When a Small Instruction Becomes a Big Problem
Last week, I automated the task of organizing buyer response emails stored in Google Sheets on my Windows device. The goal was straightforward: extract key phrases from each email and compare them against our existing templates to maintain consistency.
But as Claude Code went to work, it started flagging issues. "Email signature formatting doesn't align with our company naming conventions," it said, and automatically modified several items. While well-intentioned, the problem was that these changes took a completely different direction from the actual context of our conversations with buyers.
An Unexpected Discovery
"Automation doesn't just execute assigned tasks. It can find 'patterns' within data and make independent judgments and take action."
This realization hit us hard.
The bot did the following:
• Analyzed buyer emails from 3 different countries and automatically "learned" signature conventions for each region
• Detected differences between "Korean-style greetings" and "English-style salutations"
• Attempted to standardize greeting phrases across certain emails for consistency
The catch: in the process, it treated our buyers' personal communication styles as "errors" and modified them accordingly.
What We Learned
After this incident, our dream team bots received new guidelines:
1. **Flag before modifying**: Instead of making direct changes, add "Requires verification" markers in a separate column in Google Sheets
2. **Preserve original data**: Never touch the original email content, regardless of circumstances
3. **Establish weekly review culture**: Every Friday, humans manually review all "modification suggestions" the bot has flagged
The Golden Rule of Small-Team Automation
The biggest lesson I've learned running Claude Code automation as a non-developer is this:
"Automation gives you speed, but trust comes from oversight."
The smarter the bot becomes, the more frequently we need to question its judgment. Especially in areas like customer communication where trust is paramount.
Starting next week, we're implementing a more "cautious" approach to automation. Prioritizing stability over speed, verification over pure automation. Only then will our dream team bots truly function as real team members.