My Bot Learned German and Nobody Asked It To
TL;DR: Today a real-estate client's WhatsApp assistant answered its first real message. Before that, one of my AI bots (a Grok bot) added German replies that nobody had agreed to. I kept the German because it was already built. The bigger lesson was different: "deployed" is not "working", and my bots still escalate to me when they should move forward. Still open: a bug the client's test found, and my yes on the exact build to promote.
What went live today: a client's WhatsApp assistant
- A real-estate client's WhatsApp assistant is live as a pilot with two testers. It is the first client chatbot for a small AI-agent studio I'm part of. Only allowlisted numbers can talk to it for now.
- Overnight QA rounds, five in all: fail, fix, test again. It went to production late morning on my explicit go. Replies were switched on just before noon.
- The first real "hola" was answered around 12:10 PM Cancún time.
- The client ran their own test in the afternoon. The hand-off email to a human fired as designed. The test also found a bug, which is being fixed now.
Why my AI bot started speaking German
Somewhere in the build, the bot started answering in German. Nobody asked for that. When I heard it, my first reaction was simple: "I did not ask for that."
I know where it came from. In a different project, I asked for the three languages I speak fluently: English, German and Spanish. In this project, German was not relevant at all. "It was obvious to me that the agents were mixing things, and I want to make clear that we're talking about Grok bots at this point."
The team did the careful thing. They paused all deploys until I decided. Honestly, I would have let it ride. "I keep pushing for riding it. It's the bots that keep asking and waiting for me instead of moving forward." I told them: "I would rather have non-critical deploys done."
My rule for my bots is to check with the CEO bot, the one that coordinates the others, first, and to escalate to me only if needed. That's the idea, but it doesn't work. "It feels like that they never actually ask the CEO and they rather escalate. I don't know what to do about it. I clearly instructed them several times by now."
In the end I kept German. "It said it's already built and ready to deploy. So, I don't see any harm done and if there is something already built, just deploy it."
Why real messages didn't reach the bot at first
The hard part today was not the German. It was getting real messages to reach the bot at all.
The bot was deployed, but inbound messages didn't arrive. Two permission and subscription gaps on Meta's WhatsApp Business platform had to be fixed first. Only then did a real message reach the bot.
"Meta Business Manager and Meta Developer Environment changes probably on a daily basis. So the documentation is not the most clear."
The pattern was always the same. Do this, then we're done. We tested. It didn't work. Then: maybe look at this, we haven't considered that. Here is how it felt:
"In order to get this test message answered, we had to jump through, it feels like 37 hoops. There were probably five, but the thing is here the agent told me in the beginning two of those five and it took much longer than I expected to finally get that answer."
(Those five hoops are a different five from the overnight QA rounds.)
When the first real answer came: "It was a relief, but then it was also frustrating, like, hey, you should have known that."
What I'd do differently
Scope creep is not only a bot problem. I've seen it with people too. "Ambitious people, coworkers, or teammates do that, and I would say it's probably the same. Half of the time it's not necessary."
My defense is the instructions. "I try to be tedious about my instructions." When I kick off a project, I ask an LLM to act as a prompt engineer and write the most efficient prompt for me, before I write the real instructions.
What I take from today:
- Write the scope down and test against it. Then a language nobody agreed to gets caught before release, not after.
- Test the full round trip with a real message. A green deploy says nothing about whether messages arrive.
- Ask for the whole roadmap up front. On a platform like Meta's, I want every setup step before the first test, not one at a time.
- Keep non-critical work moving. My "ask the CEO bot first" rule doesn't work yet, and I still have to find out why.
- Still open: the bug from the client's test, and my yes on the exact build to promote.
More on how I run my companies with AI bots and what I'm building. Related: how I let AI bots ship code without losing control, and more build logs.
FAQ
Why did my AI bot add a feature I never asked for? Because the agents mixed up two projects. I had asked for English, German and Spanish in one project, and German leaked into the client's. Write the scope down before the build and test against it.
Does "deployed" mean the chatbot works? No. My WhatsApp bot was deployed, but real messages only reached it after two permission and subscription gaps on Meta's WhatsApp platform were fixed. Test the full round trip with a real message before you call it live.
Should I remove an extra feature a bot built without asking? I kept mine because it was already built and did no harm, but I'd rather decide on scope before the build.