Let me tell you about the time we built something "genius" and it completely flopped.
The Feature We Were So Proud Of
In early January, we shipped what we called "Predictive Prompts."
The idea was clever: Ferni would analyze your conversation patterns and proactively suggest topics to discuss. Feeling stressed on Sundays? We'd detect that and offer to talk about your week ahead. Mentioned your mom being sick three times? We'd gently bring it up.
On paper, it was brilliant. "AI that anticipates your needs!"
We were so proud.
What Actually Happened
Within a week, we got feedback like:
"It feels like Ferni is reading my diary."
"I mentioned my divorce ONCE and now it keeps bringing it up. Please make it stop."
"This feels less like a friend and more like a stalker with a great memory."
Ouch.
Where We Went Wrong
We confused being helpful with being intrusive.
Here's the thing: when a human friend brings up something from your past, there's social context. They read the room. They sense whether you're open to that conversation.
Our "smart" feature had no such sense. It just... brought stuff up. Regardless of whether you wanted to go there.
Lesson learned: Memory isn't enough. You need emotional intelligence about when to use it.
The Fix
We didn't remove the feature entirely. We changed how it works:
-
Ferni now asks permission. "I remember you mentioned X. Want to talk about that, or is today about something else?"
-
Users control callback frequency. Some people want Ferni to bring up past topics. Others don't. Now it's their choice.
-
We added "Please don't mention this again." If something's painful and you've processed it, you can tell Ferni to let it go.
The Deeper Lesson
This failure taught us something crucial about building AI companions:
The goal isn't to be impressive. It's to be helpful.
Predictive prompts were impressive. They showed off our tech. They made demos look cool.
But they didn't make users feel cared for. They made users feel watched.
Now, every feature we build passes one test: "Does this serve the relationship, or just show off?"
If it's the latter, we cut it.
What We're Doing Differently
Since "Predictive Prompts," we've adopted a new rule: Ship shy, then listen.
Instead of building elaborate features and releasing them fully-formed, we:
- Build the simplest possible version
- Release it quietly
- Watch what users actually do
- Iterate based on behavior, not assumptions
It's slower. It's less dramatic. But it means we fail smaller and learn faster.
Why We're Sharing This
Most companies hide their failures. They show you the polished version and pretend the messy parts never happened.
We're not doing that.
Ferni is about honesty. About being real. If we expect users to share their messy stuff with us, the least we can do is share ours.
So: we messed up. We made something intrusive. We learned. We fixed it.
And now we're better.
This is part of our ongoing "Building Ferni" series. Next: How We Designed the Six Personas—the story of how Maya, Peter, Alex, Jordan, and Nayan came to be.