The Problem Begins: Tracking Numbers Arrive, But Organization Is Manual
In B2B operations, dozens of tracking numbers arrive daily. The original process meant manually pasting them into Google Sheets, checking delivery status by hand, and coloring them with markers.
「We do the same thing every day. Can't AI handle this?」
That question became the starting point.
First Meeting with Claude Code: Beyond OCR
It started simply. Extract digits from PDF tracking number images. Claude Code received the image, digits appeared. Success, or so I thought.
But the next question was different.
「Can this automatically tell us if it's in transit or delivered?」
Claude Code proposed something interesting. Without needing a tracking API, we could use Conditional Formatting in spreadsheets to automatically color-code status based on number length, issue date patterns, and digit ranges.
「Is this... magic?」
Implementation: Not 30 Minutes, But 2 Hours
The plan was 30 minutes. Copy, paste code, done.
Reality differed.
First error: Windows environment API permission issues connecting to Google Sheets.
Second error: Conditional formatting rules were too complex, causing the spreadsheet to lag.
Third error (success): I told Claude Code 「This is too heavy, simplify it.」 AI reduced rules from 30 to 5. Speed returned.
After 2 hours, the tracking list transformed.
Red = Delivery scheduled
Yellow = In transit
Green = Delivered
Gray = Anomalous data
Unexpected Discovery: AI "Learned" from Mistakes
Something interesting happened.
Day 3: An unusual pattern appeared. Normally 13-15 digits, but one 14-digit number kept marking yellow (in transit). A human would think "probably an error."
Claude Code thought differently. Tracing that number revealed it was actually due to a shipping company system update delay. AI's rules broke human assumptions ("14 digits = error").
The next day I refined the automation rules. Three weeks later, coloring accuracy reached 96%.
Bigger Changes: Dream Team Bots Collaborating
What started as solo automation now expanded across multiple agents.
• First bot: Extract numbers
• Second bot: Predict delivery status and color
• Third bot: Flag anomalous data
• Fourth bot: Auto-generate weekly reports
Interestingly, these bots began "talking" to each other. One bot's flagged data gets received and re-verified by another, with results recorded in the spreadsheet. The non-developer team lead only checks results each morning.
Conclusion: Automation's Next Stage
Initially I thought 「just extract tracking numbers."
Now it's different.
Automation evolves through 3 stages.
Stage 1: Eliminate repetitive work (extract numbers)
Stage 2: Automate decision criteria (coloring)
Stage 3: Pattern learning and exception handling (96% accuracy)
What Claude Code combined with a Windows environment created wasn't just "time saved." It was the possibility of a small team operating like a large one.
Next experiment? Voice input for adding tracking numbers. A non-developer can definitely do it.