GitHub Copilot developers did not simply produce more code after gaining access to generative AI. They changed how they spent their working time. A Harvard Business School working paper examining open-source developers argues that AI can reorganise the mix of tasks inside a job, pushing people towards core production work and away from project management and peer interaction.
That matters because the workplace debate is often framed around a simpler question: does AI increase productivity or replace jobs? The study points to a third outcome. A role can remain recognisably the same while its internal rhythm changes, with less time spent coordinating, reviewing and dealing with peripheral work.
What changed for GitHub Copilot developers
MIT Sloan’s summary of the research says the team followed about 187,000 developers on GitHub before and after access to Copilot expanded among maintainers of highly ranked open-source projects. The researchers compared people who received free access with those who did not, using the platform’s visible activity to examine how their work changed over time.
The developers were not doing one homogeneous task. Their activity included writing code, merging contributions, reviewing changes, handling issues and managing projects. That makes the setting useful for studying work composition. Instead of relying on a survey asking whether people felt more productive, the researchers could observe changes in the kinds of activity appearing on the platform.
More coding came with much less project management
According to MIT Sloan, developers with access to Copilot increased the share of coding activity by 12.4% and reduced project-management activity by 24.9%. The study also reports a nearly 80% reduction in peer collaboration among Copilot users. Those figures are large enough to raise a question that goes beyond software development: when AI removes friction from core work, what happens to the surrounding human interactions?
One interpretation is positive. Maintainers of busy projects can be overwhelmed by reviews, issue handling and coordination. If AI helps people produce more accurate code or solve routine problems independently, they may need less back-and-forth. That can leave more time for the part of the job they value and reduce the administrative burden that often grows around expertise.
Less collaboration is not automatically better
There is another side. Collaboration is not only overhead. It is also how knowledge travels through an organisation. A code review can catch mistakes, but it can also teach a junior developer why a decision was made. A discussion around an issue can expose assumptions that would otherwise stay hidden. If AI lets individuals work more independently, some of those informal learning moments may disappear.
The study does not establish that an 80% reduction in peer collaboration is harmful. It shows a behavioural shift in this particular setting. The effect could be beneficial for overloaded maintainers while creating different trade-offs for a company where teamwork, mentoring and shared understanding are central to long-term performance.
AI can change the shape of a job before it changes headcount
This is consistent with a broader theme in workplace research. LiveAIWire has covered evidence that AI adoption can change job enrichment and employee experience rather than simply producing a yes-or-no automation outcome. Jobs are collections of tasks, and those tasks do not all respond to technology in the same way.
A coding assistant makes code production cheaper in time. That changes the relative cost of everything around it. If a developer can produce a working change quickly, waiting for a review or coordinating across a team becomes a larger share of the remaining process. Organisations may then redesign those surrounding steps, sometimes deliberately and sometimes simply because behaviour drifts.
The biggest effect may be on less-experienced workers
MIT Sloan reports that less-experienced developers in the sample showed the largest increase in time devoted to coding. That is important because AI is sometimes presented as a reason to reduce entry-level hiring. The alternative interpretation is that tools such as Copilot can help newer workers spend more time practising the central skill of the job.
LiveAIWire has also examined research suggesting that AI can narrow some workplace performance gaps. If junior staff can tackle tasks sooner, the organisation may gain productivity while accelerating skill development. But that benefit depends on people still receiving feedback, context and opportunities to understand why the generated answer is right or wrong.
A flatter hierarchy may have hidden costs
The working paper suggests generative AI could flatten organisational hierarchies by reducing coordination and management demands. That possibility is attractive to firms that want faster decisions and fewer layers. It also raises a difficult question: which managerial activities are waste, and which are the mechanisms that maintain quality, accountability and shared direction?
AI can make a person more self-sufficient without making them more aware of every consequence of their work. A developer may need fewer routine questions answered by a colleague but still benefit from architectural review, security oversight or product context. The risk is treating every decline in coordination as proof of efficiency.
Open-source GitHub is a useful setting, but not every workplace
The evidence comes from developers whose activity can be observed on GitHub, including maintainers of prominent open-source projects. That gives the research unusually detailed behavioural data, but it also limits how confidently the result can be applied to other jobs. A sales team, legal department or hospital does not leave the same trail of visible task activity.
Copilot use in open-source development also differs from a company-wide rollout where managers can redesign workflows, mandate tools or measure performance differently. The study is best read as strong evidence that access to generative AI can change task allocation, not as a universal estimate that every AI user will collaborate 80% less.
The hidden-work problem becomes more visible
AI projects frequently create or move work that is easy to miss. LiveAIWire has covered the hidden work that appears when AI projects scale, including review, checking and exception handling. In the Copilot case, some surrounding work appears to shrink. In other settings, new verification duties may grow instead.
Managers therefore need to measure more than output. If code volume rises, they should also watch review quality, defect rates, mentoring, knowledge sharing and the distribution of expertise. A productivity gain can be genuine while still weakening a capability the organisation will need later.
GitHub Copilot developers show why job design is the next AI question
The most useful lesson is not that coding assistants are good or bad. It is that AI changes the price of individual tasks inside a job. Once one activity becomes much faster, people naturally reorganise their time around the new constraint. That can make work more satisfying, reduce busywork and let skilled staff concentrate on the activity they value most.
It can also make collaboration, mentoring or oversight easier to neglect because those tasks do not accelerate at the same rate. The next generation of workplace AI decisions will therefore be about job design as much as model choice. Companies that only count hours saved may miss the more important change: what their people stop doing once the tool becomes part of the job.
Managers should watch what disappears as well as what speeds up
The easiest AI metric is usually output: more commits, faster drafts, shorter handling time. The harder measurement is the activity that vanishes. Some disappearing work is genuinely wasteful, such as repetitive status checking or mechanical formatting. Other activities carry information that only becomes visible when they are gone, including informal review, coaching and the conversations that expose a misunderstanding early.
For software teams, that means combining productivity data with measures of quality and resilience. Are defects rising? Are fewer people able to explain a critical part of the codebase? Are junior staff becoming independent because they learned faster, or because they stopped asking questions? The same reduction in collaboration can represent healthy autonomy in one team and dangerous isolation in another.
The study is about task allocation, not a verdict on Copilot
The working-paper status also matters. The research is substantial, but it is not a final universal judgement on AI coding assistants. The observed changes followed a particular access programme on GitHub and a particular period in the tool’s development. Newer models, different organisational rules and paid commercial environments could produce a different balance of coding, coordination and review.
What survives those caveats is the central mechanism: when a technology makes one category of work cheaper in time, people reallocate effort. That mechanism is likely to appear well beyond software. Accountants, designers, analysts and customer-service teams may all discover that the most important AI effect is not a vanished occupation but a job whose internal proportions have been quietly rewritten.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.