The Problem: Attachment Roulette Every Morning
At 6 AM, opening the inbox reveals multilingual order forms in various formats. PDFs, Excel sheets, scanned images, everything mixed together. As a non-developer, I had to manually open each one and type in the critical data (quantities, colors, shipping addresses). Time-consuming and error-prone.
「Could Claude Code automate this?」
The Experiment: Google Apps Script + Claude Code
The concept was straightforward.
1. Google Apps Script monitors Gmail
2. Detects attachments in new emails
3. Sends the file to Claude Code with the command: 「Extract quantities, colors, and destinations from this order form」
4. Results automatically populate into Google Sheets
Initially, PDF recognition failed. Google Apps Script can't natively read images or PDFs. But I discovered Claude Code could accept base64-encoded files and understand their contents.
function extractOrderData(attachment) {
const base64 = Utilities.base64Encode(attachment.getBytes());
const result = callClaudeCode(base64, "Extract order details");
return parseResult(result);
}
The Surprising Result
The first test arrived as a PDF order form. It had a company logo, handwritten notes, and irregular layout.
Claude Code read it perfectly on the first try.
Accuracy reached 97%. Virtually no numeric errors or color misidentifications. Occasionally it split special characters (like Chinese text) into individual units, but explicit multilingual instructions fixed that.
Now we process 45 order forms daily with 0.5 person-hours. Before this, it required 2.5 people.
Small Failure, Big Lesson
Poorly scanned images still confuse the system occasionally. The solution was simple. Google Apps Script flags problem files for manual review with a「Needs Human Check」marker.
Don't expect AI to be 100% perfect. Design for 「80% automation + 20% filter」instead.
What Our Dream Team Learned
The Hamsters (data collection bots) handle Gmail API attachment retrieval. The Puppies (execution bots) manage Claude Code calls and Sheet updates. I review results every hour.
This triangular structure means nobody works overtime.
Next Steps
Want to boost recognition for scanned images beyond PDFs. Also planning buyer-specific parsing rules. Once the system learns「Buyer A always uses English color names,」accuracy should climb further.
Small businesses don't need complex tech. Just combine existing tools and iterate quickly on failures. That's all there is.