The Trap of Automation: Translation Isn't Technology, It's Relationship
Last week, we added auto-translation for multilingual emails to Claude Code. The idea was simple. Since buyer inquiries arrived in a mix of English, Chinese, and Korean, we wanted to automatically organize and match the tone in replies.
"We were confident. Since Claude Code understands context and has learned reply patterns, we thought translation would be manageable too."
The First Signal: Buyer Replies Started Dropping
By day three, the data looked odd. Previously our 24-hour reply rate was 87%, but it suddenly fell to 62%. We even got three bouncebacks.
When we read the emails directly, the cause became clear. The auto-translated Chinese version we sent to English-speaking buyers felt "stiff and overly formal." Our typical tone is "warm but efficient," but Claude Code's translation read "like a legal contract drafted by the compliance team."
The Root Problem: Translation Is Trust, Not Vocabulary
As non-developers, we learned something critical:
Automation tools can convey information, but maintaining relationships is a different game. Especially in B2B trade, email isn't just an information channel; it's the medium where trust is built.
What could we do?
Step 1: Shift from "Auto-Translate Every Email" to "Localize Core Templates"
❌ Don't do: Auto-translate every single email
✅ Do: Manually translate only 5 core templates beforehand
(Greeting, Product Description, Shipping Info, Payment Terms, Thank You)
Step 2: Strengthen Claude Code's "Tone Learning"
We fed Claude Code 10 previous emails from each buyer so it could learn their preferred communication culture and pace. Only then did we let it generate reply drafts.
Step 3: Always End with "Human Review"
We took 5 minutes to review and approve each draft Claude Code generated. This kept automation's speed while protecting trust.
Three Weeks Later: Reply Rate Climbed Back to 91%
When we stopped auto-translating blindly and instead first understood each market's "communication culture" before designing automation, the system finally worked.
What We Learned
Automation fails not usually because of technology. It fails because the mission we assigned to the tool was wrong. It should have been "maintain trust, but use translation as a support tool," not "automate translation."
Claude Code is remarkably smart, but it can't read the feelings behind an email. That's our job.