The aha moment is no longer enough for AI products

I spent a few evenings recently testing AI video tools. Every one of them gave me the same experience. I typed a prompt, waited, and something appeared that genuinely surprised me. For a few seconds it felt like magic.

Then I tried to actually use what it made.

That's where things fell apart. The output looked impressive in the preview, but it wasn't right for the job I had in mind.

I couldn't easily tell why it was wrong, I couldn't fix the specific part that bothered me, and starting over meant gambling another prompt and hoping. After three or four rounds of this, I closed the tab.

Not because the tool couldn't generate video. It clearly could. But because I had no confident path from "it generated something" to "I can use this."

This made me realise the concept we've relied on for years in product design, the aha moment, is quietly breaking down in AI products.

The aha moment used to be hard to fake

For most of software history, the aha moment was a reasonable proxy for value.

If a user got their first report out of an analytics tool, or sent their first invoice, or saw their first design render, they had genuinely done the thing.

The moment of perceived value and the moment of actual value were roughly the same moment.

Generative AI has split those two apart.

You type something. It generates something back. The perceived value arrives instantly and it arrives every single time, regardless of whether the output is any good.

AI products have made the aha moment easier to fake than at any point I can remember.

A model can produce something that looks 90% right in ten seconds, and that first impression is powerful enough that teams celebrate it as activation.

But the real test begins after the wow. Can the user judge whether the output is correct? Can they adapt it when it's close but not right?

Can they recover when it's wrong without losing their work? Can they export it, act on it, put it in front of a client or a colleague without embarrassment?

That's a different moment entirely. I've started calling it the trust moment, and I don't think it's an established industry framework.

It's just the most honest name I've found for the gap I keep running into: the point where a user decides they can rely on what the product gives them, not just admire it.

Where does the failure actually live?

When an AI output disappoints, there's a lazy answer available: the user wrote a bad prompt. Better prompting would have produced a better result. Skill issue.

I don't buy it. If someone has to become a prompt engineer before they see real value, the product has not done enough.

The more useful question is where responsibility should sit in the design of the workflow. When I look at my own failed sessions with those video tools, there were at least four different places the product could have intervened, and none of them required a better model:

Before generation, the product could have constrained my input.

A blank prompt box is the AI equivalent of a blank canvas, and blank canvases transfer all the hard decisions to the person with the least context about what the system does well.

Templates, structured fields and worked examples would have shaped my request into something the model could actually deliver on.

During generation, the product could have asked me questions. Half my failures came from ambiguity I didn't know I'd left in the prompt.

A model that asks "is this for social or for a presentation?" before spending thirty seconds rendering is doing design work, not adding friction.

After generation, the product could have made the output judgeable. Show me what assumptions it made. Let me see which parts of my prompt it leaned on.

Give me some way of assessing "is this right?" beyond squinting at it.

And when the output was close but wrong, the product could have supported repair instead of restart. Regenerating from scratch is the most expensive possible edit.

The tools that let me change one element while keeping the rest are the ones I went back to.

None of this is model capability. All of it is product design.

High stakes make this impossible to ignore

I can afford to shrug off a bad AI video. It cost me a few minutes and some curiosity.

I'm currently building Omi, an AI companion for first-time home buyers in the UK, and that project has made the trust problem concrete for me in a way the creative tools never did.

When someone asks what's holding back their mortgage readiness, the output isn't decoration. They might act on it.

They might delay a purchase, move savings around, or walk into a broker conversation carrying whatever the product told them.

In that context, an impressive-sounding answer is actively dangerous if the user can't tell how much to trust it. So the design questions become sharper.

Where did this guidance come from? What does the product know about my situation and what is it assuming? What happens when it doesn't know?

I'd rather Omi say "I can't tell from what you've shared" than generate a confident paragraph, even though the confident paragraph would demo better.

That's the trade-off the aha-moment mindset hides. Optimising for the wow pushes you towards confident, fast, impressive output.

Optimising for trust pushes you towards bounded, explainable, correctable output.

In anything high-stakes, the second one wins, and I suspect it wins in low-stakes tools too once the novelty wears off and people need to actually ship work.

What I'd measure instead

If the first generation event is your activation metric, you're measuring the thing that's now easiest to achieve. Almost everyone who types a prompt gets an output. Congratulations, your model works.

The signals I'd watch instead:

Did the user do anything with the output? Export, copy, save, share, insert into their actual work. Usage after generation is where value lives.

Did they come back and generate again for a real task, not just to play? Repeat use for work is the closest thing to a trust signal you can read from behaviour.

How often does a session end in abandonment after multiple regenerations? That pattern is a user trying to make your product work and giving up.

It's the most important failure mode in generative products and most teams aren't tracking it at all.

A quick audit for your own product

If you're building or designing anything generative, take your first-output experience and walk through four questions:

Can the user judge it? Is there anything in the interface that helps them assess whether the output is right for their purpose, or are they on their own?

Can they improve it? Is there a path from 80% right to done that doesn't mean regenerating and praying?

Can they recover from it? When it's wrong, what do they lose? Their input? Their edits? Their confidence?

Can they use it? How many steps sit between "generated" and "in my actual workflow"?

If the honest answer to most of these is no, you have a demo, not a product. The demo will still get the wow. People will still post about it.

But the users you keep will be the ones who got past the aha moment and found something they could depend on.

Generating something impressive is now the easy part. Earning the confidence to act on it is the real product problem, and it's a design problem before it's a model problem.

I'm still working out what the full answer looks like. But I'm certain the first moment of magic isn't it anymore.