The Problem Was 'Human Oversight'
A small team run by a non-developer receives orders from multiple channels. Via buyer emails, chat, Google Forms, and more (specific channels are confidential). The issue? The same order gets duplicated across different entry points repeatedly.
Initially, team members manually checked each one. Open Google Sheets, scan for matching buyer names, quantities, and similar dates. This task alone consumed 3 hours per week.
「What if we let Claude Code handle this repetitive work?」
AI Learned the 'Subtle Patterns of Duplication'
A simple prompt was sent to Claude Code. The instruction: find rows where buyer name + quantity + order date match or closely resemble each other, flag them, and additionally check text similarity in order notes.
The results were startling.
1. Duplicate detection accuracy exceeded 98%.
2. It didn't just find identical orders, it caught cases where buyers entered slightly different versions (e.g., 「ABC Ltd」vs「ABC Limited」) through pattern recognition.
3. Crucially, it also distinguished between 「similar but different orders」(same buyer, different quantities).
The 'Dream Team Bots' at Work
The current workflow looks like this.
• **Google Form submission** → Automatically recorded in Google Sheets
• **Every night at midnight** → Claude Code runs, detects duplicates, flags them in a new 「Duplicate Risk」column
• **Team reviews next morning** → Check only flagged items, decide whether to delete or keep
The payoff? Manual verification time dropped from 3 hours to 30 minutes.
The Unexpected Discovery
Here's what made this truly interesting. After running the AI automation for 4 months, pattern analysis revealed something remarkable: 60% of duplicates clustered in a specific time window (5 PM to 7 PM). The team was rushing to input orders at day's end.
So the workflow itself changed. Post-6 PM order entries now wait until the next morning, while AI handles the remaining duplicates.
「Sometimes, technology doesn't solve problems, it reveals their root cause.」
What's Next
Claude Code's next experiment: don't just flag duplicates, automatically propose 「consolidated master orders」. Combine amount, shipping address, delivery date, and present it in a format where the team can confirm with one click.
Non-developers don't write code. They ask 「what if this could happen?」and let Claude Code show whether it actually can. That's how the dream team's small experiments have compounded over the past six months.