U.S. labor productivity accelerated to 2.7 percent. Isn't that AI?

Productivity in the nonfarm business sector did rise — 2.3 percent in 2024, against a 1.4 percent average from 2014 to 2023. But economists caution the attribution to AI is unclear. The BLS and others find little evidence of economy-wide acceleration directly caused by generative AI without, in their words, deliberate structural or managerial interventions. A real number is not the same as a known cause.

Maybe the gains just take time to show up.

Possible. But the leakage is not only a lag. 95 percent of enterprises overspend on AI infrastructure, 30 to 50 percent of AI cloud spend evaporates into idle resources, and 30 percent of generative AI projects were abandoned after proof-of-concept by the end of 2025. Some of the missing gain is not waiting to arrive. It already left.

If it helps me on every task, surely it helps overall.

That is exactly the trap. Task gains and total gains are different things. Workers cut email time by 31 percent and changed nothing about meetings or total hours. The saved time did not aggregate up. It got reabsorbed into more output, one task at a time.

Episode

The intensification era

What Actually Changed?

AI did speed up your work. Then it quietly handed you more of it.

This is not a story about whether AI works. It works on tasks. It is a story about where the saved time goes, and who it goes to.

Lets start with the part everyone agrees on.

The tool is faster.

You write the email in a third of the time.

You draft the doc before the meeting starts.

You ship the function before lunch.

You finish a thing that used to take an afternoon.

Then you look up.

It is the end of the day.

You are tired in a way that does not match how much got done.

Bottom line — The speed is real. The relief is the part that never arrives.

The reframe

AI did not free up your time. It changed where your time goes.

The gains are not a myth. They are just small and local.

In controlled studies, the speedup is real and measurable.

Developers using GitHub Copilot finished a JavaScript HTTP server task in 1 hour 11 minutes instead of 2 hours 41 minutes — about 55 percent faster, with a 95 percent confidence interval of 21 to 89 percent. A separate field study of 4,867 developers found AI assistants raised completed tasks by 26 percent.

Customer support agents gained too. A field experiment with 5,172 agents found AI raised issues resolved per hour by 15 percent on average, and 36 percent for the least experienced. A two-month new hire with AI matched a six-month veteran without it.

So the question is not whether AI speeds up a task. It does. The question is what happens to the time it saves.

Bottom line — AI compresses tasks. The honest debate is about everything downstream of that.

3.6hours saved on

Workers cut email time by 31 percent — about 3.6 hours a week. Meetings did not shrink. Total hours did not shrink. The time was reabsorbed.

interrupt
loss frame

Bottom line — You did not get 3.6 hours of your life back. You got 3.6 hours of new room to fill.

How a time-saving tool becomes a workload

01
The task gets faster
AI compresses the writing, the drafting, the first version. The work that took an afternoon takes an hour.
02
The hour is now empty
But an empty hour at work is not rest. It is capacity. And capacity gets noticed.
03
You reach for what you could not do before
The analysis you would have skipped. The cross-functional task that was someone else's job. The second project, run in parallel.
04
The expectation resets to the new output
What you produced today becomes the new normal tomorrow. The floor moves up to meet your ceiling.
05
You are busier than before, and it feels like progress
More got done, so it reads as productivity. The cost — lost recovery, blurred boundaries — does not show up on any dashboard.

Bottom line — AI does not remove work. It removes the friction that used to cap how much work you took on.

The hidden mechanism

The real product of AI at work is not productivity. It is intensification.

Researchers watched it happen for eight months.

A UC Berkeley team spent eight months inside a 200-person U.S. tech company, watching how AI actually landed.

They did not find rest. They found three forms of intensification.

Scope expansion. Product managers started writing code. Designers took on engineering tasks. Researchers began managing infrastructure. Work that belonged to specialized roles dissolved into everyone's job, because AI made cross-functional work feel possible.

Boundary dissolution. Work seeped into breaks, into evenings, into the spaces that used to be off.

Parallel processing. People ran multiple AI threads at once, holding several tasks in their head where they used to hold one.

Productivity metrics went up. Burnout risk went up faster.

Bottom line — When the cost of doing more drops to nearly zero, doing more stops feeling like a choice.

Where the saved time actually goes.

Everyone sees the speedup. Almost no one tracks where the time is reabsorbed.

01Surface
The task is faster. You finish the email, the draft, the function in less time.
02Scope
You take on work that was once out of reach or someone else's — because now you can.
03Expectation
Today's higher output becomes tomorrow's baseline. The system resets faster than you adapt.
04Recovery
The breaks and edges that used to absorb the strain get filled with more work.
05Cognition
Running parallel tasks across blurred boundaries taxes attention in a way no time-tracker counts.

The speedup is the symptom people celebrate. The reabsorption is the cost no one measures.

the promise vs. the lived reality

What AI was supposed to do

Free up your time.
Cut the busywork.
Let you focus on what matters.
Reduce the load.

What AI actually does at work

Redistribute your time toward higher scope.
Backfill the saved hour with new tasks.
Blur where the workday starts and stops.
Reset the baseline before you can rest on it.

Bottom line — AI redistributes work toward higher scope and lower recovery. That is not liberation. It is intensification wearing liberation's clothes.

THE LEAKAGE

Tasks accelerate. Total productivity does not. The gain leaks out before it reaches the bottom line.

Zoom out, and the gains vanish.

Here is the part that should stop you.

McKinsey's 2025 survey found 88 percent of companies had deployed AI in at least one function. And 94 percent reported they were not seeing significant value from it. Only 6 percent were high performers attributing about 5 percent of earnings to AI.

So adoption is near-universal, and measurable firm-level payoff is rare. Tasks get faster. The total does not move. The gain leaks — into higher expectations, coordination overhead, infrastructure cost, and reskilling that never lands.

Bottom line — If the task is 55 percent faster but the company sees nothing, the time did not disappear. It was spent somewhere the firm cannot see.

Adoption vs. value

Nearly everyone adopted AI. Almost no one captured the value.

You hear that AI is transforming work. Then you look at what companies actually report, and the gap between deployment and payoff is enormous.

How to read thisEach bar is a share of companies in a 2025 enterprise survey: how many deployed AI, versus how many say they are seeing significant value from it.

Adoption is near-universal. Captured value is rare.

Notice88 percent of companies deployed AI. Only 6 percent are high performers seeing significant value.

For you

The pressure to use AI is real. The proof that it pays off at scale is mostly not there yet — which means the burden lands on you before the system rewards it.

Behind the numbers

Source: McKinsey, State of AI 2025. Reported figures: 88% of companies have deployed AI in at least one function; 94% report not seeing 'significant' value; only 6% are 'AI high performers' attributing about 5% of earnings to AI. McKinsey identifies four sources of value leakage: behavioral and cultural barriers, coordination failures, reallocation inertia, and competitive frictions.

Verify the data ↗

Bottom line — The story is not 'AI failed.' The story is that task-level gains get consumed long before they become company-level value.

But what about…

But the productivity numbers are going up

“U.S. labor productivity accelerated to 2.7 percent. Isn't that AI?”
Productivity in the nonfarm business sector did rise — 2.3 percent in 2024, against a 1.4 percent average from 2014 to 2023. But economists caution the attribution to AI is unclear. The BLS and others find little evidence of economy-wide acceleration directly caused by generative AI without, in their words, deliberate structural or managerial interventions. A real number is not the same as a known cause.
“Maybe the gains just take time to show up.”
Possible. But the leakage is not only a lag. 95 percent of enterprises overspend on AI infrastructure, 30 to 50 percent of AI cloud spend evaporates into idle resources, and 30 percent of generative AI projects were abandoned after proof-of-concept by the end of 2025. Some of the missing gain is not waiting to arrive. It already left.
“If it helps me on every task, surely it helps overall.”
That is exactly the trap. Task gains and total gains are different things. Workers cut email time by 31 percent and changed nothing about meetings or total hours. The saved time did not aggregate up. It got reabsorbed into more output, one task at a time.

Bottom line — Faster tasks and a more productive economy are two claims. The pack supports the first far more cleanly than the second.

The pivot

If the gains are real but small at the top, the next question is who actually keeps them.

The gains do not spread. They concentrate.

The wage premium for AI-skilled workers doubled from 25 percent in 2023 to 56 percent in 2024 — the fastest-growing skill premium in modern labor-market history. And it accrues almost entirely to experienced practitioners, while the college wage premium has flattened since 2010.

Adoption itself is unequal. In a survey of 18,000 workers in AI-exposed jobs, 41 percent used ChatGPT at work — ranging from 65 percent of marketing professionals to 12 percent of financial advisors. Women were 16 percentage points less likely to use it. And the people already using it tended to earn more before AI arrived.

A tool that could lift the least expert is being held mostly by the already-advantaged.

Bottom line — AI had the potential to compress inequality. In practice it is widening it, because access and skill were already unevenly held.

56percent wage

The AI wage premium doubled in a single year — from 25 percent to 56 percent — and lands almost entirely on workers who were already experienced.

interrupt
loss frame

Bottom line — The clearest thing AI changed about pay is not the average. It is the gap.

The same split runs through firms.

It is not only which worker. It is which company.

Since ChatGPT launched, productivity at large-cap S&P 500 firms is up 5.5 percent. At small-cap Russell 2000 firms it is down 12.3 percent. Forty percent of firms report zero AI investment, and 42 percent of non-adopters say the technology is still too immature.

Large firms have the engineers, the data, and the budget to absorb a tool that costs an average of $85,521 a month and still leaks half its cloud spend. Small firms get the disruption without the leverage.

Bottom line — AI rewards scale. The organizations big enough to implement it well pull further ahead of the ones that cannot.

Large vs. small firms

Big firms pulled ahead. Small firms fell behind.

AI is supposed to be a great equalizer — cheap intelligence for everyone. The productivity data since ChatGPT launched points the other way.

How to read thisEach bar is the change in productivity since the launch of ChatGPT: large-cap firms in the S&P 500 versus small-cap firms in the Russell 2000.

The same technology, opposite outcomes, depending on firm size.

NoticeLarge-cap productivity is up 5.5 percent. Small-cap productivity is down 12.3 percent.

For you

Whether AI helps you may depend less on your skill than on whether your employer is large enough to implement it well.

Behind the numbers

Source: CNBC and market analysis, 2025, drawing on NBER analysis. Reported figures: S&P 500 (large-cap) productivity up 5.5% since the ChatGPT launch; Russell 2000 (small-cap) productivity down 12.3%. 40% of firms report zero AI investment; 42% of non-adopters cite technology immaturity. High-skill services and finance see 0.8% labor productivity gains; lower-skill sectors see 0.4%.

Verify the data ↗

Bottom line — The equalizer story is the surface. The concentration story is the system underneath it.

What it actually means

The everyday signs of intensification, decoded.

These feel like personal failures of discipline. Most are the predictable output of a tool that lowered the cost of doing more.

I finished early but I feel more behind than before.
The expectation reset to your faster output. The bar moved with you.
I keep doing parts of other people's jobs now.
Scope expansion. AI made the cross-functional task feel possible, so the boundary dissolved.
I check work during dinner without deciding to.
Boundary dissolution. When work compresses, it stops staying inside the workday.
I have four things half-open at once.
Parallel processing. You are holding more threads than attention was built to carry.
I produced more, so why am I this tired?
The output is visible. The lost recovery time is not. The fatigue is the unmeasured cost.

Bottom line — The exhaustion is not a character flaw. It is the felt experience of work redistributing toward higher scope and lower recovery.

Same tool, opposite outcome — depending on where you sit.

Columns

Who you areWhat AI changes for you

Who you are	What AI changes for you
High-skill worker at a large firm	A 56% wage premium and a tool your employer can actually implement. You compound.
Less-skilled worker at a small firm	Disruption without leverage. The tool that could lift you is hardest for your employer to deploy.
A worker whose edge AI can now do	Your comparative advantage gets automated. The thing you were good at becomes the thing the tool does.
Customer service representatives	Real near-term gains — and an 80% automation risk for routine roles. The help and the threat arrive together.

The losers are specific, not abstract.

Roughly 3.9 to 5.6 million U.S. workers sit at the intersection of high AI exposure and low capacity to adapt — routine roles, limited savings, few local alternatives. Customer service representatives face an 80 percent automation risk. AI was directly linked to 77,999 tech job cuts in the first half of 2025.

Remember the support agents who gained 15 percent? The same capability that compresses a new hire's learning curve is the capability that can replace the role entirely. The Commonwealth Bank of Australia deployed voice bots to cut 45 call-centre jobs — then reversed the decision when customer experience collapsed. The efficiency that looked clean in the spreadsheet was paid for somewhere off the spreadsheet.

Bottom line — The help and the threat are not different stories. They are the same capability pointed in two directions.

Try this

So what actually changed — the amount of work, or who carries it and who keeps the reward?

The surface answer is 'work got faster.' The honest answer is about distribution.

Work got faster at the task, but did not get smaller in total.
The saved time got reabsorbed into expanded scope, not recovery.
The gains concentrated among high-skill workers at large firms.
The cost moved to less-skilled workers, smaller firms, and your own recovery time.

Bottom line — The most accurate sentence is not 'AI saves time.' It is 'AI redistributes work and reward — unevenly.'

Prediction · claim

Through the late 2020s, headline productivity stats will stay ambiguous while the visible effects of AI are mainly distributional: a widening wage premium for AI-skilled workers, a widening gap between large and small firms, and a workforce that reports feeling busier rather

Metric: AI-skilled wage premium(%)
Confidence: 70%
Resolves: Dec 31, 2028

Bottom line — If the next two years look like the last one, the clearest change AI made will be measured in inequality, not in leisure.

Once you see it, the tiredness makes sense.

You were not promised more work.

You were promised time.

The time was real for one task.

Then it was spent on the next.

Not because you lacked discipline.

Because the cost of doing more dropped, and a system that always wanted more finally could ask for it.

The speed was the visible part.

The intensification was the product.

Bottom line — The first upgrade is not working faster. It is naming where the saved time actually went.

When AI saves you an hour, who decides where that hour goes — you, or the expectation that resets the moment you produce it?

That question is the difference between using AI and being intensified by it.

Spend part of the saved time on recovery, on purpose.
Refuse scope expansion that no one actually asked for.
Protect the boundary AI quietly erased.
Treat the saved hour as yours to allocate, not as free capacity.

Bottom line — You cannot control whether the task gets faster. You can decide whether the saved time becomes rest or just more work.

Closing line

What actually changed is not that AI freed your time. It is that AI moved your work toward higher scope and lower recovery — and moved the rewards toward those who already had the most.

Sources

Every figure in this draft is drawn verbatim from the following sources. Task-level gains, firm-level leakage, distributional effects, and worker risk each rest on the studies cited inline.