax, and totals

• Payment terms and PO numbers

3) Real-Time Validation

Before submission, the tool applies business rules to ensure data integrity:

• Vendor IDs are cross-checked

• PO totals match invoice totals

• Duplicates are detected automatically

4) Automatic ERP Submission

Once validated, invoices are pushed into the ERP system via API, reducing human touchpoints to exception handling only.

Choosing the Right AI Model: Cost vs Accuracy

When benchmarking LLMs for PDF parsing, we discovered significant cost differences:

Model	Country	Approx. Monthly Cost*	PDF Parsing	Notes
Gemini 2.5 Flash-Lite	USA	~$0.24	✅ Yes	Best balance of accuracy + cost
GPT-4o	USA	~$120.00	✅ Yes	Most accurate but expensive
GPT-4o-mini	USA	~$7.20	✅ Yes	Budget GPT option
Claude 3.5 Sonnet	USA	~$27.00	✅ Yes	Strong structured parsing
DeepSeek R1/V3	China	~$0.02	❌ Needs OCR	~50× cheaper
Qwen-VL-Max	China	~$0.62	✅ Yes	Strong multimodal alternative
Yi-Lightning	China	~$0.21	❌ Needs OCR	Lightweight and low cost

*Assumes 1,500 PDFs/month (~1,000 tokens each).

Insight: Chinese LLMs like DeepSeek and Qwen deliver comparable structured parsing for a fraction of the cost.

Sample Google Apps Script with Gemini AI

Here’s an example script showing how to parse PDFs and extract structured invoice data using Gemini 2.5 Flash-Lite:

/**
 * Parse invoices from PDF using Gemini 2.5 Flash-Lite.
 * Setup: Add a Script property GEMINI_API_KEY with your API key.
 */
function parseInvoicePdf() {
  const FILE_ID = 'YOUR_PDF_FILE_ID'; // Replace with your Drive file ID
  const API_KEY = PropertiesService.getScriptProperties().getProperty('GEMINI_API_KEY');
  if (!API_KEY) throw new Error('Set GEMINI_API_KEY in Script Properties.');

  // Read PDF and encode as Base64
  const blob = DriveApp.getFileById(FILE_ID).getBlob();
  const base64 = Utilities.base64Encode(blob.getBytes());

  // Ask Gemini to extract structured invoice data
  const prompt = `
    Extract invoice data from this PDF in JSON:
    {
      "invoice_number": string,
      "vendor": string,
      "date": string,
      "due_date": string|null,
      "po_number": string|null,
      "total": number,
      "currency": string|null,
      "line_items": [
        { "description": string, "qty": number|null, "unit_price": number|null, "amount": number|null }
      ],
      "notes": string|null
    }
    Return JSON only.
  `;

  const payload = {
    contents: [{
      role: 'user',
      parts: [
        { text: prompt },
        { inline_data: { mime_type: "application/pdf", data: base64 } }
      ]
    }]
  };

  const url = `https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-lite:generateContent?key=${API_KEY}`;
  const res = UrlFetchApp.fetch(url, {
    method: 'post',
    contentType: 'application/json',
    payload: JSON.stringify(payload),
    muteHttpExceptions: true
  });

  if (res.getResponseCode() >= 300) {
    throw new Error(`Gemini API error ${res.getResponseCode()}: ${res.getContentText()}`);
  }

  const body = JSON.parse(res.getContentText());
  const text = body.candidates?.[0]?.content?.parts?.[0]?.text || '{}';

  // Parse JSON output
  let parsed = {};
  try {
    parsed = JSON.parse(text);
  } catch (e) {
    Logger.log('Invalid JSON returned. Raw output:\n' + text);
    throw e;
  }

  Logger.log(JSON.stringify(parsed, null, 2));
}

Final Thoughts

Using Google Apps Script with Gemini AI allowed us to automate a traditionally manual process from start to finish — parsing PDFs, validating invoice data, and submitting to ERP — saving a full week of work per month.

And with the rise of cost-efficient Chinese LLMs, it’s clear the AI landscape is shifting fast. Companies that embrace automation early and choose models strategically will gain a lasting competitive advantage.

#AI #Gemini #GoogleAppsScript #Automation #PDFParsing #Invoices #WorkflowAutomation #DigitalTransformation