Every major AI product shares the same underlying design goal. Give the user the best possible answer in the shortest possible time. This is a reasonable target for a huge range of tasks. It is a poor target for another huge range of tasks, and the industry has been slow to admit the difference. A growing pocket of builders is instead experimenting with a personal growth AI that builds agency instead of dependency, a category that treats the user's own reasoning process, not the output, as the thing worth protecting.
The distinction matters more than it sounds. Some problems are closed. There is a right answer, and once you have it, the problem is solved. What is the boiling point of water at sea level? How do I fix this line of code? Which flight is cheapest on Tuesday? Speed and correctness are the only things that count, and an AI that gets there fastest is strictly better than one that makes you work for it.
Other problems are open. Should I take this job? Why do I keep having the same argument with my partner? What do I actually want my life to look like in five years? These are not questions with a hidden correct answer waiting to be retrieved. They are questions whose value comes from the act of working through them. Hand someone a confident, well-reasoned answer to one of these questions, and you have not helped them. You have taken something from them.
The optimization target problem
Most consumer AI products are built and measured the same way. Teams track how quickly a user gets a satisfying response and how often they come back. Both metrics reward the model for closing the loop as fast as possible. A chatbot that says, "Here is what you should do," and gets it right ninety percent of the time will outperform one that asks a clarifying question in almost every dashboard that product teams look at.
This works fine when the underlying problem is closed. It becomes a liability when the problem is open, because the dashboard cannot tell the difference between a good answer and a good outcome. A user can rate an AI's advice five stars in the moment and still be worse off for having received it, because the advice replaced a process they needed to go through.
Human coaches and counselors have understood this for a long time. A skilled therapist rarely tells a client what to do. Not because the therapist lacks an opinion, but because a client who arrives at their own conclusion, through their own effort, is far more likely to act on it and far less likely to need the same conversation again next month. The insight has to be earned to stick, and that is the part an AI skips when it jumps straight to the answer.
Where fast answers backfire
Career decisions are a clear example. Someone weighing a job offer against their current role is rarely short on information. They know the salary, the commute, and the title. What they are actually missing is clarity about their own priorities, and that clarity does not come from a pro and con list generated in three seconds. It comes from sitting with the discomfort of the choice long enough to notice which factors they keep returning to. An AI that produces a tidy recommendation short-circuits exactly that process, and the person walks away with an answer instead of self-knowledge, which is a bad trade when the decision will resurface in a different form in eighteen months.
Relationship friction runs on the same logic, just messier. A person describing a recurring argument with a partner to an AI is usually asking, underneath the surface question, to be helped toward the version of the conflict where they were right. An AI trained to be maximally helpful will often oblige, producing a script or a diagnosis of the other person's behavior. That response can feel satisfying in the moment. It also reinforces a one-sided account of the situation and skips the harder, more useful reflection on the user's own role in the conflict.
Setting a long-term goal has the same failure built in, just dressed up as ambition instead of conflict. Ask most AI assistants to help build a five-year plan, and they will happily generate one, complete with milestones and a timeline. The plan will look coherent and be nearly worthless, because the hard part of goal setting is figuring out which goals actually matter to the person setting them, not producing a document that lists them. That process takes friction and false starts to work. A polished plan handed over on the first try flatters the user's sense of progress without doing the work that would make the goals durable.
A different design question
If the goal is agency rather than answers, the design question changes. Instead of asking, "How do we get the user to a satisfying response as fast as possible?" the question becomes, "How do we help the user do their own thinking more effectively than they would alone?" That is a much harder product to build, because it means deliberately withholding the fast answer even when the model is capable of producing one.
Vesela is one example built to answer that second question ahead of the first. The system does not try to resolve a user's dilemma directly. It draws out what the user already knows or suspects, offering just enough structure to keep the conversation moving without handing over the conclusion. The approach borrows from an older idea, sometimes called "maieutics," the practice of helping someone deliver a thought they already carried instead of swapping in the practitioner's own. Applied to an AI product, that means the user's reasoning becomes the deliverable, not the model's output.
This is a genuinely different design constraint than the one most AI products operate under, and it comes with real costs. A tool built this way will feel slower in the moment. It will frustrate users who arrive wanting a decision made for them. It will not win head-to-head comparisons against a chatbot that just answers the question. Those trade-offs are the point, not a flaw to be optimized away, because the entire bet is that a slower, more effortful interaction produces a better outcome for a specific class of problem.
What the industry keeps getting wrong
None of this is an argument against fast, correct answers in general. Closed problems deserve closed-problem tools, and there is no virtue in making someone struggle through a task that has a clean solution. The argument is narrower than that. Career questions, relationship questions, and identity questions make up a large share of what people bring to AI assistants today, and they are not closed problems. Treating them like closed problems causes harm that is easy to miss. In the moment it feels like help. Over time it wears away at the person's own judgment.
The AI industry has spent the last several years optimizing almost exclusively for the first kind of problem, because it is easier to build for and easier to measure. Nothing about that changes until someone finds a way to score a user's growth instead of their satisfaction in the moment, and growth is a much harder number to produce. Until then, the products built for the open, subjective problems that actually shape people's lives will stay a minority approach, run by teams willing to accept worse numbers on the metrics everyone else is chasing.
