When an AI feature dies, the autopsy usually blames the model. Wrong. Most AI tools are killed by their own designers, in ways that have nothing to do with intelligence and everything to do with how the tool meets the person. Pauline Kabitsis of Irrational Labs framed these as seven deadly sins of AI product development. Here they are, with the fix for each.
Neglect
A blank box that says “ask me anything.” It looks like freedom; it lands as abandonment. The user does not know what to type, so they type nothing, and leave. Fix: scaffolding. A first prompt, a starting point, a worked example. Never open on an empty field.
Obscurity
A genuinely good AI feature buried behind an icon nobody clicks. The value exists; it is just invisible. Fix: make it impossible to miss. Put it in the path of the work, not in a menu three levels down.
Gluttony
Forty-seven options when human working memory holds three to seven. Every extra choice is a tax the user pays before getting any value. Fix: remove choice. Guide the next best action instead of presenting the full menu of everything the tool could theoretically do.
Mediocrity
Promised magic, delivered meh. The first result is underwhelming, trust drops, and the user does not come back to give it a second chance. Fix: let users co-create, show justifications for the output, and be honest while the tool is still learning. A tool that admits its limits keeps more trust than one that overpromises and misses.
Vanity
“AI-powered” stamped on the box with no clarity on what it actually does for you. Nobody pays more for the label. Fix: sell the outcome, not the technology. Superhuman did not say “AI-powered email.” It said “four hours saved per user per week.” That is a reason to pay. “AI-powered” is not.
Tyranny
Automating away all human control. The tool decides everything and the person is along for the ride, which feels less like help and more like losing the wheel. Fix: give agency. Let people undo, override, and opt in. Control is what makes automation feel like leverage instead of replacement.
Envy
Copying everyone else's sidebar and blank prompt because that is what the last demo looked like. The result is a tool indistinguishable from ten others, solving nobody's specific problem. Fix: differentiate. Start from the one workflow your users actually struggle with, not from the shape of the competitor's UI.
The thread through all seven
Read them together and the pattern is obvious. Not one of these sins is about the model. Every one is about the space between the tool and the person — what they see first, how much they have to decide, whether they keep control, whether the value is legible. The intelligence was never the bottleneck. The interface to the human was.
That is the same lesson the failure statistics keep pointing at from a different angle. AI does not fail because of the model. It fails because of the human — and the seven sins are the specific, fixable ways that failure gets designed in. Fix them and the tool has a chance. Ship them and the best model in the world will not save it.
// SOURCES
- — Pauline Kabitsis, Irrational Labs, “Seven deadly sins of AI product development,” Ottawa PMMA, December 2025.
- — Superhuman outcome framing (“4 hours saved per user per week”), cited as an example of outcome-over-label messaging.
Written by
Samer Tageldin
Senior technology delivery and AI workflow design. Field notes on putting AI into real organizations.
[ View profile ↗ ]