At First, It Was Just Simple Copy and Paste
Running a B2B business, one of the daily tasks was finding shipping number information from incoming order notes and organizing it. Every message from buyers had a slightly different format. Sometimes the numbers appeared in parentheses, sometimes after a dash, sometimes mixed together without any marker.
Opening Google Sheets on the Windows device, finding them one by one and entering them. A 30-minute routine that repeated every day.
The Examples We Provided Became a Clue
When we first explained this task to Claude Code, we showed about 10 samples. We didn't explain the length of numbers, the arrangement of letters and digits, or the text patterns before and after. We simply said, "These are shipping numbers."
One week later, the bot started automatically filtering new format shipping numbers we had never explicitly shown it. By the third week, automatic extraction accuracy had reached approximately 95%.
"The bot didn't learn the rule, it learned the pattern." (AI Team Feedback)
The Most Magical Moment
One day, a buyer's message contained 5 shipping numbers mixed together. We typically expected only the first one to be automatically extracted, but Claude Code detected all of them. And it separated each into different items.
Even more surprising was that it started filtering out content that looked like shipping numbers but were actually order numbers or reference codes. We had never explicitly entered such a rule.
A Non-Developer's Observation
This experience taught us something important. AI agents don't need precise rules like code does. Given sufficient examples, context, and time, pattern recognition ability develops naturally.
The manual work that repeated on the Windows device is now almost completely automated. All we did was say "this is important" and provide examples.
Next Steps
Now the AI team is having the bot learn not just shipping numbers, but also the shipping status information around them. Based on this week's experiment results, the current pattern recognition accuracy has exceeded 90%.
Watching such small automation evolve in just 3 weeks, even as non-developers, we're confident we can turn AI into a true work partner.