What If AI Makes Us More Productive but Gives Us More Work?
Artificial intelligence has been sold to organisations with a remarkably attractive promise. It will automate repetitive tasks, reduce administrative burden, accelerate routine work and return something increasingly scarce to employees: time.
Research from UC Berkeley’s Haas School of Business suggests that the reality may be considerably more complicated.
Researchers Xingqi Maggie Ye and Aruna Ranganathan spent eight months studying how generative AI changed everyday work inside a US technology company employing around 200 people. They found that employees did become capable of completing tasks faster, but the time they saved did not necessarily become free time. Instead, people worked at a faster pace, took responsibility for a wider range of tasks and allowed work to spread into more parts of their day.
The finding raises an uncomfortable question for organisations investing heavily in AI. What happens if the productivity revolution does not reduce work at all, but simply increases the amount of work we believe we can fit into a day?
The Researchers Watched AI Change Work From the Inside
The researchers did not begin by trying to prove that artificial intelligence increased workloads. Their objective was to understand how generative AI was changing everyday working practices.
Over eight months, Ye observed employees and spoke extensively with people across the organisation. The company had made commercial generative AI tools available to staff, but employees were not required to use them and managers had not imposed new AI-driven productivity targets.
That detail is important because the intensification of work did not simply originate from a manager demanding that everybody produce twice as much because they now had ChatGPT.
Employees themselves began discovering that tasks which had previously felt difficult, time-consuming or outside their job description suddenly appeared achievable.
AI lowered the barrier to starting the work.
Once that happened, people began doing more of it.
AI Filled Knowledge Gaps and Expanded Job Boundaries
One of generative AI’s most powerful workplace capabilities is its ability to help somebody operate outside their existing expertise.
An employee who might previously have needed assistance from another team could ask an AI system to produce a first draft. Someone who would normally outsource a task could attempt it themselves. Work that might previously have been postponed because nobody possessed the necessary knowledge could suddenly appear manageable.
From an organisational perspective, this can look like an extraordinary productivity gain because the same employee can apparently perform a wider range of activities.
The Berkeley researchers found precisely this expansion in the scope of work. Employees began taking on tasks that might previously have belonged to specialists, required additional support or simply not have been attempted.
However, expanding what somebody can attempt is not necessarily the same as expanding what they can do expertly.
This creates a new organisational problem because an AI-assisted first draft can appear substantially more complete than the employee’s underlying expertise actually is.
The work may still need checking, correcting and explaining by somebody who understands the subject properly.
AI can therefore reduce the friction involved in producing the first 80 per cent of something while creating new work around verifying whether the final 20 per cent is actually correct.
AI Can Move Work Rather Than Remove It
This distinction could become increasingly important as organisations measure the benefits of generative AI.
Suppose a junior employee previously needed a specialist to create a document. With AI, the junior employee can now produce the first version themselves.
On a productivity dashboard, that can look like a saving because one person has completed work that previously required another.
However, if the specialist subsequently spends significant time correcting the AI-assisted version, explaining its weaknesses and coaching the colleague who produced it, the work has not necessarily disappeared. It has moved.
This could create what might be described as coordination debt, where artificial intelligence makes it easier for more people to enter specialised workflows while increasing the amount of reviewing, checking and knowledge-sharing required around them.
The technology can therefore democratise capability while simultaneously increasing the burden placed upon expertise.
That is not an argument against AI. It is an argument for measuring productivity across the entire workflow rather than simply celebrating the speed of the first output.
Nobody Told Employees to Work More
Perhaps the most interesting aspect of the research is that the company did not need to impose this additional work.
Employees increased their own workloads.
One engineer explained that they had initially assumed improved productivity would save time and potentially allow them to work less. Instead, they found themselves working the same amount or more.
This reveals something important about the relationship between technology and work.
When a task becomes cheaper or easier to perform, organisations do not necessarily use the saving to reduce the amount of activity taking place.
They often use it to increase output.
Email is a useful historical example. Digital communication dramatically reduced the effort required to send written messages, but most office workers did not consequently spend less time communicating. The number of messages expanded.
Generative AI could create a similar effect across knowledge work.
When reports become easier to create, organisations may produce more reports. When presentations become easier to build, employees may create more presentations. When analysis becomes cheaper, managers may request more analysis.
AI reduces the cost of doing something, but that reduction can increase demand for the thing itself.
Starting Work Became Almost Frictionless
One of the subtler findings concerned the boundaries surrounding work.
Many traditional tasks contain natural barriers to starting. Someone needs to open several documents, gather information, think through a problem, call a colleague or find an uninterrupted period of time.
Generative AI changes that starting point because beginning a task can become as simple as typing a sentence into a prompt box.
That convenience can be extremely useful.
However, it can also make work easier to insert into moments that previously acted as natural pauses.
Employees in the Berkeley study began using generative AI during lunch breaks, meetings and other small gaps in the working day because it felt possible to advance another task with very little initial effort.
The boundary between working and not working therefore became less distinct.
A ten-minute gap that once allowed somebody to make coffee, speak to a colleague or simply think could become enough time to launch another AI task.
The problem is not that any individual prompt is particularly demanding.
The problem is what happens when every empty moment becomes another opportunity to produce something.
Productivity Can Consume the Pauses That Make Productivity Possible
This has important implications for how organisations think about efficiency.
Human beings are not processors that can simply remain at maximum utilisation throughout the working day.
Periods of apparent inactivity can serve important functions. People reflect, process information, recover attention, speak informally with colleagues and allow ideas to develop.
If AI makes it possible to fill every gap with another productive action, those spaces can begin disappearing.
The result can look exceptionally efficient when measured in outputs per hour.
It may be considerably less efficient when measured in judgement, creativity, learning or the ability to sustain performance over several years.
The researchers warn that unchecked AI-driven work intensification could contribute to cognitive fatigue, burnout and weakened decision-making.
That creates an important distinction between maximum productivity and sustainable productivity.
The two are not necessarily the same thing.
Multitasking Becomes Easier When AI Can Work in Parallel
Generative AI introduces another unusual feature into knowledge work because an employee can delegate several cognitive tasks simultaneously.
Someone can ask one AI system to analyse a spreadsheet while another conversation drafts an email and a third prepares research for a presentation. That can feel enormously productive.
However, the human being remains responsible for moving between those threads, interpreting the outputs, checking their quality and deciding what happens next.
The computer may be operating several tasks in parallel, but human attention is still switching between them. This creates the possibility of a new form of multitasking in which employees continuously launch, monitor and review AI-generated work. The bottleneck therefore moves.
Before generative AI, the limiting factor may have been how quickly somebody could produce the work. After generative AI, the limiting factor may increasingly become how much output the human being can evaluate.
This could make attention one of the most important workplace resources in the AI economy.
The Faster Rhythm Can Change Expectations Without Anyone Announcing It
Workplace pressure does not always arrive through an explicit instruction from management. It can emerge through norms.
If one employee begins using AI to complete work faster, colleagues see the faster turnaround. If an entire team begins doing this, rapid responses gradually become normal. Once rapid responses become normal, taking longer can start to look slow even if nobody formally changes the deadline.
The Berkeley research suggests that AI created this kind of self-reinforcing cycle. Faster work increased expectations around speed, increased reliance on AI and encouraged employees to take on a broader scope of tasks, which subsequently expanded the amount and density of work.
This is particularly important for organisational leaders because workplace pressure can therefore increase without appearing on any policy document. Nobody has to announce that everyone should work faster. The new pace simply becomes visible.
Employees adapt to what appears to be normal. Over time, the productivity gain can turn into the productivity expectation.
Productivity Gains Can Become Workload Creep
This phenomenon might best be described as workload creep. Someone begins using AI because it makes part of their existing job easier. They then use the time saved to complete another task. Because they can complete more tasks, their responsibilities gradually expand. As their responsibilities expand, they become increasingly dependent on AI to keep pace. Once AI-assisted productivity becomes normal, the increased workload becomes difficult to reverse. The employee may technically be more productive while simultaneously becoming busier. That distinction is crucial.
Productivity measures how much output is produced relative to the resources used. It does not automatically tell us whether somebody’s working life has improved. An organisation can therefore become more productive while its employees become more exhausted.
The Research Does Not Mean AI Makes Everyone Work More
The study deserves careful interpretation because it examines one approximately 200-person technology company rather than representing every workplace. Technology companies may be particularly receptive to experimentation, employees may be more comfortable with AI tools and organisational cultures can differ substantially between industries. The researchers also describe the work as ongoing rather than presenting it as definitive evidence about the entire labour market.
AI can genuinely reduce workload in some contexts. Automating repetitive administrative processes can remove tasks entirely. AI-assisted accessibility technologies can reduce barriers for disabled workers. Better summarisation can reduce time spent searching through documents, while automation can eliminate tedious manual processes. The Berkeley study therefore should not be interpreted as proving that AI cannot save time. It reveals something subtler. Saving time and giving people more free time are not the same outcome. An organisation decides what happens to the time that technology saves.
We Have Seen This Pattern Before
The history of workplace technology contains many examples of tools that increased efficiency without necessarily making work quieter.
- Email made communication faster but contributed to enormous communication volumes.
- Smartphones made colleagues easier to reach but extended the working day beyond the office.
- Cloud software made collaboration easier but also made work accessible almost everywhere.
- Generative AI could become the latest stage in this progression because it does not merely allow work to travel with us. It makes additional work dramatically easier to initiate.
The defining feature of AI-driven work may therefore not simply be automation. It may be acceleration. The question for organisations is whether they allow acceleration to become permanent intensification.
Companies May Need an AI Practice, Not Simply an AI Strategy
The Berkeley researchers argue that organisations need what they describe as an AI practice, which means developing intentional norms around the rhythm and boundaries of AI-enabled work rather than simply encouraging adoption. This is an important distinction.
Most organisational AI strategies concentrate on tools, licences, training, governance, data protection and return on investment. An AI practice would also ask how the technology changes human behaviour.
Teams might deliberately protect uninterrupted focus periods rather than continuously responding to AI-generated outputs. Non-urgent AI work could be batched rather than monitored constantly. Important decisions could include structured pauses that require employees to reconsider assumptions before acting on machine-generated recommendations.
Organisations could also protect opportunities for human discussion, reflection and informal conversation rather than allowing work to become increasingly individualised and mediated through AI interfaces. The objective would not be to slow down AI adoption. It would be to ensure that acceleration does not destroy the conditions required for good judgement.
We Need Better Measures of AI Productivity
This research also exposes a weakness in how many organisations currently measure the return on their AI investments. If a task that once took three hours now takes one hour, it is tempting to record a two-hour productivity saving.
However, that calculation becomes misleading if the employee uses those two hours to take on four additional tasks, works through lunch to review AI outputs and spends another hour correcting a colleague’s AI-generated work. The organisation may still be producing more.
However, the productivity story is more complicated than the original time saving suggests. Companies therefore need to measure more than individual task completion. They should examine total workload, working hours, error rates, review time, attention switching, quality, employee wellbeing and how responsibility moves between teams.
The most important productivity metric may eventually become not simply how much work AI helps us produce, but how much valuable work the organisation can sustainably absorb.
AI Could Make Expertise More Important, Not Less
There is another important consequence hidden inside the Berkeley findings.
When AI allows more people to attempt specialised tasks, experts may spend less time producing everything themselves and more time evaluating the work produced by others. This could fundamentally change professional roles.
A senior developer may write less routine code but spend more time reviewing AI-assisted code written by colleagues. A lawyer may draft fewer basic documents but review more AI-generated legal material. A communications specialist may write fewer first drafts but spend more time correcting messages produced across the organisation. Expertise therefore does not necessarily disappear, Its function changes.
The scarce skill becomes the ability to recognise what is wrong when something looks convincingly right. That could make verification, judgement and quality assurance increasingly important components of knowledge work.
The Real Question Is What We Do With the Time AI Saves
The productivity debate around artificial intelligence has concentrated heavily on whether a task takes 20 minutes instead of two hours.
- That is only half the question.
- The other half concerns what happens to the remaining hour and 40 minutes.
- The employee could finish earlier.
- They could spend more time developing professionally.
- They could talk to colleagues.
- They could undertake deeper work.
- They could recover.
- Alternatively, the organisation could fill the space with five more tasks.
- Technology does not make that decision.
- Workplace culture does.
This is why the Berkeley research should matter to managers as much as technologists. The real productivity consequences of artificial intelligence may depend less on what the model is technically capable of doing and more on what organisations decide that capability should mean for people’s workloads.
The UKBT Institute Opportunity
This question sits naturally within the UKBT Institute’s Living in a Digital World campaign because understanding AI at work requires considerably more than measuring model performance.
Computer scientists can explain what the systems can automate, while organisational researchers can examine how working practices change around them. Psychologists can investigate cognitive load and attention switching, while economists can examine productivity and labour-market effects. Designers can study how interfaces encourage constant interaction, while managers and employees can help determine what sustainable AI-enabled work should actually look like. The challenge is therefore not simply to build faster artificial intelligence.
It is to build better ways of working around artificial intelligence. That requires organisations to treat employees not merely as another component within an increasingly automated production system, but as human beings whose attention, judgement and capacity to recover remain finite.
The Bigger Picture
Generative AI undoubtedly removes friction from many forms of knowledge work.
The Berkeley study suggests that removing friction produces a consequence that deserves more attention.
- When starting becomes easier, we start more things.
- When producing becomes faster, we produce more things.
- When small gaps become sufficient to complete another task, work begins occupying those gaps.
- When everyone becomes faster, faster becomes normal.
- The danger is therefore not simply that artificial intelligence might take our work away.
- There is an equally important possibility that artificial intelligence gives us the ability to work almost continuously.
The productivity revolution will not automatically produce a shorter or calmer working life. Whether AI gives people time back will depend upon whether organisations deliberately allow some of the time it saves to remain saved.
The Berkeley researchers’ warning is therefore less about the technology itself than about the way human organisations respond to abundance. Artificial intelligence can make it easier to do more, but there will always be another report to write, another analysis to run, another message to send and another idea to investigate.
If organisations measure the success of AI solely by how completely those newly available moments can be converted into additional output, productivity gains may eventually become workload gains.
AI may reduce the friction of work, but unless organisations consciously protect the time that friction once occupied, the technology may simply teach us how to fill every available space with more work.