THE COST OF NOW: AI Is About To Make This Worse
I think AI is going to save us an extraordinary amount of time.
That's precisely what worries me.
Not because I think AI is bad.
Quite the opposite.
I use it.
I find it enormously useful.
There are things I can now research, structure, analyse, compare, draft or work through dramatically faster than I could a few years ago.
Used properly, it can remove enormous amounts of tedious work.
That is brilliant.
But we've just spent an entire essay looking at what human beings tend to do when technology makes work faster.
We don't necessarily go home.
We increase output.
So let's take a simple example.
Something currently takes three hours.
AI reduces it to thirty minutes.
There are two possible headlines.
AI Gives Worker Two And A Half Hours Back.
Or:
Worker Can Now Complete Six Times As Much Work.
Same technology.
Same productivity improvement.
Completely different future.
And I have a horrible feeling I know which headline many organisations will eventually prefer.
The efficiency gains are already real
We don't need to speculate about whether generative AI can improve productivity.
It already does in some tasks.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of a generative AI assistant among 5,179 customer-support agents.
Productivity, measured as issues resolved per hour, increased by around 14 per cent on average.
The effects were much larger for less experienced and lower-performing workers, while the most experienced employees gained far less. The researchers also found evidence of improved customer sentiment and employee retention. The study has since been published in The Quarterly Journal of Economics.
Another major experiment involved 758 Boston Consulting Group consultants.
For tasks sitting within what the researchers called AI's jagged technological frontier, access to GPT-4 increased task speed by more than 25 per cent, improved human-rated performance by more than 40 per cent and increased task completion rates.
In software development, a 2026 quasi-experiment involving programmers at Ant Group found use of a generative-AI coding tool increased code output by more than 50 per cent.
These are not theoretical future benefits.
They are already measurable.
AI can make some people substantially faster at some kinds of work.
Excellent.
Now what?
That's the question I don't think we're talking about enough.
Everyone is obsessed with how much time AI saves
You see the sales pitch everywhere.
Save hours.
Work faster.
Write faster.
Analyse faster.
Automate administration.
Summarise meetings.
Draft emails.
Create presentations.
Produce reports.
Generate ideas.
Handle routine queries.
Excellent.
But almost nobody finishes the sentence.
Save hours... for what?
That is the missing conversation.
Suppose an accountant saves five hours a week.
Does the accountant work five fewer hours?
Spend five more hours with clients?
Take on another twenty clients?
Produce more detailed advice?
Lower fees?
Increase margins?
Answer faster?
Complete work earlier?
Something will happen to those five hours.
The software cannot decide what.
The organisation does.
The customer does.
The market does.
The individual does.
Sometimes all four fight over it.
There is already evidence that AI can genuinely return time
This is important, because I don't want to manufacture a dystopian story when the evidence doesn't support one.
One of the most interesting AI workplace experiments so far involved 7,137 knowledge workers across 66 companies.
Half were randomly given access to a generative AI tool integrated into the applications they already used for email, meetings and writing.
During the second half of the six-month experiment, workers who were actively using the tool spent around two fewer hours each week on email.
They also reduced the time they spent working outside normal hours.
And the researchers did not find evidence that the quantity or composition of their other tasks immediately expanded to absorb all the saving.
Read that carefully.
Because it challenges the most pessimistic version of this essay.
AI saved time.
And at least during that experiment, some of the saving actually remained saved.
People spent less time on email.
Worked less outside normal hours.
That is exactly what I want technology to do.
It is possible.
Which means the problem is not inevitable.
The danger begins later.
When the productivity gain stops being novel.
When everybody knows what the technology can do.
When estimates change.
When clients know.
When competitors know.
When management knows.
When the person pricing the work knows.
When yesterday's three-hour job is now understood to be a thirty-minute job.
That's when the interesting bit starts.
First it saves time. Then everybody learns it saves time.
Imagine you employ someone to produce a weekly report.
It takes half a day.
They start using AI.
Same quality.
Now it takes an hour.
What do you do?
A particularly enlightened manager might say:
“Fantastic. Take the afternoon off.”
Maybe.
A commercially minded manager might instead think:
“If one report takes an hour, perhaps we don't need half a day allocated anymore.”
Reasonable.
Then perhaps:
“Could you produce reports for the other departments too?”
Also reasonable.
Then:
“Could we make these daily?”
Maybe.
Then:
“Could we add more analysis?”
Probably.
Then:
“If AI writes most of it, why does this still take an hour?”
There it is.
None of these questions is individually absurd.
That's what makes the ratchet powerful.
Once a task becomes quicker, the old allowance begins to look inefficient.
Not because anybody is evil.
Because the organisation has learned.
Capability becomes the new reference point.
The estimate changes before the working day does
This is how productivity gains get absorbed.
Suppose a consultant historically charges for a day because a particular piece of work genuinely takes a day.
AI reduces the production time to three hours.
For a while, that's fantastic.
They have gained capacity.
Then competitors adopt the same technology.
Now everyone can produce similar work in three hours.
The customer eventually discovers this.
The market discovers it.
Procurement discovers it.
The fee may reduce.
The turnaround may shorten.
The scope may increase.
What used to be:
“I'll have that to you next week.”
becomes:
“Can we have it tomorrow?”
Then:
“Could you get that across this afternoon?”
The technology saved time.
But instead of the worker owning the saving, the saving became part of the service expectation.
We've seen this movie before.
Email was faster than letters.
So communication expanded.
Computers were faster than typewriters.
So documentation expanded.
Spreadsheets were faster than manual calculation.
So analysis expanded.
Digital photography made photographs effectively free.
So we take thousands.
AI could do the same thing to knowledge work at a completely different scale.
The danger isn't AI replacing work
At least not for this essay.
There is an enormous separate debate about whether AI will remove jobs, create jobs or fundamentally restructure whole professions.
That matters.
But I'm interested in something slightly different.
What happens to the people whose jobs remain?
Suppose AI doesn't replace you.
Suppose it simply makes you twice as productive.
That sounds safe.
Until your workload is recalibrated around being twice as productive.
Now your original job hasn't disappeared.
It has doubled.
That is a form of disruption we talk about far less.
You still have the career.
You still have the salary.
You still have the desk.
You just now operate at a level of throughput that would previously have required two people.
The productivity statistic looks incredible.
The human experience may depend entirely on what happened to the saved capacity.
Knowledge workers may already be reinvesting the saving
A fascinating 2025 study interviewed 21 knowledge workers who regularly used generative AI professionally.
The researchers were investigating whether people reduce their own effort when AI becomes involved.
Generally, they didn't simply sit back.
Participants described learning how to steer the AI, checking its work and strategically reallocating their effort.
Most strikingly for this essay, the researchers found that workers often reinvested saved effort into other parts of the task rather than simply keeping the saving.
They became, in the researchers' language, increasingly responsible for administering and directing the process.
That makes perfect sense.
Imagine AI gives me a good first draft in ten minutes instead of me spending ninety minutes producing one.
I now have choices.
Stop.
Or improve it.
Research another angle.
Challenge it.
Rewrite parts.
Check evidence.
Make the argument sharper.
Create another version.
Explore five alternatives instead of one.
AI has saved effort.
But it has also made previously expensive improvements much cheaper.
So quality expectations can rise.
This is another way saved time disappears.
Not into more tasks.
Into deeper ones.
That can be excellent.
But again, it isn't necessarily leisure.
AI doesn't only reduce work. It creates new work.
This is already appearing in research.
A 2026 systematic review examined 30 empirical studies on generative AI and academic workload.
The conclusion was not:
AI reduces workload.
Nor was it:
AI increases workload.
The researchers described workload transformation.
AI could reduce the time needed for routine activities such as preparing materials or basic grading.
At the same time, it created or increased other responsibilities:
checking AI output,
evaluating quality,
dealing with academic integrity,
making ethical judgements,
learning new systems,
and integrating AI safely into existing work.
Whether AI became a genuine resource or an additional burden depended partly on individual capability and partly on organisational factors such as training, policy clarity and infrastructure.
This is crucial.
AI doesn't simply delete a box from the workflow.
It redraws the workflow.
Some things disappear.
Some shrink.
Some grow.
Some completely new tasks appear.
Drafting may get easier.
Verification may become more important.
Producing options may get easier.
Choosing between them may become harder.
Generating content may get cheaper.
Determining whether any of it is correct becomes more valuable.
The job changes shape.
There is a verification tax
This is something anyone who uses AI seriously discovers quite quickly.
Fast does not mean finished.
AI can produce an answer in seconds.
That is extraordinary.
But if the output matters, somebody still needs to ask:
Is this correct?
Is that source real?
Has it misunderstood the context?
Did it miss something important?
Is the argument coherent?
Does this comply with the rules?
Is the calculation right?
Would I put my name on this?
AI reduces the cost of producing plausible material.
That can increase the importance of judgement.
A 2026 systematic review of generative AI in organisational decision-making found highly uneven effects depending on the type of task, highlighting the difficulty humans face in calibrating when AI output should be trusted. The researchers pointed particularly to the problem that AI can produce similarly confident-looking output regardless of underlying accuracy.
That is why human expertise doesn't simply vanish when drafting becomes cheaper.
Sometimes it moves downstream.
From producing to evaluating.
And evaluation is work.
There is something else weird happening: we may feel faster even when we aren't
This is one of my favourite AI findings because it is such a perfect warning against assuming.
METR ran a randomised controlled trial with experienced open-source software developers working on real issues in projects they knew extremely well.
The developers expected AI tools to make them around 24 per cent faster.
After completing the study, they still believed AI had made them roughly 20 per cent faster.
It hadn't.
In that specific early-2025 setting, allowing AI tools caused them to take 19 per cent longer.
That's extraordinary.
AI felt faster.
The developers believed it was faster.
Experts had predicted gains.
Actual measured completion time went the other way.
Before anyone turns this into “AI doesn't make programmers faster”, don't.
The researchers explicitly warned against doing that.
There were only sixteen developers.
They were highly experienced.
They were working in mature codebases they knew extremely well.
AI capabilities have also changed rapidly since.
In February 2026, METR said its newer data was too affected by participant-selection problems to provide a reliable estimate, while noting that current tools likely provide greater speedups than the earlier generation.
The lesson is not that AI makes developers slower.
The lesson is:
Measure the gain.
Don't simply assume it because the technology feels impressive.
AI productivity is jagged
That phrase from the BCG research matters.
Jagged technological frontier.
AI can be astonishingly good at one task and unexpectedly weak at another that appears similar.
That means:
“AI makes knowledge workers 30 per cent faster”
is almost meaningless without asking:
Which worker?
Doing what?
Using which model?
How experienced are they?
What quality threshold?
How much checking?
What happens when the AI is wrong?
Customer-support research found much larger gains for less experienced workers than experienced ones.
The BCG experiment found strong gains for tasks inside AI's capabilities, while also warning that reliance becomes dangerous beyond that frontier.
The developer experiment found a slowdown in a particular high-expertise real-world setting.
So we should be suspicious of blanket claims.
AI does not save “time”.
It saves time on particular activities under particular conditions.
And the surrounding work then adapts.
But suppose the gains keep getting bigger
This is where the title of the essay earns itself.
Imagine current systems improve significantly.
Not science fiction.
Just better versions of what already exists.
An hour of basic research becomes ten minutes.
A first draft becomes seconds.
A day's coding becomes two hours.
Meeting notes become automatic.
Routine customer communication becomes mostly automated.
Presentations generate themselves from source material.
Invoices classify themselves.
Data gets analysed on request.
Documents compare themselves.
Quotes draft themselves.
Reports populate automatically.
What happens?
We could genuinely create an enormous amount of human capacity.
That is exciting.
And if we aren't deliberate, we will immediately fill every fucking minute of it.
You will be compared with people using AI
This changes the incentive.
Perhaps you personally decide:
“I am going to use AI to reduce my working hours.”
Fine.
But somebody else may use it to increase output.
And now the comparison begins.
You produce three proposals.
They produce eight.
You write one considered article.
They publish five.
You analyse ten accounts.
They analyse thirty.
You respond tomorrow.
They respond in five minutes.
You retain the old service level while they use AI to exceed it.
Customers notice.
Managers notice.
Algorithms notice.
Markets notice.
That does not automatically mean they win.
Quality matters.
Trust matters.
Originality matters.
Relationships matter.
But productivity improvement creates competitive pressure even for people who would personally prefer to keep the time.
That is why this cannot be solved entirely through individual discipline.
There is a structural component.
The first organisation to keep the old workload could become very interesting
Imagine an employer says:
“We introduced AI and our people can now complete the core workload in thirty hours rather than forty.
We're keeping the same salaries.
Enjoy the ten hours.”
Would productivity collapse?
Would people become lazy?
Would they become better?
Would the company retain excellent staff?
Would sickness fall?
Would people think more clearly?
Would they use some of the capacity voluntarily to improve work?
I don't know.
Those are empirical questions.
But it is interesting how rarely that version of the AI future dominates the corporate conversation.
We hear:
Unlock productivity.
Increase output.
Scale.
Grow.
Do more with less.
Very rarely:
Keep output roughly where it is and give people some life back.
Why?
Because organisations do not generally optimise for leisure.
They optimise around their objectives.
That is not immoral.
It's simply why time saved at work will not automatically become employee time.
If we want that outcome, somebody has to choose it.
We may create a new version of presenteeism
Old presenteeism was partly about hours.
First in.
Last out.
Still at your desk.
Look committed.
AI could create a stranger version.
Output presenteeism.
How much did you produce?
How fast?
How many clients?
How many documents?
How many campaigns?
How many analyses?
How many pieces of content?
If AI makes output cheap, visible output may become almost meaningless.
But that won't necessarily stop organisations measuring it.
Someone who produces a hundred AI-assisted documents can look extraordinarily productive.
Whether anyone needed those hundred documents is another question.
We could use AI to remove pointless work.
Or use it to manufacture pointless work at unprecedented speed.
Both futures are technically available.
Cheap output can create expensive attention problems
This matters beyond the worker using AI.
Suppose AI allows me to create ten times more content.
Fantastic.
Who reads it?
Suppose every manager can generate detailed reports instantly.
Who absorbs them?
Suppose every salesperson can create highly personalised outreach at scale.
Who receives it?
Suppose every employee can generate a polished twelve-page proposal instead of a concise three-page one.
Who reviews the extra nine pages?
AI can dramatically reduce the cost of sending work into the world.
But it does not automatically increase the amount of human attention available to receive it.
That creates an asymmetry.
Production can become near-infinite.
Human attention remains stubbornly finite.
This is precisely the problem we've already created with digital communication.
Email made sending cheap.
So we sent vastly more email.
AI could reduce the cost again.
We may save five minutes writing a message and create twenty minutes of reading for everyone copied into it.
That isn't productivity.
That's externalising effort.
The person with AI can create work faster than the person downstream can consume it
Think about management.
A senior manager can use AI to produce:
three strategy documents,
six project briefs,
detailed feedback,
twenty action points,
a competitor analysis,
a new reporting framework,
and five ideas for restructuring the department.
Before lunch.
Amazing.
Their team still has one nervous system each.
This is where AI productivity becomes particularly interesting.
It can disproportionately accelerate the creation of requests.
The person assigning work may become far more productive at assigning it than the recipient becomes at completing it.
A manager can generate ten ideas in seconds.
A team may spend six months implementing one.
An executive can ask AI for fifty useful metrics.
Somebody has to collect, interpret and act on them.
A customer can generate an extremely detailed request.
A professional still has to determine what actually matters.
We could accidentally create the most sophisticated workload-generation machine ever built.
AI could industrialise “Can You Just...?”
Can you just analyse these options?
Can you just produce three versions?
Can you just benchmark it against competitors?
Can you just summarise the last five years?
Can you just create a presentation as well?
Can you just give me a detailed breakdown?
Can you just personalise it for every customer?
Historically, some requests didn't happen because they were obviously too expensive in human time.
AI changes that.
And that's mostly useful.
But cost acted as a filter.
Effort acted as a filter.
Time acted as a filter.
If those barriers collapse, we may stop asking:
Is this worth doing?
Because doing it feels nearly free.
The generation may be free.
The consequences aren't necessarily.
Someone still reviews it.
Stores it.
Discusses it.
Makes decisions from it.
Maintains it.
Receives it.
Acts on it.
The easiest work to produce may become the hardest work to escape.
AI could make bullshit spectacularly efficient
We should probably just say it.
Every organisation contains work that shouldn't exist.
Reports nobody uses.
Meetings nobody needs.
Documents produced because they've always been produced.
Updates copied between systems.
Presentations built to describe things everyone already knows.
Layers of internal communication.
AI can automate much of this.
Good.
Or it can make it so cheap that we produce more of it.
That is the choice.
If the weekly report takes three hours, eventually someone asks whether the report is worth three hours.
If AI produces it in thirty seconds, nobody asks.
Keep the report.
Add another one.
Now everyone has twelve dashboards they don't read.
The real productivity gain would have been deleting the pointless thing.
Instead we automated the pointlessness.
There is an important distinction between:
doing work faster
and
not doing unnecessary work at all.
The second still requires human judgement.
AI should increase agency, not simply throughput
This is the opportunity I find genuinely exciting.
AI can remove the cognitive rubbish that prevents people thinking.
The repetitive administration.
The formatting.
The retyping.
The searching.
The basic synthesis.
The blank-page problem.
The low-value first pass.
If that creates more capacity for:
judgement,
creativity,
relationships,
research,
thinking,
decision-making,
presence,
learning,
then we have built something extraordinary.
That aligns almost perfectly with how I've come to think about agency.
Work ethic asks:
How hard can I work inside this system?
Agency asks:
Is there a better system?
AI gives us an incredible opportunity to ask the second question.
But if we simply use it to do the existing shit faster, then increase the volume of the existing shit, we've missed most of the opportunity.
Productivity isn't the same as value
This will matter more as AI improves.
Suppose Person A produces twenty reports.
Person B produces five.
Who is more productive?
Depends.
Did anybody need the reports?
Were they correct?
Did they improve a decision?
Did they prevent a mistake?
Did they generate revenue?
Did they help somebody?
Did they change anything?
Output is easy to count.
Value is harder.
AI will make output extremely cheap in many knowledge-work domains.
That should make us less obsessed with output, not more.
Otherwise we will reward humans for operating the machine hardest.
And that will create the same productivity trap at extraordinary scale.
There is another possible future
Let's not lose sight of it.
The same 7,137-worker field experiment that should make employers excited about AI productivity should make employees excited too.
Less time on email.
Less work outside normal hours.
That is not a minor outcome.
That's life.
Two fewer hours staring at email every week is roughly a hundred hours over a working year.
Not all of that necessarily becomes leisure, and the experiment did not establish what will happen after years of organisational adaptation.
But it demonstrates something incredibly important.
Time can remain saved.
We are not condemned to fill every gain.
We can design differently.
Perhaps AI needs a productivity dividend
We understand the concept of a financial dividend.
A system performs well.
Value is created.
Some of it gets distributed.
Maybe we need a similar way of thinking about time.
AI improves productivity by 20 per cent.
How is that gain distributed?
Some to the company.
Some to customers.
Some to quality.
Some to growth.
And perhaps some to the human being.
Not necessarily through everyone working four-day weeks tomorrow.
But through deliberately refusing to consume every available efficiency gain.
Fewer pointless meetings.
Shorter administrative days.
Protected thinking time.
More realistic capacity.
Better recovery.
Earlier finishes occasionally.
Space.
The gain doesn't have to disappear simply because the organisation could theoretically convert it into additional output.
Leaders will decide a lot of this
Imagine two managers implementing exactly the same AI tool.
Manager one sees:
“My team can now do 20 per cent more.”
Manager two sees:
“My team can now achieve the same output with 20 per cent less friction.”
Both are factually correct.
Their teams may have completely different futures.
One increases targets.
The other removes pressure.
Perhaps both do a bit of each.
That is why technology does not determine working culture.
Management does.
The tool creates options.
Leadership decides which options become expectations.
And once expectations rise, they rarely fall easily.
We need to capture the baseline before AI moves it
This may be one of the most practical things organisations can do.
Before implementing an AI system, measure:
How long does the work currently take?
How much of that is genuinely valuable?
How much is repetitive?
What quality are we producing?
How much overtime exists?
How much capacity do people actually have?
What are customers currently promised?
Then implement AI.
Measure again.
Now we can see the gain.
Without a baseline, efficiency simply disappears into a new normal.
Six months later nobody remembers that the report once took half a day.
It now takes an hour.
So an hour becomes “how long the report takes”.
The saving has vanished conceptually.
Once forgotten, it becomes very easy to spend.
We need to decide what success means before the technology decides for us
This is the question.
If AI makes your team 30 per cent more productive, what would a successful outcome look like?
Thirty per cent more output?
Twenty per cent more output and slightly less pressure?
Higher quality?
More customers?
Lower prices?
Higher margins?
Shorter weeks?
Better thinking?
Less administration?
Some combination?
There isn't one correct answer.
Businesses exist to create value.
People want careers.
Customers want improvements.
Shareholders may reasonably expect returns.
But unless the desired allocation of productivity gains is discussed consciously, the easiest default is obvious:
more.
More becomes the path of least resistance.
And more has no finishing line.
AI is about to make this worse
Not because AI wants anything.
It doesn't.
Not because productivity is bad.
It isn't.
Not because every employer will become exploitative.
They won't.
AI is about to make the temptation worse.
The temptation to believe that because something can be done faster, it should always be done faster.
The temptation to raise output every time capacity appears.
The temptation to expand scope because generation is cheap.
The temptation to send more because writing is effortless.
The temptation to demand more because the old workload suddenly looks small.
The temptation to compress deadlines because AI exists somewhere inside the process.
The temptation to confuse machine speed with human capacity.
We've already spent centuries doing this with previous technologies.
AI may simply accelerate the cycle.
The question isn't whether AI saves time
It will.
It already does.
The question is whether the person whose time was saved gets to keep any of it.
That's political at the level of society.
Commercial at the level of markets.
Cultural at the level of organisations.
And personal at the level of our own behaviour.
I can use AI to create twice as much.
Or I can use it to create the same amount and go for a walk.
Neither choice is inherent in the technology.
That's the point.
Three hours becomes thirty minutes.
There are two and a half hours sitting there.
For a moment.
Before anybody notices.
Before a customer claims them.
Before management targets them.
Before another task fills them.
Before I fill them myself because apparently an empty afternoon means I'm not trying hard enough.
That may be the real opportunity AI gives us.
Not simply the ability to work faster.
The chance to decide, perhaps more consciously than ever before, what productivity is actually for.
Because if the most powerful time-saving technology we've ever created merely teaches us to expect six times as much from the same human life, then AI will have worked perfectly.
And we will still somehow have no time.
Sources And Further Reading
Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, “Generative AI At Work”, The Quarterly Journal Of Economics (2025; earlier NBER working paper).
The study followed 5,179 customer-support agents during the staggered introduction of a generative AI assistant. Access increased issues resolved per hour by around 14 per cent on average, with substantially larger gains among novice and lower-performing workers and much smaller effects among the most experienced employees.
Fabrizio Dell'Acqua et al., “Navigating The Jagged Technological Frontier” (Harvard Business School / Boston Consulting Group field experiment).
The experiment involved 758 BCG consultants. For tasks within the demonstrated capabilities of GPT-4, AI use increased speed by more than 25 per cent, increased human-rated performance by more than 40 per cent and increased completion rates. The work also established the influential idea of a “jagged technological frontier”, where AI performance varies significantly between apparently related tasks.
Eleanor W. Dillon, Sonia Jaffe, Nicole Immorlica and Christopher T. Stanton, “Shifting Work Patterns With Generative AI”, NBER Working Paper 33795 (2025, revised November 2025).
A six-month randomised field experiment across 66 firms and 7,137 knowledge workers found active AI users spent around two fewer hours per week on email in the second half of the experiment and reduced work outside regular hours. Researchers did not detect significant changes in task quantity or composition resulting from individual AI access over the study period.
Leonardo Gambacorta, Han Qiu, Shuo Shan and Daniel M. Rees, “Generative AI And Labour Productivity: A Quasi Experiment On Coding” (2026).
Using variation in exposure to Ant Group's CodeFuse assistant, the researchers estimated that use of generative AI increased code output among the studied programmers by more than 50 per cent.
Feiya Su, Wen Fen Beh and Kenny S. L. Cheah, “Efficiency Or Intensification? A Systematic Review Of Generative AI And Academic Workload In Higher Education”, Acta Psychologica (2026).
A systematic review of 30 empirical studies concluded that generative AI produces workload transformation rather than a uniform reduction or increase. It can improve efficiency in routine work while introducing additional responsibilities involving quality assurance, ethical judgement, academic integrity, learning and integration.
Lucas Memmert, Daria Soroko and Eva Bittner, “From Effort Reduction To Effort Management: An Expectancy Theory Perspective On Professionals' Work Practices With Generative AI” (2025).
Qualitative interviews with 21 knowledge workers found that professionals did not simply reduce effort when using AI. They frequently reallocated or increased effort, learned to steer AI and reinvested saved effort into other parts of their work.
Systematic review of generative AI and organisational decision-making (2026).
A synthesis of 15 studies involving 15,752 participants reported highly heterogeneous performance effects and emphasised calibration problems created by AI systems producing confident outputs even when underlying reliability differs. The review found effects varied significantly across different decision phases.
Joel Becker, Nate Rush, Beth Barnes and David Rein, METR, “Measuring The Impact Of Early-2025 AI On Experienced Open-Source Developer Productivity” (2025).
In a randomised controlled trial involving 16 experienced developers and 246 real tasks in mature open-source repositories, AI use increased completion time by 19 per cent despite participants believing it made them faster. The researchers explicitly caution against generalising the result to software development overall.
METR, “We Are Changing Our Developer Productivity Experiment Design” (2026).
METR's follow-up work suggested later AI tools may produce greater developer productivity benefits, but strong participant-selection effects prevented reliable measurement. This underlines both rapid capability change and the methodological difficulty of producing stable estimates of AI productivity.
Systematic literature review of generative AI, workload, efficiency and labour productivity (2025).
A review of 74 studies across seven industries identified recurring effects involving workflow automation, decision support, human-machine interaction and communication while also highlighting unresolved issues involving employment, ethics, privacy and information quality.
Research Gaps
There is not yet strong evidence that organisations universally respond to AI productivity gains by increasing workloads proportionately.
This is the central future risk explored in the essay, but it should remain framed as a plausible organisational response rather than an established universal effect. The 7,137-worker field experiment actually provides an important counterexample: workers saved email time and reduced after-hours work without detectable immediate increases in task quantity.
AI capabilities are changing faster than much of the research cycle.
A productivity experiment using early-2025 models may describe those systems accurately while becoming a poor estimate of tools available in late 2026. Findings should therefore be dated and tied to the technology actually tested rather than treated as permanent properties of generative AI.
Productivity effects vary enormously by task and expertise.
Research already shows strong positive effects in some customer service, consulting and coding settings alongside negligible or even negative effects elsewhere. There is no defensible universal percentage for “AI productivity improvement”.
The METR developer finding should be interpreted narrowly.
The 19 per cent slowdown involved sixteen highly experienced open-source developers working on mature repositories they knew extremely well with early-2025 AI tools. It does not establish that AI slows programmers generally. METR's subsequent experiment also encountered selection problems severe enough that the researchers did not regard the newer estimate as reliable.
Output and productivity are not necessarily the same thing.
Producing more code, documents, emails or content may increase measurable output while failing to increase economic or organisational value. Future research needs better measures of quality, downstream consequences, rework and decision impact.
The long-term organisational adaptation to AI remains largely unknown.
Most field experiments study months rather than years. Employees and managers may behave differently once AI use becomes universal, job descriptions change, staffing models adapt and customers recalibrate turnaround expectations.
Time savings may be allocated in many different ways.
An organisation could convert AI gains into greater output, improved quality, lower prices, higher profit, shorter hours, increased training, reduced staffing or some mixture. Current evidence cannot establish which distribution will dominate across economies or industries.
AI can generate new workload as well as removing old workload.
Verification, governance, model selection, training, prompt development, security, integration and quality control can all create work. Evidence from higher education already supports workload transformation, but more sector-specific research is needed.
Self-reported productivity can be unreliable.
The METR experiment demonstrates that workers may believe AI made them faster when measured completion time suggests otherwise. Conversely, objective speed measures can also miss benefits such as reduced frustration, learning, creativity or improved quality. Mixed measurement approaches are needed.
Customer expectations have not yet been studied sufficiently.
A central argument in this essay is that faster AI-enabled production may eventually shorten expected turnaround times. This is plausible given the broader literature on reference points and service expectations, but direct longitudinal research on AI-driven customer expectation changes remains limited.
AI-assisted work may change professional skill development.
If junior workers increasingly delegate drafting, research, coding or analysis to AI, they may gain capability faster through AI guidance or lose opportunities to develop foundational skills themselves. Existing research points in both directions and the long-term balance remains uncertain.
Competitive pressure will vary substantially by market.
Some sectors may rapidly pass productivity gains into lower prices and faster delivery. Others are protected by regulation, bespoke work, scarcity, trust, relationships or professional liability. The essay's competitive-ratchet mechanism should not be treated as equally strong everywhere.
There is not yet enough evidence to know whether AI will increase or reduce overall working hours.
Early evidence includes genuine reductions in after-hours work alongside concerns around intensification. Long-term outcomes will depend heavily on labour markets, regulation, organisational culture and how productivity gains are distributed.
The idea of a “time dividend” is normative rather than an established economic framework.
The proposal that some productivity gains should deliberately be retained as human time is an argument about the purpose of productivity. Research can measure consequences of different arrangements, but it cannot determine the socially correct distribution of the gains.