We built the AI feature everyone asked for. Then we watched it make things worse.
When we first launched Murror, the number one feature request was always the same: "Can I talk to the AI?"
Users wanted a chatbot. Something they could vent to, ask for advice, get instant feedback from. Every competitor had one. Every investor asked why we didn't. So we built it.
The first month looked incredible. Session times tripled. Users came back more often. The engagement charts were exactly what you'd want to show in a board meeting.
Then we read the transcripts.
Users weren't processing their emotions — they were outsourcing them. Instead of sitting with a difficult feeling and working through it in their journal, they'd ask the AI to explain why they felt that way. The AI always had an answer. And that was the problem.
Real emotional growth isn't about getting the right answer — it's about learning to sit with the question. Our AI was short-circuiting the exact process that makes journaling valuable.
We didn't remove the feature entirely, but we redesigned it completely. Now the AI asks questions instead of giving answers. It mirrors what you wrote back to you in a different light. It never tells you what you're feeling or why.
The engagement numbers dropped back down. But the outcomes we actually care about — users reporting they feel understood, users resolving recurring worries, users feeling confident enough to have hard conversations in real life — those went up.
The lesson for us was uncomfortable: sometimes the feature your users are asking for is the one that will hurt them. And sometimes the hardest product decision is choosing to be less helpful in the moment so you can be more helpful over time.
Curious if other builders have faced this — where the data clearly supported a feature, but your gut told you it was doing more harm than good. How did you decide?
Replies
Murror
@edikan_peters You nailed the distinction between relief and growth. That's exactly what we saw. The users who stayed through the redesign started writing longer entries, revisiting old ones, and reporting real shifts in how they handled tough moments. The ones who left were often the ones using the chatbot to feel better in the moment without actually sitting with anything. We don't blame them — relief is a real need. But we realized Murror wasn't the right tool for that, and trying to be both was making us worse at the thing we actually do well.
This is a really strong product lesson. Engagement can make a feature look successful while the actual outcome gets worse. I like the decision to make the AI ask better questions instead of giving easy answers — sometimes the best AI feature is the one that gives the user more agency, not less.
Murror
@alpertayfurr Thank you, Alper! "Agency, not less" really captures it. We've started thinking about it as the difference between an AI that thinks for you vs. one that helps you think. The second one is harder to build and harder to sell, but it's the one that actually changes how people relate to their own emotions. Appreciate you reading!
@monatruong_murror That distinction is really strong. “Helps you think” is much harder to package because the value is quieter, but it probably builds deeper trust over time. Especially in emotional products, giving people room to process feels more important than giving them a quick answer.
Murror
@alpertayfurr Exactly — "quieter value" is such a good way to put it. We've actually started noticing this in how users describe Murror to others. They don't say "it gives great advice." They say "it helps me think through things." That kind of word-of-mouth is slower to build, but it's the kind that actually sticks. And you're right that trust compounds — users who stay through the quieter experience tend to come back for months, not weeks.
The transcripts insight is the most important part of this. Engagement metrics told one story, actual usage told another.
Same dynamic in a totally different domain — we build PeakAI, a B2B contact finder for India. Users kept asking for more contacts, more data fields, more enrichment. What they actually needed was fewer, better-verified ones. The 'more' version would have looked great on usage dashboards and been genuinely worse for the outcome they cared about (getting a reply).
I think the lesson generalises: when users ask for a feature, they're describing their symptom, not their cure. The AI chatbot in your case gave them the feeling of processing without the actual processing. More data in ours gives the feeling of more pipeline without the actual pipeline.
The hard part is that 'we removed the feature users asked for' is almost impossible to explain to investors. Respect for doing it anyway.
Murror
@rishav_rajak Love the PeakAI parallel. "Users are describing their symptom, not their cure" is such a precise way to put it. We've started calling this the "more trap" internally — more features, more data, more options all feel like progress but can actually dilute the core value. The investor conversation was genuinely tough. But we found that showing the qualitative shift (journal depth, emotional vocabulary, real-life behavior changes) made a stronger case than any engagement chart. Thanks for sharing your experience — it's reassuring to know this pattern shows up across such different domains.