We Were Sold More Time. We Got Less.
The question dominating every Leaders Studio cohort this quarter: why AI's promise of more time is producing the opposite, what a 19th-century economist can tell us about the mechanism, and now?
There is a paradox from 1865 that no one in the AI conversation seems willing to name. William Stanley Jevons noticed that when the steam engine became more efficient, coal consumption did not fall. It rose. The more efficient the engine, the more applications people found for it, the more engines they built, the more coal the economy consumed. The savings became the seed of new demand.
He called it a paradox. Economists now call it a law.
We are living through the same thing. The resource is not coal. It is attention. And AI is the new steam engine.
The promise was simple: AI handles the routine, you get the hours back. Satya said it. Sundar said it. Sam said it. Every SaaS dashboard with a sparkle icon in the corner said it. Save 5 hours a week. Reclaim your afternoons. Finally have time to think.
People in the highest AI-exposure jobs are working 3 hours and 15 minutes more per week than before.
UC Berkeley found that 67% of workers who adopted AI tools in 2025 reported working more hours by year-end, not fewer. Only 8% of the time AI saves is reinvested in activities that benefit the worker. The rest is absorbed by expanded output expectations, or consumed by the overhead of managing AI itself.
This was predictable. Jevons told us.
I sat with Alex Bleau, my cowriter on this piece, Senior Innovation Manager at E.ON and a Deep Dive host at the Productlab Conference this September. These are some reflections from that conversation.
If 77% of workers say AI increases their workload, where does that time appear in the hours-saved calculation?
The calculation has a structural problem. It counts the time saved generating the output. It does not count the time spent deciding whether to trust it.
77% of workers say AI tools increase their workload because of reviewing outputs, fixing mistakes, and juggling platforms. Only 17% of US adults say workplace AI is reliable without human oversight. One in three workers skips the review entirely, which does not solve the problem. It hides the cost until it surfaces as an error downstream.
There is also a baseline problem nobody names. The studies that claim time savings compare input hours to output volume. They do not compare outcomes.
“Are we comparing time saved against the same outcomes we were delivering, or just more output? It is one thing to generate more slide decks. It is another to say we delivered the same value in half the time.”
Alex gave me a small example that stayed with me. He spent three minutes asking an LLM to translate a single word into another language. A dedicated translation tool would have taken ten seconds. The instinct to reach for AI first is itself a new time cost, and nobody is counting it.
AI eliminated the cognitive break. When the 20-minute task takes 20 seconds, does the worker pause?
The 20-minute task used to have a second function nobody tracked. It was a break. Not rest exactly: you were working. But a lower-intensity moment that let the mind reset before the next hard problem.
AI compressed that task to 20 seconds. The worker does not pause. They move immediately to the next thing. And the next thing is also cognitively demanding.
This sits on top of a calendar that was already broken. Microsoft’s 2025 Work Trend Index found employees receive a ping every two minutes during core hours: 275 interruptions per day. It takes 23 minutes to regain deep focus after a single interruption. AI made it easy to run four or five prompts simultaneously, multiplying the threads to review.
“The switching cost was not removed. It was multiplied. You feel like you are doing more because the output pile is bigger. The cognitive fragmentation is worse.”
MIT Media Lab coined the term cognitive debt: participants who relied on LLMs showed the weakest neural connectivity, while those who worked from memory showed the strongest. Frontiers in Psychology traced the symptoms: long-term AI use is associated with mental exhaustion, attention strain, and reduced decision-making confidence.
What AI removed were the moments between tasks. Those moments were not wasted time. They were processing time.
If the people at Klarna freed by AI are now generating 3.6x the revenue per person, who actually got the time?
Klarna is the cleanest case study in the Jevons Paradox applied to a company. Between 2022 and 2025, headcount fell 49%: from 5,527 to 2,907. Revenue grew 108%. Revenue per employee went from roughly 340,000 to 1.24 million euros, 3.6 times the prior level.
The freed time did not go to the workers who survived. It became revenue.
Glassdoor ratings fell from 3.8 to 3.0 in the same period. That is the sound of the survivors.
When a company says AI freed up time, the question nobody asks in public is: freed up for whom?
Non-work AI adoption (50%) now exceeds work adoption (41%). Did the Jevons Paradox follow you home?
Federal Reserve data from April 2026 shows non-work AI use (50% of the population) growing faster than work use (41% of the workforce), up 26% year-over-year.
People are spending evenings building things with AI. Personal tools, side projects, apps they wanted to exist. This is genuinely exciting. It is also the Jevons Paradox running on personal time.
The promise is: build anything in an afternoon. The realistic timeline for a beginner project is 1-2 focused weeks. The AI made it possible. The AI did not make it fast.
There is a more specific version of this for product managers.
“When AI makes you a builder, it also makes you an experimenter. Experiments look like work. They feel like professional development. The risk: a portfolio of AI projects nobody asked for, solving problems that do not exist, while the real product bottlenecks bleed money.”
Efficiency expands what you attempt, not how much you rest.
If the FOMO is rational, is staying current with AI itself a form of overload?
The FOMO conversation usually goes one of two directions: psychological failing to be managed, or noise from the easily distracted. Neither framing takes the situation seriously.
A 2026 study in BMC Psychology identified FOMO, AI dependency, and anxiety as antecedents that all converge to increase cognitive load. ManpowerGroup’s 2026 Global Talent Barometer: AI use up 13% in 2025, confidence in its utility down 18%.
The most diligent people, the ones reading everything, trying every new tool, are accumulating the most cognitive debt. They are right that the floor is moving. But the act of staying current is itself exhausting, and that exhaustion compounds the problem it was supposed to solve.
There is a name forming for it: learning burnout. The specific exhaustion of trying to stay current with a field whose floor moves faster than you can absorb it. You are not tired of the work. You are tired of the homework that comes with it.
And this learning sits nowhere on the books. AI spending is projected to rise 44% in 2026. Training budgets grow 5%. Average employee learning time fell from 47 to 40 hours per year. Fewer than 20% of organizations track defined KPIs for their AI initiatives.
“Both the team and the company are paying for this learning right now with way more time and cost than we are getting out of it. And it is just accepted.”
Alex sharpened the edge. The investment framing is not wrong. Both sides are treating this as an investment. The problem is the invisibility. In every ROI conversation, in every productivity headline, these costs are real, they are paid, and they are almost never reported.
The FOMO is not irrational. The response to it is unsustainable.
Production went infinite. Time did not. Attention collapsed.
Time is the only fixed unit in this equation. Compute scales. Capital scales. Headcount scales. Production capacity is effectively unbounded. But the audience, the user, the colleague, still has the same 24 hours they had in 1865.
When publishing a blog post drops from a writer’s day to four minutes, companies do not publish four times more. They publish forty times more. The hours users are willing to spend evaluating, learning, and adopting did not expand to match. They contracted.
Emilia Korczynska, VP of Marketing at Userpilot, laid out the drop rates clearly. The numbers are uncomfortable:
Average product adoption fell from 20% to 6% as shipping accelerated. Teams shipped more in the last 12 months than the previous three years combined, with core engagement metrics moving less than 2%.
Cold outreach open rates are collapsing.
Webinar attendance is down.
Email engagement is down.
Blog traffic is flat or declining, even at companies producing 5x more content. Users are routing around the content layer entirely and asking AI tools for instant answers.
Emilia named the mechanism. Vibe marketing is what happens when production cost goes to zero and discovery gets skipped. The same vibe coding pathology, one layer up. Nobody asked whether anyone wanted the ROI calculator. But it ranks for three long-tail keywords and ChatGPT might cite it, so ship it.
So what is actually working? The opposite of vibe marketing. The channels where authentic human signal cannot be faked:
Founder-led content. Real humans saying real things from real experience.
Founder sales. CEOs and VPs personally selling to prospects.
Niche creators. Trusted precisely because they cannot be vibe-produced.
Community. Slack groups, meetups, conferences. The places AI cannot follow you yet.
Scarcity 101. As the supply of automated content explodes, the value of authentic human signal climbs in lockstep.
The market is already voting with its calendar. Conferences are multiplying: invite-only summits, founder gatherings, regional product communities, peer dinners. The format that costs the most to produce and demands physical presence in a specific room on a specific day is the one the market is rewarding.
The leverage in 2026 is not producing more. It is showing up as a recognisable human, on the platforms where humans are getting rarer.
The contrarian take on AI code: it is a bifurcation, not a quality collapse
You already know the headline numbers. 48% of AI-generated code contains security vulnerabilities. Forrester projects 75% of tech decision-makers will face moderate-to-severe tech debt by 2026. Cleanup for the 8,000+ startups that built production apps with AI is estimated between 400 million and 4 billion dollars. 2025 was the year of speed. 2026 is the year of the bill.
The interesting part is not the bill. It is who is paying it.
Two patterns are emerging on the AI slope.
The accelerators. Senior engineers and designers who treat AI as a junior collaborator, not a shipper. They read every line. They refactor what does not fit. They use AI to expand the scope of what they take on, not to lower the bar of what they ship. Their craft is compounding. The gap between their output and the average is widening every quarter.
The accumulators. Mid-level builders who treat AI output as finished work. They ship faster than they understand. They lose the ability to debug what they did not write. Their first-year output looks impressive. Their second year is consumed by maintenance on systems they no longer understand. Stack Overflow developer sentiment is the early indicator: 70% positive in 2023, 60% in 2025, and the steepest drop sits with the cohort using AI most heavily.
The new role nobody put on a slide is already being hired for: the AI cleanup engineer. Salaries are above market. Someone has to read the code that was written without being read. That work is going to the people who refused to outsource their judgment.
The people who got better with AI are the people who stopped trusting it. The most diligent prompters are the most stuck.
That bifurcation is the actual story of 2026. The quality bill is the symptom. The widening gap is the diagnosis.
So is there a version of AI adoption that returns time?
The NBER surveyed nearly 6,000 executives in 2026. More than 80% reported no discernible impact from AI on either employment or productivity. Fortune called it the return of the Solow productivity paradox: computers everywhere, productivity nowhere.
If the law holds, the promise of more time was never a communication problem. It was a structural impossibility. Efficiency creates demand for more ambitious work, higher output expectations, new categories of maintenance, new categories of learning. Every time you close the gap, it reopens somewhere else.
The companies that might break the pattern would have to frame AI differently. Not more, but different. Not faster, but better. And they would have to subtract something from the roadmap every time AI adds capacity, which no roadmap I have seen is designed to do.
The promise of more time assumes that time is the constraint. The evidence suggests the constraint is something harder to name.
One practice to try this week
Alex and I landed on one thing both of us are trying. Name the subtraction.
Every time you adopt a new AI tool, write down what you are going to stop doing because of it. Not “save time.” A specific activity, a specific meeting, a specific deliverable that you are going to stop producing. If you cannot name the subtraction, you have not adopted a tool. You have added one.
It is small. It is also the only practice I have found that pushes against Jevons rather than accelerating it.
Why we write this newsletter
We run this newsletter to surface the leadership topics that are not getting talked about clearly enough, and to help product people move up the ladder.
That same mission is why, alongside the Productlab Conference, we are building Leaders Studio: 1:1 coaching, company group
coaching, and Leaders Circles. We work with product people across the full ladder, from first-time people managers stepping into leadership, up through Heads of Product, VPs, and CPOs.
In an AI world where everyone can produce more, what compounds is the opposite. The human work. Soft skills, judgment, presence, the ability to lead other humans through change. That is what we sharpen, because that is what the next decade of product leadership will actually reward.
It is also why we asked Alex to host a Deep Dive at the Productlab Conference in September. If you are wrestling with this in your org, that is where the conversation continues. September 15 to 17 in Berlin. Details at www.productlab.app.
With love from Berlin,
Daniele
Sources
The AI productivity paradox: More work, not less, Fortune (March 2026)
AI Does Not Reduce Work, It Intensifies It, Harvard Business Review (February 2026)
Thousands of CEOs admit AI had no impact on productivity, Fortune (February 2026)
Klarna CEO says AI helped company shrink workforce by 40%, CNBC
AI helps Klarna double revenues with half the staff, Computer Weekly
One in Three Workers Skip Reviewing AI Output, Allwork.space (April 2026)
FOMO, AI dependency and cognitive load, BMC Psychology / Springer Nature (2026)
Cognitive offloading or cognitive overload, Frontiers in Psychology
Monitoring AI Adoption in the US Economy, Federal Reserve (April 2026)
Your Brain on ChatGPT: Accumulation of Cognitive Debt, MIT Media Lab (2025)
Companies are pouring billions into AI and cutting training budgets, Fortune (March 2026)
Photo by Morgan Housel on Unsplash
 | A guest post by
|