"We cut design production time from four days to six or eight hours."
That's a Head of Design in cybersecurity, describing what he attributes to AI's effect on his team's clock, in a piece published in 2026. Four days to six hours. He calls optimizing his internal processes "the easy part."
Now hold that against another number from the same season. In late May 2026, Google announced at I/O that AI would become the centrepiece of Search. Within days, according to reporting compiled by REPTILEHAUS, DuckDuckGo's "No AI" search page traffic tripled overnight, iOS installs surged nearly 70% week-over-week, and the traffic didn't fall back—it sustained roughly 84% above the pre-I/O baseline.
Not a rage-click spike. A sustained migration.
Put those two facts next to each other and the shape of the problem appears. One of them is measured. The other is assumed.
The design leader knows exactly how much faster his team ships, down to the hour, because someone counted. Nobody counted whether a single user wanted the extra features that speed produces. That number doesn't exist in the loop.
The design leader knows exactly how much faster his team ships. Nobody counted whether a single user wanted the extra features that speed produces.
So here is the question the rest of this owes you an answer to: who in this chain is accountable for proving users wanted any of it?
The comfortable answer is that this is a UX problem—the AI features just aren't polished enough yet, and once they are, adoption follows. The comfortable answer is wrong.
Look at the demand signals across sectors, and they don't read like a polish gap. Per a cross-market study cited by REPTILEHAUS, 57% of consumers say AI-heavy customer service reduces their trust in a business, up from 53% just six months earlier. 73% say they'd be more loyal to companies that use real people for all service interactions. 80% of enterprise workers are reportedly avoiding AI tools outright.
That's not "the interface needs work." That's a population declining the product.
Vitaly Friedman put the mechanism plainly: AI is not a value proposition. Labeling a feature "AI-powered" reportedly doesn't make anyone happy or excited. Features get compared to other features, not to the fallibility of humans, and an unreliable one loses to a reliable one every time.
And the resistance is specific, not reflexive. The DuckDuckGo exodus wasn't triggered by AI existing—DuckDuckGo still ships AI Chat and AI-assisted answers. It was triggered by AI arriving with no toggle, chosen for the user by the company.
Every backlash example shares that one thread: a company deciding, on the user's behalf, that AI is what they wanted.
So the demand-side evidence exists. It's abundant, it's cross-market, and it's directionally unambiguous. The problem isn't that we lack the data. The problem is that the data never enters the room where the ship-faster decision gets made.
To see why, read what Figma's own leaders admitted about themselves.
Figma's Sightlines newsletter, recapping the Leadership Collective at Config—which Figma reports drew more than 1,300 attendees, senior people from Cisco, OpenAI, Airbnb—opens by listing the questions they went in hoping peers could answer. The very first one: "How do you lead a change you're still figuring out for yourself?"
The TL;DR, in their words: "Most of us are still figuring it out."
Sit with that. The people setting the acceleration mandate are on record conceding they don't know where they're steering.
That's not a scandal. It's honest, even admirable. But it has a structural consequence that nobody in that room seems to have named.
If leadership doesn't know where it's going, the only thing it can confidently demand is that teams get there faster. Direction is uncertain; velocity is legible. So velocity becomes the instruction.
You can watch this happen in the UX Design piece. The generation part of the job got quicker—research insights that took weeks now take days, a round of concepts that took an afternoon now takes twenty minutes of prompting.
Then the sentence that reveals the whole trap: "you're not the only one who noticed."
The moment leadership sees that generating designs is faster, they ask for faster shipping—even if the team hasn't adopted the tools yet. Cut the timeline in half. Ship, ship, ship. The speed of generation gets converted directly into a speed-of-delivery quota, with no intervening step where anyone asks whether the additional output was wanted.
That intervening step is exactly the one that got deleted. And it got deleted because of how accountability flows through an org chart.
Speed pressure flows downhill. It's easy to transmit—a number, a deadline, a "why is this taking so long"—and it lands on whoever is holding the work.
Evidence of demand would have to flow uphill. And nothing in the structure requires it to. No one asks the design leader to walk into the boardroom with proof that users asked for the feature before it ships.
That asymmetry is the entire disease. The pressure to go faster is mandatory and downward. The proof that anyone wanted the output is optional and upward. Guess which one survives.
There's a timing mechanism that makes this worse, and NN/g named it. Users are still forming basic mental models of site-specific AI—still figuring out whether they're talking to a human rep or a frontier-level LLM. Get the qualities wrong and users abandon the thing fast.
Those mental models take time to settle. Speed pressure guarantees the product ships before they do. So teams launch into a population that hasn't yet decided what the thing even is—and then read the low adoption as a polish problem rather than what it is: shipping ahead of demand that was never measured.
The result is an incentive structure that only instruments one variable. Generation speed is counted to the hour. Validated demand is counted nowhere.
And "faster" is the one metric that survives contact with a boardroom. It's clean, it's comparable quarter to quarter, and it doesn't require anyone to admit the mandate might be pointed at something users are actively fleeing.
Notice what this optimizes. Companies are getting extremely good at producing the exact number that tells them nothing about whether the work was wanted.
Harvard Business School researchers forecast that at least 30% of generative AI projects will be abandoned—often not because the tools fail, but because the people they were built for quietly reject them. Read against the demand data, that's not a technology failure. It's the predictable yield of a system that measures output velocity and never once measures pull.
Which brings the loop back to the accountability gap it started with. Somebody counts the four-days-to-six-hours. Somebody could have counted the 84%. Nobody was required to put the second number in front of the person demanding the first.
There's an antidote in the evidence, and it isn't "slow down." Jen Dunnam's version of the job is to refuse the premise that speed is one thing. Between fast, cheap, and good you get two—and leaders who want all three are dodging the real conversation about which two matter right now.
That conversation is where demand could re-enter the loop. Ask which speed the organization actually needs, and you're forced to ask what outcome you're speeding toward—and whether anyone on the receiving end asked for it.
Until that question is mandatory rather than optional, "ship faster" will keep meaning "ship more of what we never checked anyone wanted."
The migration is already measured. DuckDuckGo counted it. The only thing missing is a boardroom willing to look.