Artificial intelligence (AI) can now build almost anything you describe, and fast. I see it every week: a design change that used to take a sprint goes from first commit to merged in about a day. But the speed of building was never the hard part for long. The hard part is deciding what to build, and how it should feel to the person using it.
Why aren't most AI investments paying off?
McKinsey's 2026 State of AI survey of 1,719 organizations found that only 6% attribute significant financial returns to AI. That share hasn't moved from last year, even as adoption keeps climbing.
What sets the 6% apart is not the model they picked. Nearly three in four of them fundamentally redesigned their workflows around AI. About one in four of everyone else did. That's an association, not proof of cause, but the gap is hard to ignore.
Why isn't faster the same as better?
The natural move is to take the workflows you already have and make them faster: a smarter form, a triage bot, an answer box bolted onto an old flow. It feels safe. It's also where most of the value disappears.
Yesterday's workflows are still useful, mostly as a signal of what not to rebuild. There is no proven AI playbook from the old way to copy. The edge goes to teams designing what comes next.
If AI builds this fast, what does it still get wrong?
The tool makers are honest about this. In Figma's 2026 AI report, built on more than 8,400 survey responses, nearly six in ten people said design matters more now than before AI. Figma's own chief executive officer (CEO) said this year that current models still don't understand user experience (UX) or simple judgment calls. The AI researcher Andrew Ng put it plainly: the bottleneck now is deciding what to build.
AI amplifies what goes in. Good judgment in, good product out. Weak judgment in, weak product out, just faster.
That's also why the judgment has to live somewhere the machine can't forget it. Rules repeated in a prompt drift; rules written into the system hold. I wrote about that in How to Put Your AI Agent on Rails.
Where does human judgment earn its place?
Judgment is not taste for its own sake. It's knowing the people you're designing for, the context they're in, and the behavior you're trying to change. It shows up in small decisions: what the AI says, when it stops and asks, and when it hands off to a person.
That's the layer I focus on as an AI product designer. Engineering builds the system. Design decides how it behaves around people. Clear guardrails let both move quickly, which is the point of governance done for teams that move fast. The teams in the 6% treat these as different crafts and put all of them in the room.
Building AI that people actually use? I help startup and enterprise teams decide where AI belongs and design how it works around real people.
Book an advisory introFrequently asked questions
Why do most AI projects fail to deliver ROI?
Most layer AI onto existing workflows instead of redesigning the work. McKinsey's survey data shows high performers are far more likely to redesign workflows, which suggests the gap is design and decision making, not model capability.
Can AI replace product design?
Not the part that matters most. AI is strong at generating screens and code, and still weak at judgment calls about people, context and behavior. It speeds up execution. It doesn't decide what's worth executing.
What should teams do first?
Pick one workflow, redesign it around what people actually need, and test it with real users before scaling. Days or weeks of validation beats months of building the wrong thing.