Work Did Not Shrink but Grew Denser: Berkeley Study Tracks AI at Work for Eight Months

Generative AI is often expected to free up employees' time. After observing a technology company for about eight months, two researchers at the Haas School of Business, University of California, Berkeley (UC Berkeley Haas) reached the opposite conclusion: AI did not reduce work; it made work more intense.

The study was carried out by Aruna Ranganathan, associate professor of management at Haas, and Xingqi Maggie Ye, a doctoral student. It appeared in Harvard Business Review on 9 February 2026 under the title "AI Doesn't Reduce Work—It Intensifies It". The details of the study design, findings and recommendations in this article are drawn from a Haas press release dated 18 February 2026 and from reporting by the technology outlet The Register.

One company, about 200 staff, more than 40 interviews

The study looked at a US technology company with about 200 employees, who were free to use generative AI tools widely. According to The Register, the company did not require staff to use AI and simply provided enterprise subscriptions to commercially available AI tools.

The two researchers used an ethnographic approach: observing on site, sitting in on meetings and chatting with employees day to day. They also conducted more than 40 semi-structured interviews across different functional groups. The Register says the roughly 40 people observed and interviewed came from engineering, product, design, research and operations.

The limits need to be stated first. By the authors' own account, as of the Haas release in February 2026 this was still research in progress, and it covered only one company. It is a qualitative study of a single company, describing what happened in one workplace, and should not be applied directly to other industries or organisations.

Three ways work intensified

The study found that AI did not lower workloads. Instead it widened the work that employees felt able, and willing, to take on. The researchers sum up this intensification in three forms:

• Wider task scope: employees took on work that used to belong to others, or that had never been attempted.
• Work spilling into breaks: work seeped into time that used to be rest or time after hours.
• Multitasking: several AI-assisted tasks were pushed forward at the same time.

Notably, most of these changes were self-initiated. According to the Haas press release, the researchers described employees widening their remit often without anyone asking them to. Stark Insider's account likewise says no one was asked to raise output.

The Register's report adds detail: AI made tasks easier to start, so employees were more willing to pick up work from other roles. Engineers, in turn, spent time checking junior colleagues' errors, coaching colleagues who worked by "vibe coding", and finishing projects that others had left unfinished.

Energising to use, more tiring to look back on

The paradox lies in how it felt. The Haas press release says employees often felt motivated, even excited, when interacting with AI individually, but felt busier and found it harder to step away when they looked back over the whole day.

Stark Insider reports that some employees described constantly juggling several strands of work at once, and that others became quality gatekeepers for AI assistants that produce a lot but make mistakes easily.

The authors warn that expectations may reset over time, extra effort gradually becomes the norm, and output and demands push each other up in a cycle. Constant task switching and too little recovery time may weaken judgement and increase errors. They also make it harder for organisations to tell real productivity gains from unsustainable intensity.

The authors' three-part AI practice

The researchers' answer is to build an "AI practice", a set of rules for using AI, with three elements:

1. Pause: build a short, structured pause in before major decisions, to test the assumptions behind them and tie them to goals. The Register gives an example: before finalising, require one counter-argument and a clear statement of how the decision relates to company goals.
2. Sequencing: batch non-urgent AI output, protect focus time, and avoid reacting immediately to every output. The Register paraphrases this as organising projects in phases and deliberately controlling the pace instead of following the speed of AI.
3. Human grounding: hold regular short check-ins and arrange shared reflection and conversation, to offset the effect of long stretches of using AI alone.

The researchers stress that the aim of these arrangements is not to slow innovation. The Register sums up the advice as: make the team lead the AI, rather than letting the AI lead the team.

Applying it yourself: a few things to try (editor's suggestions)

The following are practices this publication has put together from the three principles above. They are suggestions, not research findings.

• Before finalising a decision, write down one counter-argument, then one sentence on which goal the decision serves.
• Handle non-urgent AI output in fixed time slots, and keep the rest of the time for work that needs focus.
• Move team projects forward in phases, with people setting the pace rather than following however fast AI delivers.
• Keep short, regular team conversations about how each person uses AI and what they have checked.
• Check rest and after-hours time regularly, to see whether work has quietly seeped in.

What the study points to is this: the efficiency AI brings does not necessarily turn into more rest. Protecting your time means setting the rules in advance.

Prices and details correct as of November 2026

Compiled by the F50 Editorial Team from public sources, for general reference only and not professional advice. Details are correct as of the publication date and may change.