We stopped trying to make our AI smarter. Here's why.
Six months ago, our team was obsessed with making Murror's AI more intelligent. Better pattern recognition, deeper emotional analysis, more insightful reflections. Every sprint, we'd ship something that made the AI sound smarter.
And our users started disengaging.
At first we thought it was a product issue — maybe the onboarding was off, maybe the prompts were wrong. But when we dug into the data, we found something we didn't expect: the smarter the AI got, the less people felt like they were doing the work themselves.
Here's the problem. When someone journals about a difficult relationship and the AI immediately identifies the pattern, names the emotion, and suggests a reframe — it feels impressive. But it also short-circuits the process. The user didn't arrive at that insight. The AI did. And insights you're handed don't stick the way insights you discover do.
So we did something counterintuitive: we made the AI dumber. Not literally — but we deliberately added restraint. Instead of analyzing, it asks. Instead of naming the pattern, it helps you circle it. Instead of suggesting what you might be feeling, it creates space for you to sit with the discomfort long enough to figure it out.
The results surprised us:
Users who reached insights through guided questions reported 3x higher satisfaction than those who received direct analysis
Journal entries got longer and more exploratory — people were thinking more deeply, not just consuming AI output
The "aha moment" rate (our internal metric for when a user explicitly names a new understanding) went up 40%
Word-of-mouth referrals increased because people felt ownership over their breakthroughs, not like they were just reading an AI's assessment
The lesson changed how we think about AI product design. In a world where every AI company is racing to be the smartest, we're learning that the best AI companion isn't the one with the best answers. It's the one that asks the best questions.
This is hard to build and even harder to sell. "Our AI is deliberately less impressive" is not exactly a pitch deck headline. But we're finding that the products people love most aren't the ones that perform for them — they're the ones that help them perform for themselves.
Anyone else building AI products where restraint is the feature, not the limitation?
Replies
We hit the same wall with Composa. Smarter outputs weren’t the problem — the problem was that users didn’t trust outputs they couldn’t feel. The shift that worked: less analysis, more recognition. People don’t want to be told something new about themselves. They want to hear something they already half-knew, said precisely.
Murror
@dani_mashael "They want to hear something they already half-knew, said precisely" -- that's a beautiful way to put it. Recognition over revelation. We found the same thing: when the AI names what someone is already sensing but can't articulate, it lands completely differently than when it introduces a new framework. Trust comes from feeling understood, not educated.
Murror
@dani_mashael Reactance is exactly the right framework. We see that pattern constantly -- the moment the AI gets ahead of the user's own understanding, they disengage or push back. "A mirror with better vocabulary, not a therapist with a diagnosis" is such a precise way to put it. We use a similar framing internally: the AI should be a half-step ahead in articulation, never in insight. The insight has to remain the user's. Sounds like Composa is navigating the same edge.
Finally, someone gets it. Smarter AI doesn't always mean better — sometimes it just means the user checks out. Asking good questions > giving good answers. Love this approach.
Murror
@adam_idress Exactly. We kept measuring "AI quality" by how impressive the output was. But impressive and useful turned out to be very different things. The moment we reframed our metric from "how good is the AI's answer" to "did the user have a breakthrough," everything changed.
@monatruong_murror Do users ever ask for more direct answers now?
Murror
@jonathan_hayes3 Great question. Yes, some do -- especially early on. The interesting thing is that it usually happens in moments of high emotional discomfort. When someone is really hurting, the last thing they want is another question. So we built a sensitivity layer: when distress signals are high, the AI shifts toward more direct validation first, then gently opens space for reflection once the person feels heard. It's not one-size-fits-all restraint -- it's contextual.
One challenge I imagine is balancing guidance and frustration. Too much analysis can feel intrusive, but too little can leave users feeling stuck. Finding that middle ground is probably where the real product work happens.
Murror
@new_user___0932026a86e905cf4b2b7f7 You nailed the core tension. We think of it as a spectrum, not a toggle. The AI reads signals -- how long someone has been writing, the emotional weight of their words, whether they're circling the same issue -- and adjusts how much space it gives versus how much it guides. Getting that calibration right is an ongoing process, not a solved problem. But the key insight was that the middle ground isn't static -- it shifts based on the person and the moment.
Most AI products are optimized for the demo moment, not the repeated use moment. An AI that wows you in the first session often becomes something you report back to rather than think alongside. The restraint you're describing is actually the harder engineering problem — building something that makes the user feel capable rather than informed.
Murror
@habibferdous "Demo moment vs. repeated use moment" is such an important framing. We literally had this realization when we noticed our most-shared journal reflections were from users' first sessions -- the AI was performing for them. But the users who stayed for months were having quieter, less shareable experiences that were actually transforming how they related to themselves. Capable > informed is exactly right.
@monatruong_murror It's interesting that less analysis created more engagement.
Murror
@sarah_butler1 It surprised us too! Our theory is that when the AI does less interpreting, users do more reflecting. And reflection is what drives real engagement -- not the kind where people scroll endlessly, but the kind where they come back because they got something meaningful out of it. Less analysis meant users spent more time thinking for themselves, which made the sessions feel more valuable.
@monatruong_murror This sounds more like a coach than an AI assistant.
Murror
@christian_west1 That's actually the comparison we use internally. A great coach doesn't tell you the answer -- they ask the question that helps you find it yourself. The difference is that coaching requires deep trust, and trust with AI is fragile. So we had to earn it by being consistently useful without being presumptuous. It's a fine line but when it works, users describe it exactly as you did.
we are also struggle to make our AI agent smarter, if you don't mind I am in need of honest feedback for it to see if anything needs to be improved or upgrade
products/ai-hive
Murror
@nolan_vu Thanks for sharing! The tension between making AI smarter and keeping users engaged is something a lot of teams are navigating right now. What we found is that "smarter" sometimes means knowing when to hold back rather than adding more capability. Happy to take a look at ai-hive when I get a chance -- always interesting to see how others are approaching this balance.
@monatruong_murror thanks you too, you share lots of valuable topic for AI Agent makers like myself. Will definitely follow your sharing and product launching in the future
Murror
@nolan_vu Really appreciate that, Nolan! Always great to connect with others navigating similar challenges in the AI space. Looking forward to seeing how ai-hive evolves.
@monatruong_murror thanks for your compliment, I wish that Murror also evolve and upgrade smoothly to deliver quality experiences to clients
I like this way of thinking abt AI. Sometimes the most helpful answer is not the fastest or smartest one. For journaling especially, I think people need space to reach their own thoughts. If AI explains everything too quickly, it can feel useful in the moment but maybe less meaningful later. Asking better questions instead of giving conclusions sounds like a healthier direction.
Murror
@busra_seker1 "Useful in the moment but less meaningful later" -- that's exactly the trap we were falling into. For journaling especially, the value isn't in what the AI says, it's in what the user discovers while writing. Our job is to protect that discovery process, not shortcut it. Questions over conclusions is exactly the shift we made, and hearing it resonate with how you think about journaling confirms we're on the right track.
This resonates. I've noticed the same with AI writing tools, the most useful ones don't write the entire piece for me. They help me think through it.
There's a big difference between getting an answer and arriving at one. The second creates understanding, confidence, and ownership.
In a world obsessed with smarter AI, building AI that knows when to stay quiet might be the real innovation.
Murror
@kamrankhan "Getting an answer and arriving at one" -- that distinction is everything. We see it in the data too: users who arrive at their own insight through Murror's prompting report higher satisfaction and come back more often than users who received a "better" analysis. Ownership of the insight is what makes it stick. Love the AI writing tools parallel -- same principle, different domain.