An elderly guest walked up to a hotel reception on Germany's Baltic coast carrying a small box of chocolates. They were for Leah, the colleague who had answered all her questions on the phone with such patience.
"Well, you know, at my age, I don't always understand everything the first time," she said softly. "I know I ask a lot of questions. But Leah answered every single one with such patience and she made me feel heard. That kind of kindness is so rare these days, especially with the young generation."
Start here, because this is the seduction in its purest form. No manual, no onboarding, no friction. Just a voice that helped, and a person who felt met.
It is tempting to read this as proof that the chat box repealed the old laws of UX. It didn't. It ran a stress test on them.
It is tempting to read this as proof that the chat box repealed the old laws of UX. It didn't. It ran a stress test on them, and the results are more interesting than the marketing.
Take Jakob's Law first, because it looks like the obvious casualty. Nielsen wrote it in 2000 as a warning: "Users spend most of their time on other sites," so make yours behave like the ones they already know. A dropdown is a dropdown is a dropdown.
For thirty years that meant matching the conventions of the wider web. It's why companies stood up design systems teams in the first place — to keep every screen aligned with what people already expected.
Now the assistant absorbs the surfaces. Tasks that each had their own destination — draft the email, pull the numbers, build the deck — collapse into one asking box, one reply, one back-and-forth.
And Jakob's Law doesn't die. It flips. The sameness of every assistant is the feature: learning one teaches you all of them, so a new tool needs no manual and gets adopted faster than any onboarding flow could manage.
That's the tell. Jakob's Law was never really about screens. It was about mental models — the habits users carry in from everywhere else. Change the surface from a webpage to a chat box, and the law survives untouched, because people still don't change nearly as fast as their tools.
So it was structural. It was load-bearing. Good.
Now watch the other pillar buckle.
For most of a decade, the primary directive of product design was frictionlessness. Smooth is invisible. At Changi Airport, passport-less gantries reportedly clear eligible travellers in 10 to 30 seconds, and the stated goal is for the whole experience to be completely forgettable.
"Make the experience forgettable," a senior airport operations leader reportedly said, dead serious. Nobody falls in love with banking; you just want your money safe. Friction equals bad, and designers spend most of their time removing it.
The AI box is the apotheosis of this creed. Like a digital genie, an LLM grasps intent from minimal input and hands back results in seconds. Users keep coming back because it feels smooth, intuitive, immediately rewarding.
And that is exactly where the principle reveals itself as a proxy metric wearing a principle's clothes.
Consider the parents' communication app built by one team, designed to be completely frictionless with standardized announcement formats. Adoption spiked. So did support tickets. Parents couldn't retain information, lost attachments, missed key details in the infinite scroll.
So they brought friction back. Interaction pauses to flag attachments. Truncated content forcing a deliberate tap to expand. AI-generated conversation starters that made parents stop and reflect.
Time in the app went up. So did satisfaction. It was slower, and it felt more respectful.
The lesson isn't that friction is good. It's that "frictionless" was measuring the wrong thing all along. As one designer's hiking companion put it: "We spend so much time making things frictionless, but by making it frictionless we make it forgettable."
Underneath the AI genie sits a business model built on continuous consumption. The smoother the experience, the more we prompt, the more compute we burn, the more tokens we spend. Optimize entirely for instant output and you trade rigorous thinking for a fast, superficial result.
Frictionlessness was scaffolding. Useful scaffolding, sometimes. But it was standing in for goals it could not actually name — comprehension, trust, memory, the effort heuristic that makes people value what they worked for — and once AI let us remove friction from everything, its emptiness showed.
Which brings us to the pillar nobody funded.
For twenty years, information architecture was the discipline whose job titles quietly disappeared. Getting an organization to agree on one way to name and arrange things is slow, political, thankless, and it loses every quarter to whatever ships a feature.
The work was invisible. So were its failures. Humans are forgiving of messy structure — faced with a vague label, a person uses judgment, works around it, moves on. That tolerance hid the cost for two decades.
Then AI arrived, and the invisible foundation showed up on the balance sheet.
The same gaps that once cost you a confused visitor now cost you a hallucination, a wrong answer, or an agent that confidently retrieves garbage and acts on it. In a 2025 survey, data quality and availability topped the list of AI adoption barriers, outranking every other obstacle.
Poor information architecture can now be measured in token costs. A lot of them.
Here's the mechanism. Retrieval-augmented generation grounds the model's answer in documents it fetches at query time. But it inherits every flaw in the pile it fetches from.
If your content store is a flat heap of unstructured, contradictory documents, retrieval finds the loudest match — the one sharing the most surface words with the query, regardless of whether it's current, authoritative, or true.
A price, a policy, a deprecated note, and a customer quote can read as similar strings and mean opposite things. A person skims the page and knows which is which. A model handed the four snippets sees four passages of plausible language and no reason to rank one above another.
Metadata, taxonomy, controlled vocabularies — the work Rosenfeld, Morville, and Arango spent careers formalizing — are what let a system tell authoritative from anecdotal, current from expired. You can buy a better model, tune a sharper prompt, stack another evaluation layer, and still serve wrong answers, because the model retrieves from a pile nobody agreed how to organize.
IA turned out to be structural. It was load-bearing the whole time. We just never got billed until the machine removed the human who'd been absorbing the mess.
So the job is inverting.
Nielsen frames the same shift from the interaction side: AI moves computing from command-based to intent-based. The user states the outcome; the system infers the procedure. You stop operating the interface and start supervising it — a role with entirely different failure modes.
Removing steps was the old craft. But when the agent absorbs the cross-surface labor and the user only states a goal, the remaining design work isn't subtraction. It's structuring meaning so retrieval finds the right thing, and calibrating trust so the human knows what to believe.
Which returns us to Leah, and the one dial we've barely touched.
We don't anthropomorphize AI because we're irrational. We do it because our brains evolved to treat language as evidence of another mind.
That makes mind attribution a design variable, not an accident. How much "mind" your box invites a user to attribute to it — how heard they feel, how much they trust the fluent voice — is now something you tune, deliberately, the way you once tuned a checkout flow.
The smoothness that made Leah lovable is the same smoothness that makes a confident hallucination land. Fluency reads as authority whether or not the answer is true.
So the AI-native brief is not "make it forgettable." It's the opposite. Structure the meaning underneath, calibrate the trust on the surface, and decide — on purpose — how much mind the box should seem to have.
The laws didn't fall. They got sorted into the ones holding up the building and the ones that were only ever paint.