AI is getting better. The experience is getting worse.
Every week, another AI company announces a better model.
More reasoning. More context. Better coding. Better agents. Better benchmarks. On paper, the progress is real. The capability curve is moving faster than most of us can keep up with.
In daily use, it feels different.
I use these tools every day. For Finz. For writing. For research. For decisions I am working through. I have a clear view of where the models are now strong, and where they are not. The strong list is genuinely impressive. The models are writing production code. They are running deep research across hundreds of sources. They are handling work that needed a team of engineers a year ago.
The weak list is harder to talk about, because it does not show up in any benchmark.
The models still struggle to understand tone. They lose the thread of a decision halfway through. They miss the one detail that mattered. They give the shape of an answer before doing the work. They over-explain simple things and under-think the hard ones. They refuse too much, or they agree too easily.
So the user starts managing the model.
You rewrite the prompt. Add more context. Repeat constraints. Correct the tone. Ask it to try again. Ask it to be more precise. Ask it to follow the instruction it was already given. Ask it not to hallucinate. Ask it not to be lazy.
That is the frustration. AI was supposed to reduce work. Too often, it creates a new kind of work. Supervising the system that was supposed to help.
This is not just a model problem. It is a product problem.
Benchmarks measure capability under controlled conditions. Users measure usefulness inside messy work. Those are not the same thing. A model can improve on coding tests and still feel worse in a normal conversation. It can perform better on long-context evaluations and still miss the one detail the user cares about. It can become safer and more polished while becoming less direct, less decisive, and less useful.
The gap between the benchmark and the experience is widening. And the people who feel it most are the people using these tools every day.
There is another part of this that does not get said enough.
Every release raises the bar. Three months ago, a response might have felt good enough because the expectation was lower. Today, the same response feels worse because the user has seen what the system can do when it is working well. The standard has moved. The promise has moved. The actual daily output has not moved at the same pace.
The user is no longer asking, can this thing write. They are asking, can I trust this thing to understand what I mean, keep context, make a judgment call, and save me time without turning me into the operator. That is a harder bar. And it is the right bar.
The strange thing about AI products today is that capability and trust do not always move together. A new model can be more powerful and still feel less trustworthy. A new interface can add features and still slow the user down. A safety layer can reduce one kind of risk and introduce another kind of friction. A warmer personality can make the product feel more human while making the answer less rigorous.
Users feel all of this. They may not know whether the issue is model routing, guardrails, memory, latency, cost optimisation, or product design. They only know the result.
It used to answer better. Now I have to work harder.
That sentence should worry every AI company.
There is a quieter cost in all of this that does not get measured. Time. Hours of it, every day, lost to prompting and re-prompting. The subscription bill keeps climbing because we keep paying for the higher tier, hoping that this version, this release, this update will close the gap. It rarely does. The promise stays just out of reach. The work stays on us.
The next phase of AI will not be won by the company with the highest benchmark. It will be won by the product that feels consistently useful in real work. Fast when the task is simple. Careful when the task is hard. Direct when the user needs a decision. Honest when the system is unsure. Consistent enough that the user does not feel like they are starting from scratch every day.
That is the product challenge. And it is bigger than the model challenge.
AI is getting better. The experience is not. Until that changes, the gap between the promise and the daily reality will keep growing. The people building these systems will keep pointing at the benchmarks. The people using them will keep pointing at the clock.
One of those views is right.