On AI
If a robot could have written it, we didn’t send it.
Everyone’s holding their breath.
ScrollIt’s software. People wrote it, people trained it on other people’s work, and people decide where it gets used. A model doesn’t have a name on the door. We do.
AI wants agency. It needs a leash.
It will not replace empathy, or creative genius in our studio. A machine can finish the sum. It can’t make the leap nobody saw coming, and that leap is most of what we’re good at.
What it can do is the machine work. What used to take a week of flat comps takes a day and a half of working screens.
We have designed and built AI into products — before anyone called it AI. It went by machine learning, and predictive analytics before that. The label changes every few years. The problem underneath does not.
Everything that comes straight out of AI looks the same — a stock photo of a stock photo. We won't hand you that and pass the buck forward.
We give the machine the parts that are machine work, and keep the parts that aren’t. The time that frees up goes back into the thinking.
So you get working, clickable screens, with interaction and motion, while other studios are still annotating a deck. No roadmap theatre. If someone arrives with a confident story about what AI will do to your industry, it’s fair to ask why they’re telling you.
Putting AI inside your product
The question we get now isn't how we use AI. It's whether AI belongs in the thing you're building. A new layer of complexity — and an advantage nobody copies in a sprint, when it's handled right.
The interface has to be wrong gracefully
Press the button, get the result — that was deterministic software. A model is a guess with a confidence attached. What the product does when it's unsure, slow or plain wrong is the design problem.
Trust is the feature
People forgive a wrong answer. They don't forgive being unable to see where it came from or fix it. Provenance and an honest signal of confidence beat raw accuracy.
Latency is a material
A four-second response is not a spinner problem. What happens in that window decides whether the feature feels powerful or broken.
Knowing where not to put it
The hardest call, and the most valuable. Plenty of features are better as a rule, a filter, or a well-made form.
Knowing whether it works
“It seems good” is not a measure. What success looks like belongs at the start of the work, not after launch.
Get these right and you have something your competitors cannot reproduce in a sprint. Get them wrong and you have a chatbot nobody opens twice.
The Ebenezer effect
Some of what you ask a model to do comes back for you later.
Named for the obvious reason. A model will do almost anything you ask, immediately, with no record of where any of it came from. That is the entire appeal. It is also what turns up at the foot of the bed in three years, holding receipts.
It will not tell you whose work it borrowed
Ask for something in the style of a company you admire and you will likely get it, with no note on which parts came from where. Somebody asks eventually, and it is rarely you. It is their counsel, usually after the thing is on a building.
Rebranding after launch is the expensive way to find out
A mark, a name, a face, a line that sits too close to somebody else’s. The cost of finding out late is not the design fee. It is the packaging, signage, store listing, domain, and the year spent unwinding it all in public.
What you paste in does not come back out
Unreleased roadmaps, someone’s records, a contract, code you do not own. In is in. Plenty of companies have a policy. Fewer have read it, and almost nobody has checked which tools their team actually has open.
“The AI made it” is not a defence
It is your product, your name on the front of it, and your liability. The tool does not sign anything and it will not be in the room.
We are not lawyers and none of this is legal advice. It is what we have watched go wrong. In practice we track where things came from, keep your material out of tools we don’t control, and say so before it ships if something looks borrowed. That is the unglamorous half of not handing you AI slop.
Next
Contact